Profile Spark attention across Qualcomm edge devices

benchmarks/edge_harness/experiments/audit_spark_qualcomm_attention.py

@@ -0,0 +1,165 @@
+#!/usr/bin/env python3
+"""Audit a real Spark attention-layer AI Hub result without claiming model E2E."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import math
+from collections import Counter
+from pathlib import Path
+
+import numpy as np
+
+
+OUTPUT_NAMES = ("hidden", "k_new_0", "v_new_0")
+SOURCE_ONNX_SHA256 = "4e2b1c8d791b1ee0e9b75f3322a20a743c0bf8c40f9b98edd963f2944d6300bb"
+QUANTIZED_ZIP_SHA256 = "703821ca8ace5672228919ea2add4901def4887748c381b12ffc8b259665c553"
+FIXTURE_SHA256 = "5a439e6c6d148f1720938770daa110039a39edde014f1c1592c823c7d7dfc8ed"
+
+
+def sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as source:
+        for block in iter(lambda: source.read(8 << 20), b""):
+            digest.update(block)
+    return digest.hexdigest()
+
+
+def read_outputs(path: Path, *, hub: bool) -> dict[str, np.ndarray]:
+    with np.load(path, allow_pickle=False) as data:
+        actual_names = tuple(data.files)
+        expected_names = tuple(f"output_{i}__0" for i in range(3)) if hub else OUTPUT_NAMES
+        if actual_names != expected_names:
+            raise ValueError(f"{path}: output names/order differ: {actual_names}")
+        values = {name: np.asarray(data[key]) for name, key in zip(OUTPUT_NAMES, expected_names)}
+    shapes = ((1, 1, 2048), (1, 2, 1, 256), (1, 2, 1, 256))
+    for name, shape in zip(OUTPUT_NAMES, shapes):
+        value = values[name]
+        if value.shape != shape or value.dtype != np.float32 or not np.isfinite(value).all():
+            raise ValueError(f"{path}: {name} violates FP32 finite layer-output contract")
+    return values
+
+
+def compare(reference: dict[str, np.ndarray], actual: dict[str, np.ndarray]) -> dict:
+    result = {}
+    for name in OUTPUT_NAMES:
+        left = reference[name].astype(np.float64).ravel()
+        right = actual[name].astype(np.float64).ravel()
+        if not np.linalg.norm(left):
+            raise ValueError(f"{name}: zero reference cannot define relative L2")
+        result[name] = {
+            "relative_l2": float(np.linalg.norm(left - right) / np.linalg.norm(left)),
+            "max_abs": float(np.max(np.abs(left - right))),
+            "cosine": float(np.dot(left, right) / (np.linalg.norm(left) * np.linalg.norm(right))),
+        }
+    return result
+
+
+def nearest_rank(values: list[int], fraction: float) -> int:
+    ordered = sorted(values)
+    return ordered[math.ceil(fraction * len(ordered)) - 1]
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--device-name", required=True)
+    for name in (
+        "source-onnx", "quantized-zip", "fixture", "source-cpu-output",
+        "quantized-cpu-output", "s25-output", "s25-inference-report", "device-output", "compile-report",
+        "inference-report", "profile-report", "output-report",
+    ):
+        parser.add_argument(f"--{name}", type=Path, required=True)
+    args = parser.parse_args()
+    compile_report = json.loads(args.compile_report.read_text(encoding="utf-8"))
+    inference = json.loads(args.inference_report.read_text(encoding="utf-8"))
+    s25_inference = json.loads(args.s25_inference_report.read_text(encoding="utf-8"))
+    profile = json.loads(args.profile_report.read_text(encoding="utf-8"))
+    if compile_report.get("status") != "SUCCESS" or inference.get("status") != "SUCCESS" or profile.get("status") != "SUCCESS":
+        raise RuntimeError("QNN DLC compile, device inference and profile must all succeed")
+    if (compile_report.get("source_quantized_model_id") != "mmd5e3v3q"
+            or compile_report.get("options") != "--target_runtime qnn_dlc --quantize_io"):
+        raise RuntimeError("compile did not use the pinned quantized Spark layer and QNN DLC layout")
+    target = compile_report["target_model_id"]
+    if inference["model_id"] != target or profile["model_id"] != target:
+        raise RuntimeError("device inference/profile used a different compiled artifact")
+    if inference["input_dataset_id"] != "d7m83gjy2":
+        raise RuntimeError("device inference did not use the pinned cross-device fixture")
+    if (s25_inference.get("status") != "SUCCESS" or s25_inference.get("model_id") != "mnowl6xpq"
+            or s25_inference.get("input_dataset_id") != inference["input_dataset_id"]
+            or s25_inference.get("device", {}).get("name") != "Samsung Galaxy S25"):
+        raise RuntimeError("historical S25 output lacks matching model/device/dataset provenance")
+    if inference["device"]["name"] != args.device_name or profile["device"]["name"] != args.device_name:
+        raise RuntimeError("jobs did not execute on the exact requested device")
+    if inference["options"] != "--compute_unit npu" or profile["options"] != "--compute_unit npu":
+        raise RuntimeError("jobs did not request NPU placement")
+
+    execution = profile["profile"]
+    times_us = execution["execution_summary"]["all_inference_times"]
+    if not times_us or any(not isinstance(value, int) or value <= 0 for value in times_us):
+        raise RuntimeError("missing or invalid device samples")
+    units = Counter(row.get("compute_unit") for row in execution["execution_detail"])
+    if units["NPU"] <= 0:
+        raise RuntimeError("profile has no attributed NPU execution rows")
+
+    source = read_outputs(args.source_cpu_output, hub=False)
+    quantized = read_outputs(args.quantized_cpu_output, hub=False)
+    s25 = read_outputs(args.s25_output, hub=True)
+    device_output = read_outputs(args.device_output, hub=True)
+    files = {
+        name: {"bytes": path.stat().st_size, "sha256": sha256(path)}
+        for name, path in {
+            "source_onnx": args.source_onnx,
+            "quantized_onnx_zip": args.quantized_zip,
+            "fixture": args.fixture,
+            "source_cpu_output": args.source_cpu_output,
+            "quantized_cpu_output": args.quantized_cpu_output,
+            "s25_output": args.s25_output,
+            "device_output": args.device_output,
+        }.items()
+    }
+    for name, expected in {
+        "source_onnx": SOURCE_ONNX_SHA256,
+        "quantized_onnx_zip": QUANTIZED_ZIP_SHA256,
+        "fixture": FIXTURE_SHA256,
+    }.items():
+        if files[name]["sha256"] != expected:
+            raise RuntimeError(f"{name} hash differs from pinned layer artifact or fixture")
+    report = {
+        "scope": "one real Spark-X2.5 full-attention decoder layer at one static 1024-token cache shape; no generation E2E",
+        "status": "component_executed_on_npu_quality_unqualified",
+        "device": profile["device"],
+        "compile_job_id": compile_report["job_id"],
+        "inference_job_id": inference["job_id"],
+        "profile_job_id": profile["job_id"],
+        "target_model_id": target,
+        "input_dataset_id": inference["input_dataset_id"],
+        "s25_prior_inference_job_id": s25_inference["job_id"],
+        "files": files,
+        "source_vs_quantized_cpu": compare(source, quantized),
+        "quantized_cpu_vs_device_npu": compare(quantized, device_output),
+        "source_cpu_vs_device_npu": compare(source, device_output),
+        "s25_npu_vs_device_npu": compare(s25, device_output),
+        "profile": {
+            "sample_count": len(times_us),
+            "sample_min_us": min(times_us),
+            "nearest_rank_p50_us": nearest_rank(times_us, 0.50),
+            "nearest_rank_p95_us": nearest_rank(times_us, 0.95),
+            "reported_estimated_inference_time_us": execution["execution_summary"]["estimated_inference_time"],
+            "reported_peak_memory_bytes": execution["execution_summary"]["estimated_inference_peak_memory"],
+            "compute_unit_row_counts": dict(units),
+        },
+        "limits": [
+            "No task-level numerical tolerance or next-token quality gate is established.",
+            "Hub component jobs do not prove device-local prefill, KV/ring update, sampling, continuous generation or Omni deployment.",
+            "Profile sample latency is one component, not complete-request or sustained thermal performance.",
+        ],
+    }
+    args.output_report.parent.mkdir(parents=True, exist_ok=True)
+    args.output_report.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()

benchmarks/edge_harness/experiments/audit_spark_qualcomm_fallback.py

@@ -0,0 +1,102 @@
+#!/usr/bin/env python3
+"""Audit an RB3 Spark attention-layer fallback against the FP32 source."""
+
+from __future__ import annotations
+
+import argparse
+import json
+from collections import Counter
+from pathlib import Path
+
+from audit_spark_qualcomm_attention import (
+    FIXTURE_SHA256,
+    SOURCE_ONNX_SHA256,
+    compare,
+    nearest_rank,
+    read_outputs,
+    sha256,
+)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--device-name", default="Dragonwing RB3 Gen 2 Vision Kit")
+    parser.add_argument("--compute-unit", choices=("cpu", "npu"), required=True)
+    parser.add_argument("--compile-options", required=True)
+    for name in (
+        "source-onnx", "fixture", "source-cpu-output", "compile-report",
+        "inference-report", "profile-report", "device-output", "output-report",
+    ):
+        parser.add_argument(f"--{name}", type=Path, required=True)
+    args = parser.parse_args()
+    compile_report = json.loads(args.compile_report.read_text(encoding="utf-8"))
+    inference = json.loads(args.inference_report.read_text(encoding="utf-8"))
+    profile = json.loads(args.profile_report.read_text(encoding="utf-8"))
+    if any(report.get("status") != "SUCCESS" for report in (compile_report, inference, profile)):
+        raise RuntimeError("compile, inference and profile must all succeed")
+    if compile_report.get("source_model_id") != "mn14dw6zn" or compile_report.get("options") != args.compile_options:
+        raise RuntimeError("compile did not use the pinned FP32 Spark source and requested route")
+    target_id = compile_report["target_model_id"]
+    if inference.get("model_id") != target_id or profile.get("model_id") != target_id:
+        raise RuntimeError("inference/profile did not use the compiled artifact")
+    if inference.get("input_dataset_id") != "d7m83gjy2":
+        raise RuntimeError("inference did not use the pinned cross-device fixture")
+    for report in (compile_report, inference, profile):
+        if report.get("device", {}).get("name") != args.device_name:
+            raise RuntimeError("job did not use the exact requested device")
+    requested = f"--compute_unit {args.compute_unit}"
+    if inference.get("options") != requested or profile.get("options") != requested:
+        raise RuntimeError("inference/profile requested a different compute unit")
+    files = {
+        name: {"bytes": path.stat().st_size, "sha256": sha256(path)}
+        for name, path in {
+            "source_onnx": args.source_onnx,
+            "fixture": args.fixture,
+            "source_cpu_output": args.source_cpu_output,
+            "device_output": args.device_output,
+        }.items()
+    }
+    if files["source_onnx"]["sha256"] != SOURCE_ONNX_SHA256 or files["fixture"]["sha256"] != FIXTURE_SHA256:
+        raise RuntimeError("source or fixture differs from the pinned experiment")
+    output = read_outputs(args.device_output, hub=True)
+    source = read_outputs(args.source_cpu_output, hub=False)
+    execution = profile["profile"]
+    times_us = execution["execution_summary"]["all_inference_times"]
+    if not times_us or any(not isinstance(value, int) or value <= 0 for value in times_us):
+        raise RuntimeError("missing or invalid device samples")
+    units = Counter(row.get("compute_unit") for row in execution["execution_detail"])
+    if units[args.compute_unit.upper()] <= 0:
+        raise RuntimeError("profile has no rows attributed to the requested compute unit")
+    report = {
+        "scope": "one real Spark-X2.5 attention layer at one static 1024-token cache shape; no generation E2E",
+        "status": f"component_executed_on_{args.compute_unit}_quality_unqualified",
+        "device": profile["device"],
+        "compile_job_id": compile_report["job_id"],
+        "inference_job_id": inference["job_id"],
+        "profile_job_id": profile["job_id"],
+        "target_model_id": target_id,
+        "input_dataset_id": inference["input_dataset_id"],
+        "compile_options": args.compile_options,
+        "files": files,
+        "source_cpu_vs_device": compare(source, output),
+        "profile": {
+            "sample_count": len(times_us),
+            "sample_min_us": min(times_us),
+            "nearest_rank_p50_us": nearest_rank(times_us, 0.50),
+            "nearest_rank_p95_us": nearest_rank(times_us, 0.95),
+            "reported_estimated_inference_time_us": execution["execution_summary"]["estimated_inference_time"],
+            "reported_peak_memory_bytes": execution["execution_summary"]["estimated_inference_peak_memory"],
+            "compute_unit_row_counts": dict(units),
+        },
+        "limits": [
+            "No next-token or task-level numerical tolerance is established.",
+            "No complete prefill, KV/ring, sampling, continuous generation or Omni deployment is tested.",
+            "Hosted component latency is not a complete request or sustained thermal profile.",
+        ],
+    }
+    args.output_report.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()

benchmarks/edge_harness/experiments/collect_spark_aihub_component.py

@@ -0,0 +1,81 @@
+#!/usr/bin/env python3
+"""Preserve raw Workbench Spark-layer jobs submitted in an evidence directory.
+
+Requires qai-hub authentication through the local client configuration. The
+submission JSON files identify existing jobs; this script never resubmits them.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from datetime import datetime, timezone
+from pathlib import Path
+
+import numpy as np
+import qai_hub as hub
+
+
+def device_record(device: object) -> dict:
+    return {
+        "name": device.name,
+        "os": device.os,
+        "attributes": list(device.attributes),
+    }
+
+
+def collect(evidence_dir: Path, kind: str, client: hub.Client) -> str:
+    submission = json.loads((evidence_dir / f"{kind}_submission.json").read_text(encoding="utf-8"))
+    report_path = evidence_dir / f"{kind}_report.json"
+    if report_path.exists():
+        existing = json.loads(report_path.read_text(encoding="utf-8"))
+        if existing.get("job_id") != submission["job_id"]:
+            raise RuntimeError(f"{report_path}: job ID differs from its submission")
+        return f"{kind}: {submission['job_id']} {existing['status']} (already collected)"
+    job = client.get_job(submission["job_id"])
+    status = job.get_status()
+    if status.code not in {"SUCCESS", "FAILED", "CANCELED"}:
+        return f"{kind}: {submission['job_id']} {status.code} (pending)"
+    report = {
+        "timestamp_utc": datetime.now(timezone.utc).isoformat(),
+        "qai_hub_version": str(hub.__version__),
+        "facility": "Qualcomm AI Hub Workbench",
+        "job_id": submission["job_id"],
+        "url": submission["url"],
+        "status": status.code,
+        "status_message": status.message or "",
+        "options": job.options,
+        "hub_version": str(job.hub_version),
+        "device": device_record(job.device),
+        "model_id": job.model.model_id,
+    }
+    if kind == "inference":
+        report["input_dataset_id"] = job.inputs.dataset_id
+        if status.code == "SUCCESS":
+            output_path = evidence_dir / "device_output.npz"
+            outputs = job.download_output_data()
+            arrays = {f"{name}__{i}": value for name, values in outputs.items() for i, value in enumerate(values)}
+            np.savez_compressed(output_path, **arrays)
+            report["outputs"] = output_path.name
+    elif kind == "profile":
+        report["shapes"] = job.shapes
+        if status.code == "SUCCESS":
+            report["profile"] = job.download_profile()
+    else:
+        raise ValueError(f"unexpected job kind {kind}")
+    report_path.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
+    return f"{kind}: {submission['job_id']} {status.code}"
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("evidence_dirs", nargs="+", type=Path)
+    args = parser.parse_args()
+    client = hub.Client()
+    for directory in args.evidence_dirs:
+        for kind in ("inference", "profile"):
+            print(f"{directory.name}: {collect(directory, kind, client)}")
+
+
+if __name__ == "__main__":
+    main()

benchmarks/edge_harness/results/e2e_expansion_20260923/README.md

@@ -20,6 +20,10 @@ The [native-storage follow-up](evidence/minicpmo_omni_native_weights_io/README.m
 
 A separate [WSL CPU Spark GGUF Omni profile](evidence/spark_omni_wsl_cpu/README.md) ran the same pinned Q4_K_M artifact through a Linux CPU-only llama.cpp server and current Omni controller. It passed one warmup and 20/20 serial inventory requests at nearest-rank complete-request wall p50/p95 **7.82/7.97 s**, a public `AsyncOmni.generate` text request, explicit context refusal and in-flight cancellation cleanup with a 4 GiB host reservation. The Linux server omitted the Windows offload-count line; the stage now requires positive CPU tensor-buffer evidence for that log layout and rejects accelerator buffers. This is independent of the earlier WSL vLLM INT8 result and does not qualify incremental streaming or general text quality.
 
