diff --git a/benchmarks/edge_harness/experiments/profile_llamacpp_spark_text.py b/benchmarks/edge_harness/experiments/profile_llamacpp_spark_text.py
new file mode 100644
index 000000000..b63ea9e96
--- /dev/null
+++ b/benchmarks/edge_harness/experiments/profile_llamacpp_spark_text.py
@@ -0,0 +1,195 @@
+#!/usr/bin/env python3
+"""Check two complete Spark text requests and profile a resident llama.cpp stage.
+
+The target llama-server must already be running on the intended local device.
+This measures HTTP request wall time, including tokenization and delivery, not
+Omni pipeline time. The caller must retain the server launch and placement log.
+"""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import math
+import platform
+import time
+import urllib.request
+from urllib.parse import urlparse
+from pathlib import Path
+
+
+def _read_json(url: str, body: dict | None = None, timeout: float = 60) -> dict:
+    payload = None if body is None else json.dumps(body).encode("utf-8")
+    request = urllib.request.Request(
+        url, data=payload,
+        headers={"Content-Type": "application/json"} if body is not None else {},
+    )
+    with urllib.request.urlopen(request, timeout=timeout) as response:
+        return json.load(response)
+
+
+def _sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as stream:
+        for chunk in iter(lambda: stream.read(8 * 1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def _percentile(samples: list[float], fraction: float) -> float:
+    ordered = sorted(samples)
+    return ordered[math.ceil(fraction * len(ordered)) - 1]
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--endpoint", default="http://127.0.0.1:18173")
+    parser.add_argument("--model-file", type=Path, required=True)
+    parser.add_argument("--expected-model-sha256", required=True)
+    parser.add_argument("--warmups", type=int, default=1)
+    parser.add_argument("--repeats", type=int, default=20)
+    parser.add_argument("--profile-case", choices=("short", "inventory_120"),
+                        default="inventory_120")
+    parser.add_argument("--cache-prompt", action="store_true",
+                        help="Permit prompt/KV reuse across serial requests")
+    parser.add_argument("--cancel-check", action="store_true",
+                        help="Close a live stream and verify the following request")
+    parser.add_argument("--output", type=Path, required=True)
+    args = parser.parse_args()
+    if args.warmups < 0 or args.repeats < 1:
+        parser.error("warmups must be nonnegative and repeats positive")
+    actual_sha = _sha256(args.model_file)
+    if actual_sha != args.expected_model_sha256.lower():
+        raise ValueError("The GGUF differs from the pinned model artifact")
+
+    base = args.endpoint.rstrip("/")
+    if urlparse(base).hostname not in {"127.0.0.1", "localhost", "::1"}:
+        raise ValueError("This single-device probe requires a local llama-server")
+    health = _read_json(base + "/health")
+    if health.get("status") != "ok":
+        raise RuntimeError(f"Server not ready: {health}")
+    props = _read_json(base + "/props")
+    if Path(props.get("model_path", "")).resolve() != args.model_file.resolve():
+        raise RuntimeError("The server did not load the pinned GGUF path")
+    if props.get("total_slots") != 1:
+        raise RuntimeError("Expected one server slot for the serial profile")
+
+    inventory = "Here is an inventory listing. " + " ".join(
+        f"Item {i}: the code for warehouse district {i} is {1000 + 7 * i}."
+        for i in range(1, 121)
+    ) + " What is the code for warehouse district 73? Answer with just the number."
+    cases = [
+        ("short", "What is the capital of France? Answer in one word.", "Paris"),
+        ("inventory_120", inventory, "1511"),
+    ]
+
+    def run(name: str, prompt: str, expected: str) -> dict:
+        started = time.perf_counter()
+        result = _read_json(base + "/v1/chat/completions", {
+            "model": props["model_alias"],
+            "messages": [{"role": "user", "content": prompt}],
+            "temperature": 0,
+            "max_tokens": 96,
+            "stream": False,
+            "cache_prompt": args.cache_prompt,
+        })
+        wall_s = time.perf_counter() - started
+        choice = result["choices"][0]
+        content = choice["message"]["content"].strip()
+        row = {
+            "case": name,
+            "wall_s": wall_s,
+            "content": content,
+            "expected": expected,
+            "correct": content == expected,
+            "finish_reason": choice["finish_reason"],
+            "usage": result.get("usage"),
+        }
+        if not row["correct"] or row["finish_reason"] != "stop":
+            raise RuntimeError(f"Complete-response gate failed: {row}")
+        return row
+
+    checks = [run(*case) for case in cases]
+    profile_case = next(case for case in cases if case[0] == args.profile_case)
+    warmups = [run(*profile_case) for _ in range(args.warmups)]
+    measured = [run(*profile_case) for _ in range(args.repeats)]
+    times = [row["wall_s"] for row in measured]
+    cancellation = None
+    if args.cancel_check:
+        body = json.dumps({
+            "model": props["model_alias"],
+            "messages": [{"role": "user", "content": "Count from 1 to 500, separated by spaces."}],
+            "temperature": 0,
+            "max_tokens": 512,
+            "stream": True,
+            "cache_prompt": False,
+        }).encode("utf-8")
+        request = urllib.request.Request(
+            base + "/v1/chat/completions", data=body,
+            headers={"Content-Type": "application/json"},
+        )
+        started = time.perf_counter()
+        first_delta = None
+        first_delta_wall_s = None
+        with urllib.request.urlopen(request, timeout=30) as response:
+            for _ in range(100):
+                line = response.readline()
+                if not line:
+                    break
+                if not line.startswith(b"data: {"):
+                    continue
+                event = json.loads(line[6:])
+                delta = event["choices"][0].get("delta", {}).get("content")
+                if delta:
+                    first_delta = delta
+                    first_delta_wall_s = time.perf_counter() - started
+                    break
+        if not first_delta:
+            raise RuntimeError("Did not observe a live token before stream cancellation")
+        recovery = run(*cases[0])
+        cancellation = {
+            "first_delta": first_delta,
+            "stream_closed_after_first_content": True,
+            "first_delta_wall_s": first_delta_wall_s,
+            "recovery": recovery,
+        }
+    report = {
+        "scope": "Standalone llama.cpp full-model text path; no Omni stage binding",
+        "platform": platform.platform(),
+        "model_path": str(args.model_file),
+        "model_sha256": actual_sha,
+        "server": {
+            "endpoint": base,
+            "build_info": props.get("build_info"),
+            "model_alias": props["model_alias"],
+            "model_ftype": props.get("model_ftype"),
+            "total_slots": props["total_slots"],
+        },
+        "quality_checks": checks,
+        "warmups": warmups,
+        "samples": measured,
+        "cancellation": cancellation,
+        "profile": {
+            "case": args.profile_case,
+            "cache_prompt": args.cache_prompt,
+            "repeats": args.repeats,
+            "concurrency": 1,
+            "p50_wall_s": _percentile(times, 0.5),
+            "p95_wall_s": _percentile(times, 0.95),
+            "min_wall_s": min(times),
+            "max_wall_s": max(times),
+            "cached_prompt_tokens": [
+                row["usage"].get("prompt_tokens_details", {}).get("cached_tokens", 0)
+                for row in measured
+            ],
+        },
+        "status": "passed",
+    }
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
+    print(json.dumps({"status": report["status"], "profile": report["profile"]}, indent=2))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/README.md b/benchmarks/edge_harness/results/e2e_expansion_20260923/README.md
index 492a3ec26..cd881ad8b 100644
--- a/benchmarks/edge_harness/results/e2e_expansion_20260923/README.md
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/README.md
@@ -1,9 +1,11 @@
 # 2026-09-23 WSL CPU and native Windows expansion
 
-This extends the [60-cell device × model matrix](../e2e_profiling_20260922/evidence/summary/README.md) with previously unrun WSL CPU and native Windows paths. The September 22 timing tables remain historical. MiniCPM-o now has separate 20-request serial text-to-speech timing runs on WSL CPU and native Windows GPU, plus one synthetic image and one synthetic audio input on each platform; each InternVLA timing remains one observation. A Spark output-head component also has a separate 20-request Omni graph-stage profile on the HX370 AMD NPU. None is a sustained or full-modality profile. The [checkpoint manifests](../e2e_recovery_20260923/README.md) pin the earlier real weights; the [Spark component record](evidence/spark_amd_npu_output_head/README.md) pins its checkpoint revision and graph hashes. No model, precision or context was silently substituted.
+This extends the [60-cell device × model matrix](../e2e_profiling_20260922/evidence/summary/README.md) with previously unrun WSL CPU and native Windows paths. The September 22 timing tables remain historical. MiniCPM-o now has separate 20-request serial text-to-speech timing runs on WSL CPU and native Windows GPU, plus one synthetic image and one synthetic audio input on each platform; each InternVLA timing remains one observation. Spark has a 20-request Omni output-head component profile on the HX370 AMD NPU and separate 20-request **complete text** profiles through standalone llama.cpp on native Windows CPU and Radeon 890M. None is a sustained or full-modality profile; the standalone Spark paths still need Omni integration. The [checkpoint manifests](../e2e_recovery_20260923/README.md) pin the earlier real weights; the [Spark NPU component record](evidence/spark_amd_npu_output_head/README.md) and [CPU/Radeon whole-model record](evidence/spark_radeon890m_llamacpp/README.md) pin their respective artifacts. No model, precision or context was silently substituted.
 
 | Device / model | Verified depth | Observed output and time | Remaining gate |
 |---|---|---|---|
+| Native Windows HX370 CPU / Spark-X2.5-1.7B Q4_K_M GGUF, revision `1f7fa33b1245c14730da39e125714ad3a327901b` | Scoped complete standalone llama.cpp text path (P for two named inputs; outside Omni) | Correct `Paris` and 120-item inventory answer `1511`; 1 warmup + 20 uncached serial full inventory requests p50/p95 6.65/6.86 s, 20/20 correct. Stream-close recovery passed. | Omni whole-model stage binding, broader quality, explicit loading-peak admission and sustained behavior; [raw evidence](evidence/spark_radeon890m_llamacpp/README.md). |
+| Native Windows HX370 Radeon 890M / Spark-X2.5-1.7B same GGUF | Scoped complete standalone llama.cpp Vulkan1 text path (P for two named inputs; outside Omni) | All 29 logged layer assignments on Vulkan1, with a separate Vulkan_Host buffer; correct `Paris` and `1511`; 1 warmup + 20 uncached serial inventory requests p50/p95 2.18/2.20 s, 20/20 correct. Stream-close recovery passed. | Omni binding, state and memory contract, broad quality, GPU-memory attribution and sustained power/thermal behavior; [raw evidence](evidence/spark_radeon890m_llamacpp/README.md). |
 | Native Windows HX370 AMD NPU / Spark-X2.5-1.7B output head, revision `448e61eb392c00f2c403185c5b56d5e0665bfaab` | Real-weight RMS norm + full-logit projection component through Omni `external.graph.v1` (B+C, not P) | One fused NPU partition and seven CPU nodes; 4/4 captured activations top-1 match, 55.43–56.81 dB full-logit SNR. One warmup + 20 serial repeated-input component requests: p50/p95 10.21/12.44 ms, versus 17.97/20.29 ms for the original FP32 CPU head through Omni. NPU startup 18.21 s and peak worker RSS 3.52 GiB. [Raw evidence](evidence/spark_amd_npu_output_head/README.md). | Actual vLLM handoff and full generation, transfer cost, quality across inputs/context, memory admission, cancellation and sustained power/thermal behavior. |
 | WSL CPU / MiniCPM-o 4.5, revision `503e754207c94da6bb26850b4469f367c9ea3582` | Scoped complete three-stage text-to-text+speech path (P for this input only) | Coherent text and a 5.2 s mono 24 kHz WAV. First request 27.09 s after 73.65 s startup; separate 1-warmup/20-measured serial run: startup 73.89 s, request wall p50 21.86 s/p95 22.97 s. | Speech intelligibility/alignment, real image/audio quality, video input, concurrency, interruption, power and thermal behavior. |
 | WSL CPU / MiniCPM-o 4.5, same revision | Scoped three-stage **image**-to-text+speech path (P for one synthetic image) | For the 448×448 red square, returned a correct description and nonzero 5.2 s WAV. Startup 112.44 s; one request 37.70 s. Text and token IDs differed from WSL GPU while describing the same visible content. | Real-image suite, speech meaning, repeated latency, memory admission and clean shutdown. |
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/.gitattributes b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/.gitattributes
new file mode 100644
index 000000000..eecd6aa72
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/.gitattributes
@@ -0,0 +1,4 @@
+# Keep measured and failed native-Windows outputs byte-for-byte.
+*.json -text -eol
+*.log -text -eol
+*.txt -text -eol
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/README.md b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/README.md
new file mode 100644
index 000000000..db0c06c61
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/README.md
@@ -0,0 +1,31 @@
+# Spark-X2.5-1.7B Q4_K_M on native Windows CPU and Radeon 890M
+
+This is a **scoped, complete text-model path in standalone llama.cpp**, with real weights, prefill, autoregressive decode, KV state and answer delivery. It is **not** yet an Omni `StageRuntime` backend, and it says nothing about other Spark variants or the TTS/MiniCPM/VLA cells. The current [device × model matrix](../../../e2e_profiling_20260922/evidence/summary/README.md) distinguishes this backend result from Omni integration.
+
+The GGUF is [XHToken/Spark-X2.5-1.7B-GGUF](https://huggingface.co/XHToken/Spark-X2.5-1.7B-GGUF/tree/1f7fa33b1245c14730da39e125714ad3a327901b) revision `1f7fa33b1245c14730da39e125714ad3a327901b`, file `Spark-X2.5-1.7B-Q4_K_M.gguf` (1,107,457,856 bytes; SHA-256 `902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4`). It remains in local model storage, not Git. The Windows Vulkan binary is [llama.cpp b11124](https://github.com/ggml-org/llama.cpp/releases/tag/b11124), build commit `1a679828f`; the release ZIP SHA-256 is `ece009eff2a18884246b4aef785ce4efd9eb36f4fc38e3d6727af6fe53bd2c75`. See the [artifact manifest](artifact_manifest.json) and [Vulkan enumeration](vulkaninfo_summary.log). The host was Windows 11 build 26200 on Ryzen AI 9 HX 370 with Radeon 890M driver `25.20.18.06 (LLPC)`. An RTX 5090 Laptop GPU was present, so the Radeon command explicitly selected `-dev Vulkan1`; the [placement log](spark_igpu_q4km_placement_20260923.log) records all 29 logged layer assignments on Vulkan1, a 1,051 MiB Vulkan1 model buffer and a separate 210 MiB Vulkan_Host buffer. The CPU command used `-dev none -ngl 0`. Both used a 4,096-token context, one slot, temperature zero, reasoning off for server requests, and default f16 KV cache.
+
+Both backends returned `Paris` for the short question and `1511` for the 120-item inventory (district 73), with `finish_reason=stop`. The inventory generated by the profile script used 2,463 prompt tokens and five completion tokens. The [Radeon profile](spark_igpu_server_inventory_profile20_nocache_20260923.json) and [native CPU profile](spark_cpu_server_inventory_profile20_nocache_20260923.json) each contain the two quality checks, one excluded warmup, and all 20 measured serial HTTP request wall times. `cache_prompt=false` was sent and every measured response reported zero cached prompt tokens. Nearest-rank timings are:
+
+| Full text request, 120-item inventory | Radeon 890M Vulkan1 | Native Windows CPU |
+|---|---:|---:|
+| p50 request wall | 2.18 s | 6.65 s |
+| p95 request wall | 2.20 s | 6.86 s |
+| Correct measured responses | 20/20 | 20/20 |
+
+The [Radeon cancellation check](spark_igpu_server_cancel_20260923.json) and CPU profile each closed a live stream after its first content delta, then submitted a new question that correctly returned `Paris`. This checks subsequent request recovery; it does not prove the exact point at which llama.cpp stopped backend compute. A [single post-profile process snapshot](process_snapshot.json) records working sets of 1.84 GB for Radeon and 2.14 GB for CPU. These are neither peaks nor complete iGPU memory attribution.
+
+The [three-repetition synthetic llama-bench](spark_igpu_q4km_bench_20260923.json) at p128/n32 recorded mean prompt throughput 1,602 tok/s and generation throughput 67.34 tok/s on Vulkan1; the [same-artifact CPU bench](spark_cpu_q4km_bench_20260923.json) recorded 402.22 and 54.61 tok/s. These are separate kernel-oriented measurements, not the complete-request times above. Their logs and the [one-repetition verbose placement check](spark_igpu_q4km_placement_20260923.log) are retained.
+
+Two unsuccessful CLI attempts are retained: [context 2,048 refused](spark_igpu_inventory_answer_20260923.log) because the prompt needed 2,463 tokens, and [context 4,096 with reasoning enabled](spark_igpu_inventory_answer_ctx4096_20260923.log) exhausted a 256-token output limit before giving a final answer. The [reasoning-off 4,096-token run](spark_igpu_inventory_answer_no_reason_20260923.log) returned `1511`. The short reasoning-enabled [CLI run](spark_igpu_q4km_answer_20260923.log) returned `Paris`. These are workload/configuration boundaries, not general model failures.
+
+To reproduce on this Windows host, download the pinned GGUF, verify its SHA-256, and use the pinned llama.cpp Vulkan release. Start one backend at a time with the full model resident:
+
+```powershell
+$bin = 'C:\Users\zhout\w2\llama-vulkan-b11124\bin'
+$model = 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf'
+& "$bin\llama-server.exe" -m $model -dev Vulkan1 -ngl 99 -c 4096 -np 1 --host 127.0.0.1 --port 18173 --reasoning off --no-webui
+```
+
+Run [the profile script](../../../../experiments/profile_llamacpp_spark_text.py) in another terminal with `--model-file $model`, `--expected-model-sha256 902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4`, `--profile-case inventory_120`, `--cancel-check`, and an `--output` path. For CPU, launch the same binary with `-dev none -ngl 0` on another port and set `--endpoint` accordingly. The fixed artifact, context, precision, prompt and cache policy are required for comparison.
+
+The next project gate is to wrap this **whole autoregressive session** as a bounded llama.cpp stage behind Omni's request/event, memory-admission, state, streaming and cancellation contracts, then compare it with the current vLLM CPU/CUDA paths. The two-answer quality check, one repeated input, absent load-peak/power/thermal data, and client-close-only cancellation check do not establish broader release qualification.
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/artifact_manifest.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/artifact_manifest.json
new file mode 100644
index 000000000..2c8cde880
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/artifact_manifest.json
@@ -0,0 +1,21 @@
+{
+  "schema_version": 1,
+  "model": "XHToken/Spark-X2.5-1.7B-GGUF",
+  "model_revision": "1f7fa33b1245c14730da39e125714ad3a327901b",
+  "model_file": "Spark-X2.5-1.7B-Q4_K_M.gguf",
+  "model_file_bytes": 1107457856,
+  "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4",
+  "backend": "llama.cpp b11124-1a679828f Windows x64 Vulkan release",
+  "backend_release_zip_sha256": "ece009eff2a18884246b4aef785ce4efd9eb36f4fc38e3d6727af6fe53bd2c75",
+  "llama_server_exe_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52",
+  "hardware": "AMD Ryzen AI 9 HX 370, Radeon 890M integrated GPU; RTX 5090 Laptop also installed but excluded from the Radeon run",
+  "os": "Windows 11 build 26200",
+  "radeon_vulkan_driver_info": "25.20.18.06 (LLPC)",
+  "radeon_device": "Vulkan1",
+  "context_tokens": 4096,
+  "parallel_slots": 1,
+  "precision": "Q4_K_M weights, default f16 KV cache",
+  "sampling": "temperature 0; llama-server --reasoning off; maximum 96 completion tokens",
+  "profile": "one warmup, 20 measured serial HTTP chat completions of 2463 prompt tokens and 5 completion tokens, cache_prompt=false",
+  "evidence_depth": "scoped complete standalone text route; not an Omni StageRuntime binding"
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/process_snapshot.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/process_snapshot.json
new file mode 100644
index 000000000..cd687ad07
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/process_snapshot.json
@@ -0,0 +1,16 @@
+{
+  "scope": "One Get-Process snapshot after each 20-request profile, before stopping that server; not a peak or GPU allocation measurement",
+  "radeon_server": {
+    "pid": 49832,
+    "device_argument": "Vulkan1",
+    "working_set_bytes": 1841168384,
+    "private_memory_bytes": 1844895744
+  },
+  "cpu_server": {
+    "pid": 53356,
+    "device_argument": "none",
+    "working_set_bytes": 2141794304,
+    "private_memory_bytes": 1061339136
+  },
+  "note": "Radeon 890M shares host RAM. Windows process working set cannot attribute all integrated-GPU allocations. No load peak, thermal or power measurement was captured."
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_cpu_q4km_bench_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_cpu_q4km_bench_20260923.json
new file mode 100644
index 000000000..e884d1d5f
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_cpu_q4km_bench_20260923.json
@@ -0,0 +1,92 @@
+[
+  {
+    "build_commit": "1a679828f",
+    "build_number": 11124,
+    "cpu_info": "AMD Ryzen AI 9 HX 370 w/ Radeon 890M           ",
+    "gpu_info": "NVIDIA GeForce RTX 5090 Laptop GPU, AMD Radeon(TM) 890M Graphics",
+    "backends": "Vulkan",
+    "model_filename": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+    "model_type": "spark2_5 1.7B Q4_K - Medium",
+    "model_size": 1102073856,
+    "model_n_params": 1707657216,
+    "n_batch": 2048,
+    "n_ubatch": 512,
+    "n_threads": 12,
+    "cpu_mask": "0x0",
+    "cpu_strict": false,
+    "poll": 50,
+    "type_k": "f16",
+    "type_v": "f16",
+    "n_gpu_layers": 0,
+    "n_cpu_moe": 0,
+    "split_mode": "layer",
+    "main_gpu": 0,
+    "no_kv_offload": false,
+    "flash_attn": -1,
+    "devices": "none",
+    "tensor_split": "0.00",
+    "tensor_buft_overrides": "none",
+    "load_mode": "auto",
+    "lazy_mode": "auto",
+    "embeddings": false,
+    "no_op_offload": 0,
+    "no_host": false,
+    "fit_target": 0,
+    "fit_min_ctx": 0,
+    "n_prompt": 128,
+    "n_gen": 0,
+    "n_depth": 0,
+    "test_time": "2026-09-23T10:41:21Z",
+    "avg_ns": 318299633,
+    "stddev_ns": 5501058,
+    "avg_ts": 402.216268,
+    "stddev_ts": 6.896283,
+    "samples_ns": [ 324510000, 314039100, 316349800 ],
+    "samples_ts": [ 394.441, 407.593, 404.615 ]
+  },
+  {
+    "build_commit": "1a679828f",
+    "build_number": 11124,
+    "cpu_info": "AMD Ryzen AI 9 HX 370 w/ Radeon 890M           ",
+    "gpu_info": "NVIDIA GeForce RTX 5090 Laptop GPU, AMD Radeon(TM) 890M Graphics",
+    "backends": "Vulkan",
+    "model_filename": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+    "model_type": "spark2_5 1.7B Q4_K - Medium",
+    "model_size": 1102073856,
+    "model_n_params": 1707657216,
+    "n_batch": 2048,
+    "n_ubatch": 512,
+    "n_threads": 12,
+    "cpu_mask": "0x0",
+    "cpu_strict": false,
+    "poll": 50,
+    "type_k": "f16",
+    "type_v": "f16",
+    "n_gpu_layers": 0,
+    "n_cpu_moe": 0,
+    "split_mode": "layer",
+    "main_gpu": 0,
+    "no_kv_offload": false,
+    "flash_attn": -1,
+    "devices": "none",
+    "tensor_split": "0.00",
+    "tensor_buft_overrides": "none",
+    "load_mode": "auto",
+    "lazy_mode": "auto",
+    "embeddings": false,
+    "no_op_offload": 0,
+    "no_host": false,
+    "fit_target": 0,
+    "fit_min_ctx": 0,
+    "n_prompt": 0,
+    "n_gen": 32,
+    "n_depth": 0,
+    "test_time": "2026-09-23T10:41:22Z",
+    "avg_ns": 586482766,
+    "stddev_ns": 20881467,
+    "avg_ts": 54.607989,
+    "stddev_ts": 1.914133,
+    "samples_ns": [ 579411400, 570055200, 609981700 ],
+    "samples_ts": [ 55.2285, 56.1349, 52.4606 ]
+  }
+]
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_cpu_q4km_bench_20260923.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_cpu_q4km_bench_20260923.log
new file mode 100644
index 000000000..ff7b2f636
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_cpu_q4km_bench_20260923.log
@@ -0,0 +1,6 @@
+load_backend: loaded RPC backend from C:\Users\zhout\w2\llama-vulkan-b11124\bin\ggml-rpc.dll
+ggml_vulkan: Found 2 Vulkan devices:
+ggml_vulkan: 0 = NVIDIA GeForce RTX 5090 Laptop GPU (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | fp4: 0 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
+ggml_vulkan: 1 = AMD Radeon(TM) 890M Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 1 | fp4: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+load_backend: loaded Vulkan backend from C:\Users\zhout\w2\llama-vulkan-b11124\bin\ggml-vulkan.dll
+load_backend: loaded CPU backend from C:\Users\zhout\w2\llama-vulkan-b11124\bin\ggml-cpu-zen4.dll
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_cpu_server_20260923.err.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_cpu_server_20260923.err.log
new file mode 100644
index 000000000..2984176b0
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_cpu_server_20260923.err.log
@@ -0,0 +1,227 @@
+0.00.000.753 I srv  llama_server: initializing ...
