From 6d0b758dbef8e17d8330d9c4ffb50476835a6735 Mon Sep 17 00:00:00 2001 From: tzhouam Date: Thu, 24 Sep 2026 15:21:03 +0800 Subject: [PATCH] Verify Spark Windows recovery after server-started cancellation --- .../audit_spark_windows_restart.py | 114 +++++++ .../experiments/probe_omni_llamacpp_spark.py | 76 ++++- .../evidence/spark_omni_llamacpp/README.md | 10 +- .../results/e2e_expansion_20260924/README.md | 8 +- .../spark_omni_windows_restart/README.md | 42 +++ .../audit_report.json | 63 ++++ .../cpu_report.json | 247 +++++++++++++++ .../spark_omni_windows_restart/cpu_server.log | 270 ++++++++++++++++ .../cpu_server_restart.log | 208 ++++++++++++ .../cpu_v3_report.json | 156 +++++++++ .../cpu_v3_server.log | 295 ++++++++++++++++++ .../cpu_v4_report.json | 250 +++++++++++++++ .../cpu_v4_server.log | 289 +++++++++++++++++ .../cpu_v4_server_restart.log | 207 ++++++++++++ .../spark_omni_windows_restart/hardware.json | 10 + .../radeon_report.json | 250 +++++++++++++++ .../radeon_server.log | 211 +++++++++++++ .../radeon_v2_report.json | 251 +++++++++++++++ .../radeon_v2_server.log | 272 ++++++++++++++++ .../radeon_v2_server_restart.log | 211 +++++++++++++ .../radeon_v3_report.json | 253 +++++++++++++++ .../radeon_v3_server.log | 293 +++++++++++++++++ .../radeon_v3_server_restart.log | 211 +++++++++++++ .../radeon_v4_report.json | 254 +++++++++++++++ .../radeon_v4_server.log | 289 +++++++++++++++++ .../radeon_v4_server_restart.log | 210 +++++++++++++ .../evidence/summary/README.md | 8 +- docs/design/omni_edge_e2e_roadmap.md | 2 +- 28 files changed, 4950 insertions(+), 10 deletions(-) create mode 100644 benchmarks/edge_harness/experiments/audit_spark_windows_restart.py create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/README.md create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/audit_report.json create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_report.json create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_server.log create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_server_restart.log create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v3_report.json create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v3_server.log create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v4_report.json create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v4_server.log create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v4_server_restart.log create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/hardware.json create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_report.json create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_server.log create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v2_report.json create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v2_server.log create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v2_server_restart.log create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v3_report.json create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v3_server.log create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v3_server_restart.log create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v4_report.json create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v4_server.log create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v4_server_restart.log diff --git a/benchmarks/edge_harness/experiments/audit_spark_windows_restart.py b/benchmarks/edge_harness/experiments/audit_spark_windows_restart.py new file mode 100644 index 000000000..d4caaf48e --- /dev/null +++ b/benchmarks/edge_harness/experiments/audit_spark_windows_restart.py @@ -0,0 +1,114 @@ +#!