+The [hosted Galaxy S24 Spark experiment](evidence/spark_s24_full_attention/README.md) moved one real full-attention decoder layer with a fixed 1,024-token cache shape to the phone NPU. Its W8A16 QDQ QNN DLC inference matched the earlier S25 NPU outputs exactly on one pinned dataset; 100 NPU-attributed component samples had nearest-rank p50/p95 **1.109/1.205 ms**. Separate local FP32 and quantized CPU runs measured the numerical change, with 1.351% hidden-state relative L2 from original FP32 source to S24 output and no task-quality tolerance. AI Hub Workbench runs components on hosted physical devices, not the stateful local Spark generation loop; this cell remains NOT E2E.
+
+The same layer artifact then ran on hosted [Snapdragon X Elite CRD](evidence/spark_xelite_full_attention/README.md) and [SA8775P ADP](evidence/spark_sa8775_full_attention/README.md) NPUs, with exact same-fixture outputs as S24/S25 and 100-sample component p50/p95 **1.433/1.606 ms** and **1.837/2.079 ms**, respectively. On [RB3 Gen 2 / QCS6490](evidence/spark_rb3_full_attention/README.md), both the reused and exact-device-compiled W8A16 QNN DLCs failed at graph composition. An FP32 ONNX CPU layer passed with source-relative L2 below 9e-7; calibrated W8A8 ran on NPU but had 20.9% hidden-state relative L2 and is numerically unqualified. These are separate device components, not full-model deployment or cross-device speed comparisons.
+
 | Device / model | Verified depth | Observed output and time | Remaining gate |
 |---|---|---|---|
 | Native Windows HX370 Radeon 890M / InternVLA-A1 Cosmos CI8x8 encoder | Component C; unchanged Omni encoder source in separate DirectML/PyTorch 2.4 environment, not an Omni full-policy stage | One warmup + 20 serial patterned 256×256 image encodes, including upload/readback: Radeon p50/p95 26.56/27.49 ms versus CPU 156.49/158.71 ms; relative L2 error 4.73e-07. Encoder weights and output device checked. | Complete policy/action and real-observation parity; policy/controller runtime compatibility, bounded handoff, shared-RAM admission and end-to-end benefit. [Raw evidence](evidence/internvla_cosmos_radeon890m/README.md). |

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/README.md

@@ -0,0 +1,18 @@
+# Spark-X2.5 attention layer on Dragonwing RB3 Gen 2 / QCS6490
+
+**The hosted RB3 kit (QC Linux 1.6, Hexagon v68) ran one real Spark-X2.5-1.7B attention layer on CPU with near-FP32 numerical parity. A separate W8A8 NPU variant executed but has 20.9% hidden-state relative L2 error on the pinned fixture. Neither variant qualifies full Spark generation.** The [S24 provenance](../spark_s24_full_attention/README.md) pins the FP32 source ONNX, its W8A16 QDQ form, calibration and exact 1,024-token-cache input. The [artifact manifest](artifact_manifest.json) pins downloaded compile outputs outside Git.
+
+Four routes were tested on exact device **Dragonwing RB3 Gen 2 Vision Kit**, using AI Hub Workbench client 0.55.0 and service build in each raw job report:
+
+| Route | Jobs and outcome | One-layer evidence |
+|---|---|---|
+| Reuse S24-built W8A16 QNN DLC `mq9y70k0n` | [Inference](inference_report.json) `j57e4k4lp` and [profile](profile_report.json) `jp4y1m1vp` both failed: `QnnModel_composeGraphsFromDlc: MODEL_GRAPH_ERROR`. | No outputs or timing samples. |
+| Compile same QDQ source for RB3 as W8A16 QNN DLC `mn0g3ye8m` | [Compile](compile_rb3_report.json) `jgnznozrg` passed. Targeted [inference](targeted/inference_report.json) `jglyxoy85` and [profile](targeted/profile_report.json) `j5687r80g` both failed with the same graph error. | Device-specific recompilation did not resolve this graph/runtime failure. It does not establish that all W8A16 Spark exports are unsupported. |
+| Compile original FP32 source to ONNX `mmxj02ljq`, request CPU | [Compile](compile_cpu_report.json) `jglyxowj5`, [inference](cpu/inference_report.json) `jgolrodqg` and [profile](cpu/profile_report.json) `jpvldemk5` passed. All 56 execution-detail rows were CPU. | [Output](cpu/device_output.npz) hidden/K/V relative L2 versus original ONNX Runtime 1.29 CPU was 8.73e-7/2.98e-7/3.37e-7. [Audit](cpu/audit_report.json): 100 samples, nearest-rank p50/p95 19.304/21.113 ms (minimum 19.104 ms). This is explicitly CPU, not NPU fallback masquerading as acceleration. |
+| Compile original source with calibration dataset `d7x8e3d59` using `--quantize_full_type int8 --quantize_io` to QNN DLC `mqkp10ezm`, request NPU | [Compile](compile_w8a8_report.json) `jpyoxq885`, [inference](w8a8/inference_report.json) `jgzlm8yo5` and [profile](w8a8/profile_report.json) `j5wl78z3p` passed. All 61 execution-detail rows were NPU. | [Output](w8a8/device_output.npz) hidden/K/V relative L2 versus FP32 CPU was **20.914%/1.702%/4.413%**. [Audit](w8a8/audit_report.json): 100 samples, p50/p95 5.610/5.761 ms (minimum 5.089 ms). This precision is **numerically unqualified**; no next-token/task tolerance is established. |
+
+All successful jobs used the same fixture dataset `d7m83gjy2` and returned finite FP32 hidden/new-key/new-value tensors. The service `estimated_inference_time` equals each sample minimum, not the median. The [CPU](cpu/audit_report.json) and [W8A8 NPU](w8a8/audit_report.json) audits pin source and fixture hashes, exact job/device/model IDs, output hashes, placement rows and full precision differences. [Submission records](compile_w8a8_submission.json), [initial failed submissions](inference_submission.json), [RB3 W8A16 compile submission](compile_rb3_submission.json), and [CPU compile submission](compile_cpu_submission.json) preserve attempted settings. The scripts [collect existing jobs](../../../../experiments/collect_spark_aihub_component.py) and [audit fallback variants](../../../../experiments/audit_spark_qualcomm_fallback.py) without embedded credentials.
+
+The CPU and NPU numbers are different precisions and outputs, not a qualified speedup comparison. Neither job includes 27 other decoder layers, prefill, continuous decode, KV/ring, sampling, a resident Omni controller, complete-request timing or sustained power/thermal behavior. Workbench is a test facility, not a deployment dependency.
+
+**Matrix status: CPU component numerical pass; W8A8 NPU component execution with unresolved numerical quality; NOT E2E.** The next NPU experiment needs representative multi-prompt calibration and layer/next-token parity gates, then a compatible full-model artifact. The full embedded route also needs device-local state, memory admission, cancellation and sustained profiling. The two W8A16 graph errors remain specific to the tested QDQ export and SDK/device combination.

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/artifact_manifest.json

@@ -0,0 +1,33 @@
+{
+  "scope": "Pinned Spark attention-layer variants for RB3 remediation; binaries retained outside Git",
+  "artifacts": {
+    "s24_compiled_w8a16_dlc": {
+      "model_id": "mq9y70k0n",
+      "compile_job_id": "jpe7z8l15",
+      "downloaded_path": "/home/zhout/project/edge_infer/models/spark_s24_qnn_dlc_20260924.dlc",
+      "bytes": 52036972,
+      "sha256": "8f1518f1f392434306dba11fd7115fe7629657ab7bfc834c0f78ee6b684e5154"
+    },
+    "rb3_targeted_w8a16_dlc": {
+      "model_id": "mn0g3ye8m",
+      "compile_job_id": "jgnznozrg",
+      "downloaded_path": "/home/zhout/project/edge_infer/models/spark_rb3_qnn_dlc_20260924.dlc",
+      "bytes": 52036972,
+      "sha256": "614ea0b07d9119581c563d9479181390fb007efb39b538381b971ce12ef39161"
+    },
+    "rb3_w8a8_dlc": {
+      "model_id": "mqkp10ezm",
+      "compile_job_id": "jpyoxq885",
+      "downloaded_path": "/home/zhout/project/edge_infer/models/spark_rb3_w8a8_dlc_20260924.dlc",
+      "bytes": 51452356,
+      "sha256": "637f4dd52ddf86e53c46c1ae5c1ef3b030b822f866575835aa66e6b86ad3cfd2"
+    },
+    "rb3_fp32_onnx_cpu": {
+      "model_id": "mmxj02ljq",
+      "compile_job_id": "jglyxowj5",
+      "downloaded_path": "/home/zhout/project/edge_infer/models/spark_rb3_onnx_cpu_20260924.onnx.onnx.zip",
+      "bytes": 107095517,
+      "sha256": "a64c220f32f7a4a5652594014e39e8ead2f688b7dabf0ed878620b61d94d7ec2"
+    }
+  }
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/compile_cpu_report.json

@@ -0,0 +1,31 @@
+{
+  "timestamp_utc": "2026-09-23T21:56:04.464159+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jglyxowj5",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jglyxowj5/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "source_model_id": "mn14dw6zn",
+  "target_model_id": "mmxj02ljq",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "options": "--target_runtime onnx",
+  "hub_version": "aihub-2026.09.11.0"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/compile_cpu_submission.json

@@ -0,0 +1,11 @@
+{
+  "timestamp_utc": "2026-09-23T21:54:10.264314+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "job_id": "jglyxowj5",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jglyxowj5/",
+  "source_model_id": "mn14dw6zn",
+  "device": "Dragonwing RB3 Gen 2 Vision Kit",
+  "options": "--target_runtime onnx",
+  "reason": "Explicit CPU-only component fallback to separate graph compatibility from NPU capability"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/compile_rb3_report.json

@@ -0,0 +1,31 @@
+{
+  "timestamp_utc": "2026-09-23T21:47:29.642901+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jgnznozrg",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jgnznozrg/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "source_quantized_model_id": "mmd5e3v3q",
+  "target_model_id": "mn0g3ye8m",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "options": "--target_runtime qnn_dlc --quantize_io",
+  "hub_version": "aihub-2026.09.11.0"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/compile_rb3_submission.json

@@ -0,0 +1,11 @@
+{
+  "timestamp_utc": "2026-09-23T21:45:01.402565+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "job_id": "jgnznozrg",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jgnznozrg/",
+  "source_quantized_model_id": "mmd5e3v3q",
+  "device": "Dragonwing RB3 Gen 2 Vision Kit",
+  "options": "--target_runtime qnn_dlc --quantize_io",
+  "reason": "S24-targeted QNN DLC failed composeGraphsFromDlc on RB3"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/compile_w8a8_report.json

@@ -0,0 +1,32 @@
+{
+  "timestamp_utc": "2026-09-23T21:56:39.060765+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jpyoxq885",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jpyoxq885/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "source_model_id": "mn14dw6zn",
+  "calibration_dataset_id": "d7x8e3d59",
+  "target_model_id": "mqkp10ezm",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "options": "--target_runtime qnn_dlc --quantize_full_type int8 --quantize_io",
+  "hub_version": "aihub-2026.09.11.0"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/compile_w8a8_submission.json

@@ -0,0 +1,12 @@
+{
+  "timestamp_utc": "2026-09-23T21:53:55.558562+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "job_id": "jpyoxq885",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jpyoxq885/",
+  "source_model_id": "mn14dw6zn",
+  "calibration_dataset_id": "d7x8e3d59",
+  "device": "Dragonwing RB3 Gen 2 Vision Kit",
+  "options": "--target_runtime qnn_dlc --quantize_full_type int8 --quantize_io",
+  "reason": "W8A16 QNN DLC failed graph composition on RB3, including exact-device compile"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/cpu/audit_report.json

@@ -0,0 +1,79 @@
+{
+  "scope": "one real Spark-X2.5 attention layer at one static 1024-token cache shape; no generation E2E",
+  "status": "component_executed_on_cpu_quality_unqualified",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "compile_job_id": "jglyxowj5",
+  "inference_job_id": "jgolrodqg",
+  "profile_job_id": "jpvldemk5",
+  "target_model_id": "mmxj02ljq",
+  "input_dataset_id": "d7m83gjy2",
+  "compile_options": "--target_runtime onnx",
+  "files": {
+    "source_onnx": {
+      "bytes": 205618089,
+      "sha256": "4e2b1c8d791b1ee0e9b75f3322a20a743c0bf8c40f9b98edd963f2944d6300bb"
+    },
+    "fixture": {
+      "bytes": 3901238,
+      "sha256": "5a439e6c6d148f1720938770daa110039a39edde014f1c1592c823c7d7dfc8ed"
+    },
+    "source_cpu_output": {
+      "bytes": 12103,
+      "sha256": "6bbfb8e6ccb09475391ce1dbd2e855205720f4367c038a0d5be35e7dfeb2f768"
+    },
+    "device_output": {
+      "bytes": 12128,
+      "sha256": "442d3a8fef165584c853daf87592e4dc7ed3bc9788eee93b2fc6c67cbc6770b4"
+    }
+  },
+  "source_cpu_vs_device": {
+    "hidden": {
+      "relative_l2": 8.726789979684891e-07,
+      "max_abs": 3.2782554626464844e-07,
+      "cosine": 0.9999999999996264
+    },
+    "k_new_0": {
+      "relative_l2": 2.984418086993196e-07,
+      "max_abs": 2.384185791015625e-06,
+      "cosine": 0.9999999999999881
+    },
+    "v_new_0": {
+      "relative_l2": 3.3724649115031866e-07,
+      "max_abs": 8.344650268554688e-07,
+      "cosine": 0.9999999999999739
+    }
+  },
+  "profile": {
+    "sample_count": 100,
+    "sample_min_us": 19104,
+    "nearest_rank_p50_us": 19304,
+    "nearest_rank_p95_us": 21113,
+    "reported_estimated_inference_time_us": 19104,
+    "reported_peak_memory_bytes": 227733504,
+    "compute_unit_row_counts": {
+      "CPU": 56
+    }
+  },
+  "limits": [
+    "No next-token or task-level numerical tolerance is established.",
+    "No complete prefill, KV/ring, sampling, continuous generation or Omni deployment is tested.",
+    "Hosted component latency is not a complete request or sustained thermal profile."
+  ]
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/cpu/inference_report.json

@@ -0,0 +1,32 @@
+{
+  "timestamp_utc": "2026-09-23T22:01:39.257777+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jgolrodqg",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jgolrodqg/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit cpu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mmxj02ljq",
+  "input_dataset_id": "d7m83gjy2",
+  "outputs": "device_output.npz"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/cpu/inference_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T21:55:49.636711+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jglyxowj5",
+  "target_model_id": "mmxj02ljq",
+  "input_dataset_id": "d7m83gjy2",
+  "device": "Dragonwing RB3 Gen 2 Vision Kit",
+  "options": "--compute_unit cpu",
+  "kind": "inference",
+  "status": "SUBMITTED",
+  "job_id": "jgolrodqg",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jgolrodqg/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/cpu/profile_report.json