+0.00.680.017 I cmn  common_param: common_params_print_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
+0.00.680.147 I srv  init_listene: The UI is disabled
+0.00.680.148 I srv  init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable
+0.00.680.206 W srv  llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655)
+0.00.698.255 I srv    load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf'
+0.01.429.499 I cmn          init: llama threadpool init, n_threads = 12
+0.01.556.894 I srv    load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false'
+0.01.575.516 W srv          init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve)
+0.01.575.549 I srv  llama_server: model loaded
+0.01.575.553 I srv  llama_server: listening on http://127.0.0.1:18174
+0.13.249.422 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
+0.13.249.475 I slot launch_slot_: id  0 | task 0 | processing task, is_child = 0
+0.13.377.851 I slot print_timing: id  0 | task 0 | prompt eval time =     109.36 ms /    28 tokens (    3.91 ms per token,   256.05 tokens per second)
+0.13.377.861 I slot print_timing: id  0 | task 0 |        eval time =      18.98 ms /     2 tokens (   18.98 ms per token,    52.69 tokens per second)
+0.13.377.863 I slot print_timing: id  0 | task 0 |       total time =     128.33 ms /    30 tokens
+0.13.377.864 I slot print_timing: id  0 | task 0 |    graphs reused =          1
+0.13.377.907 I slot      release: id  0 | task 0 | stop processing: n_tokens = 29, truncated = 0
+0.13.408.295 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = 12699867
+0.13.408.973 I slot launch_slot_: id  0 | task 5 | processing task, is_child = 0
+0.18.412.680 I slot print_timing: id  0 | task 5 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.00 s / 389.11 tokens per second
+0.19.797.975 I slot print_timing: id  0 | task 5 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.39 s / 384.88 tokens per second
+0.19.922.141 I slot print_timing: id  0 | task 5 | prompt eval time =    6425.45 ms /  2463 tokens (    2.61 ms per token,   383.32 tokens per second)
+0.19.922.151 I slot print_timing: id  0 | task 5 |        eval time =      87.69 ms /     5 tokens (   21.92 ms per token,    45.62 tokens per second)
+0.19.922.152 I slot print_timing: id  0 | task 5 |       total time =    6513.14 ms /  2468 tokens
+0.19.922.153 I slot print_timing: id  0 | task 5 |    graphs reused =          4
+0.19.922.201 I slot      release: id  0 | task 5 | stop processing: n_tokens = 2467, truncated = 0
+0.19.956.838 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+0.19.957.093 I slot launch_slot_: id  0 | task 14 | processing task, is_child = 0
+0.24.951.867 I slot print_timing: id  0 | task 14 | prompt processing, n_tokens =   1947, progress = 0.79, t =   4.99 s / 389.81 tokens per second
+0.26.384.023 I slot print_timing: id  0 | task 14 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.43 s / 382.61 tokens per second
+0.26.511.285 I slot print_timing: id  0 | task 14 | prompt eval time =    6462.69 ms /  2463 tokens (    2.62 ms per token,   381.11 tokens per second)
+0.26.511.296 I slot print_timing: id  0 | task 14 |        eval time =      91.46 ms /     5 tokens (   22.87 ms per token,    43.73 tokens per second)
+0.26.511.298 I slot print_timing: id  0 | task 14 |       total time =    6554.15 ms /  2468 tokens
+0.26.511.299 I slot print_timing: id  0 | task 14 |    graphs reused =          7
+0.26.511.360 I slot      release: id  0 | task 14 | stop processing: n_tokens = 2467, truncated = 0
+0.26.526.973 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+0.26.527.237 I slot launch_slot_: id  0 | task 23 | processing task, is_child = 0
+0.31.527.649 I slot print_timing: id  0 | task 23 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.00 s / 389.37 tokens per second
+0.32.950.045 I slot print_timing: id  0 | task 23 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.42 s / 382.86 tokens per second
+0.33.078.228 I slot print_timing: id  0 | task 23 | prompt eval time =    6459.73 ms /  2463 tokens (    2.62 ms per token,   381.29 tokens per second)
+0.33.078.241 I slot print_timing: id  0 | task 23 |        eval time =      91.23 ms /     5 tokens (   22.81 ms per token,    43.85 tokens per second)
+0.33.078.243 I slot print_timing: id  0 | task 23 |       total time =    6550.96 ms /  2468 tokens
+0.33.078.244 I slot print_timing: id  0 | task 23 |    graphs reused =         10
+0.33.078.305 I slot      release: id  0 | task 23 | stop processing: n_tokens = 2467, truncated = 0
+0.33.095.368 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+0.33.095.610 I slot launch_slot_: id  0 | task 32 | processing task, is_child = 0
+0.38.208.106 I slot print_timing: id  0 | task 32 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.11 s / 380.84 tokens per second
+0.39.662.252 I slot print_timing: id  0 | task 32 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.57 s / 374.47 tokens per second
+0.39.789.390 I slot print_timing: id  0 | task 32 | prompt eval time =    6603.22 ms /  2463 tokens (    2.68 ms per token,   373.00 tokens per second)
+0.39.789.408 I slot print_timing: id  0 | task 32 |        eval time =      90.53 ms /     5 tokens (   22.63 ms per token,    44.18 tokens per second)
+0.39.789.409 I slot print_timing: id  0 | task 32 |       total time =    6693.75 ms /  2468 tokens
+0.39.789.411 I slot print_timing: id  0 | task 32 |    graphs reused =         13
+0.39.789.503 I slot      release: id  0 | task 32 | stop processing: n_tokens = 2467, truncated = 0
+0.39.824.522 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+0.39.824.746 I slot launch_slot_: id  0 | task 41 | processing task, is_child = 0
+0.45.094.702 I slot print_timing: id  0 | task 41 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.27 s / 369.46 tokens per second
+0.46.568.773 I slot print_timing: id  0 | task 41 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.74 s / 364.62 tokens per second
+0.46.695.855 I slot print_timing: id  0 | task 41 | prompt eval time =    6780.80 ms /  2463 tokens (    2.75 ms per token,   363.23 tokens per second)
+0.46.695.865 I slot print_timing: id  0 | task 41 |        eval time =      90.29 ms /     5 tokens (   22.57 ms per token,    44.30 tokens per second)
+0.46.695.866 I slot print_timing: id  0 | task 41 |       total time =    6871.08 ms /  2468 tokens
+0.46.695.868 I slot print_timing: id  0 | task 41 |    graphs reused =         16
+0.46.695.925 I slot      release: id  0 | task 41 | stop processing: n_tokens = 2467, truncated = 0
+0.46.740.369 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+0.46.740.649 I slot launch_slot_: id  0 | task 50 | processing task, is_child = 0
+0.51.966.492 I slot print_timing: id  0 | task 50 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.23 s / 372.57 tokens per second
+0.53.425.786 I slot print_timing: id  0 | task 50 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.69 s / 367.83 tokens per second
+0.53.557.468 I slot print_timing: id  0 | task 50 | prompt eval time =    6723.11 ms /  2463 tokens (    2.73 ms per token,   366.35 tokens per second)
+0.53.557.479 I slot print_timing: id  0 | task 50 |        eval time =      93.68 ms /     5 tokens (   23.42 ms per token,    42.70 tokens per second)
+0.53.557.481 I slot print_timing: id  0 | task 50 |       total time =    6816.79 ms /  2468 tokens
+0.53.557.482 I slot print_timing: id  0 | task 50 |    graphs reused =         19
+0.53.557.541 I slot      release: id  0 | task 50 | stop processing: n_tokens = 2467, truncated = 0
+0.53.579.546 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+0.53.579.780 I slot launch_slot_: id  0 | task 59 | processing task, is_child = 0
+0.58.828.506 I slot print_timing: id  0 | task 59 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.25 s / 370.95 tokens per second
+1.00.247.999 I slot print_timing: id  0 | task 59 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.67 s / 368.77 tokens per second
+1.00.375.273 I slot print_timing: id  0 | task 59 | prompt eval time =    6705.84 ms /  2463 tokens (    2.72 ms per token,   367.29 tokens per second)
+1.00.375.284 I slot print_timing: id  0 | task 59 |        eval time =      89.63 ms /     5 tokens (   22.41 ms per token,    44.63 tokens per second)
+1.00.375.286 I slot print_timing: id  0 | task 59 |       total time =    6795.47 ms /  2468 tokens
+1.00.375.287 I slot print_timing: id  0 | task 59 |    graphs reused =         22
+1.00.375.346 I slot      release: id  0 | task 59 | stop processing: n_tokens = 2467, truncated = 0
+1.00.394.386 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+1.00.394.629 I slot launch_slot_: id  0 | task 68 | processing task, is_child = 0
+1.05.468.999 I slot print_timing: id  0 | task 68 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.07 s / 383.70 tokens per second
+1.06.899.462 I slot print_timing: id  0 | task 68 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.50 s / 378.03 tokens per second
+1.07.024.826 I slot print_timing: id  0 | task 68 | prompt eval time =    6541.89 ms /  2463 tokens (    2.66 ms per token,   376.50 tokens per second)
+1.07.024.837 I slot print_timing: id  0 | task 68 |        eval time =      88.28 ms /     5 tokens (   22.07 ms per token,    45.31 tokens per second)
+1.07.024.839 I slot print_timing: id  0 | task 68 |       total time =    6630.17 ms /  2468 tokens
+1.07.024.840 I slot print_timing: id  0 | task 68 |    graphs reused =         25
+1.07.024.901 I slot      release: id  0 | task 68 | stop processing: n_tokens = 2467, truncated = 0
+1.07.041.049 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+1.07.041.234 I slot launch_slot_: id  0 | task 77 | processing task, is_child = 0
+1.12.106.202 I slot print_timing: id  0 | task 77 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.06 s / 384.41 tokens per second
+1.13.548.801 I slot print_timing: id  0 | task 77 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.51 s / 377.87 tokens per second
+1.13.676.825 I slot print_timing: id  0 | task 77 | prompt eval time =    6543.94 ms /  2463 tokens (    2.66 ms per token,   376.38 tokens per second)
+1.13.676.839 I slot print_timing: id  0 | task 77 |        eval time =      91.63 ms /     5 tokens (   22.91 ms per token,    43.65 tokens per second)
+1.13.676.841 I slot print_timing: id  0 | task 77 |       total time =    6635.57 ms /  2468 tokens
+1.13.676.843 I slot print_timing: id  0 | task 77 |    graphs reused =         28
+1.13.676.908 I slot      release: id  0 | task 77 | stop processing: n_tokens = 2467, truncated = 0
+1.13.694.166 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+1.13.694.409 I slot launch_slot_: id  0 | task 86 | processing task, is_child = 0
+1.18.741.246 I slot print_timing: id  0 | task 86 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.05 s / 385.79 tokens per second
+1.20.173.554 I slot print_timing: id  0 | task 86 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.48 s / 379.53 tokens per second
+1.20.306.665 I slot print_timing: id  0 | task 86 | prompt eval time =    6517.54 ms /  2463 tokens (    2.65 ms per token,   377.90 tokens per second)
+1.20.306.674 I slot print_timing: id  0 | task 86 |        eval time =      94.69 ms /     5 tokens (   23.67 ms per token,    42.24 tokens per second)
+1.20.306.676 I slot print_timing: id  0 | task 86 |       total time =    6612.23 ms /  2468 tokens
+1.20.306.676 I slot print_timing: id  0 | task 86 |    graphs reused =         31
+1.20.306.722 I slot      release: id  0 | task 86 | stop processing: n_tokens = 2467, truncated = 0
+1.20.323.832 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+1.20.324.144 I slot launch_slot_: id  0 | task 95 | processing task, is_child = 0
+1.25.380.378 I slot print_timing: id  0 | task 95 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.06 s / 385.07 tokens per second
+1.26.815.774 I slot print_timing: id  0 | task 95 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.49 s / 378.80 tokens per second
+1.26.983.293 I slot print_timing: id  0 | task 95 | prompt eval time =    6540.21 ms /  2463 tokens (    2.66 ms per token,   376.59 tokens per second)
+1.26.983.306 I slot print_timing: id  0 | task 95 |        eval time =     118.91 ms /     5 tokens (   29.73 ms per token,    33.64 tokens per second)
+1.26.983.308 I slot print_timing: id  0 | task 95 |       total time =    6659.12 ms /  2468 tokens
+1.26.983.312 I slot print_timing: id  0 | task 95 |    graphs reused =         34
+1.26.983.404 I slot      release: id  0 | task 95 | stop processing: n_tokens = 2467, truncated = 0
+1.27.008.527 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+1.27.008.800 I slot launch_slot_: id  0 | task 104 | processing task, is_child = 0
+1.32.116.018 I slot print_timing: id  0 | task 104 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.11 s / 381.23 tokens per second
+1.33.568.327 I slot print_timing: id  0 | task 104 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.56 s / 374.88 tokens per second
+1.33.695.121 I slot print_timing: id  0 | task 104 | prompt eval time =    6596.75 ms /  2463 tokens (    2.68 ms per token,   373.37 tokens per second)
+1.33.695.137 I slot print_timing: id  0 | task 104 |        eval time =      89.55 ms /     5 tokens (   22.39 ms per token,    44.67 tokens per second)
+1.33.695.139 I slot print_timing: id  0 | task 104 |       total time =    6686.29 ms /  2468 tokens
+1.33.695.140 I slot print_timing: id  0 | task 104 |    graphs reused =         37
+1.33.695.199 I slot      release: id  0 | task 104 | stop processing: n_tokens = 2467, truncated = 0
+1.33.714.810 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+1.33.715.040 I slot launch_slot_: id  0 | task 113 | processing task, is_child = 0
+1.38.833.727 I slot print_timing: id  0 | task 113 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.12 s / 380.37 tokens per second
+1.40.285.078 I slot print_timing: id  0 | task 113 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.57 s / 374.28 tokens per second
+1.40.414.623 I slot print_timing: id  0 | task 113 | prompt eval time =    6607.07 ms /  2463 tokens (    2.68 ms per token,   372.78 tokens per second)
+1.40.414.635 I slot print_timing: id  0 | task 113 |        eval time =      92.49 ms /     5 tokens (   23.12 ms per token,    43.25 tokens per second)
+1.40.414.636 I slot print_timing: id  0 | task 113 |       total time =    6699.55 ms /  2468 tokens
+1.40.414.638 I slot print_timing: id  0 | task 113 |    graphs reused =         40
+1.40.414.698 I slot      release: id  0 | task 113 | stop processing: n_tokens = 2467, truncated = 0
+1.40.457.999 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+1.40.458.332 I slot launch_slot_: id  0 | task 122 | processing task, is_child = 0
+1.45.515.244 I slot print_timing: id  0 | task 122 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.06 s / 385.02 tokens per second
+1.46.934.527 I slot print_timing: id  0 | task 122 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.48 s / 379.70 tokens per second
+1.47.059.907 I slot print_timing: id  0 | task 122 | prompt eval time =    6511.84 ms /  2463 tokens (    2.64 ms per token,   378.23 tokens per second)
+1.47.059.920 I slot print_timing: id  0 | task 122 |        eval time =      89.70 ms /     5 tokens (   22.43 ms per token,    44.59 tokens per second)
+1.47.059.921 I slot print_timing: id  0 | task 122 |       total time =    6601.55 ms /  2468 tokens
+1.47.059.923 I slot print_timing: id  0 | task 122 |    graphs reused =         43
+1.47.059.982 I slot      release: id  0 | task 122 | stop processing: n_tokens = 2467, truncated = 0
+1.47.075.104 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+1.47.075.451 I slot launch_slot_: id  0 | task 131 | processing task, is_child = 0
+1.52.162.919 I slot print_timing: id  0 | task 131 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.09 s / 382.71 tokens per second
+1.53.588.986 I slot print_timing: id  0 | task 131 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.51 s / 377.52 tokens per second
+1.53.714.560 I slot print_timing: id  0 | task 131 | prompt eval time =    6549.01 ms /  2463 tokens (    2.66 ms per token,   376.09 tokens per second)
+1.53.714.568 I slot print_timing: id  0 | task 131 |        eval time =      90.07 ms /     5 tokens (   22.52 ms per token,    44.41 tokens per second)
+1.53.714.569 I slot print_timing: id  0 | task 131 |       total time =    6639.08 ms /  2468 tokens
+1.53.714.570 I slot print_timing: id  0 | task 131 |    graphs reused =         46
+1.53.714.617 I slot      release: id  0 | task 131 | stop processing: n_tokens = 2467, truncated = 0
+1.53.736.012 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+1.53.736.343 I slot launch_slot_: id  0 | task 140 | processing task, is_child = 0
+1.58.820.464 I slot print_timing: id  0 | task 140 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.08 s / 382.96 tokens per second
+2.00.229.100 I slot print_timing: id  0 | task 140 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.49 s / 378.73 tokens per second
+2.00.354.335 I slot print_timing: id  0 | task 140 | prompt eval time =    6528.77 ms /  2463 tokens (    2.65 ms per token,   377.25 tokens per second)
+2.00.354.349 I slot print_timing: id  0 | task 140 |        eval time =      89.20 ms /     5 tokens (   22.30 ms per token,    44.84 tokens per second)
+2.00.354.351 I slot print_timing: id  0 | task 140 |       total time =    6617.96 ms /  2468 tokens
+2.00.354.352 I slot print_timing: id  0 | task 140 |    graphs reused =         49
+2.00.354.411 I slot      release: id  0 | task 140 | stop processing: n_tokens = 2467, truncated = 0
+2.00.373.667 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+2.00.373.920 I slot launch_slot_: id  0 | task 149 | processing task, is_child = 0
+2.05.440.828 I slot print_timing: id  0 | task 149 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.07 s / 384.26 tokens per second
+2.06.876.459 I slot print_timing: id  0 | task 149 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.50 s / 378.16 tokens per second
+2.07.005.166 I slot print_timing: id  0 | task 149 | prompt eval time =    6539.13 ms /  2463 tokens (    2.65 ms per token,   376.66 tokens per second)
+2.07.005.179 I slot print_timing: id  0 | task 149 |        eval time =      92.08 ms /     5 tokens (   23.02 ms per token,    43.44 tokens per second)
+2.07.005.180 I slot print_timing: id  0 | task 149 |       total time =    6631.21 ms /  2468 tokens
+2.07.005.181 I slot print_timing: id  0 | task 149 |    graphs reused =         52
+2.07.005.246 I slot      release: id  0 | task 149 | stop processing: n_tokens = 2467, truncated = 0
+2.07.025.193 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+2.07.025.411 I slot launch_slot_: id  0 | task 158 | processing task, is_child = 0
+2.12.110.161 I slot print_timing: id  0 | task 158 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.08 s / 382.91 tokens per second
+2.13.551.229 I slot print_timing: id  0 | task 158 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.53 s / 376.81 tokens per second
+2.13.692.806 I slot print_timing: id  0 | task 158 | prompt eval time =    6563.20 ms /  2463 tokens (    2.66 ms per token,   375.27 tokens per second)
+2.13.692.820 I slot print_timing: id  0 | task 158 |        eval time =     104.17 ms /     5 tokens (   26.04 ms per token,    38.40 tokens per second)
+2.13.692.822 I slot print_timing: id  0 | task 158 |       total time =    6667.37 ms /  2468 tokens
+2.13.692.823 I slot print_timing: id  0 | task 158 |    graphs reused =         55
+2.13.692.884 I slot      release: id  0 | task 158 | stop processing: n_tokens = 2467, truncated = 0
+2.13.717.566 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+2.13.717.836 I slot launch_slot_: id  0 | task 167 | processing task, is_child = 0
+2.18.883.955 I slot print_timing: id  0 | task 167 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.17 s / 376.88 tokens per second
+2.20.332.089 I slot print_timing: id  0 | task 167 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.61 s / 371.78 tokens per second
+2.20.455.551 I slot print_timing: id  0 | task 167 | prompt eval time =    6650.57 ms /  2463 tokens (    2.70 ms per token,   370.34 tokens per second)
+2.20.455.565 I slot print_timing: id  0 | task 167 |        eval time =      87.12 ms /     5 tokens (   21.78 ms per token,    45.91 tokens per second)
+2.20.455.567 I slot print_timing: id  0 | task 167 |       total time =    6737.69 ms /  2468 tokens
+2.20.455.568 I slot print_timing: id  0 | task 167 |    graphs reused =         58
+2.20.455.627 I slot      release: id  0 | task 167 | stop processing: n_tokens = 2467, truncated = 0
+2.20.493.286 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+2.20.493.630 I slot launch_slot_: id  0 | task 176 | processing task, is_child = 0
+2.25.590.307 I slot print_timing: id  0 | task 176 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.10 s / 382.02 tokens per second
+2.27.010.965 I slot print_timing: id  0 | task 176 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.52 s / 377.30 tokens per second
+2.27.147.550 I slot print_timing: id  0 | task 176 | prompt eval time =    6558.48 ms /  2463 tokens (    2.66 ms per token,   375.54 tokens per second)
+2.27.147.562 I slot print_timing: id  0 | task 176 |        eval time =      95.41 ms /     5 tokens (   23.85 ms per token,    41.93 tokens per second)
+2.27.147.564 I slot print_timing: id  0 | task 176 |       total time =    6653.89 ms /  2468 tokens
+2.27.147.565 I slot print_timing: id  0 | task 176 |    graphs reused =         61
+2.27.147.624 I slot      release: id  0 | task 176 | stop processing: n_tokens = 2467, truncated = 0
+2.27.189.739 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+2.27.190.004 I slot launch_slot_: id  0 | task 185 | processing task, is_child = 0
+2.32.259.050 I slot print_timing: id  0 | task 185 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.07 s / 384.10 tokens per second
+2.33.664.868 I slot print_timing: id  0 | task 185 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.47 s / 379.78 tokens per second
+2.33.793.530 I slot print_timing: id  0 | task 185 | prompt eval time =    6511.36 ms /  2463 tokens (    2.64 ms per token,   378.26 tokens per second)
+2.33.793.542 I slot print_timing: id  0 | task 185 |        eval time =      92.14 ms /     5 tokens (   23.03 ms per token,    43.41 tokens per second)
+2.33.793.545 I slot print_timing: id  0 | task 185 |       total time =    6603.49 ms /  2468 tokens
+2.33.793.548 I slot print_timing: id  0 | task 185 |    graphs reused =         64
+2.33.793.645 I slot      release: id  0 | task 185 | stop processing: n_tokens = 2467, truncated = 0
+2.33.813.127 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+2.33.813.397 I slot launch_slot_: id  0 | task 194 | processing task, is_child = 0
+2.38.866.970 I slot print_timing: id  0 | task 194 | prompt processing, n_tokens =   1947, progress = 0.79, t =   5.05 s / 385.28 tokens per second
+2.40.308.797 I slot print_timing: id  0 | task 194 | prompt processing, n_tokens =   2459, progress = 1.00, t =   6.50 s / 378.58 tokens per second
+2.40.439.852 I slot print_timing: id  0 | task 194 | prompt eval time =    6531.97 ms /  2463 tokens (    2.65 ms per token,   377.07 tokens per second)
+2.40.439.862 I slot print_timing: id  0 | task 194 |        eval time =      94.45 ms /     5 tokens (   23.61 ms per token,    42.35 tokens per second)
+2.40.439.864 I slot print_timing: id  0 | task 194 |       total time =    6626.43 ms /  2468 tokens
+2.40.439.865 I slot print_timing: id  0 | task 194 |    graphs reused =         67
+2.40.439.926 I slot      release: id  0 | task 194 | stop processing: n_tokens = 2467, truncated = 0
+2.40.461.612 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 0.400 (> 0.100 thold), f_keep = 0.005
+2.40.474.832 I slot launch_slot_: id  0 | task 203 | processing task, is_child = 0
+2.40.602.632 W srv          stop: cancel task, id_task = 203
+2.40.622.755 I slot      release: id  0 | task 203 | stop processing: n_tokens = 32, truncated = 0
+2.40.622.798 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 0.429 (> 0.100 thold), f_keep = 0.375
+2.40.624.131 I slot launch_slot_: id  0 | task 210 | processing task, is_child = 0
+2.40.755.107 I slot print_timing: id  0 | task 210 | prompt eval time =     110.42 ms /    28 tokens (    3.94 ms per token,   253.57 tokens per second)
+2.40.755.120 I slot print_timing: id  0 | task 210 |        eval time =      20.52 ms /     2 tokens (   20.52 ms per token,    48.73 tokens per second)
+2.40.755.122 I slot print_timing: id  0 | task 210 |       total time =     130.95 ms /    30 tokens
+2.40.755.123 I slot print_timing: id  0 | task 210 |    graphs reused =         68
+2.40.755.154 I slot      release: id  0 | task 210 | stop processing: n_tokens = 29, truncated = 0
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_cpu_server_inventory_profile20_nocache_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_cpu_server_inventory_profile20_nocache_20260923.json
new file mode 100644
index 000000000..a48337101
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_cpu_server_inventory_profile20_nocache_20260923.json
@@ -0,0 +1,441 @@
+{
+  "scope": "Standalone llama.cpp full-model text path; no Omni stage binding",
+  "platform": "Windows-11-10.0.26200-SP0",
+  "model_path": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+  "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4",
+  "server": {
+    "endpoint": "http://127.0.0.1:18174",
+    "build_info": "b11124-1a679828f",
+    "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+    "model_ftype": "Q4_K - Medium",
+    "total_slots": 1
+  },
+  "quality_checks": [
+    {
+      "case": "short",
+      "wall_s": 0.14644759998191148,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.544336999999359,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    }
+  ],
+  "warmups": [
+    {
+      "case": "inventory_120",
+      "wall_s": 6.589070199988782,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    }
+  ],
+  "samples": [
+    {
+      "case": "inventory_120",
+      "wall_s": 6.566914899973199,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.711130399955437,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.906680799904279,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.861513999989256,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.817918199929409,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.649508300004527,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.65189790003933,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.629713200032711,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.676726399920881,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.711790299974382,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.719679599977098,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.645102599984966,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.654706100001931,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.639681099914014,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.651021700003184,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.687377099995501,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.762809899984859,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.692078499938361,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.645896199974231,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 6.646268200012855,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    }
+  ],
+  "cancellation": {
+    "first_delta": "1",
+    "stream_closed_after_first_content": true,
+    "first_delta_wall_s": 0.14118409994989634,
+    "recovery": {
+      "case": "short",
+      "wall_s": 0.17412430001422763,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    }
+  },
+  "profile": {
+    "case": "inventory_120",
+    "cache_prompt": false,
+    "repeats": 20,
+    "concurrency": 1,
+    "p50_wall_s": 6.654706100001931,
+    "p95_wall_s": 6.861513999989256,
+    "min_wall_s": 6.566914899973199,
+    "max_wall_s": 6.906680799904279,
+    "cached_prompt_tokens": [
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0
+    ]
+  },
+  "status": "passed"
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_inventory_answer_20260923.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_inventory_answer_20260923.log
new file mode 100644
index 000000000..919af7a9c
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_inventory_answer_20260923.log
@@ -0,0 +1,34 @@
+0.02.289.996 E srv    send_error: task id = 0, error: request (2463 tokens) exceeds the available context size (2048 tokens), try increasing it
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b11124-1a679828f
+model      : C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Here is an inventory listing. Item 1: the code for warehouse district 1 is 1007. Item 2: the code for warehouse district 2 is 1014. Item 3: the code for warehouse district 3 is 1021. Item 4: the code for warehouse district 4 is 1028. Item 5: the code for warehouse district 5 is 1035. Item 6: the code for warehouse district 6 is 1042. Item 7: the code for warehouse district 7 is 1049. Item 8: the code for warehouse district 8 is 1056. Item 9: the code for warehouse district 9 is 1063. Item 10: th ... (truncated)
+Error: request (2463 tokens) exceeds the available context size (2048 tokens), try increasing it
+
+[ Prompt: 0.0 t/s | Generation: 0.0 t/s ]
+
+
+Exiting...