/usr/bin/env python3 +"""Audit native-Windows Spark cancellation/restart reports and server logs.""" + +from __future__ import annotations + +import argparse +import hashlib +import json +import re +from pathlib import Path + + +MODEL_SHA256 = "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4" +SERVER_SHA256 = "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52" +TASK_START = re.compile(r"slot launch_slot_: id\s+\d+ \| task \d+ \| processing task") + + +def sha256(path: Path) -> str: + with path.open("rb") as stream: + return hashlib.file_digest(stream, "sha256").hexdigest() + + +def clean_ledger(value: dict) -> bool: + return (value["reserved"] == {"host_ram": 0} + and value["owners"] == [] and value["quarantined"] == []) + + +def audit(root: Path, route: str) -> dict: + report_path = root / f"{route}_v4_report.json" + first_log = root / f"{route}_v4_server.log" + second_log = root / f"{route}_v4_server_restart.log" + run = json.loads(report_path.read_text(encoding="utf-8")) + before = run["execution_plan"] + after = run["restart_check"]["fresh_execution_plan"] + abort = run["abort_check"] + restart = run["restart_check"] + if run["status"] != "passed" or run["device"] != ("cpu" if route == "cpu" else "Vulkan1"): + raise ValueError(f"{route}: runtime outcome or device differs") + if any(plan["model_sha256"] != MODEL_SHA256 or plan["server_sha256"] != SERVER_SHA256 + for plan in (before, after)): + raise ValueError(f"{route}: model or server binary changed between sessions") + expected_offload = ["0", "29"] if route == "cpu" else ["29", "29"] + expected_buffers = [] if route == "cpu" else ["Vulkan1"] + if any(plan["offloaded_layers"] != expected_offload + or plan["model_buffer_devices"] != expected_buffers for plan in (before, after)): + raise ValueError(f"{route}: execution placement differs from the intended hardware") + if route == "cpu" and any(plan["all_model_buffers"] != ["CPU_Mapped", "CPU_REPACK"] + for plan in (before, after)): + raise ValueError("CPU route had a non-CPU model buffer") + if route == "radeon" and any(plan["expected_device_name"] != "AMD Radeon(TM) 890M Graphics" + for plan in (before, after)): + raise ValueError("Radeon route did not pin the 890M by name") + if ([row["answer"] for row in run["checks"]] != ["Paris", "1511"] + or not all(row["correct"] for row in run["checks"] + run["measured"]) + or restart["fresh_request"]["answer"] != "Paris" + or not restart["fresh_request"]["correct"]): + raise ValueError(f"{route}: named complete-request checks failed") + if (abort["abort_case"] != "inventory_120" + or abort["server_tasks_after_abort_request_start"] + <= abort["server_tasks_before_abort_request"] + or abort["stale_output"] or not abort["worker_exited"] + or restart["late_output"]): + raise ValueError(f"{route}: cancellation was not verified after server task start") + if not all(clean_ledger(value) for value in ( + abort["ledger_after_abort"], restart["first_runtime_ledger_after_shutdown"], + run["ledger_after_shutdown"], + )): + raise ValueError(f"{route}: reservation or quarantine remained") + if (before["worker_generation"] == after["worker_generation"] + or restart["fresh_request"]["stage_event"]["worker_generation"] + != after["worker_generation"]): + raise ValueError(f"{route}: resumed request did not use a fresh worker") + first_starts = len(TASK_START.findall(first_log.read_text(encoding="utf-8", errors="replace"))) + second_starts = len(TASK_START.findall(second_log.read_text(encoding="utf-8", errors="replace"))) + if first_starts < abort["server_tasks_after_abort_request_start"] or second_starts < 1: + raise ValueError(f"{route}: server logs do not contain the recorded tasks") + return { + "status": "passed", + "device": run["device"], + "old_generation": before["worker_generation"], + "new_generation": after["worker_generation"], + "first_server_task_starts": first_starts, + "second_server_task_starts": second_starts, + "started_before_cancel": True, + "fresh_answer": restart["fresh_request"]["answer"], + "ledger_cleared": True, + "files": {p.name: {"bytes": p.stat().st_size, "sha256": sha256(p)} + for p in (report_path, first_log, second_log)}, + } + + +def main() -> None: + parser = argparse.ArgumentParser(description=__doc__) + parser.add_argument("--evidence-dir", required=True, type=Path) + args = parser.parse_args() + result = { + "scope": "one in-flight cancel and fresh-stage restart per native Windows Spark placement", + "checkpoint": "Spark-X2.5-1.7B Q4_K_M", + "model_sha256": MODEL_SHA256, + "llama_server_sha256": SERVER_SHA256, + "routes": {route: audit(args.evidence_dir, route) for route in ("cpu", "radeon")}, + "limits": [ + "A fresh StageRuntime is created after cancellation; no same-session state migration is claimed.", + "The backend returns complete requests rather than incremental token events.", + "This is one restart per placement, not concurrent or sustained recovery qualification.", + ], + } + path = args.evidence_dir / "audit_report.json" + path.write_text(json.dumps(result, indent=2) + "\n", encoding="utf-8") + print(json.dumps({route: row["status"] for route, row in result["routes"].items()}, indent=2)) + + +if __name__ == "__main__": + main() diff --git a/benchmarks/edge_harness/experiments/probe_omni_llamacpp_spark.py b/benchmarks/edge_harness/experiments/probe_omni_llamacpp_spark.py index 506a86ed9..145ed366b 100644 --- a/benchmarks/edge_harness/experiments/probe_omni_llamacpp_spark.py +++ b/benchmarks/edge_harness/experiments/probe_omni_llamacpp_spark.py @@ -11,6 +11,7 @@ import argparse import asyncio import json import math +import re import time from pathlib import Path from types import SimpleNamespace @@ -38,9 +39,13 @@ async def main() -> None: parser.add_argument("--reserve-gib", type=int, default=4) parser.add_argument("--abort-check", action="store_true") parser.add_argument("--context-check", action="store_true") + parser.add_argument("--restart-check", action="store_true", + help="After in-flight cancellation, start a fresh stage and verify a request") args = parser.parse_args() if args.warmups < 0 or args.repeats <= 0 or args.capacity_gib < args.reserve_gib or args.reserve_gib < 4: parser.error("invalid serial profile or explicit memory budget") + if args.restart_check and not args.abort_check: + parser.error("--restart-check requires --abort-check") from vllm_omni.config.stage_config import ( DeployConfig, @@ -158,22 +163,89 @@ async def main() -> None: raise RuntimeError("llama.cpp accepted a prompt beyond the declared context") report["context_check"]["recovery"] = await request_one(*cases[0]) if args.abort_check: + def started_server_tasks() -> int: + log = args.server_log.read_text(encoding="utf-8", errors="replace") + return len(re.findall( + r"slot launch_slot_: id\s+\d+ \| task \d+ \| processing task", + log, + )) + + tasks_before_abort = started_server_tasks() request_id = f"abort-{time.monotonic_ns()}" await pool.submit_initial( request_id, state, - {"text": "Count from one upward and continue until the token limit.", "max_tokens": 96}, + {"text": inventory, "max_tokens": 96}, ) - await asyncio.sleep(0.01) + deadline = time.monotonic() + 10 + while started_server_tasks() <= tasks_before_abort: + if time.monotonic() >= deadline: + raise TimeoutError("abort request did not reach the owned llama.cpp server") + await asyncio.sleep(0.02) + premature = pool.poll_graph_output(0) + if premature is not None: + premature.release_stage_buffers() + raise RuntimeError("abort request completed before in-flight cancellation") await pool.abort_requests([request_id]) await asyncio.sleep(0.1) stale = pool.poll_graph_output(0) report["abort_check"] = { + "abort_case": "inventory_120", + "server_tasks_before_abort_request": tasks_before_abort, + "server_tasks_after_abort_request_start": started_server_tasks(), "stale_output": stale is not None, "worker_exited": pool.stage_client._proc.poll() is not None, "ledger_after_abort": runtime.resource_ledger.snapshot(), } + if stale is not None: + stale.release_stage_buffers() if stale is not None or not report["abort_check"]["worker_exited"]: raise RuntimeError("llama.cpp cancel did not drain worker and stale output") + if args.restart_check: + old_generation = report["execution_plan"]["worker_generation"] + runtime.shutdown() + released = runtime.resource_ledger.snapshot() + if released["reserved"]["host_ram"] or released["quarantined"]: + raise RuntimeError("cancelled stage retained a host-RAM reservation") + restart_log = args.server_log.with_name( + args.server_log.stem + "_restart" + args.server_log.suffix + ) + restart_deploy = DeployConfig( + async_chunk=False, + stages=[StageDeployConfig( + stage_id=0, + backend={**backend, "log_file": str(restart_log)}, + resource_budget=deploy.stages[0].resource_budget, + )], + ) + restart_configs = [ + stage.to_omegaconf() + for stage in merge_pipeline_deploy(pipeline, restart_deploy) + ] + runtime = StageRuntime( + restart_configs, "local-spark-gguf-restart", "", + stage_init_timeout=120, async_chunk=False, + ) + restarted_at = time.perf_counter() + runtime.initialize() + fresh_startup_s = time.perf_counter() - restarted_at + pool = runtime.stage_pools[0] + state = SimpleNamespace(sampling_params_list=[None]) + new_plan = pool.stage_client.execution_plan + if new_plan["worker_generation"] == old_generation: + raise RuntimeError("restart reused the cancelled worker generation") + restarted = await request_one(*cases[0]) + late = pool.poll_graph_output(0) + report["restart_check"] = { + "first_runtime_ledger_after_shutdown": released, + "fresh_startup_s": fresh_startup_s, + "fresh_server_log": str(restart_log), + "fresh_execution_plan": new_plan, + "fresh_request": restarted, + "late_output": late is not None, + } + if late is not None: + late.release_stage_buffers() + raise RuntimeError("restart emitted an extra output") report["status"] = "passed" except BaseException as exc: report["status"] = "failed" diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md index 8356d8721..7f6924355 100644 --- a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md +++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md @@ -1,9 +1,15 @@ # Spark GGUF through Omni on native Windows CPU and Radeon 890M -This is a **scoped complete-request text path** for one pinned Spark-X2.5-1.7B Q4_K_M GGUF. The new `external.llamacpp.text.v1` client runs inside Omni `StageRuntime`/`StagePool`; a separate public `AsyncOmni.generate` request also passed with a real Spark HF config directory and the registered `spark2_5_gguf_text` pipeline. llama.cpp owns weights, KV and internal sampling. Omni owns the process lifetime, one-slot admission, declared 4 GiB host-RAM reservation, request identity, terminal text event, acknowledgement and cancellation cleanup. It uses the existing complete-request graph control path, not an ONNX graph. Incremental token streaming, restart after cancelling an active request, model-wide quality and thermal qualification are open. +This is a **scoped complete-request text path** for one pinned Spark-X2.5-1.7B Q4_K_M GGUF. The new `external.llamacpp.text.v1` client runs inside Omni `StageRuntime`/`StagePool`; a separate public `AsyncOmni.generate` request also passed with a real Spark HF config directory and the registered `spark2_5_gguf_text` pipeline. llama.cpp owns weights, KV and internal sampling. Omni owns the process lifetime, one-slot admission, declared 4 GiB host-RAM reservation, request identity, terminal text event, acknowledgement and cancellation cleanup. It uses the existing complete-request graph control path, not an ONNX graph. At the time of this profile, incremental token streaming, restart after cancelling an active request, model-wide quality and thermal qualification were open; the later fresh-stage restart result is linked below. The pinned GGUF is [XHToken/Spark-X2.5-1.7B-GGUF](https://huggingface.co/XHToken/Spark-X2.5-1.7B-GGUF/tree/1f7fa33b1245c14730da39e125714ad3a327901b) revision `1f7fa33b1245c14730da39e125714ad3a327901b`, file `Spark-X2.5-1.7B-Q4_K_M.gguf` (1,107,457,856 bytes; SHA-256 `902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4`). The executable is llama.cpp [b11124](https://github.com/ggml-org/llama.cpp/releases/tag/b11124), build commit `1a679828f`, `llama-server.exe` SHA-256 `9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52`. The artifact remains in local model storage. The model and executable hashes are checked before launch. The source/checkpoint and earlier standalone tests are in the [prior record](../spark_radeon890m_llamacpp/README.md). +A later [2026-09-24 recovery experiment](../../../e2e_expansion_20260924/evidence/spark_omni_windows_restart/README.md) +confirmed that the server had started the cancellation workload on both CPU +and Radeon 890M, then a fresh Omni stage