@@ -0,0 +1,632 @@
+{
+  "timestamp_utc": "2026-09-23T22:01:44.553558+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jpvldemk5",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jpvldemk5/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit cpu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mmxj02ljq",
+  "shapes": {
+    "x": [
+      [
+        1,
+        1,
+        2048
+      ],
+      "float32"
+    ],
+    "cos_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "float32"
+    ],
+    "sin_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "float32"
+    ],
+    "mask_full": [
+      [
+        1,
+        1,
+        1,
+        1025
+      ],
+      "float32"
+    ],
+    "k_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "float32"
+    ],
+    "v_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "float32"
+    ]
+  },
+  "profile": {
+    "execution_summary": {
+      "estimated_inference_time": 19104,
+      "estimated_inference_peak_memory": 227733504,
+      "first_load_time": 311945,
+      "first_load_peak_memory": 221429760,
+      "warm_load_time": 319951,
+      "warm_load_peak_memory": 217300992,
+      "compile_time": 0,
+      "compile_peak_memory": 0,
+      "compile_memory_increase_range": [
+        0,
+        0
+      ],
+      "compile_memory_peak_range": [
+        0,
+        0
+      ],
+      "first_load_memory_increase_range": [
+        10790272,
+        11590944
+      ],
+      "first_load_memory_peak_range": [
+        218533888,
+        219334560
+      ],
+      "warm_load_memory_increase_range": [
+        6655248,
+        7457200
+      ],
+      "warm_load_memory_peak_range": [
+        214401024,
+        215202976
+      ],
+      "inference_memory_increase_range": [
+        0,
+        0
+      ],
+      "inference_memory_peak_range": [
+        4120576,
+        216040736
+      ],
+      "all_compile_times": [],
+      "all_first_load_times": [
+        311945
+      ],
+      "all_warm_load_times": [
+        319951
+      ],
+      "all_inference_times": [
+        22116,
+        19121,
+        21897,
+        19651,
+        21601,
+        19307,
+        19264,
+        21113,
+        19333,
+        19307,
+        21370,
+        19104,
+        19275,
+        19320,
+        19207,
+        19300,
+        19644,
+        19384,
+        19192,
+        19301,
+        19300,
+        19221,
+        19480,
+        19284,
+        19332,
+        19419,
+        19308,
+        19228,
+        19176,
+        19409,
+        19309,
+        19287,
+        19256,
+        19221,
+        19337,
+        19359,
+        19250,
+        19694,
+        19316,
+        19240,
+        19270,
+        19262,
+        19271,
+        19198,
+        19311,
+        25366,
+        19323,
+        19277,
+        19319,
+        19415,
+        19243,
+        19588,
+        19252,
+        19318,
+        19292,
+        19307,
+        19237,
+        19352,
+        19246,
+        19242,
+        19322,
+        19395,
+        19347,
+        19456,
+        19315,
+        19256,
+        19286,
+        19329,
+        19337,
+        19336,
+        19898,
+        19727,
+        19401,
+        19254,
+        19255,
+        19526,
+        19328,
+        19306,
+        19268,
+        19209,
+        19263,
+        19177,
+        19336,
+        19241,
+        19235,
+        19256,
+        19428,
+        19248,
+        19398,
+        19584,
+        19191,
+        19304,
+        19218,
+        19296,
+        19222,
+        19225,
+        19245,
+        19332,
+        19245,
+        19292
+      ]
+    },
+    "execution_detail": [
+      {
+        "name": "node_Mul_0",
+        "type": "Mul",
+        "compute_unit": "CPU",
+        "execution_time": 51,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/ReduceMean",
+        "type": "ReduceMean",
+        "compute_unit": "CPU",
+        "execution_time": 22,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add",
+        "type": "Add",
+        "compute_unit": "CPU",
+        "execution_time": 12,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Sqrt",
+        "type": "Sqrt",
+        "compute_unit": "CPU",
+        "execution_time": 13,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Div",
+        "type": "Div",
+        "compute_unit": "CPU",
+        "execution_time": 10,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_1",
+        "type": "Mul",
+        "compute_unit": "CPU",
+        "execution_time": 18,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/q_k_v_proj/MatMul",
+        "type": "MatMul",
+        "compute_unit": "CPU",
+        "execution_time": 2100,
+        "execution_cycles": 0
+      },
+      {
+        "name": "node_Split_20",
+        "type": "Split",
+        "compute_unit": "CPU",
+        "execution_time": 81,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape",
+        "type": "Reshape",
+        "compute_unit": "CPU",
+        "execution_time": 42,
+        "execution_cycles": 0
+      },
+      {
+        "name": "fused_/layers.0/Slice_3_/layers.0/Slice_6",
+        "type": "Slice",
+        "compute_unit": "CPU",
+        "execution_time": 68,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Neg",
+        "type": "Neg",
+        "compute_unit": "CPU",
+        "execution_time": 38,
+        "execution_cycles": 0
+      },
+      {
+        "name": "fused_/layers.0/Slice_3_/layers.0/Slice_5",
+        "type": "Slice",
+        "compute_unit": "CPU",
+        "execution_time": 52,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Concat",
+        "type": "Concat",
+        "compute_unit": "CPU",
+        "execution_time": 48,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_3",
+        "type": "Mul",
+        "compute_unit": "CPU",
+        "execution_time": 48,
+        "execution_cycles": 0
+      },
+      {
+        "name": "node_Split_22",
+        "type": "Split",
+        "compute_unit": "CPU",
+        "execution_time": 46,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_2",
+        "type": "Mul",
+        "compute_unit": "CPU",
+        "execution_time": 41,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_1",
+        "type": "Add",
+        "compute_unit": "CPU",
+        "execution_time": 42,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Concat_1",
+        "type": "Concat",
+        "compute_unit": "CPU",
+        "execution_time": 44,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_4",
+        "type": "Reshape",
+        "compute_unit": "CPU",
+        "execution_time": 38,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_1",
+        "type": "Reshape",
+        "compute_unit": "CPU",
+        "execution_time": 36,
+        "execution_cycles": 0
+      },
+      {
+        "name": "fused_/layers.0/Slice_7_/layers.0/Slice_10",
+        "type": "Slice",
+        "compute_unit": "CPU",
+        "execution_time": 50,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Neg_1",
+        "type": "Neg",
+        "compute_unit": "CPU",
+        "execution_time": 37,
+        "execution_cycles": 0
+      },
+      {
+        "name": "fused_/layers.0/Slice_7_/layers.0/Slice_9",
+        "type": "Slice",
+        "compute_unit": "CPU",
+        "execution_time": 46,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Concat_2",
+        "type": "Concat",
+        "compute_unit": "CPU",
+        "execution_time": 43,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_5",
+        "type": "Mul",
+        "compute_unit": "CPU",
+        "execution_time": 44,
+        "execution_cycles": 0
+      },
+      {
+        "name": "node_Split_24",
+        "type": "Split",
+        "compute_unit": "CPU",
+        "execution_time": 45,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_4",
+        "type": "Mul",
+        "compute_unit": "CPU",
+        "execution_time": 38,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_2",
+        "type": "Add",
+        "compute_unit": "CPU",
+        "execution_time": 39,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Concat_3",
+        "type": "Concat",
+        "compute_unit": "CPU",
+        "execution_time": 44,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Concat_4",
+        "type": "Concat",
+        "compute_unit": "CPU",
+        "execution_time": 309,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/MatMul/MatmulTransposeFusion//MatMulScaleFusion/",
+        "type": "FusedMatMul",
+        "compute_unit": "CPU",
+        "execution_time": 578,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_3",
+        "type": "Add",
+        "compute_unit": "CPU",
+        "execution_time": 78,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Softmax",
+        "type": "Softmax",
+        "compute_unit": "CPU",
+        "execution_time": 95,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_2",
+        "type": "Reshape",
+        "compute_unit": "CPU",
+        "execution_time": 45,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Concat_5",
+        "type": "Concat",
+        "compute_unit": "CPU",
+        "execution_time": 324,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/MatMul_1",
+        "type": "MatMul",
+        "compute_unit": "CPU",
+        "execution_time": 616,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_5",
+        "type": "Reshape",
+        "compute_unit": "CPU",
+        "execution_time": 45,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/g_proj/MatMul",
+        "type": "MatMul",
+        "compute_unit": "CPU",
+        "execution_time": 71,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Sigmoid",
+        "type": "Sigmoid",
+        "compute_unit": "CPU",
+        "execution_time": 40,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_3",
+        "type": "Reshape",
+        "compute_unit": "CPU",
+        "execution_time": 36,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_7",
+        "type": "Mul",
+        "compute_unit": "CPU",
+        "execution_time": 51,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_6",
+        "type": "Reshape",
+        "compute_unit": "CPU",
+        "execution_time": 38,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/out_proj/MatMul",
+        "type": "MatMul",
+        "compute_unit": "CPU",
+        "execution_time": 1716,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_4",
+        "type": "Add",
+        "compute_unit": "CPU",
+        "execution_time": 64,
+        "execution_cycles": 0
+      },
+      {
+        "name": "node_Mul_1",
+        "type": "Mul",
+        "compute_unit": "CPU",
+        "execution_time": 41,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/ReduceMean_1",
+        "type": "ReduceMean",
+        "compute_unit": "CPU",
+        "execution_time": 51,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_5",
+        "type": "Add",
+        "compute_unit": "CPU",
+        "execution_time": 38,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Sqrt_1",
+        "type": "Sqrt",
+        "compute_unit": "CPU",
+        "execution_time": 36,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Div_1",
+        "type": "Div",
+        "compute_unit": "CPU",
+        "execution_time": 43,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_9",
+        "type": "Mul",
+        "compute_unit": "CPU",
+        "execution_time": 52,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/up_proj/MatMul",
+        "type": "MatMul",
+        "compute_unit": "CPU",
+        "execution_time": 4437,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/gate_proj/MatMul",
+        "type": "MatMul",
+        "compute_unit": "CPU",
+        "execution_time": 4404,
+        "execution_cycles": 0
+      },
+      {
+        "name": "node_Gelu_2",
+        "type": "Gelu",
+        "compute_unit": "CPU",
+        "execution_time": 147,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_12",
+        "type": "Mul",
+        "compute_unit": "CPU",
+        "execution_time": 62,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/down_proj/MatMul",
+        "type": "MatMul",
+        "compute_unit": "CPU",
+        "execution_time": 4396,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_7",
+        "type": "Add",
+        "compute_unit": "CPU",
+        "execution_time": 74,
+        "execution_cycles": 0
+      }
+    ]
+  }
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/cpu/profile_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T21:55:51.569947+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jglyxowj5",
+  "target_model_id": "mmxj02ljq",
+  "input_dataset_id": null,
+  "device": "Dragonwing RB3 Gen 2 Vision Kit",
+  "options": "--compute_unit cpu",
+  "kind": "profile",
+  "status": "SUBMITTED",
+  "job_id": "jpvldemk5",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jpvldemk5/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/inference_report.json

@@ -0,0 +1,31 @@
+{
+  "timestamp_utc": "2026-09-23T21:44:28.625840+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "j57e4k4lp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j57e4k4lp/",
+  "status": "FAILED",
+  "status_message": "Failed to profile the model: Failed to call QnnModel_composeGraphsFromDlc: MODEL_GRAPH_ERROR.",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mq9y70k0n",
+  "input_dataset_id": "d7m83gjy2"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/inference_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T21:27:00.903338+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jpe7z8l15",
+  "target_model_id": "mq9y70k0n",
+  "input_dataset_id": "d7m83gjy2",
+  "device": "Dragonwing RB3 Gen 2 Vision Kit",
+  "options": "--compute_unit npu",
+  "kind": "inference",
+  "status": "SUBMITTED",
+  "job_id": "j57e4k4lp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j57e4k4lp/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/profile_report.json

@@ -0,0 +1,85 @@
+{
+  "timestamp_utc": "2026-09-23T21:44:29.795793+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jp4y1m1vp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jp4y1m1vp/",
+  "status": "FAILED",
+  "status_message": "Failed to profile the model: Failed to call QnnModel_composeGraphsFromDlc: MODEL_GRAPH_ERROR.",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mq9y70k0n",
+  "shapes": {
+    "x": [
+      [
+        1,
+        1,
+        2048
+      ],
+      "uint16"
+    ],
+    "cos_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint16"
+    ],
+    "sin_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint16"
+    ],
+    "mask_full": [
+      [
+        1,
+        1,
+        1,
+        1025
+      ],
+      "uint16"
+    ],
+    "k_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint16"
+    ],
+    "v_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint16"
+    ]
+  }
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/profile_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T21:27:02.068005+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jpe7z8l15",
+  "target_model_id": "mq9y70k0n",
+  "input_dataset_id": null,
+  "device": "Dragonwing RB3 Gen 2 Vision Kit",
+  "options": "--compute_unit npu",
+  "kind": "profile",
+  "status": "SUBMITTED",
+  "job_id": "jp4y1m1vp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jp4y1m1vp/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/targeted/inference_report.json

@@ -0,0 +1,31 @@
+{
+  "timestamp_utc": "2026-09-23T21:53:06.828554+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jglyxoy85",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jglyxoy85/",
+  "status": "FAILED",
+  "status_message": "Failed to profile the model: Failed to call QnnModel_composeGraphsFromDlc: MODEL_GRAPH_ERROR.",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mn0g3ye8m",
+  "input_dataset_id": "d7m83gjy2"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/targeted/inference_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T21:47:50.296704+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jgnznozrg",
+  "target_model_id": "mn0g3ye8m",
+  "input_dataset_id": "d7m83gjy2",
+  "device": "Dragonwing RB3 Gen 2 Vision Kit",
+  "options": "--compute_unit npu",
+  "kind": "inference",
+  "status": "SUBMITTED",
+  "job_id": "jglyxoy85",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jglyxoy85/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/targeted/profile_report.json

@@ -0,0 +1,85 @@
+{
+  "timestamp_utc": "2026-09-23T21:53:07.997274+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "j5687r80g",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j5687r80g/",
+  "status": "FAILED",
+  "status_message": "Failed to profile the model: Failed to call QnnModel_composeGraphsFromDlc: MODEL_GRAPH_ERROR.",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mn0g3ye8m",
+  "shapes": {
+    "x": [
+      [
+        1,
+        1,
+        2048
+      ],
+      "uint16"
+    ],
+    "cos_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint16"
+    ],
+    "sin_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint16"
+    ],
+    "mask_full": [
+      [
+        1,
+        1,
+        1,
+        1025
+      ],
+      "uint16"
+    ],
+    "k_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint16"
+    ],
+    "v_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint16"
+    ]
+  }
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/targeted/profile_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T21:47:51.950992+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jgnznozrg",
+  "target_model_id": "mn0g3ye8m",
+  "input_dataset_id": null,
+  "device": "Dragonwing RB3 Gen 2 Vision Kit",
+  "options": "--compute_unit npu",
+  "kind": "profile",
+  "status": "SUBMITTED",
+  "job_id": "j5687r80g",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j5687r80g/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/w8a8/audit_report.json

@@ -0,0 +1,79 @@
+{
+  "scope": "one real Spark-X2.5 attention layer at one static 1024-token cache shape; no generation E2E",
+  "status": "component_executed_on_npu_quality_unqualified",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "compile_job_id": "jpyoxq885",
+  "inference_job_id": "jgzlm8yo5",
+  "profile_job_id": "j5wl78z3p",
+  "target_model_id": "mqkp10ezm",
+  "input_dataset_id": "d7m83gjy2",
+  "compile_options": "--target_runtime qnn_dlc --quantize_full_type int8 --quantize_io",
+  "files": {
+    "source_onnx": {
+      "bytes": 205618089,
+      "sha256": "4e2b1c8d791b1ee0e9b75f3322a20a743c0bf8c40f9b98edd963f2944d6300bb"
+    },
+    "fixture": {
+      "bytes": 3901238,
+      "sha256": "5a439e6c6d148f1720938770daa110039a39edde014f1c1592c823c7d7dfc8ed"
+    },
+    "source_cpu_output": {
+      "bytes": 12103,
+      "sha256": "6bbfb8e6ccb09475391ce1dbd2e855205720f4367c038a0d5be35e7dfeb2f768"
+    },
+    "device_output": {
+      "bytes": 4865,
+      "sha256": "f927685b0398451268464f36bb0fa00e96507bd693302a3540e4aee0d7d8233c"
+    }
+  },
+  "source_cpu_vs_device": {
+    "hidden": {
+      "relative_l2": 0.20913809283274212,
+      "max_abs": 0.035594239830970764,
+      "cosine": 0.9800697623054245
+    },
+    "k_new_0": {
+      "relative_l2": 0.01701873991186353,
+      "max_abs": 0.15612030029296875,
+      "cosine": 0.9998555562334065
+    },
+    "v_new_0": {
+      "relative_l2": 0.04413272608978388,
+      "max_abs": 0.07510209828615189,
+      "cosine": 0.9990311183867252
+    }
+  },
+  "profile": {
+    "sample_count": 100,
+    "sample_min_us": 5089,
+    "nearest_rank_p50_us": 5610,
+    "nearest_rank_p95_us": 5761,
+    "reported_estimated_inference_time_us": 5089,
+    "reported_peak_memory_bytes": 5951488,
+    "compute_unit_row_counts": {
+      "NPU": 61
+    }
+  },
+  "limits": [
+    "No next-token or task-level numerical tolerance is established.",
+    "No complete prefill, KV/ring, sampling, continuous generation or Omni deployment is tested.",
+    "Hosted component latency is not a complete request or sustained thermal profile."
+  ]
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/w8a8/inference_report.json

@@ -0,0 +1,32 @@
+{
+  "timestamp_utc": "2026-09-23T22:06:32.836584+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jgzlm8yo5",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jgzlm8yo5/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mqkp10ezm",
+  "input_dataset_id": "d7m83gjy2",
+  "outputs": "device_output.npz"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/w8a8/inference_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T21:56:26.436621+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jpyoxq885",
+  "target_model_id": "mqkp10ezm",
+  "input_dataset_id": "d7m83gjy2",
+  "device": "Dragonwing RB3 Gen 2 Vision Kit",
+  "options": "--compute_unit npu",
+  "kind": "inference",
+  "status": "SUBMITTED",
+  "job_id": "jgzlm8yo5",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jgzlm8yo5/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/w8a8/profile_report.json