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_inventory_answer_ctx4096_20260923.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_inventory_answer_ctx4096_20260923.log
new file mode 100644
index 000000000..3dbd81d54
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_inventory_answer_ctx4096_20260923.log
@@ -0,0 +1,56 @@
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b11124-1a679828f
+model      : C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Here is an inventory listing. Item 1: the code for warehouse district 1 is 1007. Item 2: the code for warehouse district 2 is 1014. Item 3: the code for warehouse district 3 is 1021. Item 4: the code for warehouse district 4 is 1028. Item 5: the code for warehouse district 5 is 1035. Item 6: the code for warehouse district 6 is 1042. Item 7: the code for warehouse district 7 is 1049. Item 8: the code for warehouse district 8 is 1056. Item 9: the code for warehouse district 9 is 1063. Item 10: th ... (truncated)
+
+[Start thinking]
+
+We are given an inventory listing with item 1 to item 120, each with a code for warehouse district. The question asks: "What is the code for warehouse district 73?" We need to find the code corresponding to item 73.
+
+Looking at the listing:
+Item 1: code 1007 (district 1)
+Item 2: 1014 (district 2)
+...
+We need to find the code for district 73. The listing goes from item 1 to item 120. Item 73 is the 73rd item.
+
+Let's count the items to find the code for item 73.
+
+The listing is given as:
+Item 1: 1007
+Item 2: 1014
+Item 3: 1021
+Item 4: 1028
+Item 5: 1035
+Item 6: 1042
+Item 7: 1049
+Item 8: 1056
+Item 9: 1063
+Item 10:
+
+[ Prompt: 925.3 t/s | Generation: 54.2 t/s ]
+
+
+Exiting...
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_inventory_answer_no_reason_20260923.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_inventory_answer_no_reason_20260923.log
new file mode 100644
index 000000000..5b3a9c0cf
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_inventory_answer_no_reason_20260923.log
@@ -0,0 +1,33 @@
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b11124-1a679828f
+model      : C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Here is an inventory listing. Item 1: the code for warehouse district 1 is 1007. Item 2: the code for warehouse district 2 is 1014. Item 3: the code for warehouse district 3 is 1021. Item 4: the code for warehouse district 4 is 1028. Item 5: the code for warehouse district 5 is 1035. Item 6: the code for warehouse district 6 is 1042. Item 7: the code for warehouse district 7 is 1049. Item 8: the code for warehouse district 8 is 1056. Item 9: the code for warehouse district 9 is 1063. Item 10: th ... (truncated)
+1511
+
+[ Prompt: 1193.4 t/s | Generation: 52.1 t/s ]
+
+
+Exiting...
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_inventory_prompt_20260923.txt b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_inventory_prompt_20260923.txt
new file mode 100644
index 000000000..50c8a583a
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_inventory_prompt_20260923.txt
@@ -0,0 +1 @@
+Here is an inventory listing. Item 1: the code for warehouse district 1 is 1007. Item 2: the code for warehouse district 2 is 1014. Item 3: the code for warehouse district 3 is 1021. Item 4: the code for warehouse district 4 is 1028. Item 5: the code for warehouse district 5 is 1035. Item 6: the code for warehouse district 6 is 1042. Item 7: the code for warehouse district 7 is 1049. Item 8: the code for warehouse district 8 is 1056. Item 9: the code for warehouse district 9 is 1063. Item 10: the code for warehouse district 10 is 1070. Item 11: the code for warehouse district 11 is 1077. Item 12: the code for warehouse district 12 is 1084. Item 13: the code for warehouse district 13 is 1091. Item 14: the code for warehouse district 14 is 1098. Item 15: the code for warehouse district 15 is 1105. Item 16: the code for warehouse district 16 is 1112. Item 17: the code for warehouse district 17 is 1119. Item 18: the code for warehouse district 18 is 1126. Item 19: the code for warehouse district 19 is 1133. Item 20: the code for warehouse district 20 is 1140. Item 21: the code for warehouse district 21 is 1147. Item 22: the code for warehouse district 22 is 1154. Item 23: the code for warehouse district 23 is 1161. Item 24: the code for warehouse district 24 is 1168. Item 25: the code for warehouse district 25 is 1175. Item 26: the code for warehouse district 26 is 1182. Item 27: the code for warehouse district 27 is 1189. Item 28: the code for warehouse district 28 is 1196. Item 29: the code for warehouse district 29 is 1203. Item 30: the code for warehouse district 30 is 1210. Item 31: the code for warehouse district 31 is 1217. Item 32: the code for warehouse district 32 is 1224. Item 33: the code for warehouse district 33 is 1231. Item 34: the code for warehouse district 34 is 1238. Item 35: the code for warehouse district 35 is 1245. Item 36: the code for warehouse district 36 is 1252. Item 37: the code for warehouse district 37 is 1259. Item 38: the code for warehouse district 38 is 1266. Item 39: the code for warehouse district 39 is 1273. Item 40: the code for warehouse district 40 is 1280. Item 41: the code for warehouse district 41 is 1287. Item 42: the code for warehouse district 42 is 1294. Item 43: the code for warehouse district 43 is 1301. Item 44: the code for warehouse district 44 is 1308. Item 45: the code for warehouse district 45 is 1315. Item 46: the code for warehouse district 46 is 1322. Item 47: the code for warehouse district 47 is 1329. Item 48: the code for warehouse district 48 is 1336. Item 49: the code for warehouse district 49 is 1343. Item 50: the code for warehouse district 50 is 1350. Item 51: the code for warehouse district 51 is 1357. Item 52: the code for warehouse district 52 is 1364. Item 53: the code for warehouse district 53 is 1371. Item 54: the code for warehouse district 54 is 1378. Item 55: the code for warehouse district 55 is 1385. Item 56: the code for warehouse district 56 is 1392. Item 57: the code for warehouse district 57 is 1399. Item 58: the code for warehouse district 58 is 1406. Item 59: the code for warehouse district 59 is 1413. Item 60: the code for warehouse district 60 is 1420. Item 61: the code for warehouse district 61 is 1427. Item 62: the code for warehouse district 62 is 1434. Item 63: the code for warehouse district 63 is 1441. Item 64: the code for warehouse district 64 is 1448. Item 65: the code for warehouse district 65 is 1455. Item 66: the code for warehouse district 66 is 1462. Item 67: the code for warehouse district 67 is 1469. Item 68: the code for warehouse district 68 is 1476. Item 69: the code for warehouse district 69 is 1483. Item 70: the code for warehouse district 70 is 1490. Item 71: the code for warehouse district 71 is 1497. Item 72: the code for warehouse district 72 is 1504. Item 73: the code for warehouse district 73 is 1511. Item 74: the code for warehouse district 74 is 1518. Item 75: the code for warehouse district 75 is 1525. Item 76: the code for warehouse district 76 is 1532. Item 77: the code for warehouse district 77 is 1539. Item 78: the code for warehouse district 78 is 1546. Item 79: the code for warehouse district 79 is 1553. Item 80: the code for warehouse district 80 is 1560. Item 81: the code for warehouse district 81 is 1567. Item 82: the code for warehouse district 82 is 1574. Item 83: the code for warehouse district 83 is 1581. Item 84: the code for warehouse district 84 is 1588. Item 85: the code for warehouse district 85 is 1595. Item 86: the code for warehouse district 86 is 1602. Item 87: the code for warehouse district 87 is 1609. Item 88: the code for warehouse district 88 is 1616. Item 89: the code for warehouse district 89 is 1623. Item 90: the code for warehouse district 90 is 1630. Item 91: the code for warehouse district 91 is 1637. Item 92: the code for warehouse district 92 is 1644. Item 93: the code for warehouse district 93 is 1651. Item 94: the code for warehouse district 94 is 1658. Item 95: the code for warehouse district 95 is 1665. Item 96: the code for warehouse district 96 is 1672. Item 97: the code for warehouse district 97 is 1679. Item 98: the code for warehouse district 98 is 1686. Item 99: the code for warehouse district 99 is 1693. Item 100: the code for warehouse district 100 is 1700. Item 101: the code for warehouse district 101 is 1707. Item 102: the code for warehouse district 102 is 1714. Item 103: the code for warehouse district 103 is 1721. Item 104: the code for warehouse district 104 is 1728. Item 105: the code for warehouse district 105 is 1735. Item 106: the code for warehouse district 106 is 1742. Item 107: the code for warehouse district 107 is 1749. Item 108: the code for warehouse district 108 is 1756. Item 109: the code for warehouse district 109 is 1763. Item 110: the code for warehouse district 110 is 1770. Item 111: the code for warehouse district 111 is 1777. Item 112: the code for warehouse district 112 is 1784. Item 113: the code for warehouse district 113 is 1791. Item 114: the code for warehouse district 114 is 1798. Item 115: the code for warehouse district 115 is 1805. Item 116: the code for warehouse district 116 is 1812. Item 117: the code for warehouse district 117 is 1819. Item 118: the code for warehouse district 118 is 1826. Item 119: the code for warehouse district 119 is 1833. Item 120: the code for warehouse district 120 is 1840. What is the code for warehouse district 73? Answer with just the number.
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_answer_20260923.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_answer_20260923.log
new file mode 100644
index 000000000..1fa3ce58b
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_answer_20260923.log
@@ -0,0 +1,39 @@
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b11124-1a679828f
+model      : C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> What is the capital of France? Answer in one word.
+
+[Start thinking]
+
+We are asked: "What is the capital of France? Answer in one word." So the answer should be a single word: "Paris". But we need to output just the word? The instruction says "Answer in one word." So we should output "Paris". But let's make sure: The capital of France is Paris. So the answer is Paris.
+[End thinking]
+
+Paris
+
+[ Prompt: 267.3 t/s | Generation: 62.9 t/s ]
+
+
+Exiting...
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_bench_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_bench_20260923.json
new file mode 100644
index 000000000..d0f14468f
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_bench_20260923.json
@@ -0,0 +1,92 @@
+[
+  {
+    "build_commit": "1a679828f",
+    "build_number": 11124,
+    "cpu_info": "AMD Ryzen AI 9 HX 370 w/ Radeon 890M           ",
+    "gpu_info": "NVIDIA GeForce RTX 5090 Laptop GPU, AMD Radeon(TM) 890M Graphics",
+    "backends": "Vulkan",
+    "model_filename": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+    "model_type": "spark2_5 1.7B Q4_K - Medium",
+    "model_size": 1102073856,
+    "model_n_params": 1707657216,
+    "n_batch": 2048,
+    "n_ubatch": 512,
+    "n_threads": 12,
+    "cpu_mask": "0x0",
+    "cpu_strict": false,
+    "poll": 50,
+    "type_k": "f16",
+    "type_v": "f16",
+    "n_gpu_layers": 99,
+    "n_cpu_moe": 0,
+    "split_mode": "layer",
+    "main_gpu": 0,
+    "no_kv_offload": false,
+    "flash_attn": -1,
+    "devices": "Vulkan1",
+    "tensor_split": "0.00",
+    "tensor_buft_overrides": "none",
+    "load_mode": "auto",
+    "lazy_mode": "auto",
+    "embeddings": false,
+    "no_op_offload": 0,
+    "no_host": false,
+    "fit_target": 0,
+    "fit_min_ctx": 0,
+    "n_prompt": 128,
+    "n_gen": 0,
+    "n_depth": 0,
+    "test_time": "2026-09-23T10:40:08Z",
+    "avg_ns": 79903533,
+    "stddev_ns": 642773,
+    "avg_ts": 1602.000451,
+    "stddev_ts": 12.827994,
+    "samples_ns": [ 79585000, 80643300, 79482300 ],
+    "samples_ts": [ 1608.34, 1587.24, 1610.42 ]
+  },
+  {
+    "build_commit": "1a679828f",
+    "build_number": 11124,
+    "cpu_info": "AMD Ryzen AI 9 HX 370 w/ Radeon 890M           ",
+    "gpu_info": "NVIDIA GeForce RTX 5090 Laptop GPU, AMD Radeon(TM) 890M Graphics",
+    "backends": "Vulkan",
+    "model_filename": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+    "model_type": "spark2_5 1.7B Q4_K - Medium",
+    "model_size": 1102073856,
+    "model_n_params": 1707657216,
+    "n_batch": 2048,
+    "n_ubatch": 512,
+    "n_threads": 12,
+    "cpu_mask": "0x0",
+    "cpu_strict": false,
+    "poll": 50,
+    "type_k": "f16",
+    "type_v": "f16",
+    "n_gpu_layers": 99,
+    "n_cpu_moe": 0,
+    "split_mode": "layer",
+    "main_gpu": 0,
+    "no_kv_offload": false,
+    "flash_attn": -1,
+    "devices": "Vulkan1",
+    "tensor_split": "0.00",
+    "tensor_buft_overrides": "none",
+    "load_mode": "auto",
+    "lazy_mode": "auto",
+    "embeddings": false,
+    "no_op_offload": 0,
+    "no_host": false,
+    "fit_target": 0,
+    "fit_min_ctx": 0,
+    "n_prompt": 0,
+    "n_gen": 32,
+    "n_depth": 0,
+    "test_time": "2026-09-23T10:40:09Z",
+    "avg_ns": 475194133,
+    "stddev_ns": 2768818,
+    "avg_ts": 67.342423,
+    "stddev_ts": 0.392295,
+    "samples_ns": [ 472460200, 477996400, 475125800 ],
+    "samples_ts": [ 67.7306, 66.9461, 67.3506 ]
+  }
+]
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_bench_20260923.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_bench_20260923.log
new file mode 100644
index 000000000..ff7b2f636
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_bench_20260923.log
@@ -0,0 +1,6 @@
+load_backend: loaded RPC backend from C:\Users\zhout\w2\llama-vulkan-b11124\bin\ggml-rpc.dll
+ggml_vulkan: Found 2 Vulkan devices:
+ggml_vulkan: 0 = NVIDIA GeForce RTX 5090 Laptop GPU (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | fp4: 0 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
+ggml_vulkan: 1 = AMD Radeon(TM) 890M Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 1 | fp4: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+load_backend: loaded Vulkan backend from C:\Users\zhout\w2\llama-vulkan-b11124\bin\ggml-vulkan.dll
+load_backend: loaded CPU backend from C:\Users\zhout\w2\llama-vulkan-b11124\bin\ggml-cpu-zen4.dll
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_placement_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_placement_20260923.json
new file mode 100644
index 000000000..b2acbfd0c
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_placement_20260923.json
@@ -0,0 +1,92 @@
+[
+  {
+    "build_commit": "1a679828f",
+    "build_number": 11124,
+    "cpu_info": "AMD Ryzen AI 9 HX 370 w/ Radeon 890M           ",
+    "gpu_info": "NVIDIA GeForce RTX 5090 Laptop GPU, AMD Radeon(TM) 890M Graphics",
+    "backends": "Vulkan",
+    "model_filename": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+    "model_type": "spark2_5 1.7B Q4_K - Medium",
+    "model_size": 1102073856,
+    "model_n_params": 1707657216,
+    "n_batch": 2048,
+    "n_ubatch": 512,
+    "n_threads": 12,
+    "cpu_mask": "0x0",
+    "cpu_strict": false,
+    "poll": 50,
+    "type_k": "f16",
+    "type_v": "f16",
+    "n_gpu_layers": 99,
+    "n_cpu_moe": 0,
+    "split_mode": "layer",
+    "main_gpu": 0,
+    "no_kv_offload": false,
+    "flash_attn": -1,
+    "devices": "Vulkan1",
+    "tensor_split": "0.00",
+    "tensor_buft_overrides": "none",
+    "load_mode": "auto",
+    "lazy_mode": "auto",
+    "embeddings": false,
+    "no_op_offload": 0,
+    "no_host": false,
+    "fit_target": 0,
+    "fit_min_ctx": 0,
+    "n_prompt": 128,
+    "n_gen": 0,
+    "n_depth": 0,
+    "test_time": "2026-09-23T10:40:43Z",
+    "avg_ns": 87737000,
+    "stddev_ns": 0,
+    "avg_ts": 1458.905593,
+    "stddev_ts": 0.000000,
+    "samples_ns": [ 87737000 ],
+    "samples_ts": [ 1458.91 ]
+  },
+  {
+    "build_commit": "1a679828f",
+    "build_number": 11124,
+    "cpu_info": "AMD Ryzen AI 9 HX 370 w/ Radeon 890M           ",
+    "gpu_info": "NVIDIA GeForce RTX 5090 Laptop GPU, AMD Radeon(TM) 890M Graphics",
+    "backends": "Vulkan",
+    "model_filename": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+    "model_type": "spark2_5 1.7B Q4_K - Medium",
+    "model_size": 1102073856,
+    "model_n_params": 1707657216,
+    "n_batch": 2048,
+    "n_ubatch": 512,
+    "n_threads": 12,
+    "cpu_mask": "0x0",
+    "cpu_strict": false,
+    "poll": 50,
+    "type_k": "f16",
+    "type_v": "f16",
+    "n_gpu_layers": 99,
+    "n_cpu_moe": 0,
+    "split_mode": "layer",
+    "main_gpu": 0,
+    "no_kv_offload": false,
+    "flash_attn": -1,
+    "devices": "Vulkan1",
+    "tensor_split": "0.00",
+    "tensor_buft_overrides": "none",
+    "load_mode": "auto",
+    "lazy_mode": "auto",
+    "embeddings": false,
+    "no_op_offload": 0,
+    "no_host": false,
+    "fit_target": 0,
+    "fit_min_ctx": 0,
+    "n_prompt": 0,
+    "n_gen": 32,
+    "n_depth": 0,
+    "test_time": "2026-09-23T10:40:43Z",
+    "avg_ns": 507961300,
+    "stddev_ns": 0,
+    "avg_ts": 62.996925,
+    "stddev_ts": 0.000000,
+    "samples_ns": [ 507961300 ],
+    "samples_ts": [ 62.9969 ]
+  }
+]
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_placement_20260923.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_placement_20260923.log
new file mode 100644
index 000000000..b661fb847
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_placement_20260923.log
@@ -0,0 +1,3080 @@
+load_backend: loaded RPC backend from C:\Users\zhout\w2\llama-vulkan-b11124\bin\ggml-rpc.dll
+ggml_vulkan: Found 2 Vulkan devices:
+ggml_vulkan: 0 = NVIDIA GeForce RTX 5090 Laptop GPU (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | fp4: 0 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
+ggml_vulkan: 1 = AMD Radeon(TM) 890M Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 1 | fp4: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+load_backend: loaded Vulkan backend from C:\Users\zhout\w2\llama-vulkan-b11124\bin\ggml-vulkan.dll
+load_backend: loaded CPU backend from C:\Users\zhout\w2\llama-vulkan-b11124\bin\ggml-cpu-zen4.dll
+llama_model_loader: loaded meta data with 37 key-value pairs and 226 tensors from C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf (version GGUF V3 (latest))
+llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
+llama_model_loader: - kv   0:                       general.architecture str              = spark2_5
+llama_model_loader: - kv   1:                               general.type str              = model
+llama_model_loader: - kv   2:                     general.sampling.top_k i32              = 0
+llama_model_loader: - kv   3:                     general.sampling.top_p f32              = 0.950000
+llama_model_loader: - kv   4:                      general.sampling.temp f32              = 1.000000
+llama_model_loader: - kv   5:                               general.name str              = Hf_Format
+llama_model_loader: - kv   6:                         general.size_label str              = 1.7B
+llama_model_loader: - kv   7:                       spark2_5.block_count u32              = 28
+llama_model_loader: - kv   8:                    spark2_5.context_length u32              = 1048576
+llama_model_loader: - kv   9:                  spark2_5.embedding_length u32              = 2048
+llama_model_loader: - kv  10:               spark2_5.feed_forward_length u32              = 6656
+llama_model_loader: - kv  11:              spark2_5.attention.head_count u32              = 8
+llama_model_loader: - kv  12:           spark2_5.attention.head_count_kv u32              = 2
+llama_model_loader: - kv  13:                    spark2_5.rope.freq_base f32              = 5000000.000000
+llama_model_loader: - kv  14:                spark2_5.rope.freq_base_swa f32              = 10000.000000
+llama_model_loader: - kv  15:  spark2_5.attention.layer_norm_rms_epsilon f32              = 0.000001
+llama_model_loader: - kv  16:              spark2_5.attention.key_length u32              = 256
+llama_model_loader: - kv  17:            spark2_5.attention.value_length u32              = 256
+llama_model_loader: - kv  18:                        spark2_5.vocab_size u32              = 131072
+llama_model_loader: - kv  19:          spark2_5.attention.sliding_window u32              = 512
+llama_model_loader: - kv  20:  spark2_5.attention.sliding_window_pattern arr[bool,28]     = [true, true, true, false, true, true,...
+llama_model_loader: - kv  21:              spark2_5.rope.dimension_count u32              = 64
+llama_model_loader: - kv  22:          spark2_5.rope.dimension_count_swa u32              = 256
+llama_model_loader: - kv  23:                       tokenizer.ggml.model str              = gpt2
+llama_model_loader: - kv  24:                         tokenizer.ggml.pre str              = spark2_5
+llama_model_loader: - kv  25:                      tokenizer.ggml.tokens arr[str,131072]  = ["<｜start▁of▁sentence｜>", "<�...
+llama_model_loader: - kv  26:                  tokenizer.ggml.token_type arr[i32,131072]  = [3, 3, 3, 4, 4, 3, 3, 3, 4, 4, 4, 3, ...
+llama_model_loader: - kv  27:                      tokenizer.ggml.merges arr[str,130716]  = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "Ġ�...
+llama_model_loader: - kv  28:                tokenizer.ggml.bos_token_id u32              = 0
+llama_model_loader: - kv  29:                tokenizer.ggml.eos_token_id u32              = 1
+llama_model_loader: - kv  30:            tokenizer.ggml.unknown_token_id u32              = 5
+llama_model_loader: - kv  31:            tokenizer.ggml.padding_token_id u32              = 2
+llama_model_loader: - kv  32:               tokenizer.ggml.add_bos_token bool             = false
+llama_model_loader: - kv  33:               tokenizer.ggml.add_eos_token bool             = false
+llama_model_loader: - kv  34:                    tokenizer.chat_template str              = {%- if not messages %}\n    {{- raise_...
+llama_model_loader: - kv  35:               general.quantization_version u32              = 2
+llama_model_loader: - kv  36:                          general.file_type u32              = 15
+llama_model_loader: - type  f32:   57 tensors
+llama_model_loader: - type q4_K:  140 tensors
+llama_model_loader: - type q6_K:   29 tensors
+print_info: file format = GGUF V3 (latest)
+print_info: file type   = Q4_K - Medium
+print_info: file size   = 1.03 GiB (5.16 BPW) 
+llama_prepare_model_devices: using device Vulkan1 (AMD Radeon(TM) 890M Graphics) (unknown id) - 46080 MiB free
+init_tokenizer: initializing tokenizer for type 2
+load: 86 unused tokens
+load: control token:      0 '<｜start▁of▁sentence｜>' is not marked as EOG
+load: control token:      5 '<unk>' is not marked as EOG
+load: control token:     11 '<｜fim▁begin｜>' is not marked as EOG
+load: control token:      7 '<｜end▁of▁text｜>' is not marked as EOG
+load: control token:      2 '<｜▁pad▁｜>' is not marked as EOG
+load: control token:      6 '<｜start▁of▁text｜>' is not marked as EOG
+load: control token:     12 '<｜fim▁hole｜>' is not marked as EOG
+load: control token:     13 '<｜fim▁end｜>' is not marked as EOG
+load: control token: 130975 '<|Developer|>' is not marked as EOG
+load: control token: 130976 '<|Bot|>' is not marked as EOG
+load: control token: 130972 '<|System|>' is not marked as EOG
+load: control token: 130973 '<|User|>' is not marked as EOG
+load: control token: 130974 '<|Tool|>' is not marked as EOG
+load: printing all EOG tokens:
+load:   - 1 ('<｜end▁of▁sentence｜>')
+load: special tokens cache size = 113
+load: token to piece cache size = 0.7975 MB
+print_info: arch                  = spark2_5
+print_info: vocab_only            = 0
+print_info: no_alloc              = 0
+print_info: n_ctx_train           = 1048576
+print_info: n_embd_inp            = 2048
+print_info: n_embd                = 2048
+print_info: n_embd_out            = 2048
+print_info: n_layer               = 28
+print_info: n_layer_all           = 28
+print_info: n_head                = 8
+print_info: n_head_kv             = 2
+print_info: n_rot                 = 64
+print_info: n_swa                 = 512
+print_info: is_swa_any            = 1
+print_info: non_causal_type       = 0
+print_info: n_embd_head_k         = 256
+print_info: n_embd_head_v         = 256
+print_info: n_gqa                 = 4
+print_info: n_embd_k_gqa          = 512
+print_info: n_embd_v_gqa          = 512
+print_info: f_norm_eps            = 0.0e+00
+print_info: f_norm_rms_eps        = 1.0e-06
+print_info: f_clamp_kqv           = 0.0e+00
+print_info: f_max_alibi_bias      = 0.0e+00
+print_info: f_logit_scale         = 0.0e+00
+print_info: f_attn_scale          = 0.0e+00
+print_info: f_attn_value_scale    = 0.0000
+print_info: n_ff                  = 6656
+print_info: n_expert              = 0
+print_info: n_expert_used         = 0
+print_info: n_expert_groups       = 0
+print_info: n_group_used          = 0
+print_info: causal attn           = 1
+print_info: pooling type          = -1
+print_info: rope type             = 2
+print_info: rope scaling          = linear
+print_info: freq_base_train       = 5000000.0
+print_info: freq_scale_train      = 1
+print_info: freq_base_swa         = 10000.0
+print_info: freq_scale_swa        = 1
+print_info: n_embd_head_k_swa     = 256
+print_info: n_embd_head_v_swa     = 256
+print_info: n_rot_swa             = 256
+print_info: n_ctx_orig_yarn       = 1048576
+print_info: rope_yarn_log_mul     = 0.0000
+print_info: rope_finetuned        = unknown
+print_info: model type            = 1.7B
+print_info: model params          = 1.71 B
+print_info: general.name          = Hf_Format
+print_info: vocab type            = BPE
+print_info: n_vocab               = 131072
+print_info: n_merges              = 130716
+print_info: BOS token             = 0 '<｜start▁of▁sentence｜>'
+print_info: EOS token             = 1 '<｜end▁of▁sentence｜>'
+print_info: EOT token             = 1 '<｜end▁of▁sentence｜>'
+print_info: UNK token             = 5 '<unk>'
+print_info: PAD token             = 2 '<｜▁pad▁｜>'
+print_info: LF token              = 198 'Ċ'
+print_info: FIM PRE token         = 11 '<｜fim▁begin｜>'
+print_info: FIM SUF token         = 12 '<｜fim▁hole｜>'
+print_info: FIM MID token         = 13 '<｜fim▁end｜>'
+print_info: EOG token             = 1 '<｜end▁of▁sentence｜>'
+print_info: max token length      = 210
+load_tensors: loading model tensors, this can take a while... (load_mode = none)
+load_tensors: layer   0 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer   1 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer   2 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer   3 assigned to device Vulkan1, is_swa = 0
+load_tensors: layer   4 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer   5 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer   6 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer   7 assigned to device Vulkan1, is_swa = 0
+load_tensors: layer   8 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer   9 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  10 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  11 assigned to device Vulkan1, is_swa = 0
+load_tensors: layer  12 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  13 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  14 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  15 assigned to device Vulkan1, is_swa = 0
+load_tensors: layer  16 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  17 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  18 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  19 assigned to device Vulkan1, is_swa = 0
+load_tensors: layer  20 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  21 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  22 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  23 assigned to device Vulkan1, is_swa = 0
+load_tensors: layer  24 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  25 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  26 assigned to device Vulkan1, is_swa = 1
+load_tensors: layer  27 assigned to device Vulkan1, is_swa = 0
+load_tensors: layer  28 assigned to device Vulkan1, is_swa = 0
+create_tensor: loading tensor token_embd.weight
+create_tensor: loading tensor output_norm.weight
+create_tensor: loading tensor output.weight
+create_tensor: loading tensor token_embd.weight
+create_tensor: loading tensor blk.0.attn_norm.weight
+create_tensor: loading tensor blk.0.attn_qkv.weight
+create_tensor: loading tensor blk.0.attn_qkv.bias
+create_tensor: loading tensor blk.0.attn_q.bias
+create_tensor: loading tensor blk.0.attn_k.bias
+create_tensor: loading tensor blk.0.attn_v.bias
+create_tensor: loading tensor blk.0.attn_gate.weight
+create_tensor: loading tensor blk.0.attn_output.weight
+create_tensor: loading tensor blk.0.ffn_norm.weight
+create_tensor: loading tensor blk.0.ffn_gate.weight
+create_tensor: loading tensor blk.0.ffn_up.weight
+create_tensor: loading tensor blk.0.ffn_down.weight
+create_tensor: loading tensor blk.1.attn_norm.weight
+create_tensor: loading tensor blk.1.attn_qkv.weight
+create_tensor: loading tensor blk.1.attn_qkv.bias
+create_tensor: loading tensor blk.1.attn_q.bias
+create_tensor: loading tensor blk.1.attn_k.bias
+create_tensor: loading tensor blk.1.attn_v.bias
+create_tensor: loading tensor blk.1.attn_gate.weight
+create_tensor: loading tensor blk.1.attn_output.weight
+create_tensor: loading tensor blk.1.ffn_norm.weight
+create_tensor: loading tensor blk.1.ffn_gate.weight
+create_tensor: loading tensor blk.1.ffn_up.weight
+create_tensor: loading tensor blk.1.ffn_down.weight
+create_tensor: loading tensor blk.2.attn_norm.weight
+create_tensor: loading tensor blk.2.attn_qkv.weight
+create_tensor: loading tensor blk.2.attn_qkv.bias
+create_tensor: loading tensor blk.2.attn_q.bias
+create_tensor: loading tensor blk.2.attn_k.bias
+create_tensor: loading tensor blk.2.attn_v.bias
+create_tensor: loading tensor blk.2.attn_gate.weight
+create_tensor: loading tensor blk.2.attn_output.weight
+create_tensor: loading tensor blk.2.ffn_norm.weight
+create_tensor: loading tensor blk.2.ffn_gate.weight
+create_tensor: loading tensor blk.2.ffn_up.weight
+create_tensor: loading tensor blk.2.ffn_down.weight
+create_tensor: loading tensor blk.3.attn_norm.weight
+create_tensor: loading tensor blk.3.attn_qkv.weight
+create_tensor: loading tensor blk.3.attn_qkv.bias
+create_tensor: loading tensor blk.3.attn_q.bias
+create_tensor: loading tensor blk.3.attn_k.bias
+create_tensor: loading tensor blk.3.attn_v.bias
+create_tensor: loading tensor blk.3.attn_gate.weight
+create_tensor: loading tensor blk.3.attn_output.weight
+create_tensor: loading tensor blk.3.ffn_norm.weight
+create_tensor: loading tensor blk.3.ffn_gate.weight
+create_tensor: loading tensor blk.3.ffn_up.weight
+create_tensor: loading tensor blk.3.ffn_down.weight
+create_tensor: loading tensor blk.4.attn_norm.weight
+create_tensor: loading tensor blk.4.attn_qkv.weight
+create_tensor: loading tensor blk.4.attn_qkv.bias
+create_tensor: loading tensor blk.4.attn_q.bias
+create_tensor: loading tensor blk.4.attn_k.bias
+create_tensor: loading tensor blk.4.attn_v.bias
+create_tensor: loading tensor blk.4.attn_gate.weight
+create_tensor: loading tensor blk.4.attn_output.weight
+create_tensor: loading tensor blk.4.ffn_norm.weight
+create_tensor: loading tensor blk.4.ffn_gate.weight
+create_tensor: loading tensor blk.4.ffn_up.weight
+create_tensor: loading tensor blk.4.ffn_down.weight
+create_tensor: loading tensor blk.5.attn_norm.weight
+create_tensor: loading tensor blk.5.attn_qkv.weight
+create_tensor: loading tensor blk.5.attn_qkv.bias
+create_tensor: loading tensor blk.5.attn_q.bias
+create_tensor: loading tensor blk.5.attn_k.bias
+create_tensor: loading tensor blk.5.attn_v.bias
+create_tensor: loading tensor blk.5.attn_gate.weight
+create_tensor: loading tensor blk.5.attn_output.weight
+create_tensor: loading tensor blk.5.ffn_norm.weight
+create_tensor: loading tensor blk.5.ffn_gate.weight
+create_tensor: loading tensor blk.5.ffn_up.weight
+create_tensor: loading tensor blk.5.ffn_down.weight
+create_tensor: loading tensor blk.6.attn_norm.weight
+create_tensor: loading tensor blk.6.attn_qkv.weight
+create_tensor: loading tensor blk.6.attn_qkv.bias
+create_tensor: loading tensor blk.6.attn_q.bias
+create_tensor: loading tensor blk.6.attn_k.bias
+create_tensor: loading tensor blk.6.attn_v.bias
+create_tensor: loading tensor blk.6.attn_gate.weight
+create_tensor: loading tensor blk.6.attn_output.weight
+create_tensor: loading tensor blk.6.ffn_norm.weight
+create_tensor: loading tensor blk.6.ffn_gate.weight
+create_tensor: loading tensor blk.6.ffn_up.weight
+create_tensor: loading tensor blk.6.ffn_down.weight
+create_tensor: loading tensor blk.7.attn_norm.weight
+create_tensor: loading tensor blk.7.attn_qkv.weight
+create_tensor: loading tensor blk.7.attn_qkv.bias
+create_tensor: loading tensor blk.7.attn_q.bias
+create_tensor: loading tensor blk.7.attn_k.bias
+create_tensor: loading tensor blk.7.attn_v.bias
+create_tensor: loading tensor blk.7.attn_gate.weight
+create_tensor: loading tensor blk.7.attn_output.weight
+create_tensor: loading tensor blk.7.ffn_norm.weight
+create_tensor: loading tensor blk.7.ffn_gate.weight
+create_tensor: loading tensor blk.7.ffn_up.weight
+create_tensor: loading tensor blk.7.ffn_down.weight
+create_tensor: loading tensor blk.8.attn_norm.weight
+create_tensor: loading tensor blk.8.attn_qkv.weight
+create_tensor: loading tensor blk.8.attn_qkv.bias
+create_tensor: loading tensor blk.8.attn_q.bias
+create_tensor: loading tensor blk.8.attn_k.bias
+create_tensor: loading tensor blk.8.attn_v.bias
+create_tensor: loading tensor blk.8.attn_gate.weight
+create_tensor: loading tensor blk.8.attn_output.weight
+create_tensor: loading tensor blk.8.ffn_norm.weight
+create_tensor: loading tensor blk.8.ffn_gate.weight
+create_tensor: loading tensor blk.8.ffn_up.weight
+create_tensor: loading tensor blk.8.ffn_down.weight
+create_tensor: loading tensor blk.9.attn_norm.weight
+create_tensor: loading tensor blk.9.attn_qkv.weight
+create_tensor: loading tensor blk.9.attn_qkv.bias
+create_tensor: loading tensor blk.9.attn_q.bias
+create_tensor: loading tensor blk.9.attn_k.bias
+create_tensor: loading tensor blk.9.attn_v.bias
+create_tensor: loading tensor blk.9.attn_gate.weight
+create_tensor: loading tensor blk.9.attn_output.weight
+create_tensor: loading tensor blk.9.ffn_norm.weight
+create_tensor: loading tensor blk.9.ffn_gate.weight
+create_tensor: loading tensor blk.9.ffn_up.weight
+create_tensor: loading tensor blk.9.ffn_down.weight
+create_tensor: loading tensor blk.10.attn_norm.weight
+create_tensor: loading tensor blk.10.attn_qkv.weight
+create_tensor: loading tensor blk.10.attn_qkv.bias
+create_tensor: loading tensor blk.10.attn_q.bias
+create_tensor: loading tensor blk.10.attn_k.bias
+create_tensor: loading tensor blk.10.attn_v.bias
+create_tensor: loading tensor blk.10.attn_gate.weight
+create_tensor: loading tensor blk.10.attn_output.weight
+create_tensor: loading tensor blk.10.ffn_norm.weight
+create_tensor: loading tensor blk.10.ffn_gate.weight
+create_tensor: loading tensor blk.10.ffn_up.weight
+create_tensor: loading tensor blk.10.ffn_down.weight
+create_tensor: loading tensor blk.11.attn_norm.weight
+create_tensor: loading tensor blk.11.attn_qkv.weight
+create_tensor: loading tensor blk.11.attn_qkv.bias
+create_tensor: loading tensor blk.11.attn_q.bias
+create_tensor: loading tensor blk.11.attn_k.bias
+create_tensor: loading tensor blk.11.attn_v.bias
+create_tensor: loading tensor blk.11.attn_gate.weight
+create_tensor: loading tensor blk.11.attn_output.weight
+create_tensor: loading tensor blk.11.ffn_norm.weight
+create_tensor: loading tensor blk.11.ffn_gate.weight
+create_tensor: loading tensor blk.11.ffn_up.weight
+create_tensor: loading tensor blk.11.ffn_down.weight
+create_tensor: loading tensor blk.12.attn_norm.weight
+create_tensor: loading tensor blk.12.attn_qkv.weight
+create_tensor: loading tensor blk.12.attn_qkv.bias
+create_tensor: loading tensor blk.12.attn_q.bias
+create_tensor: loading tensor blk.12.attn_k.bias
+create_tensor: loading tensor blk.12.attn_v.bias
+create_tensor: loading tensor blk.12.attn_gate.weight
+create_tensor: loading tensor blk.12.attn_output.weight
+create_tensor: loading tensor blk.12.ffn_norm.weight
+create_tensor: loading tensor blk.12.ffn_gate.weight
+create_tensor: loading tensor blk.12.ffn_up.weight
+create_tensor: loading tensor blk.12.ffn_down.weight
+create_tensor: loading tensor blk.13.attn_norm.weight
+create_tensor: loading tensor blk.13.attn_qkv.weight
+create_tensor: loading tensor blk.13.attn_qkv.bias
+create_tensor: loading tensor blk.13.attn_q.bias
+create_tensor: loading tensor blk.13.attn_k.bias
+create_tensor: loading tensor blk.13.attn_v.bias
+create_tensor: loading tensor blk.13.attn_gate.weight
+create_tensor: loading tensor blk.13.attn_output.weight
+create_tensor: loading tensor blk.13.ffn_norm.weight
+create_tensor: loading tensor blk.13.ffn_gate.weight
+create_tensor: loading tensor blk.13.ffn_up.weight
+create_tensor: loading tensor blk.13.ffn_down.weight
+create_tensor: loading tensor blk.14.attn_norm.weight
+create_tensor: loading tensor blk.14.attn_qkv.weight
+create_tensor: loading tensor blk.14.attn_qkv.bias
+create_tensor: loading tensor blk.14.attn_q.bias
+create_tensor: loading tensor blk.14.attn_k.bias
+create_tensor: loading tensor blk.14.attn_v.bias
+create_tensor: loading tensor blk.14.attn_gate.weight
+create_tensor: loading tensor blk.14.attn_output.weight
+create_tensor: loading tensor blk.14.ffn_norm.weight
+create_tensor: loading tensor blk.14.ffn_gate.weight
+create_tensor: loading tensor blk.14.ffn_up.weight
+create_tensor: loading tensor blk.14.ffn_down.weight
+create_tensor: loading tensor blk.15.attn_norm.weight
+create_tensor: loading tensor blk.15.attn_qkv.weight
+create_tensor: loading tensor blk.15.attn_qkv.bias
+create_tensor: loading tensor blk.15.attn_q.bias
+create_tensor: loading tensor blk.15.attn_k.bias
+create_tensor: loading tensor blk.15.attn_v.bias
+create_tensor: loading tensor blk.15.attn_gate.weight
+create_tensor: loading tensor blk.15.attn_output.weight
+create_tensor: loading tensor blk.15.ffn_norm.weight
+create_tensor: loading tensor blk.15.ffn_gate.weight
+create_tensor: loading tensor blk.15.ffn_up.weight
+create_tensor: loading tensor blk.15.ffn_down.weight
+create_tensor: loading tensor blk.16.attn_norm.weight
+create_tensor: loading tensor blk.16.attn_qkv.weight
+create_tensor: loading tensor blk.16.attn_qkv.bias
+create_tensor: loading tensor blk.16.attn_q.bias
+create_tensor: loading tensor blk.16.attn_k.bias
+create_tensor: loading tensor blk.16.attn_v.bias
+create_tensor: loading tensor blk.16.attn_gate.weight
+create_tensor: loading tensor blk.16.attn_output.weight
+create_tensor: loading tensor blk.16.ffn_norm.weight
+create_tensor: loading tensor blk.16.ffn_gate.weight
+create_tensor: loading tensor blk.16.ffn_up.weight
+create_tensor: loading tensor blk.16.ffn_down.weight
+create_tensor: loading tensor blk.17.attn_norm.weight
+create_tensor: loading tensor blk.17.attn_qkv.weight
+create_tensor: loading tensor blk.17.attn_qkv.bias
+create_tensor: loading tensor blk.17.attn_q.bias
+create_tensor: loading tensor blk.17.attn_k.bias
+create_tensor: loading tensor blk.17.attn_v.bias
+create_tensor: loading tensor blk.17.attn_gate.weight
+create_tensor: loading tensor blk.17.attn_output.weight
+create_tensor: loading tensor blk.17.ffn_norm.weight
+create_tensor: loading tensor blk.17.ffn_gate.weight
+create_tensor: loading tensor blk.17.ffn_up.weight
+create_tensor: loading tensor blk.17.ffn_down.weight
+create_tensor: loading tensor blk.18.attn_norm.weight
+create_tensor: loading tensor blk.18.attn_qkv.weight
+create_tensor: loading tensor blk.18.attn_qkv.bias
+create_tensor: loading tensor blk.18.attn_q.bias
+create_tensor: loading tensor blk.18.attn_k.bias
+create_tensor: loading tensor blk.18.attn_v.bias
+create_tensor: loading tensor blk.18.attn_gate.weight
+create_tensor: loading tensor blk.18.attn_output.weight
+create_tensor: loading tensor blk.18.ffn_norm.weight
+create_tensor: loading tensor blk.18.ffn_gate.weight
+create_tensor: loading tensor blk.18.ffn_up.weight
+create_tensor: loading tensor blk.18.ffn_down.weight
+create_tensor: loading tensor blk.19.attn_norm.weight
+create_tensor: loading tensor blk.19.attn_qkv.weight
+create_tensor: loading tensor blk.19.attn_qkv.bias
+create_tensor: loading tensor blk.19.attn_q.bias
+create_tensor: loading tensor blk.19.attn_k.bias
+create_tensor: loading tensor blk.19.attn_v.bias
+create_tensor: loading tensor blk.19.attn_gate.weight
+create_tensor: loading tensor blk.19.attn_output.weight
+create_tensor: loading tensor blk.19.ffn_norm.weight
+create_tensor: loading tensor blk.19.ffn_gate.weight
+create_tensor: loading tensor blk.19.ffn_up.weight
+create_tensor: loading tensor blk.19.ffn_down.weight
+create_tensor: loading tensor blk.20.attn_norm.weight
+create_tensor: loading tensor blk.20.attn_qkv.weight
+create_tensor: loading tensor blk.20.attn_qkv.bias
+create_tensor: loading tensor blk.20.attn_q.bias
+create_tensor: loading tensor blk.20.attn_k.bias
+create_tensor: loading tensor blk.20.attn_v.bias
+create_tensor: loading tensor blk.20.attn_gate.weight
+create_tensor: loading tensor blk.20.attn_output.weight
+create_tensor: loading tensor blk.20.ffn_norm.weight
+create_tensor: loading tensor blk.20.ffn_gate.weight
+create_tensor: loading tensor blk.20.ffn_up.weight
+create_tensor: loading tensor blk.20.ffn_down.weight
+create_tensor: loading tensor blk.21.attn_norm.weight
+create_tensor: loading tensor blk.21.attn_qkv.weight
+create_tensor: loading tensor blk.21.attn_qkv.bias
+create_tensor: loading tensor blk.21.attn_q.bias
+create_tensor: loading tensor blk.21.attn_k.bias
+create_tensor: loading tensor blk.21.attn_v.bias
+create_tensor: loading tensor blk.21.attn_gate.weight
+create_tensor: loading tensor blk.21.attn_output.weight
+create_tensor: loading tensor blk.21.ffn_norm.weight
+create_tensor: loading tensor blk.21.ffn_gate.weight
+create_tensor: loading tensor blk.21.ffn_up.weight
+create_tensor: loading tensor blk.21.ffn_down.weight
+create_tensor: loading tensor blk.22.attn_norm.weight
+create_tensor: loading tensor blk.22.attn_qkv.weight
+create_tensor: loading tensor blk.22.attn_qkv.bias
+create_tensor: loading tensor blk.22.attn_q.bias
+create_tensor: loading tensor blk.22.attn_k.bias
+create_tensor: loading tensor blk.22.attn_v.bias
+create_tensor: loading tensor blk.22.attn_gate.weight
+create_tensor: loading tensor blk.22.attn_output.weight
+create_tensor: loading tensor blk.22.ffn_norm.weight
+create_tensor: loading tensor blk.22.ffn_gate.weight
+create_tensor: loading tensor blk.22.ffn_up.weight
+create_tensor: loading tensor blk.22.ffn_down.weight
+create_tensor: loading tensor blk.23.attn_norm.weight
+create_tensor: loading tensor blk.23.attn_qkv.weight
+create_tensor: loading tensor blk.23.attn_qkv.bias
+create_tensor: loading tensor blk.23.attn_q.bias
+create_tensor: loading tensor blk.23.attn_k.bias
+create_tensor: loading tensor blk.23.attn_v.bias
+create_tensor: loading tensor blk.23.attn_gate.weight
+create_tensor: loading tensor blk.23.attn_output.weight
+create_tensor: loading tensor blk.23.ffn_norm.weight
+create_tensor: loading tensor blk.23.ffn_gate.weight
+create_tensor: loading tensor blk.23.ffn_up.weight
+create_tensor: loading tensor blk.23.ffn_down.weight
+create_tensor: loading tensor blk.24.attn_norm.weight
+create_tensor: loading tensor blk.24.attn_qkv.weight
+create_tensor: loading tensor blk.24.attn_qkv.bias
+create_tensor: loading tensor blk.24.attn_q.bias
+create_tensor: loading tensor blk.24.attn_k.bias
+create_tensor: loading tensor blk.24.attn_v.bias
+create_tensor: loading tensor blk.24.attn_gate.weight
+create_tensor: loading tensor blk.24.attn_output.weight
+create_tensor: loading tensor blk.24.ffn_norm.weight
+create_tensor: loading tensor blk.24.ffn_gate.weight
+create_tensor: loading tensor blk.24.ffn_up.weight
+create_tensor: loading tensor blk.24.ffn_down.weight
+create_tensor: loading tensor blk.25.attn_norm.weight
+create_tensor: loading tensor blk.25.attn_qkv.weight
+create_tensor: loading tensor blk.25.attn_qkv.bias
+create_tensor: loading tensor blk.25.attn_q.bias
+create_tensor: loading tensor blk.25.attn_k.bias
+create_tensor: loading tensor blk.25.attn_v.bias
+create_tensor: loading tensor blk.25.attn_gate.weight
+create_tensor: loading tensor blk.25.attn_output.weight
+create_tensor: loading tensor blk.25.ffn_norm.weight
+create_tensor: loading tensor blk.25.ffn_gate.weight
+create_tensor: loading tensor blk.25.ffn_up.weight
+create_tensor: loading tensor blk.25.ffn_down.weight
+create_tensor: loading tensor blk.26.attn_norm.weight
+create_tensor: loading tensor blk.26.attn_qkv.weight
+create_tensor: loading tensor blk.26.attn_qkv.bias
+create_tensor: loading tensor blk.26.attn_q.bias
+create_tensor: loading tensor blk.26.attn_k.bias
+create_tensor: loading tensor blk.26.attn_v.bias
+create_tensor: loading tensor blk.26.attn_gate.weight
+create_tensor: loading tensor blk.26.attn_output.weight
+create_tensor: loading tensor blk.26.ffn_norm.weight
+create_tensor: loading tensor blk.26.ffn_gate.weight
+create_tensor: loading tensor blk.26.ffn_up.weight
+create_tensor: loading tensor blk.26.ffn_down.weight
+create_tensor: loading tensor blk.27.attn_norm.weight
+create_tensor: loading tensor blk.27.attn_qkv.weight
+create_tensor: loading tensor blk.27.attn_qkv.bias