@@ -0,0 +1,667 @@
+{
+  "timestamp_utc": "2026-09-23T22:06:37.770571+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "j5wl78z3p",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j5wl78z3p/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mqkp10ezm",
+  "shapes": {
+    "x": [
+      [
+        1,
+        1,
+        2048
+      ],
+      "uint8"
+    ],
+    "cos_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint8"
+    ],
+    "sin_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint8"
+    ],
+    "k_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint8"
+    ],
+    "v_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint8"
+    ],
+    "mask_full": [
+      [
+        1,
+        1,
+        1,
+        1025
+      ],
+      "uint8"
+    ]
+  },
+  "profile": {
+    "execution_summary": {
+      "estimated_inference_time": 5089,
+      "estimated_inference_peak_memory": 5951488,
+      "first_load_time": 1056655,
+      "first_load_peak_memory": 299769856,
+      "warm_load_time": 237131,
+      "warm_load_peak_memory": 8380416,
+      "compile_time": 0,
+      "compile_peak_memory": 0,
+      "compile_memory_increase_range": [
+        0,
+        0
+      ],
+      "compile_memory_peak_range": [
+        0,
+        0
+      ],
+      "first_load_memory_increase_range": [
+        87453872,
+        88254496
+      ],
+      "first_load_memory_peak_range": [
+        296873984,
+        297674608
+      ],
+      "warm_load_memory_increase_range": [
+        3804784,
+        4606864
+      ],
+      "warm_load_memory_peak_range": [
+        5480448,
+        6282528
+      ],
+      "inference_memory_increase_range": [
+        0,
+        59568
+      ],
+      "inference_memory_peak_range": [
+        565248,
+        3277312
+      ],
+      "all_compile_times": [],
+      "all_first_load_times": [
+        1056655
+      ],
+      "all_warm_load_times": [
+        237131
+      ],
+      "all_inference_times": [
+        6294,
+        5914,
+        5642,
+        5540,
+        5501,
+        5839,
+        5646,
+        5125,
+        5152,
+        5089,
+        5143,
+        5610,
+        5301,
+        5596,
+        5612,
+        5710,
+        5709,
+        5444,
+        5465,
+        5572,
+        5674,
+        5678,
+        5774,
+        5458,
+        5323,
+        5424,
+        5673,
+        5682,
+        5705,
+        5601,
+        5668,
+        5701,
+        5761,
+        5711,
+        5367,
+        5426,
+        5717,
+        5544,
+        5446,
+        5522,
+        5472,
+        5453,
+        5642,
+        5754,
+        5495,
+        5653,
+        5722,
+        5362,
+        5677,
+        5475,
+        5647,
+        5680,
+        5737,
+        5440,
+        5511,
+        5725,
+        5452,
+        5465,
+        5656,
+        5436,
+        5843,
+        5700,
+        5492,
+        5421,
+        5414,
+        5689,
+        5702,
+        5277,
+        5428,
+        5434,
+        5502,
+        5648,
+        5437,
+        5727,
+        5698,
+        5440,
+        5412,
+        5705,
+        5722,
+        5704,
+        5424,
+        5439,
+        5427,
+        5353,
+        5698,
+        5432,
+        5652,
+        5692,
+        5490,
+        5400,
+        5399,
+        5635,
+        5650,
+        5708,
+        5642,
+        5704,
+        5709,
+        5694,
+        5749,
+        5443
+      ]
+    },
+    "execution_detail": [
+      {
+        "name": "Input",
+        "type": "Input",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 356410
+      },
+      {
+        "name": "rms_norm_/layers.0/Mul_1_output_0",
+        "type": "rms",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 54755
+      },
+      {
+        "name": "/layers.0/q_k_v_proj/MatMul",
+        "type": "/layers.0/q",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/q_k_v_proj/MatMul_post_reshape",
+        "type": "/layers.0/q",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "node_Split_20_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 442481
+      },
+      {
+        "name": "node_Split_20_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 112903
+      },
+      {
+        "name": "node_Split_20_to_slice_2",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 121150
+      },
+      {
+        "name": "/layers.0/Reshape",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_1",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_2",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/g_proj/MatMul",
+        "type": "/layers.0/g",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 7832
+      },
+      {
+        "name": "/layers.0/g_proj/MatMul_post_reshape",
+        "type": "/layers.0/g",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Sigmoid",
+        "type": "/layers.0/Sigmoid",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1767
+      },
+      {
+        "name": "/layers.0/Reshape_3",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 677
+      },
+      {
+        "name": "node_Split_22_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 282
+      },
+      {
+        "name": "node_Split_22_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 323
+      },
+      {
+        "name": "fused_/layers.0/Slice_3_/layers.0/Slice_5",
+        "type": "fused",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 337
+      },
+      {
+        "name": "fused_/layers.0/Slice_3_/layers.0/Slice_6",
+        "type": "fused",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 451
+      },
+      {
+        "name": "/layers.0/Mul_2",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 602
+      },
+      {
+        "name": "/layers.0/Neg",
+        "type": "/layers.0/Neg",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 172
+      },
+      {
+        "name": "/layers.0/Concat",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 798
+      },
+      {
+        "name": "/layers.0/Mul_3",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1364
+      },
+      {
+        "name": "/layers.0/Add_1",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 961
+      },
+      {
+        "name": "/layers.0/Concat_1",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 660
+      },
+      {
+        "name": "node_Split_24_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 275
+      },
+      {
+        "name": "node_Split_24_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1825
+      },
+      {
+        "name": "fused_/layers.0/Slice_7_/layers.0/Slice_9",
+        "type": "fused",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 232
+      },
+      {
+        "name": "fused_/layers.0/Slice_7_/layers.0/Slice_10",
+        "type": "fused",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3753
+      },
+      {
+        "name": "/layers.0/Mul_4",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 581
+      },
+      {
+        "name": "/layers.0/Neg_1",
+        "type": "/layers.0/Neg",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 653
+      },
+      {
+        "name": "/layers.0/Concat_2",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1227
+      },
+      {
+        "name": "/layers.0/Mul_5",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1114
+      },
+      {
+        "name": "/layers.0/Add_2",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3012
+      },
+      {
+        "name": "/layers.0/Concat_3",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2091
+      },
+      {
+        "name": "/layers.0/Concat_4",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 9979
+      },
+      {
+        "name": "/layers.0/Concat_5",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_4",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2759
+      },
+      {
+        "name": "/layers.0/Transpose",
+        "type": "/layers.0/Transpose",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/MatMul",
+        "type": "/layers.0/MatMul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 414649
+      },
+      {
+        "name": "/layers.0/Mul_6",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3908
+      },
+      {
+        "name": "/layers.0/Add_3",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 6454
+      },
+      {
+        "name": "/layers.0/Softmax",
+        "type": "/layers.0/Softmax",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 10636
+      },
+      {
+        "name": "/layers.0/MatMul_1",
+        "type": "/layers.0/MatMul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 212095
+      },
+      {
+        "name": "/layers.0/Reshape_5",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_7",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 17068
+      },
+      {
+        "name": "/layers.0/out_proj/MatMul",
+        "type": "/layers.0/out",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 443675
+      },
+      {
+        "name": "/layers.0/out_proj/MatMul_post_reshape",
+        "type": "/layers.0/out",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_4",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 7210
+      },
+      {
+        "name": "rms_norm_/layers.0/Mul_9_output_0",
+        "type": "rms",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 24034
+      },
+      {
+        "name": "/layers.0/gate_proj/MatMul",
+        "type": "/layers.0/gate",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1434127
+      },
+      {
+        "name": "/layers.0/gate_proj/MatMul_post_reshape",
+        "type": "/layers.0/gate",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "node_Gelu_2",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 10469
+      },
+      {
+        "name": "/layers.0/up_proj/MatMul",
+        "type": "/layers.0/up",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1430626
+      },
+      {
+        "name": "/layers.0/up_proj/MatMul_post_reshape",
+        "type": "/layers.0/up",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_12",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 38676
+      },
+      {
+        "name": "/layers.0/down_proj/MatMul",
+        "type": "/layers.0/down",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1440368
+      },
+      {
+        "name": "/layers.0/down_proj/MatMul_post_reshape",
+        "type": "/layers.0/down",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_7",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1363
+      },
+      {
+        "name": "/layers.0/Concat_3_identity",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_2_identity",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "Output",
+        "type": "Output",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 10720
+      }
+    ]
+  }
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/w8a8/profile_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T21:56:28.293907+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jpyoxq885",
+  "target_model_id": "mqkp10ezm",
+  "input_dataset_id": null,
+  "device": "Dragonwing RB3 Gen 2 Vision Kit",
+  "options": "--compute_unit npu",
+  "kind": "profile",
+  "status": "SUBMITTED",
+  "job_id": "j5wl78z3p",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j5wl78z3p/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/README.md

@@ -0,0 +1,11 @@
+# Spark-X2.5 full-attention layer on Galaxy S24 NPU
+
+**One real Spark-X2.5-1.7B full-attention decoder layer ran on a hosted Samsung Galaxy S24 NPU; complete Spark generation did not.** This is a fixed decode-step component with a 1,024-token full-attention cache, not the 28-layer model, prefill, sampling, a resident KV/ring session or a device-local Omni controller. Qualcomm AI Hub Workbench provisions and releases a physical device for each component job; it is a test facility, not the deployment runtime. ADB still lists no attached device.
+
+The source is the existing FP32 `spark_full_ctx1024.onnx` export from real Spark weights. Its SHA256 is `4e2b1c8d791b1ee0e9b75f3322a20a743c0bf8c40f9b98edd963f2944d6300bb` (205,618,089 bytes); downloading uploaded source model `mn14dw6zn` yielded the same byte-for-byte hash. Workbench [quantization job](quantization_provenance.json) `jg9zne6wp` used calibration dataset `d7x8e3d59` and produced QDQ ONNX source model `mmd5e3v3q`. Its downloaded external-data ZIP SHA256 is `703821ca8ace5672228919ea2add4901def4887748c381b12ffc8b259665c553`; the graph has 92 QuantizeLinear/92 DequantizeLinear nodes, INT8 weight zero-points and UINT16 activation zero-points (W8A16). The original and quantized graphs and the exact inference fixture are stored outside Git under `/home/zhout/project/edge_infer/models/`; the [audit](audit_report.json) pins their sizes and hashes.
+
+The [compile job](compile_report.json) succeeded for exact device **Samsung Galaxy S24, Android 14**, using Workbench client `0.55.0`, service build recorded in the job JSON, target model `mq9y70k0n` and `--target_runtime qnn_dlc --quantize_io`. The old context-binary target was [rejected at submission](compile_legacy_option_rejection.json) by the current API; [Qualcomm's compile options](https://workbench.aihub.qualcomm.com/docs/hub/api.html) list `qnn_dlc` as a current target. The [inference job](inference_report.json) used the pinned dataset `d7m83gjy2` and requested `--compute_unit npu`; its three finite FP32 outputs are [retained](s24_output.npz). The [profile job](profile_report.json) also requested NPU and attributed all 73 execution-detail rows to NPU. Its 100 `all_inference_times` samples had nearest-rank p50/p95 **1,109/1,205 microseconds** (minimum 1,098 microseconds); the separate `estimated_inference_time` field equals the minimum and is not a median. Reported inference peak memory was 210,259,968 bytes. These are component-job figures; they exclude controller, transfers between stages, prefill, all other layers, sampling, loading and sustained heat.
+
+The fixed fixture has FP32 input `[1,1,2048]`, cache tensors `[1,2,1024,256]`, rotary angles and mask. Local ONNX Runtime 1.29 CPU executions of the original FP32 source and downloaded QDQ source produced separate pinned outputs. Relative L2 errors for `hidden` / new key / new value were **1.236% / 0.376% / 0.651%** for QDQ CPU versus FP32 CPU, then **0.555% / 0.0065% / 0.0126%** for S24 NPU versus QDQ CPU. S24 and the prior S25 [NPU output](s25_output_prior.npz) matched exactly on the same dataset ([S25 inference provenance](s25_inference_prior.json), [S25 profile](s25_profile_prior.json)), although they used different compiled target formats and Android versions. These measurements suggest quantization contributes substantially to the source-to-device difference; they do **not** establish next-token or task-quality tolerance. The [audit report](audit_report.json) retains full precision metrics, cosines, artifact hashes, sample statistics and job IDs. The [audit script](../../../../experiments/audit_spark_qualcomm_attention.py) reproduces those comparisons with the downloaded artifacts and fixture.
+
+The matrix remains **NOT E2E**. The next gate is a device-resident, complete Spark prefill/continuous-decode/KV-ring/sampling loop with explicit state lifecycle, numeric token-quality checks, 128-token generation across 512/1024 context transitions, cancellation, memory admission and sustained thermal/power profiling. Workbench component jobs cannot substitute for that local runtime.

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/audit_report.json

@@ -0,0 +1,143 @@
+{
+  "scope": "one real Spark-X2.5 full-attention decoder layer at one static 1024-token cache shape; no generation E2E",
+  "status": "component_executed_on_npu_quality_unqualified",
+  "device": {
+    "name": "Samsung Galaxy S24",
+    "os": "14",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "abi:aarch64-android",
+      "vendor:samsung",
+      "format:phone",
+      "chipset:qualcomm-snapdragon-8gen3",
+      "chipset:sm8650",
+      "hexagon:v75",
+      "soc-model:57",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "compile_job_id": "jpe7z8l15",
+  "inference_job_id": "j57e4k9lp",
+  "profile_job_id": "jp4y1m3vp",
+  "target_model_id": "mq9y70k0n",
+  "input_dataset_id": "d7m83gjy2",
+  "s25_prior_inference_job_id": "jp0mvkj6g",
+  "files": {
+    "source_onnx": {
+      "bytes": 205618089,
+      "sha256": "4e2b1c8d791b1ee0e9b75f3322a20a743c0bf8c40f9b98edd963f2944d6300bb"
+    },
+    "quantized_onnx_zip": {
+      "bytes": 186772390,
+      "sha256": "703821ca8ace5672228919ea2add4901def4887748c381b12ffc8b259665c553"
+    },
+    "fixture": {
+      "bytes": 3901238,
+      "sha256": "5a439e6c6d148f1720938770daa110039a39edde014f1c1592c823c7d7dfc8ed"
+    },
+    "source_cpu_output": {
+      "bytes": 12103,
+      "sha256": "6bbfb8e6ccb09475391ce1dbd2e855205720f4367c038a0d5be35e7dfeb2f768"
+    },
+    "quantized_cpu_output": {
+      "bytes": 11574,
+      "sha256": "061654773166d102a11228c4a2e4cd06de6c2482c7e4e7860315b7b9a95fa8a1"
+    },
+    "s25_output": {
+      "bytes": 13072,
+      "sha256": "2dea9f4b478cb02b9be74a8d3ea4ad42e562e4de72413967d22706a1b6c6b9c0"
+    },
+    "device_output": {
+      "bytes": 11576,
+      "sha256": "9381fccee14b303edc5b432edf92422f2948e111ea6848df22cba3cd289a7073"
+    }
+  },
+  "source_vs_quantized_cpu": {
+    "hidden": {
+      "relative_l2": 0.012364819167814612,
+      "max_abs": 0.0028282999992370605,
+      "cosine": 0.9999236499353131
+    },
+    "k_new_0": {
+      "relative_l2": 0.0037627070087598766,
+      "max_abs": 0.03641480207443237,
+      "cosine": 0.9999929226489408
+    },
+    "v_new_0": {
+      "relative_l2": 0.006507432119585315,
+      "max_abs": 0.011739224195480347,
+      "cosine": 0.9999789079393114
+    }
+  },
+  "quantized_cpu_vs_device_npu": {
+    "hidden": {
+      "relative_l2": 0.005550082388164841,
+      "max_abs": 0.002016603946685791,
+      "cosine": 0.9999847274053486
+    },
+    "k_new_0": {
+      "relative_l2": 6.533750243101402e-05,
+      "max_abs": 0.00033855438232421875,
+      "cosine": 0.9999999978819449
+    },
+    "v_new_0": {
+      "relative_l2": 0.0001257028618212242,
+      "max_abs": 0.0002911090850830078,
+      "cosine": 0.9999999921460195
+    }
+  },
+  "source_cpu_vs_device_npu": {
+    "hidden": {
+      "relative_l2": 0.01350919207884596,
+      "max_abs": 0.0048449039459228516,
+      "cosine": 0.9999091967737982
+    },
+    "k_new_0": {
+      "relative_l2": 0.003766040678719007,
+      "max_abs": 0.03641480207443237,
+      "cosine": 0.9999929097855774
+    },
+    "v_new_0": {
+      "relative_l2": 0.006504874746412689,
+      "max_abs": 0.011739224195480347,
+      "cosine": 0.9999789285134126
+    }
+  },
+  "s25_npu_vs_device_npu": {
+    "hidden": {
+      "relative_l2": 0.0,
+      "max_abs": 0.0,
+      "cosine": 0.9999999999999998
+    },
+    "k_new_0": {
+      "relative_l2": 0.0,
+      "max_abs": 0.0,
+      "cosine": 0.9999999999999998
+    },
+    "v_new_0": {
+      "relative_l2": 0.0,
+      "max_abs": 0.0,
+      "cosine": 1.0
+    }
+  },
+  "profile": {
+    "sample_count": 100,
+    "sample_min_us": 1098,
+    "nearest_rank_p50_us": 1109,
+    "nearest_rank_p95_us": 1205,
+    "reported_estimated_inference_time_us": 1098,
+    "reported_peak_memory_bytes": 210259968,
+    "compute_unit_row_counts": {
+      "NPU": 73
+    }
+  },
+  "limits": [
+    "No task-level numerical tolerance or next-token quality gate is established.",
+    "Hub component jobs do not prove device-local prefill, KV/ring update, sampling, continuous generation or Omni deployment.",
+    "Profile sample latency is one component, not complete-request or sustained thermal performance."
+  ]
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/compile_legacy_option_rejection.json

@@ -0,0 +1,9 @@
+{
+  "timestamp_utc": "2026-09-23T21:12:36.038442+00:00",
+  "qai_hub_version": "0.55.0",
+  "device": "Samsung Galaxy S24",
+  "source_quantized_model_id": "mmd5e3v3q",
+  "requested_options": "--target_runtime qnn_context_binary --quantize_io",
+  "status": "REJECTED_AT_SUBMISSION",
+  "message": "Unsupported target runtime: qnn_context_binary. Supported values reported by API: onnx, precompiled_qnn_onnx, qnn_dlc, tflite."
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/compile_report.json

@@ -0,0 +1,32 @@
+{
+  "timestamp_utc": "2026-09-23T21:15:44.477548+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jpe7z8l15",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jpe7z8l15/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "source_quantized_model_id": "mmd5e3v3q",
+  "target_model_id": "mq9y70k0n",
+  "device": {
+    "name": "Samsung Galaxy S24",
+    "os": "14",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "abi:aarch64-android",
+      "vendor:samsung",
+      "format:phone",
+      "chipset:qualcomm-snapdragon-8gen3",
+      "chipset:sm8650",
+      "hexagon:v75",
+      "soc-model:57",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "options": "--target_runtime qnn_dlc --quantize_io",
+  "hub_version": "aihub-2026.09.11.0"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/compile_submission.json

@@ -0,0 +1,11 @@
+{
+  "timestamp_utc": "2026-09-23T21:12:39.099981+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "job_id": "jpe7z8l15",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jpe7z8l15/",
+  "source_quantized_model_id": "mmd5e3v3q",
+  "source_onnx_sha256": "4e2b1c8d791b1ee0e9b75f3322a20a743c0bf8c40f9b98edd963f2944d6300bb",
+  "device": "Samsung Galaxy S24",
+  "options": "--target_runtime qnn_dlc --quantize_io"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/inference_report.json

@@ -0,0 +1,33 @@
+{
+  "timestamp_utc": "2026-09-23T21:19:44.223245+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "j57e4k9lp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j57e4k9lp/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Samsung Galaxy S24",
+    "os": "14",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "abi:aarch64-android",
+      "vendor:samsung",
+      "format:phone",
+      "chipset:qualcomm-snapdragon-8gen3",
+      "chipset:sm8650",
+      "hexagon:v75",
+      "soc-model:57",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mq9y70k0n",
+  "input_dataset_id": "d7m83gjy2",
+  "outputs": "s24_output.npz"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/inference_submission.json

@@ -0,0 +1,9 @@
+{
+  "timestamp_utc": "2026-09-23T21:15:46.475323+00:00",
+  "job_id": "j57e4k9lp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j57e4k9lp/",
+  "target_model_id": "mq9y70k0n",
+  "device": "Samsung Galaxy S24",
+  "input_dataset_id": "d7m83gjy2",
+  "options": "--compute_unit npu"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/profile_report.json