+create_tensor: loading tensor blk.27.attn_q.bias
+create_tensor: loading tensor blk.27.attn_k.bias
+create_tensor: loading tensor blk.27.attn_v.bias
+create_tensor: loading tensor blk.27.attn_gate.weight
+create_tensor: loading tensor blk.27.attn_output.weight
+create_tensor: loading tensor blk.27.ffn_norm.weight
+create_tensor: loading tensor blk.27.ffn_gate.weight
+create_tensor: loading tensor blk.27.ffn_up.weight
+create_tensor: loading tensor blk.27.ffn_down.weight
+create_tensor: loading tensor blk.0.attn_output.scale
+create_tensor: loading tensor blk.0.attn_qkv.scale
+create_tensor: loading tensor blk.0.attn_gate.scale
+create_tensor: loading tensor blk.0.ffn_gate.scale
+create_tensor: loading tensor blk.0.ffn_down.scale
+create_tensor: loading tensor blk.0.ffn_up.scale
+create_tensor: loading tensor blk.0.attn_output.input_scale
+create_tensor: loading tensor blk.0.attn_qkv.input_scale
+create_tensor: loading tensor blk.0.attn_gate.input_scale
+create_tensor: loading tensor blk.0.ffn_gate.input_scale
+create_tensor: loading tensor blk.0.ffn_down.input_scale
+create_tensor: loading tensor blk.0.ffn_up.input_scale
+create_tensor: loading tensor blk.1.attn_output.scale
+create_tensor: loading tensor blk.1.attn_qkv.scale
+create_tensor: loading tensor blk.1.attn_gate.scale
+create_tensor: loading tensor blk.1.ffn_gate.scale
+create_tensor: loading tensor blk.1.ffn_down.scale
+create_tensor: loading tensor blk.1.ffn_up.scale
+create_tensor: loading tensor blk.1.attn_output.input_scale
+create_tensor: loading tensor blk.1.attn_qkv.input_scale
+create_tensor: loading tensor blk.1.attn_gate.input_scale
+create_tensor: loading tensor blk.1.ffn_gate.input_scale
+create_tensor: loading tensor blk.1.ffn_down.input_scale
+create_tensor: loading tensor blk.1.ffn_up.input_scale
+create_tensor: loading tensor blk.2.attn_output.scale
+create_tensor: loading tensor blk.2.attn_qkv.scale
+create_tensor: loading tensor blk.2.attn_gate.scale
+create_tensor: loading tensor blk.2.ffn_gate.scale
+create_tensor: loading tensor blk.2.ffn_down.scale
+create_tensor: loading tensor blk.2.ffn_up.scale
+create_tensor: loading tensor blk.2.attn_output.input_scale
+create_tensor: loading tensor blk.2.attn_qkv.input_scale
+create_tensor: loading tensor blk.2.attn_gate.input_scale
+create_tensor: loading tensor blk.2.ffn_gate.input_scale
+create_tensor: loading tensor blk.2.ffn_down.input_scale
+create_tensor: loading tensor blk.2.ffn_up.input_scale
+create_tensor: loading tensor blk.3.attn_output.scale
+create_tensor: loading tensor blk.3.attn_qkv.scale
+create_tensor: loading tensor blk.3.attn_gate.scale
+create_tensor: loading tensor blk.3.ffn_gate.scale
+create_tensor: loading tensor blk.3.ffn_down.scale
+create_tensor: loading tensor blk.3.ffn_up.scale
+create_tensor: loading tensor blk.3.attn_output.input_scale
+create_tensor: loading tensor blk.3.attn_qkv.input_scale
+create_tensor: loading tensor blk.3.attn_gate.input_scale
+create_tensor: loading tensor blk.3.ffn_gate.input_scale
+create_tensor: loading tensor blk.3.ffn_down.input_scale
+create_tensor: loading tensor blk.3.ffn_up.input_scale
+create_tensor: loading tensor blk.4.attn_output.scale
+create_tensor: loading tensor blk.4.attn_qkv.scale
+create_tensor: loading tensor blk.4.attn_gate.scale
+create_tensor: loading tensor blk.4.ffn_gate.scale
+create_tensor: loading tensor blk.4.ffn_down.scale
+create_tensor: loading tensor blk.4.ffn_up.scale
+create_tensor: loading tensor blk.4.attn_output.input_scale
+create_tensor: loading tensor blk.4.attn_qkv.input_scale
+create_tensor: loading tensor blk.4.attn_gate.input_scale
+create_tensor: loading tensor blk.4.ffn_gate.input_scale
+create_tensor: loading tensor blk.4.ffn_down.input_scale
+create_tensor: loading tensor blk.4.ffn_up.input_scale
+create_tensor: loading tensor blk.5.attn_output.scale
+create_tensor: loading tensor blk.5.attn_qkv.scale
+create_tensor: loading tensor blk.5.attn_gate.scale
+create_tensor: loading tensor blk.5.ffn_gate.scale
+create_tensor: loading tensor blk.5.ffn_down.scale
+create_tensor: loading tensor blk.5.ffn_up.scale
+create_tensor: loading tensor blk.5.attn_output.input_scale
+create_tensor: loading tensor blk.5.attn_qkv.input_scale
+create_tensor: loading tensor blk.5.attn_gate.input_scale
+create_tensor: loading tensor blk.5.ffn_gate.input_scale
+create_tensor: loading tensor blk.5.ffn_down.input_scale
+create_tensor: loading tensor blk.5.ffn_up.input_scale
+create_tensor: loading tensor blk.6.attn_output.scale
+create_tensor: loading tensor blk.6.attn_qkv.scale
+create_tensor: loading tensor blk.6.attn_gate.scale
+create_tensor: loading tensor blk.6.ffn_gate.scale
+create_tensor: loading tensor blk.6.ffn_down.scale
+create_tensor: loading tensor blk.6.ffn_up.scale
+create_tensor: loading tensor blk.6.attn_output.input_scale
+create_tensor: loading tensor blk.6.attn_qkv.input_scale
+create_tensor: loading tensor blk.6.attn_gate.input_scale
+create_tensor: loading tensor blk.6.ffn_gate.input_scale
+create_tensor: loading tensor blk.6.ffn_down.input_scale
+create_tensor: loading tensor blk.6.ffn_up.input_scale
+create_tensor: loading tensor blk.7.attn_output.scale
+create_tensor: loading tensor blk.7.attn_qkv.scale
+create_tensor: loading tensor blk.7.attn_gate.scale
+create_tensor: loading tensor blk.7.ffn_gate.scale
+create_tensor: loading tensor blk.7.ffn_down.scale
+create_tensor: loading tensor blk.7.ffn_up.scale
+create_tensor: loading tensor blk.7.attn_output.input_scale
+create_tensor: loading tensor blk.7.attn_qkv.input_scale
+create_tensor: loading tensor blk.7.attn_gate.input_scale
+create_tensor: loading tensor blk.7.ffn_gate.input_scale
+create_tensor: loading tensor blk.7.ffn_down.input_scale
+create_tensor: loading tensor blk.7.ffn_up.input_scale
+create_tensor: loading tensor blk.8.attn_output.scale
+create_tensor: loading tensor blk.8.attn_qkv.scale
+create_tensor: loading tensor blk.8.attn_gate.scale
+create_tensor: loading tensor blk.8.ffn_gate.scale
+create_tensor: loading tensor blk.8.ffn_down.scale
+create_tensor: loading tensor blk.8.ffn_up.scale
+create_tensor: loading tensor blk.8.attn_output.input_scale
+create_tensor: loading tensor blk.8.attn_qkv.input_scale
+create_tensor: loading tensor blk.8.attn_gate.input_scale
+create_tensor: loading tensor blk.8.ffn_gate.input_scale
+create_tensor: loading tensor blk.8.ffn_down.input_scale
+create_tensor: loading tensor blk.8.ffn_up.input_scale
+create_tensor: loading tensor blk.9.attn_output.scale
+create_tensor: loading tensor blk.9.attn_qkv.scale
+create_tensor: loading tensor blk.9.attn_gate.scale
+create_tensor: loading tensor blk.9.ffn_gate.scale
+create_tensor: loading tensor blk.9.ffn_down.scale
+create_tensor: loading tensor blk.9.ffn_up.scale
+create_tensor: loading tensor blk.9.attn_output.input_scale
+create_tensor: loading tensor blk.9.attn_qkv.input_scale
+create_tensor: loading tensor blk.9.attn_gate.input_scale
+create_tensor: loading tensor blk.9.ffn_gate.input_scale
+create_tensor: loading tensor blk.9.ffn_down.input_scale
+create_tensor: loading tensor blk.9.ffn_up.input_scale
+create_tensor: loading tensor blk.10.attn_output.scale
+create_tensor: loading tensor blk.10.attn_qkv.scale
+create_tensor: loading tensor blk.10.attn_gate.scale
+create_tensor: loading tensor blk.10.ffn_gate.scale
+create_tensor: loading tensor blk.10.ffn_down.scale
+create_tensor: loading tensor blk.10.ffn_up.scale
+create_tensor: loading tensor blk.10.attn_output.input_scale
+create_tensor: loading tensor blk.10.attn_qkv.input_scale
+create_tensor: loading tensor blk.10.attn_gate.input_scale
+create_tensor: loading tensor blk.10.ffn_gate.input_scale
+create_tensor: loading tensor blk.10.ffn_down.input_scale
+create_tensor: loading tensor blk.10.ffn_up.input_scale
+create_tensor: loading tensor blk.11.attn_output.scale
+create_tensor: loading tensor blk.11.attn_qkv.scale
+create_tensor: loading tensor blk.11.attn_gate.scale
+create_tensor: loading tensor blk.11.ffn_gate.scale
+create_tensor: loading tensor blk.11.ffn_down.scale
+create_tensor: loading tensor blk.11.ffn_up.scale
+create_tensor: loading tensor blk.11.attn_output.input_scale
+create_tensor: loading tensor blk.11.attn_qkv.input_scale
+create_tensor: loading tensor blk.11.attn_gate.input_scale
+create_tensor: loading tensor blk.11.ffn_gate.input_scale
+create_tensor: loading tensor blk.11.ffn_down.input_scale
+create_tensor: loading tensor blk.11.ffn_up.input_scale
+create_tensor: loading tensor blk.12.attn_output.scale
+create_tensor: loading tensor blk.12.attn_qkv.scale
+create_tensor: loading tensor blk.12.attn_gate.scale
+create_tensor: loading tensor blk.12.ffn_gate.scale
+create_tensor: loading tensor blk.12.ffn_down.scale
+create_tensor: loading tensor blk.12.ffn_up.scale
+create_tensor: loading tensor blk.12.attn_output.input_scale
+create_tensor: loading tensor blk.12.attn_qkv.input_scale
+create_tensor: loading tensor blk.12.attn_gate.input_scale
+create_tensor: loading tensor blk.12.ffn_gate.input_scale
+create_tensor: loading tensor blk.12.ffn_down.input_scale
+create_tensor: loading tensor blk.12.ffn_up.input_scale
+create_tensor: loading tensor blk.13.attn_output.scale
+create_tensor: loading tensor blk.13.attn_qkv.scale
+create_tensor: loading tensor blk.13.attn_gate.scale
+create_tensor: loading tensor blk.13.ffn_gate.scale
+create_tensor: loading tensor blk.13.ffn_down.scale
+create_tensor: loading tensor blk.13.ffn_up.scale
+create_tensor: loading tensor blk.13.attn_output.input_scale
+create_tensor: loading tensor blk.13.attn_qkv.input_scale
+create_tensor: loading tensor blk.13.attn_gate.input_scale
+create_tensor: loading tensor blk.13.ffn_gate.input_scale
+create_tensor: loading tensor blk.13.ffn_down.input_scale
+create_tensor: loading tensor blk.13.ffn_up.input_scale
+create_tensor: loading tensor blk.14.attn_output.scale
+create_tensor: loading tensor blk.14.attn_qkv.scale
+create_tensor: loading tensor blk.14.attn_gate.scale
+create_tensor: loading tensor blk.14.ffn_gate.scale
+create_tensor: loading tensor blk.14.ffn_down.scale
+create_tensor: loading tensor blk.14.ffn_up.scale
+create_tensor: loading tensor blk.14.attn_output.input_scale
+create_tensor: loading tensor blk.14.attn_qkv.input_scale
+create_tensor: loading tensor blk.14.attn_gate.input_scale
+create_tensor: loading tensor blk.14.ffn_gate.input_scale
+create_tensor: loading tensor blk.14.ffn_down.input_scale
+create_tensor: loading tensor blk.14.ffn_up.input_scale
+create_tensor: loading tensor blk.15.attn_output.scale
+create_tensor: loading tensor blk.15.attn_qkv.scale
+create_tensor: loading tensor blk.15.attn_gate.scale
+create_tensor: loading tensor blk.15.ffn_gate.scale
+create_tensor: loading tensor blk.15.ffn_down.scale
+create_tensor: loading tensor blk.15.ffn_up.scale
+create_tensor: loading tensor blk.15.attn_output.input_scale
+create_tensor: loading tensor blk.15.attn_qkv.input_scale
+create_tensor: loading tensor blk.15.attn_gate.input_scale
+create_tensor: loading tensor blk.15.ffn_gate.input_scale
+create_tensor: loading tensor blk.15.ffn_down.input_scale
+create_tensor: loading tensor blk.15.ffn_up.input_scale
+create_tensor: loading tensor blk.16.attn_output.scale
+create_tensor: loading tensor blk.16.attn_qkv.scale
+create_tensor: loading tensor blk.16.attn_gate.scale
+create_tensor: loading tensor blk.16.ffn_gate.scale
+create_tensor: loading tensor blk.16.ffn_down.scale
+create_tensor: loading tensor blk.16.ffn_up.scale
+create_tensor: loading tensor blk.16.attn_output.input_scale
+create_tensor: loading tensor blk.16.attn_qkv.input_scale
+create_tensor: loading tensor blk.16.attn_gate.input_scale
+create_tensor: loading tensor blk.16.ffn_gate.input_scale
+create_tensor: loading tensor blk.16.ffn_down.input_scale
+create_tensor: loading tensor blk.16.ffn_up.input_scale
+create_tensor: loading tensor blk.17.attn_output.scale
+create_tensor: loading tensor blk.17.attn_qkv.scale
+create_tensor: loading tensor blk.17.attn_gate.scale
+create_tensor: loading tensor blk.17.ffn_gate.scale
+create_tensor: loading tensor blk.17.ffn_down.scale
+create_tensor: loading tensor blk.17.ffn_up.scale
+create_tensor: loading tensor blk.17.attn_output.input_scale
+create_tensor: loading tensor blk.17.attn_qkv.input_scale
+create_tensor: loading tensor blk.17.attn_gate.input_scale
+create_tensor: loading tensor blk.17.ffn_gate.input_scale
+create_tensor: loading tensor blk.17.ffn_down.input_scale
+create_tensor: loading tensor blk.17.ffn_up.input_scale
+create_tensor: loading tensor blk.18.attn_output.scale
+create_tensor: loading tensor blk.18.attn_qkv.scale
+create_tensor: loading tensor blk.18.attn_gate.scale
+create_tensor: loading tensor blk.18.ffn_gate.scale
+create_tensor: loading tensor blk.18.ffn_down.scale
+create_tensor: loading tensor blk.18.ffn_up.scale
+create_tensor: loading tensor blk.18.attn_output.input_scale
+create_tensor: loading tensor blk.18.attn_qkv.input_scale
+create_tensor: loading tensor blk.18.attn_gate.input_scale
+create_tensor: loading tensor blk.18.ffn_gate.input_scale
+create_tensor: loading tensor blk.18.ffn_down.input_scale
+create_tensor: loading tensor blk.18.ffn_up.input_scale
+create_tensor: loading tensor blk.19.attn_output.scale
+create_tensor: loading tensor blk.19.attn_qkv.scale
+create_tensor: loading tensor blk.19.attn_gate.scale
+create_tensor: loading tensor blk.19.ffn_gate.scale
+create_tensor: loading tensor blk.19.ffn_down.scale
+create_tensor: loading tensor blk.19.ffn_up.scale
+create_tensor: loading tensor blk.19.attn_output.input_scale
+create_tensor: loading tensor blk.19.attn_qkv.input_scale
+create_tensor: loading tensor blk.19.attn_gate.input_scale
+create_tensor: loading tensor blk.19.ffn_gate.input_scale
+create_tensor: loading tensor blk.19.ffn_down.input_scale
+create_tensor: loading tensor blk.19.ffn_up.input_scale
+create_tensor: loading tensor blk.20.attn_output.scale
+create_tensor: loading tensor blk.20.attn_qkv.scale
+create_tensor: loading tensor blk.20.attn_gate.scale
+create_tensor: loading tensor blk.20.ffn_gate.scale
+create_tensor: loading tensor blk.20.ffn_down.scale
+create_tensor: loading tensor blk.20.ffn_up.scale
+create_tensor: loading tensor blk.20.attn_output.input_scale
+create_tensor: loading tensor blk.20.attn_qkv.input_scale
+create_tensor: loading tensor blk.20.attn_gate.input_scale
+create_tensor: loading tensor blk.20.ffn_gate.input_scale
+create_tensor: loading tensor blk.20.ffn_down.input_scale
+create_tensor: loading tensor blk.20.ffn_up.input_scale
+create_tensor: loading tensor blk.21.attn_output.scale
+create_tensor: loading tensor blk.21.attn_qkv.scale
+create_tensor: loading tensor blk.21.attn_gate.scale
+create_tensor: loading tensor blk.21.ffn_gate.scale
+create_tensor: loading tensor blk.21.ffn_down.scale
+create_tensor: loading tensor blk.21.ffn_up.scale
+create_tensor: loading tensor blk.21.attn_output.input_scale
+create_tensor: loading tensor blk.21.attn_qkv.input_scale
+create_tensor: loading tensor blk.21.attn_gate.input_scale
+create_tensor: loading tensor blk.21.ffn_gate.input_scale
+create_tensor: loading tensor blk.21.ffn_down.input_scale
+create_tensor: loading tensor blk.21.ffn_up.input_scale
+create_tensor: loading tensor blk.22.attn_output.scale
+create_tensor: loading tensor blk.22.attn_qkv.scale
+create_tensor: loading tensor blk.22.attn_gate.scale
+create_tensor: loading tensor blk.22.ffn_gate.scale
+create_tensor: loading tensor blk.22.ffn_down.scale
+create_tensor: loading tensor blk.22.ffn_up.scale
+create_tensor: loading tensor blk.22.attn_output.input_scale
+create_tensor: loading tensor blk.22.attn_qkv.input_scale
+create_tensor: loading tensor blk.22.attn_gate.input_scale
+create_tensor: loading tensor blk.22.ffn_gate.input_scale
+create_tensor: loading tensor blk.22.ffn_down.input_scale
+create_tensor: loading tensor blk.22.ffn_up.input_scale
+create_tensor: loading tensor blk.23.attn_output.scale
+create_tensor: loading tensor blk.23.attn_qkv.scale
+create_tensor: loading tensor blk.23.attn_gate.scale
+create_tensor: loading tensor blk.23.ffn_gate.scale
+create_tensor: loading tensor blk.23.ffn_down.scale
+create_tensor: loading tensor blk.23.ffn_up.scale
+create_tensor: loading tensor blk.23.attn_output.input_scale
+create_tensor: loading tensor blk.23.attn_qkv.input_scale
+create_tensor: loading tensor blk.23.attn_gate.input_scale
+create_tensor: loading tensor blk.23.ffn_gate.input_scale
+create_tensor: loading tensor blk.23.ffn_down.input_scale
+create_tensor: loading tensor blk.23.ffn_up.input_scale
+create_tensor: loading tensor blk.24.attn_output.scale
+create_tensor: loading tensor blk.24.attn_qkv.scale
+create_tensor: loading tensor blk.24.attn_gate.scale
+create_tensor: loading tensor blk.24.ffn_gate.scale
+create_tensor: loading tensor blk.24.ffn_down.scale
+create_tensor: loading tensor blk.24.ffn_up.scale
+create_tensor: loading tensor blk.24.attn_output.input_scale
+create_tensor: loading tensor blk.24.attn_qkv.input_scale
+create_tensor: loading tensor blk.24.attn_gate.input_scale
+create_tensor: loading tensor blk.24.ffn_gate.input_scale
+create_tensor: loading tensor blk.24.ffn_down.input_scale
+create_tensor: loading tensor blk.24.ffn_up.input_scale
+create_tensor: loading tensor blk.25.attn_output.scale
+create_tensor: loading tensor blk.25.attn_qkv.scale
+create_tensor: loading tensor blk.25.attn_gate.scale
+create_tensor: loading tensor blk.25.ffn_gate.scale
+create_tensor: loading tensor blk.25.ffn_down.scale
+create_tensor: loading tensor blk.25.ffn_up.scale
+create_tensor: loading tensor blk.25.attn_output.input_scale
+create_tensor: loading tensor blk.25.attn_qkv.input_scale
+create_tensor: loading tensor blk.25.attn_gate.input_scale
+create_tensor: loading tensor blk.25.ffn_gate.input_scale
+create_tensor: loading tensor blk.25.ffn_down.input_scale
+create_tensor: loading tensor blk.25.ffn_up.input_scale
+create_tensor: loading tensor blk.26.attn_output.scale
+create_tensor: loading tensor blk.26.attn_qkv.scale
+create_tensor: loading tensor blk.26.attn_gate.scale
+create_tensor: loading tensor blk.26.ffn_gate.scale
+create_tensor: loading tensor blk.26.ffn_down.scale
+create_tensor: loading tensor blk.26.ffn_up.scale
+create_tensor: loading tensor blk.26.attn_output.input_scale
+create_tensor: loading tensor blk.26.attn_qkv.input_scale
+create_tensor: loading tensor blk.26.attn_gate.input_scale
+create_tensor: loading tensor blk.26.ffn_gate.input_scale
+create_tensor: loading tensor blk.26.ffn_down.input_scale
+create_tensor: loading tensor blk.26.ffn_up.input_scale
+create_tensor: loading tensor blk.27.attn_output.scale
+create_tensor: loading tensor blk.27.attn_qkv.scale
+create_tensor: loading tensor blk.27.attn_gate.scale
+create_tensor: loading tensor blk.27.ffn_gate.scale
+create_tensor: loading tensor blk.27.ffn_down.scale
+create_tensor: loading tensor blk.27.ffn_up.scale
+create_tensor: loading tensor blk.27.attn_output.input_scale
+create_tensor: loading tensor blk.27.attn_qkv.input_scale
+create_tensor: loading tensor blk.27.attn_gate.input_scale
+create_tensor: loading tensor blk.27.ffn_gate.input_scale
+create_tensor: loading tensor blk.27.ffn_down.input_scale
+create_tensor: loading tensor blk.27.ffn_up.input_scale
+load_tensors: offloading output layer to GPU
+load_tensors: offloading 27 repeating layers to GPU
+load_tensors: offloaded 29/29 layers to GPU
+load_tensors:      Vulkan1 model buffer size =  1051.02 MiB
+load_tensors:  Vulkan_Host model buffer size =   210.00 MiB
+load_all_data: using async uploads for device Vulkan1, buffer type Vulkan1, backend Vulkan1
+....................................................................load_all_data: buffer type Vulkan_Host is not the default buffer type for device Vulkan0 for async uploads
+.
+llama_context: constructing llama_context
+llama_context: n_seq_max             = 1
+llama_context: n_ctx                 = 256
+llama_context: n_ctx_seq             = 256
+llama_context: n_batch               = 128
+llama_context: n_ubatch              = 128
+llama_context: causal_attn           = 1
+llama_context: flash_attn            = auto
+llama_context: kv_unified            = false
+llama_context: freq_base             = 5000000.0
+llama_context: freq_scale            = 1
+llama_context: n_rs_seq              = 0
+llama_context: n_outputs_max         = 128
+llama_context: n_outputs_max_per_seq = 1
+llama_context: n_ctx_seq (256) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized
+set_abort_callback: call
+llama_context: Vulkan_Host  output buffer size =     0.50 MiB
+llama_kv_cache_iswa: creating non-SWA KV cache, size = 256 cells
+llama_kv_cache: layer   0: filtered
+llama_kv_cache: layer   1: filtered
+llama_kv_cache: layer   2: filtered
+llama_kv_cache: layer   3: dev = Vulkan1
+llama_kv_cache: layer   4: filtered
+llama_kv_cache: layer   5: filtered
+llama_kv_cache: layer   6: filtered
+llama_kv_cache: layer   7: dev = Vulkan1
+llama_kv_cache: layer   8: filtered
+llama_kv_cache: layer   9: filtered
+llama_kv_cache: layer  10: filtered
+llama_kv_cache: layer  11: dev = Vulkan1
+llama_kv_cache: layer  12: filtered
+llama_kv_cache: layer  13: filtered
+llama_kv_cache: layer  14: filtered
+llama_kv_cache: layer  15: dev = Vulkan1
+llama_kv_cache: layer  16: filtered
+llama_kv_cache: layer  17: filtered
+llama_kv_cache: layer  18: filtered
+llama_kv_cache: layer  19: dev = Vulkan1
+llama_kv_cache: layer  20: filtered
+llama_kv_cache: layer  21: filtered
+llama_kv_cache: layer  22: filtered
+llama_kv_cache: layer  23: dev = Vulkan1
+llama_kv_cache: layer  24: filtered
+llama_kv_cache: layer  25: filtered
+llama_kv_cache: layer  26: filtered
+llama_kv_cache: layer  27: dev = Vulkan1
+llama_kv_cache:    Vulkan1 KV buffer size =     3.50 MiB
+llama_kv_cache: size =    3.50 MiB (   256 cells,   7 layers,  1/1 seqs), K (f16):    1.75 MiB, V (f16):    1.75 MiB
+llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256
+llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256
+llama_kv_cache_iswa: creating     SWA KV cache, size = 256 cells
+llama_kv_cache: layer   0: dev = Vulkan1
+llama_kv_cache: layer   1: dev = Vulkan1
+llama_kv_cache: layer   2: dev = Vulkan1
+llama_kv_cache: layer   3: filtered
+llama_kv_cache: layer   4: dev = Vulkan1
+llama_kv_cache: layer   5: dev = Vulkan1
+llama_kv_cache: layer   6: dev = Vulkan1
+llama_kv_cache: layer   7: filtered
+llama_kv_cache: layer   8: dev = Vulkan1
+llama_kv_cache: layer   9: dev = Vulkan1
+llama_kv_cache: layer  10: dev = Vulkan1
+llama_kv_cache: layer  11: filtered
+llama_kv_cache: layer  12: dev = Vulkan1
+llama_kv_cache: layer  13: dev = Vulkan1
+llama_kv_cache: layer  14: dev = Vulkan1
+llama_kv_cache: layer  15: filtered
+llama_kv_cache: layer  16: dev = Vulkan1
+llama_kv_cache: layer  17: dev = Vulkan1
+llama_kv_cache: layer  18: dev = Vulkan1
+llama_kv_cache: layer  19: filtered
+llama_kv_cache: layer  20: dev = Vulkan1
+llama_kv_cache: layer  21: dev = Vulkan1
+llama_kv_cache: layer  22: dev = Vulkan1
+llama_kv_cache: layer  23: filtered
+llama_kv_cache: layer  24: dev = Vulkan1
+llama_kv_cache: layer  25: dev = Vulkan1
+llama_kv_cache: layer  26: dev = Vulkan1
+llama_kv_cache: layer  27: filtered
+llama_kv_cache:    Vulkan1 KV buffer size =    10.50 MiB
+llama_kv_cache: size =   10.50 MiB (   256 cells,  21 layers,  1/1 seqs), K (f16):    5.25 MiB, V (f16):    5.25 MiB
+llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256
+llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256
+llama_context: enumerating backends
+llama_context: backend_ptrs.size() = 2
+sched_reserve: reserving ...
+sched_reserve: max_nodes = 1816
+sched_reserve: reserving full memory module
+sched_reserve: worst-case: n_tokens = 128, n_seqs = 1, n_outputs = 1
+graph_reserve: reserving a graph for ubatch with n_tokens =    1, n_seqs =  1, n_outputs =    1
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {     1,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+resolve_fused_ops: Flash Attention enabled
+resolve_fused_ops: resolving fused DeepSeek V4 HC support:
+graph_reserve: reserving a graph for ubatch with n_tokens =    1, n_seqs =  1, n_outputs =    1
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {     1,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+resolve_fused_ops: fused DeepSeek V4 HC pre enabled
+graph_reserve: reserving a graph for ubatch with n_tokens =    1, n_seqs =  1, n_outputs =    1
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {     1,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+resolve_fused_ops: fused DeepSeek V4 HC comb enabled
+graph_reserve: reserving a graph for ubatch with n_tokens =    1, n_seqs =  1, n_outputs =    1
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {     1,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+resolve_fused_ops: fused DeepSeek V4 HC post enabled
+graph_reserve: reserving a graph for ubatch with n_tokens =  128, n_seqs =  1, n_outputs =  128
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {   128,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {   128,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {   128,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+graph_reserve: reserving a graph for ubatch with n_tokens =    1, n_seqs =  1, n_outputs =    1
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {     1,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+graph_reserve: reserving a graph for ubatch with n_tokens =  128, n_seqs =  1, n_outputs =  128
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {   128,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {   128,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {   128,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,   128,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {   128,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {   128,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+sched_reserve:    Vulkan1 compute buffer size =    65.00 MiB
+sched_reserve: Vulkan_Host compute buffer size =     2.13 MiB
+sched_reserve: graph: nodes = 986, splits = 2, input objects = 4, input tensors = 9
+sched_reserve: reserve took 14.50 ms, sched copies = 1
+attach_threadpool: call
+set_n_threads: n_threads = 12, n_threads_batch = 12
+set_n_threads: n_threads = 12, n_threads_batch = 12
+llama_perf_context_print:        load time =    1192.43 ms
+llama_perf_context_print: prompt eval time =       0.00 ms /   256 tokens (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =    1280.33 ms /   257 tokens
+llama_perf_context_print:    graphs reused =          1
+~llama_context:    Vulkan1 compute buffer size is  65.0000 MiB, matches expectation of  65.0000 MiB
+~llama_context: Vulkan_Host compute buffer size is   2.1304 MiB, matches expectation of   2.1304 MiB
+llama_context: constructing llama_context
+llama_context: n_seq_max             = 1
+llama_context: n_ctx                 = 256
+llama_context: n_ctx_seq             = 256
+llama_context: n_batch               = 32
+llama_context: n_ubatch              = 32
+llama_context: causal_attn           = 1
+llama_context: flash_attn            = auto
+llama_context: kv_unified            = false
+llama_context: freq_base             = 5000000.0
+llama_context: freq_scale            = 1
+llama_context: n_rs_seq              = 0
+llama_context: n_outputs_max         = 32
+llama_context: n_outputs_max_per_seq = 1
+llama_context: n_ctx_seq (256) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized
+set_abort_callback: call
+llama_context: Vulkan_Host  output buffer size =     0.50 MiB
+llama_kv_cache_iswa: creating non-SWA KV cache, size = 256 cells
+llama_kv_cache: layer   0: filtered
+llama_kv_cache: layer   1: filtered
+llama_kv_cache: layer   2: filtered
+llama_kv_cache: layer   3: dev = Vulkan1
+llama_kv_cache: layer   4: filtered
+llama_kv_cache: layer   5: filtered
+llama_kv_cache: layer   6: filtered
+llama_kv_cache: layer   7: dev = Vulkan1
+llama_kv_cache: layer   8: filtered
+llama_kv_cache: layer   9: filtered
+llama_kv_cache: layer  10: filtered
+llama_kv_cache: layer  11: dev = Vulkan1
+llama_kv_cache: layer  12: filtered
+llama_kv_cache: layer  13: filtered
+llama_kv_cache: layer  14: filtered
+llama_kv_cache: layer  15: dev = Vulkan1
+llama_kv_cache: layer  16: filtered
+llama_kv_cache: layer  17: filtered
+llama_kv_cache: layer  18: filtered
+llama_kv_cache: layer  19: dev = Vulkan1
+llama_kv_cache: layer  20: filtered
+llama_kv_cache: layer  21: filtered
+llama_kv_cache: layer  22: filtered
+llama_kv_cache: layer  23: dev = Vulkan1
+llama_kv_cache: layer  24: filtered
+llama_kv_cache: layer  25: filtered
+llama_kv_cache: layer  26: filtered
+llama_kv_cache: layer  27: dev = Vulkan1
+llama_kv_cache:    Vulkan1 KV buffer size =     3.50 MiB
+llama_kv_cache: size =    3.50 MiB (   256 cells,   7 layers,  1/1 seqs), K (f16):    1.75 MiB, V (f16):    1.75 MiB
+llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256
+llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256
+llama_kv_cache_iswa: creating     SWA KV cache, size = 256 cells
+llama_kv_cache: layer   0: dev = Vulkan1
+llama_kv_cache: layer   1: dev = Vulkan1
+llama_kv_cache: layer   2: dev = Vulkan1
+llama_kv_cache: layer   3: filtered
+llama_kv_cache: layer   4: dev = Vulkan1
+llama_kv_cache: layer   5: dev = Vulkan1
+llama_kv_cache: layer   6: dev = Vulkan1
+llama_kv_cache: layer   7: filtered
+llama_kv_cache: layer   8: dev = Vulkan1
+llama_kv_cache: layer   9: dev = Vulkan1
+llama_kv_cache: layer  10: dev = Vulkan1
+llama_kv_cache: layer  11: filtered
+llama_kv_cache: layer  12: dev = Vulkan1
+llama_kv_cache: layer  13: dev = Vulkan1
+llama_kv_cache: layer  14: dev = Vulkan1
+llama_kv_cache: layer  15: filtered
+llama_kv_cache: layer  16: dev = Vulkan1
+llama_kv_cache: layer  17: dev = Vulkan1
+llama_kv_cache: layer  18: dev = Vulkan1
+llama_kv_cache: layer  19: filtered
+llama_kv_cache: layer  20: dev = Vulkan1
+llama_kv_cache: layer  21: dev = Vulkan1
+llama_kv_cache: layer  22: dev = Vulkan1
+llama_kv_cache: layer  23: filtered
+llama_kv_cache: layer  24: dev = Vulkan1
+llama_kv_cache: layer  25: dev = Vulkan1
+llama_kv_cache: layer  26: dev = Vulkan1
+llama_kv_cache: layer  27: filtered
+llama_kv_cache:    Vulkan1 KV buffer size =    10.50 MiB
+llama_kv_cache: size =   10.50 MiB (   256 cells,  21 layers,  1/1 seqs), K (f16):    5.25 MiB, V (f16):    5.25 MiB
+llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256
+llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256
+llama_context: enumerating backends
+llama_context: backend_ptrs.size() = 2
+sched_reserve: reserving ...
+sched_reserve: max_nodes = 1816
+sched_reserve: reserving full memory module
+sched_reserve: worst-case: n_tokens = 32, n_seqs = 1, n_outputs = 1
+graph_reserve: reserving a graph for ubatch with n_tokens =    1, n_seqs =  1, n_outputs =    1
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {     1,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+resolve_fused_ops: Flash Attention enabled
+resolve_fused_ops: resolving fused DeepSeek V4 HC support:
+graph_reserve: reserving a graph for ubatch with n_tokens =    1, n_seqs =  1, n_outputs =    1
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {     1,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+resolve_fused_ops: fused DeepSeek V4 HC pre enabled
+graph_reserve: reserving a graph for ubatch with n_tokens =    1, n_seqs =  1, n_outputs =    1
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {     1,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+resolve_fused_ops: fused DeepSeek V4 HC comb enabled
+graph_reserve: reserving a graph for ubatch with n_tokens =    1, n_seqs =  1, n_outputs =    1
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {     1,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+resolve_fused_ops: fused DeepSeek V4 HC post enabled
+graph_reserve: reserving a graph for ubatch with n_tokens =   32, n_seqs =  1, n_outputs =   32
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {    32,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {    32,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {    32,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+graph_reserve: reserving a graph for ubatch with n_tokens =    1, n_seqs =  1, n_outputs =    1
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {     1,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {     1,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,     1,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {     1,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {     1,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+graph_reserve: reserving a graph for ubatch with n_tokens =   32, n_seqs =  1, n_outputs =   32
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_124' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_264' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_404' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_544' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_684' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_824' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_964' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                      inp_tokens' [i32, ne = {    32,     1,     1,     1 }] is used by node 'embd' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-0' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-1' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-2' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-3' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-4' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-5' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-6' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-7' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-8' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-9' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-10' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-11' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-12' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-13' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-14' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-15' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-16' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-17' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-18' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-19' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-20' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-21' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-22' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-23' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-24' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-25' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-26' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Qcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                         inp_pos' [i32, ne = {    32,     1,     1,     1 }] is used by node 'Kcur_rope-27' (ROPE)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l0 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l1 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l2 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l4 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l5 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l6 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l8 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l9 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l10 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l12 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l13 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l14 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l16 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l17 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l18 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l20 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l21 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l22 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l24 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l25 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l26 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_k_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_k_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_19' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_54' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_89' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_159' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_194' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_229' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_299' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_334' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_369' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_439' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_474' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_509' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_579' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_614' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_649' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_719' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_754' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_789' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_859' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_894' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                attn_inp_kq_mask' [f16, ne = {   256,    32,     1,     1 }] is used by node 'node_929' (FLASH_ATTN_EXT)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l3 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l7 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l11 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l15 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l19 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l23 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                 attn_inp_v_idxs' [i64, ne = {    32,     1,     1,     1 }] is used by node 'cache_v_l27 (view)' (SET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {    32,     1,     1,     1 }] is used by node 'node_973' (GET_ROWS)
+llama_graph_n_input_tensors: input tensor '                         out_ids' [i32, ne = {    32,     1,     1,     1 }] is used by node 'node_974' (GET_ROWS)
+sched_reserve:    Vulkan1 compute buffer size =    16.25 MiB
+sched_reserve: Vulkan_Host compute buffer size =     0.53 MiB
+sched_reserve: graph: nodes = 986, splits = 2, input objects = 4, input tensors = 9
+sched_reserve: reserve took 29.43 ms, sched copies = 1
+attach_threadpool: call
+set_n_threads: n_threads = 12, n_threads_batch = 12
+set_n_threads: n_threads = 12, n_threads_batch = 12
+llama_perf_context_print:        load time =    1341.05 ms
+llama_perf_context_print: prompt eval time =       0.00 ms /     1 tokens (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /    33 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =    1849.11 ms /    34 tokens
+llama_perf_context_print:    graphs reused =         32
+~llama_context:    Vulkan1 compute buffer size is  16.2500 MiB, matches expectation of  16.2500 MiB
+~llama_context: Vulkan_Host compute buffer size is   0.5326 MiB, matches expectation of   0.5326 MiB
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_single_20260923.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_single_20260923.log
new file mode 100644
index 000000000..045dd1272
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_q4km_single_20260923.log
@@ -0,0 +1,36 @@
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b11124-1a679828f
+model      : C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> What is the capital of France? Answer in one word.
+
+[Start thinking]
+
+We are asked: "What is the capital of France? Answer in one word." So the answer should be a single word: "Paris". But we need to output just the word? The instruction says "Answer in one word." So we should output "Paris". But let's make sure: The capital of France
+
+[ Prompt: 13.5 t/s | Generation: 57.7 t/s ]
+
+
+Exiting...
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_20260923.err.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_20260923.err.log
new file mode 100644
index 000000000..e00561be3
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_20260923.err.log
@@ -0,0 +1,388 @@
+0.00.000.801 I srv  llama_server: initializing ...
+0.00.588.250 I cmn  common_param: common_params_print_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
+0.00.588.396 I srv  init_listene: The UI is disabled
+0.00.588.397 I srv  init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable
+0.00.588.458 W srv  llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655)
+0.00.603.434 I srv    load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf'
+0.02.545.243 I cmn          init: llama threadpool init, n_threads = 12
+0.05.505.041 I srv    load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false'
+0.05.529.021 W srv          init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve)
+0.05.529.060 I srv  llama_server: model loaded
+0.05.529.101 I srv  llama_server: listening on http://127.0.0.1:18173
+0.16.873.186 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
+0.16.873.261 I slot launch_slot_: id  0 | task 0 | processing task, is_child = 0
+0.19.089.882 I slot print_timing: id  0 | task 0 | prompt eval time =    2196.02 ms /    28 tokens (   78.43 ms per token,    12.75 tokens per second)
+0.19.089.896 I slot print_timing: id  0 | task 0 |        eval time =      20.53 ms /     2 tokens (   20.53 ms per token,    48.70 tokens per second)
+0.19.089.898 I slot print_timing: id  0 | task 0 |       total time =    2216.55 ms /    30 tokens
+0.19.089.900 I slot print_timing: id  0 | task 0 |    graphs reused =          1
+0.19.089.950 I slot      release: id  0 | task 0 | stop processing: n_tokens = 29, truncated = 0
+0.26.456.765 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = 18538781
+0.26.560.412 I slot launch_slot_: id  0 | task 5 | processing task, is_child = 0
+0.30.378.510 I slot print_timing: id  0 | task 5 | prompt processing, n_tokens =   1940, progress = 0.79, t =   3.82 s / 508.12 tokens per second
+0.30.854.818 I slot print_timing: id  0 | task 5 | prompt processing, n_tokens =   2452, progress = 1.00, t =   4.29 s / 570.99 tokens per second
+0.31.482.753 I slot print_timing: id  0 | task 5 | prompt eval time =    4845.13 ms /  2456 tokens (    1.97 ms per token,   506.90 tokens per second)
+0.31.482.766 I slot print_timing: id  0 | task 5 |        eval time =      77.15 ms /     5 tokens (   19.29 ms per token,    51.85 tokens per second)
+0.31.482.768 I slot print_timing: id  0 | task 5 |       total time =    4922.28 ms /  2461 tokens
+0.31.482.770 I slot print_timing: id  0 | task 5 |    graphs reused =          4
+0.31.482.900 I slot      release: id  0 | task 5 | stop processing: n_tokens = 2469, truncated = 0
+2.30.021.779 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 0.429 (> 0.100 thold), f_keep = 0.005
+2.30.410.526 I slot launch_slot_: id  0 | task 14 | processing task, is_child = 0
+2.30.484.767 I slot print_timing: id  0 | task 14 | prompt eval time =      55.17 ms /     5 tokens (   11.04 ms per token,    90.62 tokens per second)
+2.30.484.773 I slot print_timing: id  0 | task 14 |        eval time =      19.03 ms /     2 tokens (   19.03 ms per token,    52.54 tokens per second)
+2.30.484.775 I slot print_timing: id  0 | task 14 |       total time =      74.21 ms /     7 tokens
+2.30.484.776 I slot print_timing: id  0 | task 14 |    graphs reused =          4
+2.30.484.806 I slot      release: id  0 | task 14 | stop processing: n_tokens = 29, truncated = 0
+2.30.506.319 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = 149933637
+2.30.525.022 I slot launch_slot_: id  0 | task 18 | processing task, is_child = 0
+2.31.263.314 I slot print_timing: id  0 | task 18 | prompt eval time =     664.25 ms /   515 tokens (    1.29 ms per token,   775.32 tokens per second)
+2.31.263.323 I slot print_timing: id  0 | task 18 |        eval time =      74.02 ms /     5 tokens (   18.50 ms per token,    54.04 tokens per second)
+2.31.263.324 I slot print_timing: id  0 | task 18 |       total time =     738.26 ms /   520 tokens
+2.31.263.325 I slot print_timing: id  0 | task 18 |    graphs reused =          7
+2.31.263.387 I slot      release: id  0 | task 18 | stop processing: n_tokens = 2467, truncated = 0
+2.31.294.713 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 0.429 (> 0.100 thold), f_keep = 0.005
+2.31.676.391 I slot launch_slot_: id  0 | task 25 | processing task, is_child = 0
+2.31.748.802 I slot print_timing: id  0 | task 25 | prompt eval time =      53.94 ms /     5 tokens (   10.79 ms per token,    92.70 tokens per second)
+2.31.748.810 I slot print_timing: id  0 | task 25 |        eval time =      18.43 ms /     2 tokens (   18.43 ms per token,    54.25 tokens per second)
+2.31.748.811 I slot print_timing: id  0 | task 25 |       total time =      72.37 ms /     7 tokens
+2.31.748.812 I slot print_timing: id  0 | task 25 |    graphs reused =          7
+2.31.748.837 I slot      release: id  0 | task 25 | stop processing: n_tokens = 29, truncated = 0
+2.31.781.216 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.31.781.384 I slot launch_slot_: id  0 | task 29 | processing task, is_child = 0
+2.31.847.648 I slot print_timing: id  0 | task 29 | prompt eval time =      49.01 ms /     5 tokens (    9.80 ms per token,   102.02 tokens per second)
+2.31.847.654 I slot print_timing: id  0 | task 29 |        eval time =      17.23 ms /     2 tokens (   17.23 ms per token,    58.02 tokens per second)
+2.31.847.656 I slot print_timing: id  0 | task 29 |       total time =      66.25 ms /     7 tokens
+2.31.847.657 I slot print_timing: id  0 | task 29 |    graphs reused =          7
+2.31.847.682 I slot      release: id  0 | task 29 | stop processing: n_tokens = 29, truncated = 0
+2.31.876.772 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.31.876.937 I slot launch_slot_: id  0 | task 33 | processing task, is_child = 0
+2.31.946.556 I slot print_timing: id  0 | task 33 | prompt eval time =      50.89 ms /     5 tokens (   10.18 ms per token,    98.25 tokens per second)
+2.31.946.562 I slot print_timing: id  0 | task 33 |        eval time =      18.71 ms /     2 tokens (   18.71 ms per token,    53.44 tokens per second)
+2.31.946.563 I slot print_timing: id  0 | task 33 |       total time =      69.60 ms /     7 tokens
+2.31.946.564 I slot print_timing: id  0 | task 33 |    graphs reused =          7
+2.31.946.591 I slot      release: id  0 | task 33 | stop processing: n_tokens = 29, truncated = 0
+2.31.967.053 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.31.967.236 I slot launch_slot_: id  0 | task 37 | processing task, is_child = 0
+2.32.036.531 I slot print_timing: id  0 | task 37 | prompt eval time =      51.77 ms /     5 tokens (   10.35 ms per token,    96.58 tokens per second)
+2.32.036.539 I slot print_timing: id  0 | task 37 |        eval time =      17.51 ms /     2 tokens (   17.51 ms per token,    57.12 tokens per second)
+2.32.036.541 I slot print_timing: id  0 | task 37 |       total time =      69.28 ms /     7 tokens
+2.32.036.542 I slot print_timing: id  0 | task 37 |    graphs reused =          7
+2.32.036.571 I slot      release: id  0 | task 37 | stop processing: n_tokens = 29, truncated = 0
+2.32.060.707 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.32.060.879 I slot launch_slot_: id  0 | task 41 | processing task, is_child = 0
+2.32.127.945 I slot print_timing: id  0 | task 41 | prompt eval time =      50.24 ms /     5 tokens (   10.05 ms per token,    99.52 tokens per second)
+2.32.127.952 I slot print_timing: id  0 | task 41 |        eval time =      16.81 ms /     2 tokens (   16.81 ms per token,    59.50 tokens per second)
+2.32.127.953 I slot print_timing: id  0 | task 41 |       total time =      67.05 ms /     7 tokens
+2.32.127.954 I slot print_timing: id  0 | task 41 |    graphs reused =          7
+2.32.127.980 I slot      release: id  0 | task 41 | stop processing: n_tokens = 29, truncated = 0
+2.32.137.029 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.32.137.385 I slot launch_slot_: id  0 | task 45 | processing task, is_child = 0
+2.32.205.397 I slot print_timing: id  0 | task 45 | prompt eval time =      50.61 ms /     5 tokens (   10.12 ms per token,    98.79 tokens per second)
+2.32.205.408 I slot print_timing: id  0 | task 45 |        eval time =      17.37 ms /     2 tokens (   17.37 ms per token,    57.58 tokens per second)
+2.32.205.410 I slot print_timing: id  0 | task 45 |       total time =      67.98 ms /     7 tokens
+2.32.205.413 I slot print_timing: id  0 | task 45 |    graphs reused =          7
+2.32.205.452 I slot      release: id  0 | task 45 | stop processing: n_tokens = 29, truncated = 0
+2.32.235.765 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.32.235.955 I slot launch_slot_: id  0 | task 49 | processing task, is_child = 0
+2.32.305.291 I slot print_timing: id  0 | task 49 | prompt eval time =      51.99 ms /     5 tokens (   10.40 ms per token,    96.17 tokens per second)
+2.32.305.299 I slot print_timing: id  0 | task 49 |        eval time =      17.32 ms /     2 tokens (   17.32 ms per token,    57.72 tokens per second)
+2.32.305.300 I slot print_timing: id  0 | task 49 |       total time =      69.31 ms /     7 tokens
+2.32.305.301 I slot print_timing: id  0 | task 49 |    graphs reused =          7
+2.32.305.332 I slot      release: id  0 | task 49 | stop processing: n_tokens = 29, truncated = 0
+2.32.330.005 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.32.330.166 I slot launch_slot_: id  0 | task 53 | processing task, is_child = 0
+2.32.397.935 I slot print_timing: id  0 | task 53 | prompt eval time =      51.05 ms /     5 tokens (   10.21 ms per token,    97.94 tokens per second)
+2.32.397.943 I slot print_timing: id  0 | task 53 |        eval time =      16.70 ms /     2 tokens (   16.70 ms per token,    59.89 tokens per second)
+2.32.397.944 I slot print_timing: id  0 | task 53 |       total time =      67.75 ms /     7 tokens
+2.32.397.945 I slot print_timing: id  0 | task 53 |    graphs reused =          7
+2.32.397.974 I slot      release: id  0 | task 53 | stop processing: n_tokens = 29, truncated = 0
+2.32.426.051 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.32.426.493 I slot launch_slot_: id  0 | task 57 | processing task, is_child = 0
+2.32.494.098 I slot print_timing: id  0 | task 57 | prompt eval time =      49.13 ms /     5 tokens (    9.83 ms per token,   101.77 tokens per second)
+2.32.494.107 I slot print_timing: id  0 | task 57 |        eval time =      18.45 ms /     2 tokens (   18.45 ms per token,    54.21 tokens per second)
+2.32.494.108 I slot print_timing: id  0 | task 57 |       total time =      67.58 ms /     7 tokens
+2.32.494.109 I slot print_timing: id  0 | task 57 |    graphs reused =          7
+2.32.494.137 I slot      release: id  0 | task 57 | stop processing: n_tokens = 29, truncated = 0
+2.32.523.126 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.32.523.318 I slot launch_slot_: id  0 | task 61 | processing task, is_child = 0
+2.32.594.390 I slot print_timing: id  0 | task 61 | prompt eval time =      52.82 ms /     5 tokens (   10.56 ms per token,    94.66 tokens per second)
+2.32.594.399 I slot print_timing: id  0 | task 61 |        eval time =      18.23 ms /     2 tokens (   18.23 ms per token,    54.85 tokens per second)
+2.32.594.400 I slot print_timing: id  0 | task 61 |       total time =      71.05 ms /     7 tokens
+2.32.594.401 I slot print_timing: id  0 | task 61 |    graphs reused =          7
+2.32.594.427 I slot      release: id  0 | task 61 | stop processing: n_tokens = 29, truncated = 0
+2.32.605.360 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.32.605.532 I slot launch_slot_: id  0 | task 65 | processing task, is_child = 0
+2.32.672.476 I slot print_timing: id  0 | task 65 | prompt eval time =      48.62 ms /     5 tokens (    9.72 ms per token,   102.83 tokens per second)
+2.32.672.483 I slot print_timing: id  0 | task 65 |        eval time =      18.30 ms /     2 tokens (   18.30 ms per token,    54.63 tokens per second)