@@ -0,0 +1,752 @@
+{
+  "timestamp_utc": "2026-09-23T21:19:48.832365+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jp4y1m3vp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jp4y1m3vp/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Samsung Galaxy S24",
+    "os": "14",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "abi:aarch64-android",
+      "vendor:samsung",
+      "format:phone",
+      "chipset:qualcomm-snapdragon-8gen3",
+      "chipset:sm8650",
+      "hexagon:v75",
+      "soc-model:57",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mq9y70k0n",
+  "shapes": {
+    "x": [
+      [
+        1,
+        1,
+        2048
+      ],
+      "uint16"
+    ],
+    "v_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint16"
+    ],
+    "cos_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint16"
+    ],
+    "sin_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint16"
+    ],
+    "k_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint16"
+    ],
+    "mask_full": [
+      [
+        1,
+        1,
+        1,
+        1025
+      ],
+      "uint16"
+    ]
+  },
+  "profile": {
+    "execution_summary": {
+      "estimated_inference_time": 1098,
+      "estimated_inference_peak_memory": 210259968,
+      "first_load_time": 943277,
+      "first_load_peak_memory": 340230144,
+      "warm_load_time": 174148,
+      "warm_load_peak_memory": 214511616,
+      "compile_time": 0,
+      "compile_peak_memory": 0,
+      "compile_memory_increase_range": [
+        0,
+        0
+      ],
+      "compile_memory_peak_range": [
+        0,
+        0
+      ],
+      "first_load_memory_increase_range": [
+        62276784,
+        69247552
+      ],
+      "first_load_memory_peak_range": [
+        258228224,
+        265198992
+      ],
+      "warm_load_memory_increase_range": [
+        0,
+        1646624
+      ],
+      "warm_load_memory_peak_range": [
+        606208,
+        113631744
+      ],
+      "inference_memory_increase_range": [
+        0,
+        0
+      ],
+      "inference_memory_peak_range": [
+        20480,
+        111512672
+      ],
+      "all_compile_times": [],
+      "all_first_load_times": [
+        943277
+      ],
+      "all_warm_load_times": [
+        174148
+      ],
+      "all_inference_times": [
+        1834,
+        1201,
+        1153,
+        1156,
+        1165,
+        1143,
+        1133,
+        1141,
+        1170,
+        1150,
+        1144,
+        1425,
+        1225,
+        1205,
+        1194,
+        1212,
+        1106,
+        1103,
+        1104,
+        1098,
+        1098,
+        1098,
+        1107,
+        1100,
+        1110,
+        1102,
+        1103,
+        1100,
+        1099,
+        1109,
+        1151,
+        1120,
+        1116,
+        1131,
+        1108,
+        1107,
+        1107,
+        1118,
+        1292,
+        1123,
+        1109,
+        1115,
+        1111,
+        1107,
+        1107,
+        1111,
+        1111,
+        1107,
+        1112,
+        1105,
+        1110,
+        1107,
+        1106,
+        1121,
+        1106,
+        1107,
+        1106,
+        1105,
+        1105,
+        1123,
+        1118,
+        1111,
+        1120,
+        1109,
+        1105,
+        1104,
+        1102,
+        1115,
+        1105,
+        1108,
+        1108,
+        1110,
+        1104,
+        1108,
+        1109,
+        1139,
+        1130,
+        1125,
+        1124,
+        1118,
+        1104,
+        1098,
+        1119,
+        1098,
+        1099,
+        1098,
+        1107,
+        1102,
+        1135,
+        1117,
+        1131,
+        1106,
+        1103,
+        1121,
+        1109,
+        1109,
+        1117,
+        1111,
+        1106,
+        1107
+      ]
+    },
+    "execution_detail": [
+      {
+        "name": "Input",
+        "type": "Input",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 42949
+      },
+      {
+        "name": "/layers.0/Pow",
+        "type": "/layers.0/Pow",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 41382
+      },
+      {
+        "name": "/layers.0/ReduceMean",
+        "type": "/layers.0/ReduceMean",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 8204
+      },
+      {
+        "name": "/layers.0/Add",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 15143
+      },
+      {
+        "name": "/layers.0/Sqrt",
+        "type": "/layers.0/Sqrt",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Div",
+        "type": "/layers.0/Div",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 82056
+      },
+      {
+        "name": "/layers.0/Mul",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 61350
+      },
+      {
+        "name": "/layers.0/Mul_1",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 17794
+      },
+      {
+        "name": "/layers.0/q_k_v_proj/MatMul",
+        "type": "/layers.0/q",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/q_k_v_proj/MatMul_post_reshape",
+        "type": "/layers.0/q",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/g_proj/MatMul",
+        "type": "/layers.0/g",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 6373
+      },
+      {
+        "name": "/layers.0/g_proj/MatMul_post_reshape",
+        "type": "/layers.0/g",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "node_Split_1_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 44219
+      },
+      {
+        "name": "node_Split_1_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 23821
+      },
+      {
+        "name": "node_Split_1_to_slice_2",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 9323
+      },
+      {
+        "name": "/layers.0/Sigmoid",
+        "type": "/layers.0/Sigmoid",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_1",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_2",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 6592
+      },
+      {
+        "name": "/layers.0/Reshape_3",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 8691
+      },
+      {
+        "name": "node_Split_3_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 456
+      },
+      {
+        "name": "node_Split_3_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1295
+      },
+      {
+        "name": "node_Split_5_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 340
+      },
+      {
+        "name": "node_Split_5_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 5592
+      },
+      {
+        "name": "/layers.0/Concat_5",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 31506
+      },
+      {
+        "name": "node_Split_7_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1508
+      },
+      {
+        "name": "node_Split_7_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 471
+      },
+      {
+        "name": "/layers.0/Mul_2",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2954
+      },
+      {
+        "name": "node_Split_9_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 4193
+      },
+      {
+        "name": "node_Split_9_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1297
+      },
+      {
+        "name": "/layers.0/Mul_4",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 6658
+      },
+      {
+        "name": "/layers.0/Neg",
+        "type": "/layers.0/Neg",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1367
+      },
+      {
+        "name": "/layers.0/Neg_1",
+        "type": "/layers.0/Neg",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 4097
+      },
+      {
+        "name": "/layers.0/Concat",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 727
+      },
+      {
+        "name": "/layers.0/Concat_2",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3507
+      },
+      {
+        "name": "/layers.0/Mul_3",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3769
+      },
+      {
+        "name": "/layers.0/Mul_5",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2749
+      },
+      {
+        "name": "/layers.0/Add_1",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1524
+      },
+      {
+        "name": "/layers.0/Add_2",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 4679
+      },
+      {
+        "name": "/layers.0/Concat_1",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 833
+      },
+      {
+        "name": "/layers.0/Concat_3",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 791
+      },
+      {
+        "name": "/layers.0/Reshape_4",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 248
+      },
+      {
+        "name": "/layers.0/Concat_4",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Transpose",
+        "type": "/layers.0/Transpose",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/MatMul",
+        "type": "/layers.0/MatMul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 685248
+      },
+      {
+        "name": "/layers.0/Mul_6",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2929
+      },
+      {
+        "name": "/layers.0/Add_3",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 26814
+      },
+      {
+        "name": "/layers.0/Softmax",
+        "type": "/layers.0/Softmax",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 17675
+      },
+      {
+        "name": "/layers.0/MatMul_1",
+        "type": "/layers.0/MatMul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 637945
+      },
+      {
+        "name": "/layers.0/Reshape_5",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_7",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 22264
+      },
+      {
+        "name": "/layers.0/out_proj/MatMul",
+        "type": "/layers.0/out",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 66743
+      },
+      {
+        "name": "/layers.0/out_proj/MatMul_post_reshape",
+        "type": "/layers.0/out",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_4",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 6574
+      },
+      {
+        "name": "/layers.0/Pow_1",
+        "type": "/layers.0/Pow",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3443
+      },
+      {
+        "name": "/layers.0/ReduceMean_1",
+        "type": "/layers.0/ReduceMean",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1522
+      },
+      {
+        "name": "/layers.0/Add_5",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 179
+      },
+      {
+        "name": "/layers.0/Sqrt_1",
+        "type": "/layers.0/Sqrt",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Div_1",
+        "type": "/layers.0/Div",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 57936
+      },
+      {
+        "name": "/layers.0/Mul_8",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 34269
+      },
+      {
+        "name": "/layers.0/Mul_9",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 7526
+      },
+      {
+        "name": "/layers.0/gate_proj/MatMul",
+        "type": "/layers.0/gate",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/gate_proj/MatMul_post_reshape",
+        "type": "/layers.0/gate",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/up_proj/MatMul",
+        "type": "/layers.0/up",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 426843
+      },
+      {
+        "name": "/layers.0/up_proj/MatMul_post_reshape",
+        "type": "/layers.0/up",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "_elementwiseneuron_0",
+        "type": "elementwiseneuron",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 419979
+      },
+      {
+        "name": "/layers.0/Mul_12",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 12346
+      },
+      {
+        "name": "/layers.0/down_proj/MatMul",
+        "type": "/layers.0/down",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 406308
+      },
+      {
+        "name": "/layers.0/down_proj/MatMul_post_reshape",
+        "type": "/layers.0/down",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_7",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 5803
+      },
+      {
+        "name": "/layers.0/Concat_3_identity",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_2_identity",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "Output",
+        "type": "Output",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 8424
+      }
+    ]
+  }
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/profile_submission.json

@@ -0,0 +1,9 @@
+{
+  "timestamp_utc": "2026-09-23T21:15:47.628043+00:00",
+  "job_id": "jp4y1m3vp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jp4y1m3vp/",
+  "target_model_id": "mq9y70k0n",
+  "device": "Samsung Galaxy S24",
+  "input_dataset_id": null,
+  "options": "--compute_unit npu"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/quantization_provenance.json

@@ -0,0 +1,17 @@
+{
+  "timestamp_utc": "2026-09-23T21:25:02.276846+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jg9zne6wp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jg9zne6wp/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "source_model_id": "mn14dw6zn",
+  "source_model_name": "spark25_full1024_src",
+  "source_download_sha256": "4e2b1c8d791b1ee0e9b75f3322a20a743c0bf8c40f9b98edd963f2944d6300bb",
+  "calibration_dataset_id": "d7x8e3d59",
+  "target_quantized_model_id": "mmd5e3v3q",
+  "downloaded_qdq_zip_sha256": "703821ca8ace5672228919ea2add4901def4887748c381b12ffc8b259665c553",
+  "options": "",
+  "hub_version": "aihub-2026.08.28.1"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/s25_inference_prior.json

@@ -0,0 +1,34 @@
+{
+  "timestamp_utc": "2026-09-22T03:52:48.169689+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "command": "collect",
+  "job_id": "jp0mvkj6g",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jp0mvkj6g/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Samsung Galaxy S25",
+    "os": "15",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "abi:aarch64-android",
+      "vendor:samsung",
+      "format:phone",
+      "chipset:qualcomm-snapdragon-8-elite-for-galaxy",
+      "chipset:sm8750-ac",
+      "hexagon:v79",
+      "soc-model:69",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mnowl6xpq",
+  "input_dataset_id": "d7m83gjy2",
+  "outputs": "aihub-s25-infer.npz"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/s25_profile_prior.json

@@ -0,0 +1,753 @@
+{
+  "timestamp_utc": "2026-09-22T03:53:16.578078+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "command": "collect",
+  "job_id": "jpyo7nx75",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jpyo7nx75/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Samsung Galaxy S25",
+    "os": "15",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "abi:aarch64-android",
+      "vendor:samsung",
+      "format:phone",
+      "chipset:qualcomm-snapdragon-8-elite-for-galaxy",
+      "chipset:sm8750-ac",
+      "hexagon:v79",
+      "soc-model:69",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mnowl6xpq",
+  "shapes": {
+    "x": [
+      [
+        1,
+        1,
+        2048
+      ],
+      "uint16"
+    ],
+    "v_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint16"
+    ],
+    "cos_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint16"
+    ],
+    "sin_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint16"
+    ],
+    "k_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint16"
+    ],
+    "mask_full": [
+      [
+        1,
+        1,
+        1,
+        1025
+      ],
+      "uint16"
+    ]
+  },
+  "profile": {
+    "execution_summary": {
+      "estimated_inference_time": 945,
+      "estimated_inference_peak_memory": 140197888,
+      "first_load_time": 113160,
+      "first_load_peak_memory": 113930240,
+      "warm_load_time": 113545,
+      "warm_load_peak_memory": 125128704,
+      "compile_time": 0,
+      "compile_peak_memory": 0,
+      "compile_memory_increase_range": [
+        0,
+        0
+      ],
+      "compile_memory_peak_range": [
+        0,
+        0
+      ],
+      "first_load_memory_increase_range": [
+        0,
+        2327312
+      ],
+      "first_load_memory_peak_range": [
+        1159168,
+        10519888
+      ],
+      "warm_load_memory_increase_range": [
+        0,
+        1550192
+      ],
+      "warm_load_memory_peak_range": [
+        0,
+        9639312
+      ],
+      "inference_memory_increase_range": [
+        0,
+        2138160
+      ],
+      "inference_memory_peak_range": [
+        0,
+        9296720
+      ],
+      "all_compile_times": [],
+      "all_first_load_times": [
+        113160
+      ],
+      "all_warm_load_times": [
+        113545
+      ],
+      "all_inference_times": [
+        1683,
+        1022,
+        967,
+        956,
+        957,
+        947,
+        973,
+        951,
+        950,
+        957,
+        953,
+        1142,
+        996,
+        980,
+        977,
+        1006,
+        973,
+        976,
+        982,
+        1005,
+        978,
+        973,
+        975,
+        980,
+        963,
+        997,
+        958,
+        973,
+        994,
+        1028,
+        1015,
+        1004,
+        1030,
+        1017,
+        1052,
+        1092,
+        971,
+        953,
+        950,
+        951,
+        947,
+        958,
+        947,
+        945,
+        970,
+        948,
+        950,
+        947,
+        954,
+        949,
+        957,
+        970,
+        953,
+        949,
+        950,
+        952,
+        949,
+        950,
+        947,
+        952,
+        957,
+        949,
+        948,
+        952,
+        948,
+        957,
+        945,
+        952,
+        961,
+        946,
+        947,
+        948,
+        947,
+        954,
+        947,
+        948,
+        947,
+        970,
+        951,
+        946,
+        990,
+        955,
+        987,
+        959,
+        952,
+        997,
+        967,
+        948,
+        947,
+        945,
+        950,
+        959,
+        947,
+        945,
+        950,
+        991,
+        990,
+        985,
+        978,
+        962
+      ]
+    },
+    "execution_detail": [
+      {
+        "name": "Input",
+        "type": "Input",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 27823
+      },
+      {
+        "name": "/layers.0/Pow",
+        "type": "/layers.0/Pow",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 39594
+      },
+      {
+        "name": "/layers.0/ReduceMean",
+        "type": "/layers.0/ReduceMean",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 9352
+      },
+      {
+        "name": "/layers.0/Add",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 5344
+      },
+      {
+        "name": "/layers.0/Sqrt",
+        "type": "/layers.0/Sqrt",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Div",
+        "type": "/layers.0/Div",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 83375
+      },
+      {
+        "name": "/layers.0/Mul",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 72302
+      },
+      {
+        "name": "/layers.0/Mul_1",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 17867
+      },
+      {
+        "name": "/layers.0/q_k_v_proj/MatMul",
+        "type": "/layers.0/q",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/q_k_v_proj/MatMul_post_reshape",
+        "type": "/layers.0/q",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/g_proj/MatMul",
+        "type": "/layers.0/g",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 9085
+      },
+      {
+        "name": "/layers.0/g_proj/MatMul_post_reshape",
+        "type": "/layers.0/g",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "node_Slice_748",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 50418
+      },
+      {
+        "name": "node_Slice_751",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 20704
+      },
+      {
+        "name": "node_Slice_754",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 6418
+      },
+      {
+        "name": "/layers.0/Sigmoid",
+        "type": "/layers.0/Sigmoid",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_1",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_2",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 5771
+      },
+      {
+        "name": "/layers.0/Reshape_3",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 18115
+      },
+      {
+        "name": "node_Slice_531",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 506
+      },
+      {
+        "name": "node_Slice_534",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1788
+      },
+      {
+        "name": "node_Slice_538",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 760
+      },
+      {
+        "name": "node_Slice_541",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 9602
+      },
+      {
+        "name": "/layers.0/Concat_5",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 26383
+      },
+      {
+        "name": "node_Slice_545",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1228
+      },
+      {
+        "name": "node_Slice_548",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 573
+      },
+      {
+        "name": "/layers.0/Mul_2",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3456
+      },
+      {
+        "name": "node_Slice_1174",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 8335
+      },
+      {
+        "name": "node_Slice_1177",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 866
+      },
+      {
+        "name": "/layers.0/Mul_4",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 19251
+      },
+      {
+        "name": "/layers.0/Neg",
+        "type": "/layers.0/Neg",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 788
+      },
+      {
+        "name": "/layers.0/Neg_1",
+        "type": "/layers.0/Neg",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 5970
+      },
+      {
+        "name": "/layers.0/Concat",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 886
+      },
+      {
+        "name": "/layers.0/Concat_2",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3254
+      },
+      {
+        "name": "/layers.0/Mul_3",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3302
+      },
+      {
+        "name": "/layers.0/Mul_5",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2889
+      },
+      {
+        "name": "/layers.0/Add_1",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1082
+      },
+      {
+        "name": "/layers.0/Add_2",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 4002
+      },
+      {
+        "name": "/layers.0/Concat_1",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 823
+      },
+      {
+        "name": "/layers.0/Concat_3",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 842
+      },
+      {
+        "name": "/layers.0/Reshape_4",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 319
+      },
+      {
+        "name": "/layers.0/Concat_4",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Transpose",
+        "type": "/layers.0/Transpose",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/MatMul",
+        "type": "/layers.0/MatMul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 867826
+      },
+      {
+        "name": "/layers.0/Mul_6",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2467
+      },
+      {
+        "name": "/layers.0/Add_3",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 50892
+      },
+      {
+        "name": "/layers.0/Softmax",
+        "type": "/layers.0/Softmax",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 13765
+      },
+      {
+        "name": "/layers.0/MatMul_1",
+        "type": "/layers.0/MatMul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 611657
+      },
+      {
+        "name": "/layers.0/Reshape_5",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_7",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 29593
+      },
+      {
+        "name": "/layers.0/out_proj/MatMul",
+        "type": "/layers.0/out",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 53626
+      },
+      {
+        "name": "/layers.0/out_proj/MatMul_post_reshape",
+        "type": "/layers.0/out",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_4",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 7114
+      },
+      {
+        "name": "/layers.0/Pow_1",
+        "type": "/layers.0/Pow",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 8995
+      },
+      {
+        "name": "/layers.0/ReduceMean_1",
+        "type": "/layers.0/ReduceMean",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1481
+      },
+      {
+        "name": "/layers.0/Add_5",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 196
+      },
+      {
+        "name": "/layers.0/Sqrt_1",
+        "type": "/layers.0/Sqrt",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Div_1",
+        "type": "/layers.0/Div",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 60120
+      },
+      {
+        "name": "/layers.0/Mul_8",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 42681
+      },
+      {
+        "name": "/layers.0/Mul_9",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 8458
+      },
+      {
+        "name": "/layers.0/gate_proj/MatMul",
+        "type": "/layers.0/gate",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/gate_proj/MatMul_post_reshape",
+        "type": "/layers.0/gate",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/up_proj/MatMul",
+        "type": "/layers.0/up",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 389054
+      },
+      {
+        "name": "/layers.0/up_proj/MatMul_post_reshape",
+        "type": "/layers.0/up",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "_elementwiseneuron_0",
+        "type": "elementwiseneuron",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 395384
+      },
+      {
+        "name": "/layers.0/Mul_12",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 18593
+      },
+      {
+        "name": "/layers.0/down_proj/MatMul",
+        "type": "/layers.0/down",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 415523
+      },
+      {
+        "name": "/layers.0/down_proj/MatMul_post_reshape",
+        "type": "/layers.0/down",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_7",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 5800
+      },
+      {
+        "name": "/layers.0/Concat_3_identity",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_2_identity",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "Output",
+        "type": "Output",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 7431
+      }
+    ]
+  }
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_sa8775_full_attention/README.md