+2.32.672.484 I slot print_timing: id  0 | task 65 |       total time =      66.93 ms /     7 tokens
+2.32.672.485 I slot print_timing: id  0 | task 65 |    graphs reused =          7
+2.32.672.510 I slot      release: id  0 | task 65 | stop processing: n_tokens = 29, truncated = 0
+2.32.695.796 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.32.696.137 I slot launch_slot_: id  0 | task 69 | processing task, is_child = 0
+2.32.766.845 I slot print_timing: id  0 | task 69 | prompt eval time =      52.34 ms /     5 tokens (   10.47 ms per token,    95.53 tokens per second)
+2.32.766.854 I slot print_timing: id  0 | task 69 |        eval time =      18.33 ms /     2 tokens (   18.33 ms per token,    54.56 tokens per second)
+2.32.766.855 I slot print_timing: id  0 | task 69 |       total time =      70.66 ms /     7 tokens
+2.32.766.856 I slot print_timing: id  0 | task 69 |    graphs reused =          7
+2.32.766.879 I slot      release: id  0 | task 69 | stop processing: n_tokens = 29, truncated = 0
+2.32.789.399 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.32.789.570 I slot launch_slot_: id  0 | task 73 | processing task, is_child = 0
+2.32.859.910 I slot print_timing: id  0 | task 73 | prompt eval time =      52.30 ms /     5 tokens (   10.46 ms per token,    95.60 tokens per second)
+2.32.859.918 I slot print_timing: id  0 | task 73 |        eval time =      18.02 ms /     2 tokens (   18.02 ms per token,    55.50 tokens per second)
+2.32.859.919 I slot print_timing: id  0 | task 73 |       total time =      70.32 ms /     7 tokens
+2.32.859.920 I slot print_timing: id  0 | task 73 |    graphs reused =          7
+2.32.859.948 I slot      release: id  0 | task 73 | stop processing: n_tokens = 29, truncated = 0
+2.32.869.835 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.32.870.031 I slot launch_slot_: id  0 | task 77 | processing task, is_child = 0
+2.32.937.715 I slot print_timing: id  0 | task 77 | prompt eval time =      49.62 ms /     5 tokens (    9.93 ms per token,   100.76 tokens per second)
+2.32.937.722 I slot print_timing: id  0 | task 77 |        eval time =      18.04 ms /     2 tokens (   18.04 ms per token,    55.44 tokens per second)
+2.32.937.723 I slot print_timing: id  0 | task 77 |       total time =      67.66 ms /     7 tokens
+2.32.937.724 I slot print_timing: id  0 | task 77 |    graphs reused =          7
+2.32.937.746 I slot      release: id  0 | task 77 | stop processing: n_tokens = 29, truncated = 0
+2.32.964.108 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.32.964.318 I slot launch_slot_: id  0 | task 81 | processing task, is_child = 0
+2.33.033.277 I slot print_timing: id  0 | task 81 | prompt eval time =      49.60 ms /     5 tokens (    9.92 ms per token,   100.80 tokens per second)
+2.33.033.312 I slot print_timing: id  0 | task 81 |        eval time =      19.33 ms /     2 tokens (   19.33 ms per token,    51.73 tokens per second)
+2.33.033.316 I slot print_timing: id  0 | task 81 |       total time =      68.93 ms /     7 tokens
+2.33.033.319 I slot print_timing: id  0 | task 81 |    graphs reused =          7
+2.33.033.376 I slot      release: id  0 | task 81 | stop processing: n_tokens = 29, truncated = 0
+2.33.063.351 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.33.063.543 I slot launch_slot_: id  0 | task 85 | processing task, is_child = 0
+2.33.132.992 I slot print_timing: id  0 | task 85 | prompt eval time =      51.96 ms /     5 tokens (   10.39 ms per token,    96.22 tokens per second)
+2.33.132.999 I slot print_timing: id  0 | task 85 |        eval time =      17.47 ms /     2 tokens (   17.47 ms per token,    57.25 tokens per second)
+2.33.133.000 I slot print_timing: id  0 | task 85 |       total time =      69.43 ms /     7 tokens
+2.33.133.001 I slot print_timing: id  0 | task 85 |    graphs reused =          7
+2.33.133.030 I slot      release: id  0 | task 85 | stop processing: n_tokens = 29, truncated = 0
+2.33.152.112 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.33.152.276 I slot launch_slot_: id  0 | task 89 | processing task, is_child = 0
+2.33.222.945 I slot print_timing: id  0 | task 89 | prompt eval time =      52.77 ms /     5 tokens (   10.55 ms per token,    94.75 tokens per second)
+2.33.222.952 I slot print_timing: id  0 | task 89 |        eval time =      17.88 ms /     2 tokens (   17.88 ms per token,    55.92 tokens per second)
+2.33.222.953 I slot print_timing: id  0 | task 89 |       total time =      70.65 ms /     7 tokens
+2.33.222.954 I slot print_timing: id  0 | task 89 |    graphs reused =          7
+2.33.222.979 I slot      release: id  0 | task 89 | stop processing: n_tokens = 29, truncated = 0
+2.33.246.947 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.33.247.331 I slot launch_slot_: id  0 | task 93 | processing task, is_child = 0
+2.33.315.202 I slot print_timing: id  0 | task 93 | prompt eval time =      50.52 ms /     5 tokens (   10.10 ms per token,    98.98 tokens per second)
+2.33.315.208 I slot print_timing: id  0 | task 93 |        eval time =      17.33 ms /     2 tokens (   17.33 ms per token,    57.69 tokens per second)
+2.33.315.210 I slot print_timing: id  0 | task 93 |       total time =      67.85 ms /     7 tokens
+2.33.315.211 I slot print_timing: id  0 | task 93 |    graphs reused =          7
+2.33.315.234 I slot      release: id  0 | task 93 | stop processing: n_tokens = 29, truncated = 0
+2.33.339.108 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.33.339.279 I slot launch_slot_: id  0 | task 97 | processing task, is_child = 0
+2.33.407.789 I slot print_timing: id  0 | task 97 | prompt eval time =      50.54 ms /     5 tokens (   10.11 ms per token,    98.92 tokens per second)
+2.33.407.798 I slot print_timing: id  0 | task 97 |        eval time =      17.95 ms /     2 tokens (   17.95 ms per token,    55.71 tokens per second)
+2.33.407.800 I slot print_timing: id  0 | task 97 |       total time =      68.49 ms /     7 tokens
+2.33.407.801 I slot print_timing: id  0 | task 97 |    graphs reused =          7
+2.33.407.830 I slot      release: id  0 | task 97 | stop processing: n_tokens = 29, truncated = 0
+2.33.432.838 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.33.433.158 I slot launch_slot_: id  0 | task 101 | processing task, is_child = 0
+2.33.505.887 I slot print_timing: id  0 | task 101 | prompt eval time =      54.42 ms /     5 tokens (   10.88 ms per token,    91.88 tokens per second)
+2.33.505.902 I slot print_timing: id  0 | task 101 |        eval time =      18.26 ms /     2 tokens (   18.26 ms per token,    54.76 tokens per second)
+2.33.505.904 I slot print_timing: id  0 | task 101 |       total time =      72.68 ms /     7 tokens
+2.33.505.905 I slot print_timing: id  0 | task 101 |    graphs reused =          7
+2.33.505.949 I slot      release: id  0 | task 101 | stop processing: n_tokens = 29, truncated = 0
+2.33.518.498 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.966
+2.33.518.797 I slot launch_slot_: id  0 | task 105 | processing task, is_child = 0
+2.33.590.480 I slot print_timing: id  0 | task 105 | prompt eval time =      53.77 ms /     5 tokens (   10.75 ms per token,    92.99 tokens per second)
+2.33.590.489 I slot print_timing: id  0 | task 105 |        eval time =      17.89 ms /     2 tokens (   17.89 ms per token,    55.89 tokens per second)
+2.33.590.490 I slot print_timing: id  0 | task 105 |       total time =      71.66 ms /     7 tokens
+2.33.590.491 I slot print_timing: id  0 | task 105 |    graphs reused =          7
+2.33.590.521 I slot      release: id  0 | task 105 | stop processing: n_tokens = 29, truncated = 0
+2.50.259.814 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = 153039352
+2.50.505.888 I slot launch_slot_: id  0 | task 109 | processing task, is_child = 0
+2.52.637.757 I slot print_timing: id  0 | task 109 | prompt eval time =    2060.26 ms /  2465 tokens (    0.84 ms per token,  1196.45 tokens per second)
+2.52.637.764 I slot print_timing: id  0 | task 109 |        eval time =      71.58 ms /     5 tokens (   17.89 ms per token,    55.88 tokens per second)
+2.52.637.765 I slot print_timing: id  0 | task 109 |       total time =    2131.83 ms /  2470 tokens
+2.52.637.766 I slot print_timing: id  0 | task 109 |    graphs reused =         10
+2.52.637.805 I slot      release: id  0 | task 109 | stop processing: n_tokens = 2469, truncated = 0
+3.16.232.385 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 0.429 (> 0.100 thold), f_keep = 0.005
+3.17.005.954 I slot launch_slot_: id  0 | task 118 | processing task, is_child = 0
+3.17.122.792 I slot print_timing: id  0 | task 118 | prompt eval time =      99.63 ms /    28 tokens (    3.56 ms per token,   281.05 tokens per second)
+3.17.122.798 I slot print_timing: id  0 | task 118 |        eval time =      17.17 ms /     2 tokens (   17.17 ms per token,    58.24 tokens per second)
+3.17.122.800 I slot print_timing: id  0 | task 118 |       total time =     116.80 ms /    30 tokens
+3.17.122.800 I slot print_timing: id  0 | task 118 |    graphs reused =         10
+3.17.122.825 I slot      release: id  0 | task 118 | stop processing: n_tokens = 29, truncated = 0
+3.17.135.603 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = 196571656
+3.17.153.668 I slot launch_slot_: id  0 | task 123 | processing task, is_child = 0
+3.19.295.506 I slot print_timing: id  0 | task 123 | prompt eval time =    2067.70 ms /  2463 tokens (    0.84 ms per token,  1191.18 tokens per second)
+3.19.295.513 I slot print_timing: id  0 | task 123 |        eval time =      74.11 ms /     5 tokens (   18.53 ms per token,    53.98 tokens per second)
+3.19.295.514 I slot print_timing: id  0 | task 123 |       total time =    2141.81 ms /  2468 tokens
+3.19.295.515 I slot print_timing: id  0 | task 123 |    graphs reused =         13
+3.19.295.577 I slot      release: id  0 | task 123 | stop processing: n_tokens = 2467, truncated = 0
+3.19.330.923 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.19.331.147 I slot launch_slot_: id  0 | task 132 | processing task, is_child = 0
+3.21.462.375 I slot print_timing: id  0 | task 132 | prompt eval time =    2059.94 ms /  2463 tokens (    0.84 ms per token,  1195.67 tokens per second)
+3.21.462.396 I slot print_timing: id  0 | task 132 |        eval time =      71.26 ms /     5 tokens (   17.81 ms per token,    56.14 tokens per second)
+3.21.462.399 I slot print_timing: id  0 | task 132 |       total time =    2131.19 ms /  2468 tokens
+3.21.462.401 I slot print_timing: id  0 | task 132 |    graphs reused =         16
+3.21.462.499 I slot      release: id  0 | task 132 | stop processing: n_tokens = 2467, truncated = 0
+3.21.487.596 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.21.487.778 I slot launch_slot_: id  0 | task 141 | processing task, is_child = 0
+3.23.661.681 I slot print_timing: id  0 | task 141 | prompt eval time =    2100.03 ms /  2463 tokens (    0.85 ms per token,  1172.84 tokens per second)
+3.23.661.688 I slot print_timing: id  0 | task 141 |        eval time =      73.85 ms /     5 tokens (   18.46 ms per token,    54.16 tokens per second)
+3.23.661.689 I slot print_timing: id  0 | task 141 |       total time =    2173.88 ms /  2468 tokens
+3.23.661.690 I slot print_timing: id  0 | task 141 |    graphs reused =         19
+3.23.661.760 I slot      release: id  0 | task 141 | stop processing: n_tokens = 2467, truncated = 0
+3.23.696.726 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.23.696.904 I slot launch_slot_: id  0 | task 150 | processing task, is_child = 0
+3.25.856.538 I slot print_timing: id  0 | task 150 | prompt eval time =    2086.02 ms /  2463 tokens (    0.85 ms per token,  1180.72 tokens per second)
+3.25.856.545 I slot print_timing: id  0 | task 150 |        eval time =      73.59 ms /     5 tokens (   18.40 ms per token,    54.36 tokens per second)
+3.25.856.546 I slot print_timing: id  0 | task 150 |       total time =    2159.61 ms /  2468 tokens
+3.25.856.548 I slot print_timing: id  0 | task 150 |    graphs reused =         22
+3.25.856.616 I slot      release: id  0 | task 150 | stop processing: n_tokens = 2467, truncated = 0
+3.25.886.677 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.25.886.852 I slot launch_slot_: id  0 | task 159 | processing task, is_child = 0
+3.28.040.278 I slot print_timing: id  0 | task 159 | prompt eval time =    2081.08 ms /  2463 tokens (    0.84 ms per token,  1183.52 tokens per second)
+3.28.040.288 I slot print_timing: id  0 | task 159 |        eval time =      72.32 ms /     5 tokens (   18.08 ms per token,    55.31 tokens per second)
+3.28.040.289 I slot print_timing: id  0 | task 159 |       total time =    2153.40 ms /  2468 tokens
+3.28.040.290 I slot print_timing: id  0 | task 159 |    graphs reused =         25
+3.28.040.358 I slot      release: id  0 | task 159 | stop processing: n_tokens = 2467, truncated = 0
+3.28.058.778 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.28.059.004 I slot launch_slot_: id  0 | task 168 | processing task, is_child = 0
+3.30.225.040 I slot print_timing: id  0 | task 168 | prompt eval time =    2089.74 ms /  2463 tokens (    0.85 ms per token,  1178.62 tokens per second)
+3.30.225.053 I slot print_timing: id  0 | task 168 |        eval time =      76.26 ms /     5 tokens (   19.07 ms per token,    52.45 tokens per second)
+3.30.225.055 I slot print_timing: id  0 | task 168 |       total time =    2166.00 ms /  2468 tokens
+3.30.225.056 I slot print_timing: id  0 | task 168 |    graphs reused =         28
+3.30.225.126 I slot      release: id  0 | task 168 | stop processing: n_tokens = 2467, truncated = 0
+3.30.249.310 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.30.249.584 I slot launch_slot_: id  0 | task 177 | processing task, is_child = 0
+3.32.416.181 I slot print_timing: id  0 | task 177 | prompt eval time =    2097.21 ms /  2463 tokens (    0.85 ms per token,  1174.42 tokens per second)
+3.32.416.188 I slot print_timing: id  0 | task 177 |        eval time =      69.37 ms /     5 tokens (   17.34 ms per token,    57.66 tokens per second)
+3.32.416.189 I slot print_timing: id  0 | task 177 |       total time =    2166.58 ms /  2468 tokens
+3.32.416.190 I slot print_timing: id  0 | task 177 |    graphs reused =         31
+3.32.416.251 I slot      release: id  0 | task 177 | stop processing: n_tokens = 2467, truncated = 0
+3.32.443.793 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.32.444.074 I slot launch_slot_: id  0 | task 186 | processing task, is_child = 0
+3.34.582.313 I slot print_timing: id  0 | task 186 | prompt eval time =    2065.17 ms /  2463 tokens (    0.84 ms per token,  1192.64 tokens per second)
+3.34.582.320 I slot print_timing: id  0 | task 186 |        eval time =      73.05 ms /     5 tokens (   18.26 ms per token,    54.76 tokens per second)
+3.34.582.321 I slot print_timing: id  0 | task 186 |       total time =    2138.21 ms /  2468 tokens
+3.34.582.322 I slot print_timing: id  0 | task 186 |    graphs reused =         34
+3.34.582.384 I slot      release: id  0 | task 186 | stop processing: n_tokens = 2467, truncated = 0
+3.34.618.944 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.34.619.210 I slot launch_slot_: id  0 | task 195 | processing task, is_child = 0
+3.36.784.049 I slot print_timing: id  0 | task 195 | prompt eval time =    2088.51 ms /  2463 tokens (    0.85 ms per token,  1179.31 tokens per second)
+3.36.784.056 I slot print_timing: id  0 | task 195 |        eval time =      76.30 ms /     5 tokens (   19.07 ms per token,    52.43 tokens per second)
+3.36.784.057 I slot print_timing: id  0 | task 195 |       total time =    2164.81 ms /  2468 tokens
+3.36.784.058 I slot print_timing: id  0 | task 195 |    graphs reused =         37
+3.36.784.122 I slot      release: id  0 | task 195 | stop processing: n_tokens = 2467, truncated = 0
+3.36.817.594 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.36.817.780 I slot launch_slot_: id  0 | task 204 | processing task, is_child = 0
+3.38.954.816 I slot print_timing: id  0 | task 204 | prompt eval time =    2063.18 ms /  2463 tokens (    0.84 ms per token,  1193.79 tokens per second)
+3.38.954.823 I slot print_timing: id  0 | task 204 |        eval time =      73.84 ms /     5 tokens (   18.46 ms per token,    54.17 tokens per second)
+3.38.954.824 I slot print_timing: id  0 | task 204 |       total time =    2137.02 ms /  2468 tokens
+3.38.954.825 I slot print_timing: id  0 | task 204 |    graphs reused =         40
+3.38.954.888 I slot      release: id  0 | task 204 | stop processing: n_tokens = 2467, truncated = 0
+3.38.987.689 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.38.987.869 I slot launch_slot_: id  0 | task 213 | processing task, is_child = 0
+3.41.149.401 I slot print_timing: id  0 | task 213 | prompt eval time =    2088.78 ms /  2463 tokens (    0.85 ms per token,  1179.16 tokens per second)
+3.41.149.410 I slot print_timing: id  0 | task 213 |        eval time =      72.73 ms /     5 tokens (   18.18 ms per token,    54.99 tokens per second)
+3.41.149.412 I slot print_timing: id  0 | task 213 |       total time =    2161.52 ms /  2468 tokens
+3.41.149.413 I slot print_timing: id  0 | task 213 |    graphs reused =         43
+3.41.149.476 I slot      release: id  0 | task 213 | stop processing: n_tokens = 2467, truncated = 0
+3.41.180.413 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.41.180.581 I slot launch_slot_: id  0 | task 222 | processing task, is_child = 0
+3.43.301.806 I slot print_timing: id  0 | task 222 | prompt eval time =    2050.16 ms /  2463 tokens (    0.83 ms per token,  1201.37 tokens per second)
+3.43.301.812 I slot print_timing: id  0 | task 222 |        eval time =      71.04 ms /     5 tokens (   17.76 ms per token,    56.30 tokens per second)
+3.43.301.814 I slot print_timing: id  0 | task 222 |       total time =    2121.21 ms /  2468 tokens
+3.43.301.815 I slot print_timing: id  0 | task 222 |    graphs reused =         46
+3.43.301.880 I slot      release: id  0 | task 222 | stop processing: n_tokens = 2467, truncated = 0
+3.43.334.231 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.43.334.497 I slot launch_slot_: id  0 | task 231 | processing task, is_child = 0
+3.45.479.197 I slot print_timing: id  0 | task 231 | prompt eval time =    2074.01 ms /  2463 tokens (    0.84 ms per token,  1187.55 tokens per second)
+3.45.479.205 I slot print_timing: id  0 | task 231 |        eval time =      70.67 ms /     5 tokens (   17.67 ms per token,    56.60 tokens per second)
+3.45.479.207 I slot print_timing: id  0 | task 231 |       total time =    2144.68 ms /  2468 tokens
+3.45.479.207 I slot print_timing: id  0 | task 231 |    graphs reused =         49
+3.45.479.251 I slot      release: id  0 | task 231 | stop processing: n_tokens = 2467, truncated = 0
+3.45.514.475 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.45.514.871 I slot launch_slot_: id  0 | task 240 | processing task, is_child = 0
+3.47.666.125 I slot print_timing: id  0 | task 240 | prompt eval time =    2079.71 ms /  2463 tokens (    0.84 ms per token,  1184.30 tokens per second)
+3.47.666.132 I slot print_timing: id  0 | task 240 |        eval time =      71.51 ms /     5 tokens (   17.88 ms per token,    55.94 tokens per second)
+3.47.666.133 I slot print_timing: id  0 | task 240 |       total time =    2151.22 ms /  2468 tokens
+3.47.666.134 I slot print_timing: id  0 | task 240 |    graphs reused =         52
+3.47.666.198 I slot      release: id  0 | task 240 | stop processing: n_tokens = 2467, truncated = 0
+3.47.702.034 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.47.702.214 I slot launch_slot_: id  0 | task 249 | processing task, is_child = 0
+3.49.850.027 I slot print_timing: id  0 | task 249 | prompt eval time =    2078.09 ms /  2463 tokens (    0.84 ms per token,  1185.22 tokens per second)
+3.49.850.033 I slot print_timing: id  0 | task 249 |        eval time =      69.71 ms /     5 tokens (   17.43 ms per token,    57.38 tokens per second)
+3.49.850.035 I slot print_timing: id  0 | task 249 |       total time =    2147.80 ms /  2468 tokens
+3.49.850.035 I slot print_timing: id  0 | task 249 |    graphs reused =         55
+3.49.850.095 I slot      release: id  0 | task 249 | stop processing: n_tokens = 2467, truncated = 0
+3.49.875.200 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.49.875.613 I slot launch_slot_: id  0 | task 258 | processing task, is_child = 0
+3.52.017.757 I slot print_timing: id  0 | task 258 | prompt eval time =    2067.63 ms /  2463 tokens (    0.84 ms per token,  1191.22 tokens per second)
+3.52.017.763 I slot print_timing: id  0 | task 258 |        eval time =      74.49 ms /     5 tokens (   18.62 ms per token,    53.70 tokens per second)
+3.52.017.765 I slot print_timing: id  0 | task 258 |       total time =    2142.11 ms /  2468 tokens
+3.52.017.766 I slot print_timing: id  0 | task 258 |    graphs reused =         58
+3.52.017.832 I slot      release: id  0 | task 258 | stop processing: n_tokens = 2467, truncated = 0
+3.52.029.474 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.52.029.683 I slot launch_slot_: id  0 | task 267 | processing task, is_child = 0
+3.54.166.906 I slot print_timing: id  0 | task 267 | prompt eval time =    2066.09 ms /  2463 tokens (    0.84 ms per token,  1192.11 tokens per second)
+3.54.166.912 I slot print_timing: id  0 | task 267 |        eval time =      71.11 ms /     5 tokens (   17.78 ms per token,    56.25 tokens per second)
+3.54.166.914 I slot print_timing: id  0 | task 267 |       total time =    2137.20 ms /  2468 tokens
+3.54.166.915 I slot print_timing: id  0 | task 267 |    graphs reused =         61
+3.54.166.975 I slot      release: id  0 | task 267 | stop processing: n_tokens = 2467, truncated = 0
+3.54.200.277 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.54.200.454 I slot launch_slot_: id  0 | task 276 | processing task, is_child = 0
+3.56.348.622 I slot print_timing: id  0 | task 276 | prompt eval time =    2073.38 ms /  2463 tokens (    0.84 ms per token,  1187.92 tokens per second)
+3.56.348.630 I slot print_timing: id  0 | task 276 |        eval time =      74.77 ms /     5 tokens (   18.69 ms per token,    53.50 tokens per second)
+3.56.348.631 I slot print_timing: id  0 | task 276 |       total time =    2148.15 ms /  2468 tokens
+3.56.348.632 I slot print_timing: id  0 | task 276 |    graphs reused =         64
+3.56.348.697 I slot      release: id  0 | task 276 | stop processing: n_tokens = 2467, truncated = 0
+3.56.384.477 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.56.384.646 I slot launch_slot_: id  0 | task 285 | processing task, is_child = 0
+3.58.539.474 I slot print_timing: id  0 | task 285 | prompt eval time =    2083.22 ms /  2463 tokens (    0.85 ms per token,  1182.30 tokens per second)
+3.58.539.482 I slot print_timing: id  0 | task 285 |        eval time =      71.59 ms /     5 tokens (   17.90 ms per token,    55.88 tokens per second)
+3.58.539.483 I slot print_timing: id  0 | task 285 |       total time =    2154.81 ms /  2468 tokens
+3.58.539.484 I slot print_timing: id  0 | task 285 |    graphs reused =         67
+3.58.539.552 I slot      release: id  0 | task 285 | stop processing: n_tokens = 2467, truncated = 0
+3.58.551.461 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+3.58.551.747 I slot launch_slot_: id  0 | task 294 | processing task, is_child = 0
+4.00.720.076 I slot print_timing: id  0 | task 294 | prompt eval time =    2091.43 ms /  2463 tokens (    0.85 ms per token,  1177.67 tokens per second)
+4.00.720.086 I slot print_timing: id  0 | task 294 |        eval time =      76.88 ms /     5 tokens (   19.22 ms per token,    52.03 tokens per second)
+4.00.720.088 I slot print_timing: id  0 | task 294 |       total time =    2168.30 ms /  2468 tokens
+4.00.720.089 I slot print_timing: id  0 | task 294 |    graphs reused =         70
+4.00.720.154 I slot      release: id  0 | task 294 | stop processing: n_tokens = 2467, truncated = 0
+4.00.732.918 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+4.00.733.158 I slot launch_slot_: id  0 | task 303 | processing task, is_child = 0
+4.02.892.100 I slot print_timing: id  0 | task 303 | prompt eval time =    2087.92 ms /  2463 tokens (    0.85 ms per token,  1179.64 tokens per second)
+4.02.892.106 I slot print_timing: id  0 | task 303 |        eval time =      71.00 ms /     5 tokens (   17.75 ms per token,    56.34 tokens per second)
+4.02.892.107 I slot print_timing: id  0 | task 303 |       total time =    2158.92 ms /  2468 tokens
+4.02.892.108 I slot print_timing: id  0 | task 303 |    graphs reused =         73
+4.02.892.169 I slot      release: id  0 | task 303 | stop processing: n_tokens = 2467, truncated = 0
+4.02.922.891 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+4.02.923.085 I slot launch_slot_: id  0 | task 312 | processing task, is_child = 0
+4.05.074.275 I slot print_timing: id  0 | task 312 | prompt eval time =    2077.90 ms /  2463 tokens (    0.84 ms per token,  1185.33 tokens per second)
+4.05.074.290 I slot print_timing: id  0 | task 312 |        eval time =      73.25 ms /     5 tokens (   18.31 ms per token,    54.61 tokens per second)
+4.05.074.292 I slot print_timing: id  0 | task 312 |       total time =    2151.16 ms /  2468 tokens
+4.05.074.294 I slot print_timing: id  0 | task 312 |    graphs reused =         76
+4.05.074.395 I slot      release: id  0 | task 312 | stop processing: n_tokens = 2467, truncated = 0
+4.49.094.433 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 0.429 (> 0.100 thold), f_keep = 0.005
+4.49.500.550 I slot launch_slot_: id  0 | task 321 | processing task, is_child = 0
+4.49.619.376 I slot print_timing: id  0 | task 321 | prompt eval time =     100.54 ms /    28 tokens (    3.59 ms per token,   278.49 tokens per second)
+4.49.619.389 I slot print_timing: id  0 | task 321 |        eval time =      18.23 ms /     2 tokens (   18.23 ms per token,    54.84 tokens per second)
+4.49.619.391 I slot print_timing: id  0 | task 321 |       total time =     118.78 ms /    30 tokens
+4.49.619.393 I slot print_timing: id  0 | task 321 |    graphs reused =         76
+4.49.619.438 I slot      release: id  0 | task 321 | stop processing: n_tokens = 29, truncated = 0
+4.49.644.166 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = 289068269
+4.49.662.392 I slot launch_slot_: id  0 | task 326 | processing task, is_child = 0
+4.51.785.841 I slot print_timing: id  0 | task 326 | prompt eval time =    2053.07 ms /  2463 tokens (    0.83 ms per token,  1199.67 tokens per second)
+4.51.785.848 I slot print_timing: id  0 | task 326 |        eval time =      70.34 ms /     5 tokens (   17.59 ms per token,    56.87 tokens per second)
+4.51.785.849 I slot print_timing: id  0 | task 326 |       total time =    2123.41 ms /  2468 tokens
+4.51.785.850 I slot print_timing: id  0 | task 326 |    graphs reused =         79
+4.51.785.912 I slot      release: id  0 | task 326 | stop processing: n_tokens = 2467, truncated = 0
+4.51.815.045 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998
+4.51.815.252 I slot launch_slot_: id  0 | task 335 | processing task, is_child = 0
+4.53.994.661 I slot print_timing: id  0 | task 335 | prompt eval time =    2104.71 ms /  2463 tokens (    0.85 ms per token,  1170.23 tokens per second)
+4.53.994.671 I slot print_timing: id  0 | task 335 |        eval time =      74.68 ms /     5 tokens (   18.67 ms per token,    53.56 tokens per second)
+4.53.994.672 I slot print_timing: id  0 | task 335 |       total time =    2179.39 ms /  2468 tokens
+4.53.994.673 I slot print_timing: id  0 | task 335 |    graphs reused =         82
+4.53.994.745 I slot      release: id  0 | task 335 | stop processing: n_tokens = 2467, truncated = 0
+4.54.027.777 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 0.400 (> 0.100 thold), f_keep = 0.005
+4.54.411.656 I slot launch_slot_: id  0 | task 344 | processing task, is_child = 0
+4.54.538.040 W srv          stop: cancel task, id_task = 344
+4.54.553.598 I slot      release: id  0 | task 344 | stop processing: n_tokens = 32, truncated = 0
+4.54.553.624 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, f_sim_best = 0.429 (> 0.100 thold), f_keep = 0.375
+4.54.568.422 I slot launch_slot_: id  0 | task 351 | processing task, is_child = 0
+4.54.685.918 I slot print_timing: id  0 | task 351 | prompt eval time =      99.32 ms /    28 tokens (    3.55 ms per token,   281.92 tokens per second)
+4.54.685.928 I slot print_timing: id  0 | task 351 |        eval time =      18.13 ms /     2 tokens (   18.13 ms per token,    55.16 tokens per second)