@@ -0,0 +1,13 @@
+# Spark-X2.5 attention layer on SA8775P ADP
+
+**One fixed-shape attention layer from the real Spark-X2.5-1.7B checkpoint executed on the hosted SA8775P ADP NPU (Android 14). This is component evidence, not complete Spark generation or an embedded Omni deployment.**
+
+The [Galaxy S24 provenance record](../spark_s24_full_attention/README.md) pins the FP32 source ONNX, W8A16 QDQ ONNX, calibration and exact input fixture. Its QNN DLC compile job `jpe7z8l15` produced target model `mq9y70k0n`; the same DLC and fixture dataset `d7m83gjy2` were reused here without recompilation. [Inference submission](inference_submission.json) job `jp1nj7j2g` and [profile submission](profile_submission.json) job `jgdd383eg` both requested `--compute_unit npu` for exact device **SA8775P ADP**. The completed [inference](inference_report.json) and [profile](profile_report.json) reports pin service build, Android version and chipset.
+
+The [three outputs](device_output.npz) were finite FP32 hidden/new-key/new-value tensors at a 1,024-token cache shape. They were bitwise identical to the earlier S24/S25 NPU outputs on the same fixture. The [audit](audit_report.json) measured hidden/key/value relative L2 of **1.351%/0.377%/0.650%** against original FP32 ONNX CPU and **0.555%/0.0065%/0.0126%** against QDQ ONNX CPU. No next-token or task-quality tolerance has been established.
+
+All 73 execution-detail rows in the profile were attributed to NPU. The 100 component samples had nearest-rank p50/p95 **1.837/2.079 ms** (minimum 1.749 ms); the service's `estimated_inference_time` was that minimum, not the median. Reported component inference peak memory was 90,877,952 bytes. These hosted component measurements exclude prefill, 27 other decoder layers, sampling, handoff, loading and a resident KV/ring/controller. They are not paired speed comparisons with other devices or OS builds.
+
+The [audit script](../../../../experiments/audit_spark_qualcomm_attention.py) reproduces the provenance and numerical checks using pinned source artifact paths in the S24 record. The [collector](../../../../experiments/collect_spark_aihub_component.py) retrieves these existing jobs via the local client configuration; neither script contains credentials.
+
+**Matrix status: component NPU inference/profiled; NOT E2E.** Full qualification needs a complete device-resident Spark artifact and Omni backend with prefill/decode/KV-ring/sampling, state, token quality, cancellation and memory admission, followed by sustained power and thermal measurement.

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_sa8775_full_attention/audit_report.json

@@ -0,0 +1,142 @@
+{
+  "scope": "one real Spark-X2.5 full-attention decoder layer at one static 1024-token cache shape; no generation E2E",
+  "status": "component_executed_on_npu_quality_unqualified",
+  "device": {
+    "name": "SA8775P ADP",
+    "os": "14",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-android",
+      "format:auto",
+      "chipset:qualcomm-sa8775p",
+      "hexagon:v73",
+      "soc-model:52",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "compile_job_id": "jpe7z8l15",
+  "inference_job_id": "jp1nj7j2g",
+  "profile_job_id": "jgdd383eg",
+  "target_model_id": "mq9y70k0n",
+  "input_dataset_id": "d7m83gjy2",
+  "s25_prior_inference_job_id": "jp0mvkj6g",
+  "files": {
+    "source_onnx": {
+      "bytes": 205618089,
+      "sha256": "4e2b1c8d791b1ee0e9b75f3322a20a743c0bf8c40f9b98edd963f2944d6300bb"
+    },
+    "quantized_onnx_zip": {
+      "bytes": 186772390,
+      "sha256": "703821ca8ace5672228919ea2add4901def4887748c381b12ffc8b259665c553"
+    },
+    "fixture": {
+      "bytes": 3901238,
+      "sha256": "5a439e6c6d148f1720938770daa110039a39edde014f1c1592c823c7d7dfc8ed"
+    },
+    "source_cpu_output": {
+      "bytes": 12103,
+      "sha256": "6bbfb8e6ccb09475391ce1dbd2e855205720f4367c038a0d5be35e7dfeb2f768"
+    },
+    "quantized_cpu_output": {
+      "bytes": 11574,
+      "sha256": "061654773166d102a11228c4a2e4cd06de6c2482c7e4e7860315b7b9a95fa8a1"
+    },
+    "s25_output": {
+      "bytes": 13072,
+      "sha256": "2dea9f4b478cb02b9be74a8d3ea4ad42e562e4de72413967d22706a1b6c6b9c0"
+    },
+    "device_output": {
+      "bytes": 11576,
+      "sha256": "9381fccee14b303edc5b432edf92422f2948e111ea6848df22cba3cd289a7073"
+    }
+  },
+  "source_vs_quantized_cpu": {
+    "hidden": {
+      "relative_l2": 0.012364819167814612,
+      "max_abs": 0.0028282999992370605,
+      "cosine": 0.9999236499353131
+    },
+    "k_new_0": {
+      "relative_l2": 0.0037627070087598766,
+      "max_abs": 0.03641480207443237,
+      "cosine": 0.9999929226489408
+    },
+    "v_new_0": {
+      "relative_l2": 0.006507432119585315,
+      "max_abs": 0.011739224195480347,
+      "cosine": 0.9999789079393114
+    }
+  },
+  "quantized_cpu_vs_device_npu": {
+    "hidden": {
+      "relative_l2": 0.005550082388164841,
+      "max_abs": 0.002016603946685791,
+      "cosine": 0.9999847274053486
+    },
+    "k_new_0": {
+      "relative_l2": 6.533750243101402e-05,
+      "max_abs": 0.00033855438232421875,
+      "cosine": 0.9999999978819449
+    },
+    "v_new_0": {
+      "relative_l2": 0.0001257028618212242,
+      "max_abs": 0.0002911090850830078,
+      "cosine": 0.9999999921460195
+    }
+  },
+  "source_cpu_vs_device_npu": {
+    "hidden": {
+      "relative_l2": 0.01350919207884596,
+      "max_abs": 0.0048449039459228516,
+      "cosine": 0.9999091967737982
+    },
+    "k_new_0": {
+      "relative_l2": 0.003766040678719007,
+      "max_abs": 0.03641480207443237,
+      "cosine": 0.9999929097855774
+    },
+    "v_new_0": {
+      "relative_l2": 0.006504874746412689,
+      "max_abs": 0.011739224195480347,
+      "cosine": 0.9999789285134126
+    }
+  },
+  "s25_npu_vs_device_npu": {
+    "hidden": {
+      "relative_l2": 0.0,
+      "max_abs": 0.0,
+      "cosine": 0.9999999999999998
+    },
+    "k_new_0": {
+      "relative_l2": 0.0,
+      "max_abs": 0.0,
+      "cosine": 0.9999999999999998
+    },
+    "v_new_0": {
+      "relative_l2": 0.0,
+      "max_abs": 0.0,
+      "cosine": 1.0
+    }
+  },
+  "profile": {
+    "sample_count": 100,
+    "sample_min_us": 1749,
+    "nearest_rank_p50_us": 1837,
+    "nearest_rank_p95_us": 2079,
+    "reported_estimated_inference_time_us": 1749,
+    "reported_peak_memory_bytes": 90877952,
+    "compute_unit_row_counts": {
+      "NPU": 73
+    }
+  },
+  "limits": [
+    "No task-level numerical tolerance or next-token quality gate is established.",
+    "Hub component jobs do not prove device-local prefill, KV/ring update, sampling, continuous generation or Omni deployment.",
+    "Profile sample latency is one component, not complete-request or sustained thermal performance."
+  ]
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_sa8775_full_attention/inference_report.json

@@ -0,0 +1,32 @@
+{
+  "timestamp_utc": "2026-09-23T21:38:12.406098+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jp1nj7j2g",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jp1nj7j2g/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "SA8775P ADP",
+    "os": "14",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-android",
+      "format:auto",
+      "chipset:qualcomm-sa8775p",
+      "hexagon:v73",
+      "soc-model:52",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mq9y70k0n",
+  "input_dataset_id": "d7m83gjy2",
+  "outputs": "device_output.npz"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_sa8775_full_attention/inference_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T21:26:58.289867+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jpe7z8l15",
+  "target_model_id": "mq9y70k0n",
+  "input_dataset_id": "d7m83gjy2",
+  "device": "SA8775P ADP",
+  "options": "--compute_unit npu",
+  "kind": "inference",
+  "status": "SUBMITTED",
+  "job_id": "jp1nj7j2g",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jp1nj7j2g/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_sa8775_full_attention/profile_report.json

@@ -0,0 +1,751 @@
+{
+  "timestamp_utc": "2026-09-23T21:38:17.144484+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jgdd383eg",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jgdd383eg/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "SA8775P ADP",
+    "os": "14",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-android",
+      "format:auto",
+      "chipset:qualcomm-sa8775p",
+      "hexagon:v73",
+      "soc-model:52",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mq9y70k0n",
+  "shapes": {
+    "x": [
+      [
+        1,
+        1,
+        2048
+      ],
+      "uint16"
+    ],
+    "v_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint16"
+    ],
+    "cos_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint16"
+    ],
+    "sin_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint16"
+    ],
+    "k_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint16"
+    ],
+    "mask_full": [
+      [
+        1,
+        1,
+        1,
+        1025
+      ],
+      "uint16"
+    ]
+  },
+  "profile": {
+    "execution_summary": {
+      "estimated_inference_time": 1749,
+      "estimated_inference_peak_memory": 90877952,
+      "first_load_time": 3647570,
+      "first_load_peak_memory": 259239936,
+      "warm_load_time": 382261,
+      "warm_load_peak_memory": 97042432,
+      "compile_time": 0,
+      "compile_peak_memory": 0,
+      "compile_memory_increase_range": [
+        0,
+        0
+      ],
+      "compile_memory_peak_range": [
+        0,
+        0
+      ],
+      "first_load_memory_increase_range": [
+        33965328,
+        40029184
+      ],
+      "first_load_memory_peak_range": [
+        161763328,
+        167827184
+      ],
+      "warm_load_memory_increase_range": [
+        0,
+        1755728
+      ],
+      "warm_load_memory_peak_range": [
+        720896,
+        87597344
+      ],
+      "inference_memory_increase_range": [
+        0,
+        2117632
+      ],
+      "inference_memory_peak_range": [
+        765952,
+        86432400
+      ],
+      "all_compile_times": [],
+      "all_first_load_times": [
+        3647570
+      ],
+      "all_warm_load_times": [
+        382261
+      ],
+      "all_inference_times": [
+        3082,
+        1834,
+        1909,
+        1877,
+        1852,
+        1835,
+        1795,
+        1784,
+        1809,
+        1816,
+        1785,
+        2063,
+        1781,
+        1805,
+        1800,
+        1833,
+        1828,
+        1808,
+        1819,
+        1836,
+        1903,
+        1841,
+        1806,
+        1913,
+        1878,
+        1873,
+        1875,
+        1908,
+        1836,
+        1911,
+        1875,
+        1859,
+        1893,
+        1812,
+        1888,
+        1825,
+        2079,
+        1934,
+        1920,
+        2208,
+        2393,
+        1855,
+        1821,
+        1882,
+        1823,
+        1827,
+        1829,
+        1929,
+        1886,
+        1865,
+        1939,
+        1855,
+        1828,
+        1778,
+        1775,
+        1792,
+        1811,
+        1779,
+        1770,
+        1810,
+        1829,
+        1842,
+        2060,
+        2285,
+        1803,
+        1801,
+        1810,
+        1790,
+        1793,
+        1780,
+        1777,
+        2510,
+        2044,
+        1871,
+        1837,
+        1897,
+        1866,
+        1845,
+        1854,
+        1809,
+        1834,
+        1776,
+        1855,
+        1882,
+        1904,
+        1847,
+        1822,
+        1822,
+        1816,
+        1965,
+        1849,
+        1785,
+        1749,
+        1848,
+        1849,
+        1871,
+        1837,
+        1840,
+        1884,
+        1798
+      ]
+    },
+    "execution_detail": [
+      {
+        "name": "Input",
+        "type": "Input",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 11247
+      },
+      {
+        "name": "/layers.0/Pow",
+        "type": "/layers.0/Pow",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 33404
+      },
+      {
+        "name": "/layers.0/ReduceMean",
+        "type": "/layers.0/ReduceMean",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 9730
+      },
+      {
+        "name": "/layers.0/Add",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 8895
+      },
+      {
+        "name": "/layers.0/Sqrt",
+        "type": "/layers.0/Sqrt",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Div",
+        "type": "/layers.0/Div",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 89939
+      },
+      {
+        "name": "/layers.0/Mul",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 65476
+      },
+      {
+        "name": "/layers.0/Mul_1",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 9537
+      },
+      {
+        "name": "/layers.0/q_k_v_proj/MatMul",
+        "type": "/layers.0/q",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/q_k_v_proj/MatMul_post_reshape",
+        "type": "/layers.0/q",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/g_proj/MatMul",
+        "type": "/layers.0/g",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 7303
+      },
+      {
+        "name": "/layers.0/g_proj/MatMul_post_reshape",
+        "type": "/layers.0/g",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "node_Split_1_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 58156
+      },
+      {
+        "name": "node_Split_1_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 21341
+      },
+      {
+        "name": "node_Split_1_to_slice_2",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 10062
+      },
+      {
+        "name": "/layers.0/Sigmoid",
+        "type": "/layers.0/Sigmoid",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_1",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_2",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 8808
+      },
+      {
+        "name": "/layers.0/Reshape_3",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 16083
+      },
+      {
+        "name": "node_Split_3_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 454
+      },
+      {
+        "name": "node_Split_3_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1329
+      },
+      {
+        "name": "node_Split_5_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 285
+      },
+      {
+        "name": "node_Split_5_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 6562
+      },
+      {
+        "name": "/layers.0/Concat_5",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 33224
+      },
+      {
+        "name": "node_Split_7_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 765
+      },
+      {
+        "name": "node_Split_7_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 520
+      },
+      {
+        "name": "/layers.0/Mul_2",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 4049
+      },
+      {
+        "name": "node_Split_9_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 5377
+      },
+      {
+        "name": "node_Split_9_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 905
+      },
+      {
+        "name": "/layers.0/Mul_4",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 5969
+      },
+      {
+        "name": "/layers.0/Neg",
+        "type": "/layers.0/Neg",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 744
+      },
+      {
+        "name": "/layers.0/Neg_1",
+        "type": "/layers.0/Neg",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 5334
+      },
+      {
+        "name": "/layers.0/Concat",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 697
+      },
+      {
+        "name": "/layers.0/Concat_2",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1658
+      },
+      {
+        "name": "/layers.0/Mul_3",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3289
+      },
+      {
+        "name": "/layers.0/Mul_5",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2271
+      },
+      {
+        "name": "/layers.0/Add_1",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 948
+      },
+      {
+        "name": "/layers.0/Add_2",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2997
+      },
+      {
+        "name": "/layers.0/Concat_1",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 803
+      },
+      {
+        "name": "/layers.0/Concat_3",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1305
+      },
+      {
+        "name": "/layers.0/Reshape_4",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 325
+      },
+      {
+        "name": "/layers.0/Concat_4",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Transpose",
+        "type": "/layers.0/Transpose",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/MatMul",
+        "type": "/layers.0/MatMul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 745909
+      },
+      {
+        "name": "/layers.0/Mul_6",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2748
+      },
+      {
+        "name": "/layers.0/Add_3",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 25552
+      },
+      {
+        "name": "/layers.0/Softmax",
+        "type": "/layers.0/Softmax",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 10411
+      },
+      {
+        "name": "/layers.0/MatMul_1",
+        "type": "/layers.0/MatMul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 651695
+      },
+      {
+        "name": "/layers.0/Reshape_5",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_7",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 14796
+      },
+      {
+        "name": "/layers.0/out_proj/MatMul",
+        "type": "/layers.0/out",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 32941
+      },
+      {
+        "name": "/layers.0/out_proj/MatMul_post_reshape",
+        "type": "/layers.0/out",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_4",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 8452
+      },
+      {
+        "name": "/layers.0/Pow_1",
+        "type": "/layers.0/Pow",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 5886
+      },
+      {
+        "name": "/layers.0/ReduceMean_1",
+        "type": "/layers.0/ReduceMean",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2501
+      },
+      {
+        "name": "/layers.0/Add_5",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 197
+      },
+      {
+        "name": "/layers.0/Sqrt_1",
+        "type": "/layers.0/Sqrt",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Div_1",
+        "type": "/layers.0/Div",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 58344
+      },
+      {
+        "name": "/layers.0/Mul_8",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 38272
+      },
+      {
+        "name": "/layers.0/Mul_9",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 7562
+      },
+      {
+        "name": "/layers.0/gate_proj/MatMul",
+        "type": "/layers.0/gate",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/gate_proj/MatMul_post_reshape",
+        "type": "/layers.0/gate",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/up_proj/MatMul",
+        "type": "/layers.0/up",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 414626
+      },
+      {
+        "name": "/layers.0/up_proj/MatMul_post_reshape",
+        "type": "/layers.0/up",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "_elementwiseneuron_0",
+        "type": "elementwiseneuron",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 396528
+      },
+      {
+        "name": "/layers.0/Mul_12",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 21386
+      },
+      {
+        "name": "/layers.0/down_proj/MatMul",
+        "type": "/layers.0/down",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 369308
+      },
+      {
+        "name": "/layers.0/down_proj/MatMul_post_reshape",
+        "type": "/layers.0/down",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_7",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 5980
+      },
+      {
+        "name": "/layers.0/Concat_3_identity",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_2_identity",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "Output",
+        "type": "Output",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 7279
+      }
+    ]
+  }
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_sa8775_full_attention/profile_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T21:26:59.698671+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jpe7z8l15",
+  "target_model_id": "mq9y70k0n",
+  "input_dataset_id": null,
+  "device": "SA8775P ADP",
+  "options": "--compute_unit npu",
+  "kind": "profile",
+  "status": "SUBMITTED",
+  "job_id": "jgdd383eg",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jgdd383eg/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_xelite_full_attention/README.md