+4.54.685.929 I slot print_timing: id  0 | task 351 |       total time =     117.45 ms /    30 tokens
+4.54.685.930 I slot print_timing: id  0 | task 351 |    graphs reused =         83
+4.54.685.959 I slot      release: id  0 | task 351 | stop processing: n_tokens = 29, truncated = 0
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_cancel_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_cancel_20260923.json
new file mode 100644
index 000000000..5ede5122d
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_cancel_20260923.json
@@ -0,0 +1,101 @@
+{
+  "scope": "Standalone llama.cpp full-model text path; no Omni stage binding",
+  "platform": "Windows-11-10.0.26200-SP0",
+  "model_path": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+  "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4",
+  "server": {
+    "endpoint": "http://127.0.0.1:18173",
+    "build_info": "b11124-1a679828f",
+    "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+    "model_ftype": "Q4_K - Medium",
+    "total_slots": 1
+  },
+  "quality_checks": [
+    {
+      "case": "short",
+      "wall_s": 0.5372719999868423,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.166346100042574,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    }
+  ],
+  "warmups": [],
+  "samples": [
+    {
+      "case": "inventory_120",
+      "wall_s": 2.2086611000122502,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    }
+  ],
+  "cancellation": {
+    "first_delta": "1",
+    "stream_closed_after_first_content": true,
+    "first_delta_wall_s": 0.5234057001071051,
+    "recovery": {
+      "case": "short",
+      "wall_s": 0.16786349995527416,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    }
+  },
+  "profile": {
+    "case": "inventory_120",
+    "cache_prompt": false,
+    "repeats": 1,
+    "concurrency": 1,
+    "p50_wall_s": 2.2086611000122502,
+    "p95_wall_s": 2.2086611000122502,
+    "min_wall_s": 2.2086611000122502,
+    "max_wall_s": 2.2086611000122502,
+    "cached_prompt_tokens": [
+      0
+    ]
+  },
+  "status": "passed"
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_inventory_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_inventory_20260923.json
new file mode 100644
index 000000000..086b2c7e8
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_inventory_20260923.json
@@ -0,0 +1,36 @@
+{
+  "choices": [
+    {
+      "finish_reason": "stop",
+      "index": 0,
+      "message": {
+        "role": "assistant",
+        "content": "1511"
+      }
+    }
+  ],
+  "created": 1790160209,
+  "model": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+  "system_fingerprint": "b11124-1a679828f",
+  "object": "chat.completion",
+  "usage": {
+    "completion_tokens": 5,
+    "prompt_tokens": 2465,
+    "total_tokens": 2470,
+    "prompt_tokens_details": {
+      "cached_tokens": 9
+    }
+  },
+  "id": "chatcmpl-mCtMsGHw7erZQESvX4Fmh29LVtKo261I",
+  "timings": {
+    "cache_n": 9,
+    "prompt_n": 2456,
+    "prompt_ms": 4845.128,
+    "prompt_per_token_ms": 1.972771986970684,
+    "prompt_per_second": 506.9009528747228,
+    "predicted_n": 5,
+    "predicted_ms": 77.149,
+    "predicted_per_token_ms": 19.28725,
+    "predicted_per_second": 51.84772323685336
+  }
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_inventory_profile20_nocache_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_inventory_profile20_nocache_20260923.json
new file mode 100644
index 000000000..143e73e0b
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_inventory_profile20_nocache_20260923.json
@@ -0,0 +1,420 @@
+{
+  "scope": "Standalone llama.cpp full-model text path; no Omni stage binding",
+  "platform": "Windows-11-10.0.26200-SP0",
+  "model_path": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+  "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4",
+  "server": {
+    "endpoint": "http://127.0.0.1:18173",
+    "build_info": "b11124-1a679828f",
+    "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+    "model_ftype": "Q4_K - Medium",
+    "total_slots": 1
+  },
+  "quality_checks": [
+    {
+      "case": "short",
+      "wall_s": 0.9138094999361783,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.172446300042793,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    }
+  ],
+  "warmups": [
+    {
+      "case": "inventory_120",
+      "wall_s": 2.1668586999876425,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    }
+  ],
+  "samples": [
+    {
+      "case": "inventory_120",
+      "wall_s": 2.199468299979344,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.194800299941562,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.183633800013922,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.18500210007187,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.19071440002881,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.1662180999992415,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.201768200029619,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.170874699950218,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.194402000051923,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.15237570006866,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.1774167999392375,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.186952000018209,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.183857300085947,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.167893800069578,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.149119200068526,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.181731199962087,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.1910894999746233,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.180448899976909,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.171951900003478,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 2.182072600000538,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 0
+        }
+      }
+    }
+  ],
+  "profile": {
+    "case": "inventory_120",
+    "cache_prompt": false,
+    "repeats": 20,
+    "concurrency": 1,
+    "p50_wall_s": 2.182072600000538,
+    "p95_wall_s": 2.199468299979344,
+    "min_wall_s": 2.149119200068526,
+    "max_wall_s": 2.201768200029619,
+    "cached_prompt_tokens": [
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0
+    ]
+  },
+  "status": "passed"
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_profile20_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_profile20_20260923.json
new file mode 100644
index 000000000..15f26b9a2
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_profile20_20260923.json
@@ -0,0 +1,418 @@
+{
+  "scope": "Standalone llama.cpp full-model text path; no Omni stage binding",
+  "platform": "Windows-11-10.0.26200-SP0",
+  "model_path": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+  "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4",
+  "server": {
+    "endpoint": "http://127.0.0.1:18173",
+    "build_info": "b11124-1a679828f",
+    "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+    "model_ftype": "Q4_K - Medium",
+    "total_slots": 1
+  },
+  "quality_checks": [
+    {
+      "case": "short",
+      "wall_s": 0.4887153999879956,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "inventory_120",
+      "wall_s": 0.7786927999695763,
+      "content": "1511",
+      "expected": "1511",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 5,
+        "prompt_tokens": 2463,
+        "total_tokens": 2468,
+        "prompt_tokens_details": {
+          "cached_tokens": 1948
+        }
+      }
+    }
+  ],
+  "warmups": [
+    {
+      "case": "short",
+      "wall_s": 0.48504260007757694,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    }
+  ],
+  "samples": [
+    {
+      "case": "short",
+      "wall_s": 0.09932070004288107,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.0984904000069946,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.09055399999488145,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.09084180009085685,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.07738210004754364,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.10014190000947565,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.09278060006909072,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.09588450007140636,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.10041170008480549,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.07804280007258058,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.09451030008494854,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.09293620008975267,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.07760550000239164,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.09585659997537732,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.09953960007987916,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.08980209997389466,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.09244449995458126,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.09291470004245639,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.09785889997147024,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    },
+    {
+      "case": "short",
+      "wall_s": 0.08484689996112138,
+      "content": "Paris",
+      "expected": "Paris",
+      "correct": true,
+      "finish_reason": "stop",
+      "usage": {
+        "completion_tokens": 2,
+        "prompt_tokens": 28,
+        "total_tokens": 30,
+        "prompt_tokens_details": {
+          "cached_tokens": 23
+        }
+      }
+    }
+  ],
+  "profile": {
+    "repeats": 20,
+    "concurrency": 1,
+    "p50_wall_s": 0.09291470004245639,
+    "p95_wall_s": 0.10014190000947565,
+    "min_wall_s": 0.07738210004754364,
+    "max_wall_s": 0.10041170008480549,
+    "cached_prompt_tokens": [
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23,
+      23
+    ]
+  },
+  "status": "passed"
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_short_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_short_20260923.json
new file mode 100644
index 000000000..707887ee1
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/spark_igpu_server_short_20260923.json
@@ -0,0 +1,36 @@
+{
+  "choices": [
+    {
+      "finish_reason": "stop",
+      "index": 0,
+      "message": {
+        "role": "assistant",
+        "content": "Paris"
+      }
+    }
+  ],
+  "created": 1790160196,
+  "model": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf",
+  "system_fingerprint": "b11124-1a679828f",
+  "object": "chat.completion",
+  "usage": {
+    "completion_tokens": 2,
+    "prompt_tokens": 28,
+    "total_tokens": 30,
+    "prompt_tokens_details": {
+      "cached_tokens": 0
+    }
+  },
+  "id": "chatcmpl-UjzKml8myD3dr9aDdN5xctZystb7iwut",
+  "timings": {
+    "cache_n": 0,
+    "prompt_n": 28,
+    "prompt_ms": 2196.018,
+    "prompt_per_token_ms": 78.42921428571428,
+    "prompt_per_second": 12.750350862333551,
+    "predicted_n": 2,
+    "predicted_ms": 20.534,
+    "predicted_per_token_ms": 20.534,
+    "predicted_per_second": 48.699717541638265
+  }
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/vulkaninfo_summary.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/vulkaninfo_summary.log
new file mode 100644
index 000000000..91041de9c
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/vulkaninfo_summary.log
@@ -0,0 +1,65 @@
+WARNING: [Loader Message] Code 0 : windows_read_data_files_in_registry: Registry lookup failed to get layer manifest files.
+==========
+VULKANINFO
+==========
+
+Vulkan Instance Version: 1.4.341
+
+
+Instance Extensions: count = 20
+-------------------------------
+VK_EXT_debug_report                    : extension revision 10
+VK_EXT_debug_utils                     : extension revision 2
+VK_EXT_direct_mode_display             : extension revision 1
+VK_EXT_surface_maintenance1            : extension revision 1
+VK_EXT_swapchain_colorspace            : extension revision 5
+VK_KHR_device_group_creation           : extension revision 1
+VK_KHR_display                         : extension revision 23
+VK_KHR_external_fence_capabilities     : extension revision 1
+VK_KHR_external_memory_capabilities    : extension revision 1
+VK_KHR_external_semaphore_capabilities : extension revision 1
+VK_KHR_get_display_properties2         : extension revision 1
+VK_KHR_get_physical_device_properties2 : extension revision 2
+VK_KHR_get_surface_capabilities2       : extension revision 1
+VK_KHR_portability_enumeration         : extension revision 1
+VK_KHR_surface                         : extension revision 25
+VK_KHR_surface_maintenance1            : extension revision 1
+VK_KHR_surface_protected_capabilities  : extension revision 1
+VK_KHR_win32_surface                   : extension revision 6
+VK_LUNARG_direct_driver_loading        : extension revision 1
+VK_NV_external_memory_capabilities     : extension revision 1
+
+Instance Layers: count = 3
+--------------------------
+VK_LAYER_AMD_switchable_graphics AMD switchable graphics layer 1.4.320  version 1
+VK_LAYER_NV_optimus              NVIDIA Optimus layer          1.4.341  version 1
+VK_LAYER_NV_present              NVIDIA Presentation Layer     1.4.341  version 1
+
+Devices:
+========
+GPU0:
+	apiVersion         = 1.4.341
+	driverVersion      = 610.71.0.0
+	vendorID           = 0x10de
+	deviceID           = 0x2c18
+	deviceType         = PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
+	deviceName         = NVIDIA GeForce RTX 5090 Laptop GPU
+	driverID           = DRIVER_ID_NVIDIA_PROPRIETARY
+	driverName         = NVIDIA
+	driverInfo         = 610.71
+	conformanceVersion = 1.4.3.3
+	deviceUUID         = bf6e910e-657c-15d7-c558-f169df134d8a
+	driverUUID         = 32da8b74-b45b-55e0-9ba2-ab437d951f07
+GPU1:
+	apiVersion         = 1.4.320
+	driverVersion      = 2.0.359
+	vendorID           = 0x1002
+	deviceID           = 0x150e
+	deviceType         = PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
+	deviceName         = AMD Radeon(TM) 890M Graphics
+	driverID           = DRIVER_ID_AMD_PROPRIETARY
+	driverName         = AMD proprietary driver
+	driverInfo         = 25.20.18.06 (LLPC)
+	conformanceVersion = 1.4.0.0
+	deviceUUID         = 00000000-6500-0000-0000-000000000000
+	driverUUID         = 414d442d-5749-4e2d-4452-560000000000
diff --git a/benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md b/benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md
index 6448e30e8..3d824a6f4 100644
--- a/benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md
+++ b/benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md
@@ -1,18 +1,18 @@
 # E2E check and profiling
 
-All 60 pairings have a disposition. The timing tables below remain the 2026-09-22 measurements. The configuration matrix and affected per-cell reasons incorporate the [2026-09-23 scoped recovery runs](../../../e2e_recovery_20260923/README.md) and [WSL CPU/native Windows expansion](../../../e2e_expansion_20260923/README.md). The expansion adds separate 20-request serial MiniCPM-o text-to-speech profiles on WSL CPU and native Windows GPU, plus a separately profiled Spark output-head component on the HX370 AMD NPU. Profiling completion does not imply full release qualification.
+All 60 pairings have a disposition. The timing tables below remain the 2026-09-22 measurements. The configuration matrix and affected per-cell reasons incorporate the [2026-09-23 scoped recovery runs](../../../e2e_recovery_20260923/README.md) and [WSL CPU/native Windows expansion](../../../e2e_expansion_20260923/README.md). The expansion adds separate 20-request serial MiniCPM-o text-to-speech profiles on WSL CPU and native Windows GPU, a Spark output-head component profile on the HX370 AMD NPU, and complete Spark text-request profiles through standalone llama.cpp on native Windows CPU and Radeon 890M. Profiling completion does not imply full release qualification or Omni integration.
 
 ## Configuration × model
 
-NOT E2E means no complete device/model workload was verified. “Scoped E2E” identifies only the named input/output path; “synthetic policy pass” does not establish a real robot-task result. See the per-cell reasons below.
+NOT E2E means no complete device/model workload was verified. “Scoped E2E” identifies only the named input/output path; “standalone” means the whole model ran outside Omni and still needs a StageRuntime binding. “Synthetic policy pass” does not establish a real robot-task result. See the per-cell reasons below.
 
 | Device | Spark-X2.5 | Qwen3-TTS 0.6B CustomVoice | Qwen3.8-27B | MiniCPM-o 4.5 | InternVLA-A1 |
 |---|---|---|---|---|---|
 | PC HX370: CPU / WSL | profile completed | scoped E2E: short stream | NOT E2E | scoped E2E: text/image/audio → speech | synthetic policy pass |
-| PC HX370: CPU / native Windows | NOT E2E | NOT E2E | NOT E2E | NOT E2E | synthetic policy pass |
+| PC HX370: CPU / native Windows | scoped E2E: standalone Spark text | NOT E2E | NOT E2E | NOT E2E | synthetic policy pass |
 | PC HX370 + RTX 5090 Laptop / WSL | profile completed | profile completed | scoped E2E: text + image | scoped E2E: text/image/audio → speech | synthetic policy pass |
 | PC HX370 + RTX 5090 Laptop / Windows | profile completed | profile completed | NOT E2E | scoped E2E: text/image/audio → speech | synthetic policy pass |
-| PC HX370 + Radeon 890M / Windows worker | NOT E2E | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
+| PC HX370 + Radeon 890M / Windows worker | scoped E2E: standalone Spark text | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
 | PC HX370 + AMD NPU / Windows worker | NOT E2E (output-head component passed) | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
 | PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | NOT E2E | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
 | PC Snapdragon X Elite CRD / AI Hub | NOT E2E | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
@@ -30,7 +30,7 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i
 | PC HX370: CPU / WSL | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No compatible CPU artifact/deployment for this model was found in the checked model roots. Available Qwen 27B FP8/NVFP4 CUDA artifacts do not establish a CPU route. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | PC HX370: CPU / WSL | MiniCPM-o 4.5 | SCOPED_E2E | The verified checkpoint completed real three-stage text, one synthetic image and one synthetic audio input to text+speech on the vLLM 0.28 CPU backend in separate bounded plans. The 448×448 red square was described correctly with a nonzero 5.2 s WAV in one 37.70 s request; the 2 s/440 Hz tone was described as “A loud and high pitched beep” with a nonzero 2.16 s WAV in one 17.30 s request. CPU wording/token IDs differ from the WSL GPU runs, with no task-quality threshold established. A separate 1-warmup/20-measured serial **text** run had request wall p50 21.86 s/p95 22.97 s. WSL swap reached 5.75 GB during that text profile; manual snapshots in the multimodal runs showed heavy swap use but do not prove peak memory. Both multimodal runs force-killed stage 0 at shutdown and reported shared-memory cleanup warnings. | Validate real image/audio suites, speech intelligibility, video input, concurrency, interruption, admission/loading peak, clean shutdown and sustained thermal behavior before broader CPU qualification; [CPU evidence](../../../e2e_expansion_20260923/README.md). |
 | PC HX370: CPU / WSL | InternVLA-A1 | SYNTHETIC_POLICY_PASS | The Place_Markpen checkpoint strict-loaded in CPU-only PyTorch and produced finite `[1, 50, 32]` actions from zero observations/noise in one 3.25 s forward. Compared with the same CUDA checkpoint/input, relative L2 error was 1.30% and cosine similarity 0.999917; no task-quality tolerance is established. | Validate real observations against reference actions, units/timestamps, repeated latency and memory before robot-task qualification; [CPU evidence](../../../e2e_expansion_20260923/README.md). |
-| PC HX370: CPU / native Windows | Spark-X2.5 | NOT_E2E_PROFILEABLE | The installed native Windows vLLM extension lacks cpu_attn_reshape_and_cache and cpu_attn_get_scheduler_metadata. The separate ORT CPU graph worker is not a whole-model text/TTS adapter; a qualified CPU build/binding is missing. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
+| PC HX370: CPU / native Windows | Spark-X2.5 | SCOPED_E2E_STANDALONE | The real Spark-X2.5-1.7B Q4_K_M GGUF ran a complete text path in native Windows llama.cpp CPU (`-dev none -ngl 0`): `Paris` and the correct `1511` answer to a 2,463-token inventory prompt, one warmup plus 20 uncached serial inventory requests with p50/p95 6.65/6.86 s and 20/20 correct, and a post-stream-close recovery request. This does not repair the missing vLLM CPU attention operators or establish an Omni stage binding. One repeated input, no loading peak or sustained thermal/power gate. | Bind the whole llama.cpp session to Omni's state, request/event, memory and cancellation contracts; validate broader text quality and sustained use; [standalone backend evidence](../../../e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/README.md). |
 | PC HX370: CPU / native Windows | Qwen3-TTS 0.6B CustomVoice | NOT_E2E_PROFILEABLE | The installed native Windows vLLM extension lacks cpu_attn_reshape_and_cache and cpu_attn_get_scheduler_metadata. The separate ORT CPU graph worker is not a whole-model text/TTS adapter; a qualified CPU build/binding is missing. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | PC HX370: CPU / native Windows | Qwen3.8-27B | NOT_E2E_PROFILEABLE | The installed native Windows vLLM extension lacks cpu_attn_reshape_and_cache and cpu_attn_get_scheduler_metadata. The separate ORT CPU graph worker is not a whole-model text/TTS adapter; a qualified CPU build/binding is missing. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | PC HX370: CPU / native Windows | MiniCPM-o 4.5 | NOT_E2E_PROFILEABLE | The exact checkpoint is verified in the local Linux model store, but the installed native Windows vLLM extension lacks required CPU attention operators; no native full-model run or memory admission exists. | Establish a native Windows CPU backend and model access, then run the full requested path; [weight evidence](../../../e2e_recovery_20260923/evidence/weights/minicpmo.json). |
@@ -45,7 +45,7 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i
 | PC HX370 + RTX 5090 Laptop / Windows | Qwen3.8-27B | NOT_E2E_PROFILEABLE | The Omni pipeline binding now exists and passed scoped text/image cases in WSL, but no native Windows Omni run has verified the NVFP4 checkpoint, memory admission or modality behavior. | Run the same checkpoint and explicit context on native Windows, then broaden quality and profiling; [WSL-only evidence](../../../e2e_recovery_20260923/README.md). |
 | PC HX370 + RTX 5090 Laptop / Windows | MiniCPM-o 4.5 | SCOPED_E2E | The verified checkpoint completed native Windows three-stage BF16 text, one synthetic image, and one synthetic audio input to text+speech in separate constrained plans with 5 GiB thinker host offload and explicit 2,048/512-token limits. Text and all 32 thinker token IDs matched WSL CUDA; a 5.2 s WAV was nonzero. The red-square image produced the same description and 27 thinker token IDs as WSL plus a nonzero 5.2 s WAV in 17.15 s. The 2 s/440 Hz tone produced the same whistling description and nine IDs as WSL plus a nonzero 2.16 s WAV in 8.26 s. The initial text run needed 900/600 s startup limits after a 300 s timeout; a separate 1-warmup/20-measured serial text run had request wall p50 18.38 s/p95 18.73 s, with sampled host available RAM as low as 0.78 GB and pagefile use up to 11.17 GB. These synthetic inputs do not establish general image/audio or speech quality. | Check real image/audio suites, video input, speech intelligibility/alignment, concurrency, interruption, loading-peak admission and sustained thermal/power behavior before broader qualification; [native evidence](../../../e2e_expansion_20260923/README.md). |
 | PC HX370 + RTX 5090 Laptop / Windows | InternVLA-A1 | SYNTHETIC_POLICY_PASS | The Place_Markpen checkpoint strict-loaded from a UNC path and produced finite `[1, 50, 32]` actions on synthetic zero observations/noise. Policy/output CUDA placement was checked; one native forward took 1.12 s, with 6.68 GB peak allocated. Relative L2 difference from WSL CUDA was 0.83% (no task tolerance). | Validate real observations/reference actions, units/timestamps, memory and repeated latency before task qualification; [native evidence](../../../e2e_expansion_20260923/README.md). |
-| PC HX370 + Radeon 890M / Windows worker | Spark-X2.5 | NOT_E2E_PROFILEABLE | No placement- and quality-verified whole-model graph bundle/stateful adapter for this model on this AMD route. Small backend graphs do not establish model support. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
+| PC HX370 + Radeon 890M / Windows worker | Spark-X2.5 | SCOPED_E2E_STANDALONE | The same real Spark-X2.5-1.7B Q4_K_M GGUF completed text prefill/decode via llama.cpp Vulkan1 on Radeon 890M; the placement log assigns Spark layers to Vulkan1, not the installed RTX 5090. It returned `Paris` and `1511` on the named short/2,463-token inventory inputs. One warmup plus 20 uncached serial inventory requests were 20/20 correct, p50/p95 wall 2.18/2.20 s. Closing a stream after its first content delta was followed by a correct new request. This is outside Omni and has only a post-profile process-memory snapshot, no loading peak, device-memory attribution or sustained power/thermal check. | Add a bounded llama.cpp whole-model stage to Omni, then verify placement, state, memory admission, token quality, cancellation and full-request benefit against CPU; [standalone backend evidence](../../../e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/README.md). |
 | PC HX370 + Radeon 890M / Windows worker | Qwen3-TTS 0.6B CustomVoice | NOT_E2E_PROFILEABLE | No placement- and quality-verified whole-model graph bundle/stateful adapter for this model on this AMD route. Small backend graphs do not establish model support. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | PC HX370 + Radeon 890M / Windows worker | Qwen3.8-27B | NOT_E2E_PROFILEABLE | FP32 448x448 vision tower: DirectML 85/113 nodes; normalized max error 9.095e-06 versus torch CPU. Historical component path, not full 27B or current graph-stage image-to-text integration. | Recover the hashed tower bundle and qualify its current Omni graph handoff with the language stage. |
 | PC HX370 + Radeon 890M / Windows worker | MiniCPM-o 4.5 | NOT_E2E_PROFILEABLE | No placement- and quality-verified whole-model graph bundle/stateful adapter for this model on this AMD route. Small backend graphs do not establish model support. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
diff --git a/docs/design/omni_edge_backends.md b/docs/design/omni_edge_backends.md
index 3afeb7bf9..068ff6191 100644
--- a/docs/design/omni_edge_backends.md
+++ b/docs/design/omni_edge_backends.md
@@ -51,7 +51,7 @@ memory-accounting and unavailable-device gaps remain documented repair items.
 
 ### Mobile and PC × model support matrix
 
-This section preserves the 2026-09-22 audit. For the latest per-cell disposition and 2026-09-23 follow-up evidence, use the [current 60-cell E2E matrix](../../benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md). In particular, the HX370 AMD NPU now has a placement- and quality-verified **Spark 1.7B output-head component** through Omni, while complete Spark generation on that route remains unverified; see the [component record](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_amd_npu_output_head/README.md). This advances the M3 component gate, not its whole-model or same-device handoff gate.
+This section preserves the 2026-09-22 audit. For the latest per-cell disposition and 2026-09-23 follow-up evidence, use the [current 60-cell E2E matrix](../../benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md). The HX370 AMD NPU now has a placement- and quality-verified **Spark 1.7B output-head component** through Omni, while complete Spark generation on that route remains unverified; see the [component record](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_amd_npu_output_head/README.md). Native Windows CPU and Radeon 890M also completed [scoped whole-model Spark text requests](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/README.md) with the same Q4_K_M GGUF in standalone llama.cpp. An Omni stage binding, explicit admission and broader quality gates remain open. These results advance the M3 hardware/backend evidence without claiming a joint multi-accelerator plan.
 
 **All 60 named pairings were checked on 2026-09-22** through actual runs,
 raw-record review or artifact/runtime preflight. This is not 60 successful