@@ -0,0 +1,13 @@
+# Spark-X2.5 attention layer on Snapdragon X Elite CRD
+
+**One fixed-shape decoder attention layer from the real Spark-X2.5-1.7B checkpoint executed on the hosted Snapdragon X Elite CRD NPU (Windows 11 ARM64). This is component evidence, not complete Spark generation or a local Omni deployment.**
+
+The [Galaxy S24 provenance record](../spark_s24_full_attention/README.md) pins the FP32 source ONNX, W8A16 QDQ ONNX, calibration and exact input fixture. Its QNN DLC compile job `jpe7z8l15` produced target model `mq9y70k0n`; the same DLC and fixture dataset `d7m83gjy2` were reused here without recompilation. [Inference submission](inference_submission.json) job `j5wl7876p` and [profile submission](profile_submission.json) job `jg9zmkmlp` both requested `--compute_unit npu` for the exact device. The [raw inference job](inference_report.json) and [raw profile job](profile_report.json) completed. Their service build, OS, chipset and target IDs are preserved in those reports.
+
+The [three outputs](device_output.npz) were finite FP32 hidden/new-key/new-value tensors at a 1,024-token cache shape. They were bitwise identical to the earlier Galaxy S25 and S24 NPU outputs on the same fixture. The [audit](audit_report.json) measured hidden/key/value relative L2 of **1.351%/0.377%/0.650%** against original FP32 ONNX CPU, and **0.555%/0.0065%/0.0126%** against QDQ ONNX CPU. No next-token or task-level quality tolerance has been established.
+
+All 73 execution-detail rows in the device profile were attributed to NPU. The 100 component inference-time samples had nearest-rank p50/p95 **1.433/1.606 ms** (minimum 1.346 ms). The service's `estimated_inference_time` was 1.346 ms, matching the minimum rather than the median. The reported component inference peak memory was 18,423,808 bytes. These are hosted component-job measurements; they exclude prefill, 27 other decoder layers, sampling, transfer, loading and a resident KV/ring/controller. Hardware, OS and hosted power state differ from the mobile runs, so these are not paired speed comparisons.
+
+Re-audit with `benchmarks/edge_harness/experiments/audit_spark_qualcomm_attention.py` and the [pinned source artifact paths](../spark_s24_full_attention/README.md). `benchmarks/edge_harness/experiments/collect_spark_aihub_component.py` retrieves the existing jobs using the local AI Hub client configuration. Neither script contains credentials.
+
+**Matrix status: component NPU inference/profiled; NOT E2E.** Full qualification still needs an X Elite machine running the complete 28-layer Spark model through a resident Omni backend with prefill/decode/KV-ring/sampling, state and cancellation, token-quality checks, memory admission and sustained power/thermal measurements. Workbench is a test facility, not that deployment runtime.

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_xelite_full_attention/audit_report.json

@@ -0,0 +1,142 @@
+{
+  "scope": "one real Spark-X2.5 full-attention decoder layer at one static 1024-token cache shape; no generation E2E",
+  "status": "component_executed_on_npu_quality_unqualified",
+  "device": {
+    "name": "Snapdragon X Elite CRD",
+    "os": "11",
+    "attributes": [
+      "os:windows",
+      "format:compute",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-windows",
+      "chipset:qualcomm-snapdragon-x-elite",
+      "chipset:sc8380xp",
+      "hexagon:v73",
+      "soc-model:60",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "compile_job_id": "jpe7z8l15",
+  "inference_job_id": "j5wl7876p",
+  "profile_job_id": "jg9zmkmlp",
+  "target_model_id": "mq9y70k0n",
+  "input_dataset_id": "d7m83gjy2",
+  "s25_prior_inference_job_id": "jp0mvkj6g",
+  "files": {
+    "source_onnx": {
+      "bytes": 205618089,
+      "sha256": "4e2b1c8d791b1ee0e9b75f3322a20a743c0bf8c40f9b98edd963f2944d6300bb"
+    },
+    "quantized_onnx_zip": {
+      "bytes": 186772390,
+      "sha256": "703821ca8ace5672228919ea2add4901def4887748c381b12ffc8b259665c553"
+    },
+    "fixture": {
+      "bytes": 3901238,
+      "sha256": "5a439e6c6d148f1720938770daa110039a39edde014f1c1592c823c7d7dfc8ed"
+    },
+    "source_cpu_output": {
+      "bytes": 12103,
+      "sha256": "6bbfb8e6ccb09475391ce1dbd2e855205720f4367c038a0d5be35e7dfeb2f768"
+    },
+    "quantized_cpu_output": {
+      "bytes": 11574,
+      "sha256": "061654773166d102a11228c4a2e4cd06de6c2482c7e4e7860315b7b9a95fa8a1"
+    },
+    "s25_output": {
+      "bytes": 13072,
+      "sha256": "2dea9f4b478cb02b9be74a8d3ea4ad42e562e4de72413967d22706a1b6c6b9c0"
+    },
+    "device_output": {
+      "bytes": 11576,
+      "sha256": "9381fccee14b303edc5b432edf92422f2948e111ea6848df22cba3cd289a7073"
+    }
+  },
+  "source_vs_quantized_cpu": {
+    "hidden": {
+      "relative_l2": 0.012364819167814612,
+      "max_abs": 0.0028282999992370605,
+      "cosine": 0.9999236499353131
+    },
+    "k_new_0": {
+      "relative_l2": 0.0037627070087598766,
+      "max_abs": 0.03641480207443237,
+      "cosine": 0.9999929226489408
+    },
+    "v_new_0": {
+      "relative_l2": 0.006507432119585315,
+      "max_abs": 0.011739224195480347,
+      "cosine": 0.9999789079393114
+    }
+  },
+  "quantized_cpu_vs_device_npu": {
+    "hidden": {
+      "relative_l2": 0.005550082388164841,
+      "max_abs": 0.002016603946685791,
+      "cosine": 0.9999847274053486
+    },
+    "k_new_0": {
+      "relative_l2": 6.533750243101402e-05,
+      "max_abs": 0.00033855438232421875,
+      "cosine": 0.9999999978819449
+    },
+    "v_new_0": {
+      "relative_l2": 0.0001257028618212242,
+      "max_abs": 0.0002911090850830078,
+      "cosine": 0.9999999921460195
+    }
+  },
+  "source_cpu_vs_device_npu": {
+    "hidden": {
+      "relative_l2": 0.01350919207884596,
+      "max_abs": 0.0048449039459228516,
+      "cosine": 0.9999091967737982
+    },
+    "k_new_0": {
+      "relative_l2": 0.003766040678719007,
+      "max_abs": 0.03641480207443237,
+      "cosine": 0.9999929097855774
+    },
+    "v_new_0": {
+      "relative_l2": 0.006504874746412689,
+      "max_abs": 0.011739224195480347,
+      "cosine": 0.9999789285134126
+    }
+  },
+  "s25_npu_vs_device_npu": {
+    "hidden": {
+      "relative_l2": 0.0,
+      "max_abs": 0.0,
+      "cosine": 0.9999999999999998
+    },
+    "k_new_0": {
+      "relative_l2": 0.0,
+      "max_abs": 0.0,
+      "cosine": 0.9999999999999998
+    },
+    "v_new_0": {
+      "relative_l2": 0.0,
+      "max_abs": 0.0,
+      "cosine": 1.0
+    }
+  },
+  "profile": {
+    "sample_count": 100,
+    "sample_min_us": 1346,
+    "nearest_rank_p50_us": 1433,
+    "nearest_rank_p95_us": 1606,
+    "reported_estimated_inference_time_us": 1346,
+    "reported_peak_memory_bytes": 18423808,
+    "compute_unit_row_counts": {
+      "NPU": 73
+    }
+  },
+  "limits": [
+    "No task-level numerical tolerance or next-token quality gate is established.",
+    "Hub component jobs do not prove device-local prefill, KV/ring update, sampling, continuous generation or Omni deployment.",
+    "Profile sample latency is one component, not complete-request or sustained thermal performance."
+  ]
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_xelite_full_attention/inference_report.json

@@ -0,0 +1,32 @@
+{
+  "timestamp_utc": "2026-09-23T21:33:23.832588+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "j5wl7876p",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j5wl7876p/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Snapdragon X Elite CRD",
+    "os": "11",
+    "attributes": [
+      "os:windows",
+      "format:compute",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-windows",
+      "chipset:qualcomm-snapdragon-x-elite",
+      "chipset:sc8380xp",
+      "hexagon:v73",
+      "soc-model:60",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mq9y70k0n",
+  "input_dataset_id": "d7m83gjy2",
+  "outputs": "device_output.npz"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_xelite_full_attention/inference_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T21:26:55.958764+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jpe7z8l15",
+  "target_model_id": "mq9y70k0n",
+  "input_dataset_id": "d7m83gjy2",
+  "device": "Snapdragon X Elite CRD",
+  "options": "--compute_unit npu",
+  "kind": "inference",
+  "status": "SUBMITTED",
+  "job_id": "j5wl7876p",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j5wl7876p/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_xelite_full_attention/profile_report.json

@@ -0,0 +1,751 @@
+{
+  "timestamp_utc": "2026-09-23T21:33:27.915381+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jg9zmkmlp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jg9zmkmlp/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit npu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Snapdragon X Elite CRD",
+    "os": "11",
+    "attributes": [
+      "os:windows",
+      "format:compute",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-windows",
+      "chipset:qualcomm-snapdragon-x-elite",
+      "chipset:sc8380xp",
+      "hexagon:v73",
+      "soc-model:60",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mq9y70k0n",
+  "shapes": {
+    "x": [
+      [
+        1,
+        1,
+        2048
+      ],
+      "uint16"
+    ],
+    "v_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint16"
+    ],
+    "cos_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint16"
+    ],
+    "sin_full": [
+      [
+        1,
+        1,
+        1,
+        64
+      ],
+      "uint16"
+    ],
+    "k_cache_0": [
+      [
+        1,
+        2,
+        1024,
+        256
+      ],
+      "uint16"
+    ],
+    "mask_full": [
+      [
+        1,
+        1,
+        1,
+        1025
+      ],
+      "uint16"
+    ]
+  },
+  "profile": {
+    "execution_summary": {
+      "estimated_inference_time": 1346,
+      "estimated_inference_peak_memory": 18423808,
+      "first_load_time": 2429828,
+      "first_load_peak_memory": 318615552,
+      "warm_load_time": 242574,
+      "warm_load_peak_memory": 14589952,
+      "compile_time": 0,
+      "compile_peak_memory": 0,
+      "compile_memory_increase_range": [
+        0,
+        0
+      ],
+      "compile_memory_peak_range": [
+        0,
+        0
+      ],
+      "first_load_memory_increase_range": [
+        207298560,
+        207298560
+      ],
+      "first_load_memory_peak_range": [
+        313659392,
+        313659392
+      ],
+      "warm_load_memory_increase_range": [
+        9183232,
+        9183232
+      ],
+      "warm_load_memory_peak_range": [
+        9629696,
+        9629696
+      ],
+      "inference_memory_increase_range": [
+        20480,
+        20480
+      ],
+      "inference_memory_peak_range": [
+        2109440,
+        2109440
+      ],
+      "all_compile_times": [],
+      "all_first_load_times": [
+        2429828
+      ],
+      "all_warm_load_times": [
+        242574
+      ],
+      "all_inference_times": [
+        2684,
+        1456,
+        1479,
+        1446,
+        1378,
+        1375,
+        1481,
+        1412,
+        1430,
+        1431,
+        1540,
+        1399,
+        1398,
+        1526,
+        1439,
+        1455,
+        1599,
+        1416,
+        1352,
+        1433,
+        1440,
+        1507,
+        1472,
+        1685,
+        1546,
+        1512,
+        1531,
+        1468,
+        1353,
+        1367,
+        1454,
+        1434,
+        1489,
+        1514,
+        1421,
+        1364,
+        1387,
+        1414,
+        1390,
+        1412,
+        1398,
+        1413,
+        1429,
+        1549,
+        1467,
+        1425,
+        1525,
+        1420,
+        1523,
+        1606,
+        1504,
+        1490,
+        1494,
+        1458,
+        1521,
+        1410,
+        1733,
+        1548,
+        1398,
+        1494,
+        1426,
+        1449,
+        1389,
+        1352,
+        1415,
+        1501,
+        1564,
+        1407,
+        1548,
+        1461,
+        1531,
+        1505,
+        1514,
+        1690,
+        1451,
+        1488,
+        3416,
+        1418,
+        1375,
+        1411,
+        1416,
+        1545,
+        1425,
+        1409,
+        1406,
+        1419,
+        1405,
+        1383,
+        1429,
+        1381,
+        1467,
+        1396,
+        1418,
+        1361,
+        1346,
+        1365,
+        1452,
+        1408,
+        1390,
+        1409
+      ]
+    },
+    "execution_detail": [
+      {
+        "name": "Input",
+        "type": "Input",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 15300
+      },
+      {
+        "name": "/layers.0/Pow",
+        "type": "/layers.0/Pow",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 41331
+      },
+      {
+        "name": "/layers.0/ReduceMean",
+        "type": "/layers.0/ReduceMean",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 6753
+      },
+      {
+        "name": "/layers.0/Add",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 7094
+      },
+      {
+        "name": "/layers.0/Sqrt",
+        "type": "/layers.0/Sqrt",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Div",
+        "type": "/layers.0/Div",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 66952
+      },
+      {
+        "name": "/layers.0/Mul",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 59984
+      },
+      {
+        "name": "/layers.0/Mul_1",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 12187
+      },
+      {
+        "name": "/layers.0/q_k_v_proj/MatMul",
+        "type": "/layers.0/q",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/q_k_v_proj/MatMul_post_reshape",
+        "type": "/layers.0/q",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/g_proj/MatMul",
+        "type": "/layers.0/g",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 5297
+      },
+      {
+        "name": "/layers.0/g_proj/MatMul_post_reshape",
+        "type": "/layers.0/g",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "node_Split_1_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 34704
+      },
+      {
+        "name": "node_Split_1_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 17816
+      },
+      {
+        "name": "node_Split_1_to_slice_2",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 8512
+      },
+      {
+        "name": "/layers.0/Sigmoid",
+        "type": "/layers.0/Sigmoid",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_1",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_2",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 8560
+      },
+      {
+        "name": "/layers.0/Reshape_3",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 7452
+      },
+      {
+        "name": "node_Split_3_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 319
+      },
+      {
+        "name": "node_Split_3_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1721
+      },
+      {
+        "name": "node_Split_5_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 278
+      },
+      {
+        "name": "node_Split_5_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 6044
+      },
+      {
+        "name": "/layers.0/Concat_5",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 29616
+      },
+      {
+        "name": "node_Split_7_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 989
+      },
+      {
+        "name": "node_Split_7_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 525
+      },
+      {
+        "name": "/layers.0/Mul_2",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3468
+      },
+      {
+        "name": "node_Split_9_to_slice_0",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1018
+      },
+      {
+        "name": "node_Split_9_to_slice_1",
+        "type": "node",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1371
+      },
+      {
+        "name": "/layers.0/Mul_4",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 9373
+      },
+      {
+        "name": "/layers.0/Neg",
+        "type": "/layers.0/Neg",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 266
+      },
+      {
+        "name": "/layers.0/Neg_1",
+        "type": "/layers.0/Neg",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3136
+      },
+      {
+        "name": "/layers.0/Concat",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 817
+      },
+      {
+        "name": "/layers.0/Concat_2",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2424
+      },
+      {
+        "name": "/layers.0/Mul_3",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3893
+      },
+      {
+        "name": "/layers.0/Mul_5",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2650
+      },
+      {
+        "name": "/layers.0/Add_1",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 361
+      },
+      {
+        "name": "/layers.0/Add_2",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 2670
+      },
+      {
+        "name": "/layers.0/Concat_1",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1245
+      },
+      {
+        "name": "/layers.0/Concat_3",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1124
+      },
+      {
+        "name": "/layers.0/Reshape_4",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 362
+      },
+      {
+        "name": "/layers.0/Concat_4",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Transpose",
+        "type": "/layers.0/Transpose",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/MatMul",
+        "type": "/layers.0/MatMul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 746665
+      },
+      {
+        "name": "/layers.0/Mul_6",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 3374
+      },
+      {
+        "name": "/layers.0/Add_3",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 17560
+      },
+      {
+        "name": "/layers.0/Softmax",
+        "type": "/layers.0/Softmax",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 14098
+      },
+      {
+        "name": "/layers.0/MatMul_1",
+        "type": "/layers.0/MatMul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 658338
+      },
+      {
+        "name": "/layers.0/Reshape_5",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Mul_7",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 32722
+      },
+      {
+        "name": "/layers.0/out_proj/MatMul",
+        "type": "/layers.0/out",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 35771
+      },
+      {
+        "name": "/layers.0/out_proj/MatMul_post_reshape",
+        "type": "/layers.0/out",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_4",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 9773
+      },
+      {
+        "name": "/layers.0/Pow_1",
+        "type": "/layers.0/Pow",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 4443
+      },
+      {
+        "name": "/layers.0/ReduceMean_1",
+        "type": "/layers.0/ReduceMean",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 1724
+      },
+      {
+        "name": "/layers.0/Add_5",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 184
+      },
+      {
+        "name": "/layers.0/Sqrt_1",
+        "type": "/layers.0/Sqrt",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Div_1",
+        "type": "/layers.0/Div",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 50440
+      },
+      {
+        "name": "/layers.0/Mul_8",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 36313
+      },
+      {
+        "name": "/layers.0/Mul_9",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 8408
+      },
+      {
+        "name": "/layers.0/gate_proj/MatMul",
+        "type": "/layers.0/gate",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/gate_proj/MatMul_post_reshape",
+        "type": "/layers.0/gate",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/up_proj/MatMul",
+        "type": "/layers.0/up",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 301316
+      },
+      {
+        "name": "/layers.0/up_proj/MatMul_post_reshape",
+        "type": "/layers.0/up",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "_elementwiseneuron_0",
+        "type": "elementwiseneuron",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 319204
+      },
+      {
+        "name": "/layers.0/Mul_12",
+        "type": "/layers.0/Mul",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 20213
+      },
+      {
+        "name": "/layers.0/down_proj/MatMul",
+        "type": "/layers.0/down",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 312506
+      },
+      {
+        "name": "/layers.0/down_proj/MatMul_post_reshape",
+        "type": "/layers.0/down",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Add_7",
+        "type": "/layers.0/Add",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 5976
+      },
+      {
+        "name": "/layers.0/Concat_3_identity",
+        "type": "/layers.0/Concat",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "/layers.0/Reshape_2_identity",
+        "type": "/layers.0/Reshape",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 0
+      },
+      {
+        "name": "Output",
+        "type": "Output",
+        "compute_unit": "NPU",
+        "execution_time": 0,
+        "execution_cycles": 15482
+      }
+    ]
+  }
+}

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_xelite_full_attention/profile_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T21:26:57.092534+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jpe7z8l15",
+  "target_model_id": "mq9y70k0n",
+  "input_dataset_id": null,
+  "device": "Snapdragon X Elite CRD",
+  "options": "--compute_unit npu",
+  "kind": "profile",
+  "status": "SUBMITTED",
+  "job_id": "jg9zmkmlp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jg9zmkmlp/"
+}

benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md

@@ -7,6 +7,10 @@ The WSL CPU and CPU+Radeon InternVLA cells also have [bounded Omni complete-poli
 A [follow-up MiniCPM-o storage experiment](../../../e2e_expansion_20260923/evidence/minicpmo_omni_native_weights_io/README.md) copied the same ten GGUF weights to native NTFS. One warmup plus 20 complete requests passed on Windows CPU (p50/p95 23.62/23.88 s) and Radeon+CPU (16.35/17.20 s), with no prefetch warnings. Separate one-request UNC controls were much slower; the original UNC 20-request Radeon profile remains valid for that path. Run order and OS cache were not controlled, so no causal or cross-placement speedup claim follows. A [real InternVLA Cosmos encoder NPU artifact attempt](../../../e2e_expansion_20260923/evidence/internvla_cosmos_amd_npu/README.md) passed fixed-shape CPU export parity but its tested A16W8 candidates failed the CPU numerical gate or VitisAI compilation; the AMD NPU cell remains NOT E2E.
 
 The [WSL CPU Spark GGUF Omni profile](../../../e2e_expansion_20260923/evidence/spark_omni_wsl_cpu/README.md) adds a separate current-Omni complete-request path to the earlier vLLM INT8 CPU result. It passed two named prompts, 20/20 serial measured inventory requests at p50/p95 7.82/7.97 s, a public text request, context refusal and cancellation cleanup. The CPU-only llama.cpp build required explicit CPU-buffer placement evidence because it omits the Windows offload-count log line. This does not establish token streaming or broad quality.
+
+On a hosted [Galaxy S24 Spark attention-layer experiment](../../../e2e_expansion_20260923/evidence/spark_s24_full_attention/README.md), a real one-layer W8A16 QDQ artifact executed on NPU and matched prior S25 component outputs exactly on one dataset. The 100-sample layer p50/p95 was 1.109/1.205 ms; the service's 1.098 ms estimated-time field was the sample minimum. FP32-source versus S24 output hidden-state relative L2 was 1.351% without a downstream token-quality gate. This adds component C evidence only; device-local generation remains absent.
+
+The same W8A16 layer also ran on hosted [Snapdragon X Elite CRD](../../../e2e_expansion_20260923/evidence/spark_xelite_full_attention/README.md) and [SA8775P ADP](../../../e2e_expansion_20260923/evidence/spark_sa8775_full_attention/README.md) NPUs, with pinned-fixture outputs identical to S24/S25; the 100-sample component p50/p95 was 1.433/1.606 ms and 1.837/2.079 ms, respectively. On [RB3 Gen 2 / QCS6490](../../../e2e_expansion_20260923/evidence/spark_rb3_full_attention/README.md), that W8A16 graph failed QNN DLC loading even after target-specific recompilation. An FP32 ONNX CPU layer then passed with source-relative L2 below 9e-7, while a calibrated W8A8 NPU layer executed but had 20.9% hidden-state relative L2 and remains numerically unqualified. These hosted component outcomes do not establish full generation or comparable device speed.
 
 ## Configuration × model
 
@@ -27,11 +31,11 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i
 | PC HX370 + Radeon 890M / Windows worker | scoped E2E: Omni Spark text | scoped E2E: Omni Radeon talker/codec + CPU predictor | NOT E2E | scoped E2E: Omni Radeon+CPU image+audio → text+speech (native-storage 20-run profile) | synthetic policy pass: Omni CPU+Radeon Cosmos (WSL/native Windows) |
 | 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 |
+| PC Snapdragon X Elite CRD / AI Hub | NOT E2E (one Spark attention layer on NPU) | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
 | Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | NOT E2E | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
-| Mobile Galaxy S24 / Snapdragon 8 Gen 3 | NOT E2E | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
-| Embedded SA8775P ADP / AI Hub | NOT E2E | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
-| Embedded RB3 Gen 2 / QCS6490 / AI Hub | NOT E2E | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
+| Mobile Galaxy S24 / Snapdragon 8 Gen 3 | NOT E2E (one Spark attention layer on NPU) | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
+| Embedded SA8775P ADP / AI Hub | NOT E2E (one Spark attention layer on NPU) | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
+| Embedded RB3 Gen 2 / QCS6490 / AI Hub | NOT E2E (Spark CPU layer passed; W8A8 NPU quality open) | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
 
 ## Per-cell reasons and next steps
 
@@ -72,7 +76,7 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i
 | PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No qualifying same-model multi-accelerator deployment/artifact plan or overlap/recovery evidence. Individual routes are not a joint pipeline. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | MiniCPM-o 4.5 | NOT_E2E_PROFILEABLE | No qualifying same-model multi-accelerator deployment/artifact plan or overlap/recovery evidence. Individual routes are not a joint pipeline. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | InternVLA-A1 | NOT_E2E_PROFILEABLE | No qualifying same-model multi-accelerator deployment/artifact plan or overlap/recovery evidence. Individual routes are not a joint pipeline. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
-| PC Snapdragon X Elite CRD / AI Hub | Spark-X2.5 | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
+| PC Snapdragon X Elite CRD / AI Hub | Spark-X2.5 | COMPONENT_NPU_INFERENCE_PROFILED; NOT_E2E_PROFILEABLE | A real Spark-X2.5-1.7B one-layer W8A16 QDQ attention artifact, compiled as QNN DLC, ran on the exact hosted Snapdragon X Elite CRD (Windows 11 ARM64). Inference returned finite hidden/new-K/new-V tensors bitwise identical to the prior S24/S25 NPU outputs on the same fixture; the profile attributed all 73 execution-detail rows to NPU. Across 100 component samples, nearest-rank p50/p95 was 1.433/1.606 ms; the 1.346 ms estimated-time field is the minimum. Hidden/K/V relative L2 versus original FP32 ONNX CPU was 1.351%/0.377%/0.650%, without a next-token quality tolerance. There is no complete 28-layer prefill/decode/KV-ring/sampling loop, resident Omni controller or complete-request timing. | Qualify a complete X Elite resident artifact/controller with state, token quality, memory admission/cancel and sustained power; [X Elite NPU raw evidence](../../../e2e_expansion_20260923/evidence/spark_xelite_full_attention/README.md). |
 | PC Snapdragon X Elite CRD / AI Hub | Qwen3-TTS 0.6B CustomVoice | NOT_E2E_PROFILEABLE | Historical QNN talker profile re-fetched successfully; no numerical/full-pipeline gate. Historical TFLite vocoder target was refused on this Windows ARM device. | Validate target-specific numerical quality, complete local execution, memory and sustained behavior. |
 | PC Snapdragon X Elite CRD / AI Hub | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | PC Snapdragon X Elite CRD / AI Hub | MiniCPM-o 4.5 | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
@@ -82,17 +86,17 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i
 | Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | MiniCPM-o 4.5 | NOT_E2E_PROFILEABLE | Fresh FP16 speech-head inference passed; all 41 outputs exactly match the same artifact/dataset reference. Historical FP16 numerical parity passed; tested W8A16 top-1 parity failed. Thinker/encoders/vocoder are not qualified together. | Supply a qualified complete mobile checkpoint/export and validate the full multimodal/stateful pipeline. |
 | Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | InternVLA-A1 | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
-| Mobile Galaxy S24 / Snapdragon 8 Gen 3 | Spark-X2.5 | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
+| Mobile Galaxy S24 / Snapdragon 8 Gen 3 | Spark-X2.5 | COMPONENT_NPU_INFERENCE_PROFILED; NOT_E2E_PROFILEABLE | The real 1.7B checkpoint's fixed-shape full-attention decoder **one-layer** W8A16 QDQ artifact compiled as QNN DLC and ran on an exact hosted Galaxy S24 (Android 14). A same-dataset inference returned finite hidden/new-K/new-V outputs identical to the prior S25 NPU outputs; the profile attributed all 73 execution-detail rows to NPU. Across 100 component samples, nearest-rank p50/p95 was 1.109/1.205 ms; the 1.098 ms `estimated_inference_time` field is the minimum, not a median. Relative L2 versus original FP32 ONNX CPU was 1.351%/0.377%/0.650% for hidden/K/V; no token-quality tolerance is known. This is one cache bucket, not 28-layer prefill/decode, KV-ring, sampling, a resident local controller or complete-request timing. ADB has no attached device. | Qualify a complete device-resident artifact/controller with state and token-quality gates, 128-token generation across 512/1024 context transitions, admission/cancel and sustained memory/power; [S24 NPU raw evidence](../../../e2e_expansion_20260923/evidence/spark_s24_full_attention/README.md). |
 | Mobile Galaxy S24 / Snapdragon 8 Gen 3 | Qwen3-TTS 0.6B CustomVoice | NOT_E2E_PROFILEABLE | Historical FP16 decoder parity has SNR 39.13 dB. Tested calibrated INT8 decoder parity failed with negative SNR. Neither result is a full local TTS stream. | Validate the complete local pipeline and a quality-passing artifact set. |
 | Mobile Galaxy S24 / Snapdragon 8 Gen 3 | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | Mobile Galaxy S24 / Snapdragon 8 Gen 3 | MiniCPM-o 4.5 | NOT_E2E_PROFILEABLE | Historical FP16 speech-head job re-fetched successfully; SNR 49.22 dB/top-1 match in the recorded parity test. W8A16 top-1 failed. Not complete MiniCPM-o. | Qualify thinker/encoders/vocoder and stateful local integration. |
 | Mobile Galaxy S24 / Snapdragon 8 Gen 3 | InternVLA-A1 | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
-| Embedded SA8775P ADP / AI Hub | Spark-X2.5 | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
+| Embedded SA8775P ADP / AI Hub | Spark-X2.5 | COMPONENT_NPU_INFERENCE_PROFILED; NOT_E2E_PROFILEABLE | A real Spark-X2.5-1.7B one-layer W8A16 QDQ attention artifact, compiled as QNN DLC, ran on the exact hosted SA8775P ADP (Android 14). Inference returned finite hidden/new-K/new-V tensors bitwise identical to the prior S24/S25 NPU outputs on the same fixture; the profile attributed all 73 execution-detail rows to NPU. Across 100 component samples, nearest-rank p50/p95 was 1.837/2.079 ms; the 1.749 ms estimated-time field is the minimum. Hidden/K/V relative L2 versus original FP32 ONNX CPU was 1.351%/0.377%/0.650%, without a next-token quality tolerance. There is no complete 28-layer prefill/decode/KV-ring/sampling loop, resident Omni controller or complete-request timing. | Qualify a complete embedded resident artifact/controller with state, token quality, memory admission/cancel and sustained power; [SA8775P NPU raw evidence](../../../e2e_expansion_20260923/evidence/spark_sa8775_full_attention/README.md). |
 | Embedded SA8775P ADP / AI Hub | Qwen3-TTS 0.6B CustomVoice | NOT_E2E_PROFILEABLE | Historical NPU talker and GPU vocoder profiles re-fetched successfully; no complete pipeline, parity or co-residency qualification. | Validate target-specific numerical quality, complete local execution, memory and sustained behavior. |
 | Embedded SA8775P ADP / AI Hub | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | Embedded SA8775P ADP / AI Hub | MiniCPM-o 4.5 | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | Embedded SA8775P ADP / AI Hub | InternVLA-A1 | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
-| Embedded RB3 Gen 2 / QCS6490 / AI Hub | Spark-X2.5 | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
+| Embedded RB3 Gen 2 / QCS6490 / AI Hub | Spark-X2.5 | COMPONENT_CPU_PARITY_PASS; COMPONENT_NPU_NUMERIC_UNQUALIFIED; NOT_E2E_PROFILEABLE | On the exact hosted Dragonwing RB3 Gen 2 Vision Kit (QCS6490, QC Linux 1.6), the reused S24 W8A16 QNN DLC and a separately RB3-compiled W8A16 DLC each failed inference/profile at QnnModel_composeGraphsFromDlc: MODEL_GRAPH_ERROR. An FP32 ONNX one-layer CPU route passed with all 56 profile rows on CPU, source-relative hidden/K/V L2 of 8.73e-7/2.98e-7/3.37e-7 and 100-sample component p50/p95 19.304/21.113 ms. A calibrated W8A8 QNN DLC ran with all 61 rows on NPU and p50/p95 5.610/5.761 ms, but hidden/K/V relative L2 was 20.914%/1.702%/4.413% versus FP32 CPU, without a next-token quality gate. These are different precisions, not a qualified speedup; no complete 28-layer prefill/decode/KV-ring/sampling, local Omni controller or complete-request timing exists. | Improve representative W8A8 calibration or find another compatible quality-passing artifact; validate token parity, then full device-resident generation, admission/cancel and sustained power; [RB3 raw evidence](../../../e2e_expansion_20260923/evidence/spark_rb3_full_attention/README.md). |
 | Embedded RB3 Gen 2 / QCS6490 / AI Hub | Qwen3-TTS 0.6B CustomVoice | NOT_E2E_PROFILEABLE | Historical GPU vocoder profile re-fetched successfully. The attempted NPU talker build was rejected for floating-point inputs. No full model pipeline. | Validate target-specific numerical quality, complete local execution, memory and sustained behavior. |
 | Embedded RB3 Gen 2 / QCS6490 / AI Hub | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | Embedded RB3 Gen 2 / QCS6490 / AI Hub | MiniCPM-o 4.5 | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |

docs/design/omni_edge_e2e_roadmap.md

@@ -13,6 +13,8 @@ three-stage text-to-speech functional pass after correcting request embeddings,
 and an InternVLA real-weight synthetic action forward. Model-quality and device gates
 remain open; the 2026-09-22 audit remains the historical baseline.
 
+The [Spark hosted Qualcomm component expansion](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/README.md) advances M1 only to one real fixed-shape decoder attention layer. The same W8A16 QNN DLC ran on S24, Snapdragon X Elite CRD and SA8775P ADP NPUs with identical outputs on one pinned fixture; S25 had prior separate component evidence. On [RB3 Gen 2](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/README.md), the tested W8A16 graph failed QNN loading even after an exact-device compile. FP32 ONNX ran on CPU with near-source parity, while calibrated W8A8 ran on NPU but had 20.9% hidden-state relative L2 and remains numerically unqualified. These are component C results, not full M1: device-local 28-layer prefill/continuous decode, KV/ring, sampling, token quality, admission/cancellation, complete-request timing and sustained power/thermal gates remain open. AI Hub is the test facility, not a deployment dependency.
+
 The [WSL CPU expansion](../../benchmarks/edge_harness/results/e2e_expansion_20260923/README.md)
 adds constrained three-stage MiniCPM-o text-to-speech requests and a CPU-only
 InternVLA synthetic policy forward. A subsequent 20-request serial MiniCPM-o