using the same artifact answered +correctly with a new worker generation and no stale output or reservation. +The original profiles below remain separate measurements. + Host: Windows 11 build 26200, Ryzen AI 9 HX 370, Radeon 890M driver `32.0.22018.6001`, Performance power scheme. An RTX 5090 Laptop GPU was installed but was not selected. The native Python environment was vLLM 0.29.0 with the current Omni source. The server used one slot, 4,096 context tokens, FP16 KV, temperature zero, reasoning off, `--fit off`, `--cache-ram 0`, and per-request `cache_prompt=false`. The backend tokenizes each prompt before submission and refuses a prompt plus maximum generation bound that exceeds the declared context. It refuses a partial generation `finish_reason`. The 4 GiB admission includes the 1.11 GB model plus declared 2 GiB KV/workspace/transfer/headroom; this is an explicit bound, **not** a measured loading or shared GPU-memory peak. | Complete 120-item inventory request | Native CPU | Radeon 890M Vulkan1 | @@ -21,4 +27,4 @@ The Radeon run rejected an over-context prompt before generation, then answered Reproduce with the [stage profile script](../../../../experiments/probe_omni_llamacpp_spark.py) or [public entrypoint script](../../../../experiments/probe_omni_llamacpp_entrypoint.py) from the native Windows environment. Set `PYTHONPATH` to this checkout, `PYTHONUTF8=1`, and `VLLM_ENABLE_V1_MULTIPROCESSING=0`. Supply the pinned `--model-file`, `--server-bin`, their SHA-256 arguments, `--device cpu` or `--device Vulkan1 --expected-device-name 'AMD Radeon(TM) 890M Graphics'`, and output paths. Use `--warmups 1 --repeats 20 --context-check --abort-check` for the Radeon stage profile; the CPU profile used the same warmup/repeats and abort check. The public probe additionally accepts `--model-config-dir` for the real Spark HF config. Raw driver logs retain the exact environment setup and failures. -The acceptance boundary remains two named text inputs, one repeated serial profile, a single local worker and one short public API request. There is no broad text-quality suite, concurrency, load peak, device-memory attribution, restart-after-cancel, token streaming, or sustained power/thermal measurement. The next experiment should establish those before declaring this CPU or Radeon path generally qualified; the other 58 device × model cells are unchanged by this backend. +The acceptance boundary of these original runs remains two named text inputs, one repeated serial profile, a single local worker and one short public API request. The later linked experiment covers one fresh-stage restart after confirmed server-start cancellation per device. Broad text quality, concurrency, load peak, device-memory attribution, token streaming and sustained power/thermal behavior remain unverified; the other device × model cells are unchanged by this backend. diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/README.md b/benchmarks/edge_harness/results/e2e_expansion_20260924/README.md index e17fc8cec..2c622a123 100644 --- a/benchmarks/edge_harness/results/e2e_expansion_20260924/README.md +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/README.md @@ -21,7 +21,13 @@ complete-stream behavior remain open. An additional [Spark BF16 restart probe](evidence/spark_bf16_wsl_cpu/restart_branch_report.json) cancelled after eight token events, rejected the retired session handle and verified that a fresh session produced the same 128 greedy token IDs with no -late events. For S25, the historical [25-frame TFLite GPU-requested vocoder](evidence/qwen_tts_vocoder_qualcomm/s25/full_tflite_gpu_profile/README.md) +late events. A [native Windows Spark GGUF restart probe](evidence/spark_omni_windows_restart/README.md) +confirmed server task start before cancellation on both HX370 CPU and Radeon +890M Vulkan1, then loaded the same artifact in a fresh Omni stage, returned +Paris and cleared the ledger without stale output. This is fresh-stage +recovery, not same-session token continuation. + +For S25, the historical [25-frame TFLite GPU-requested vocoder](evidence/qwen_tts_vocoder_qualcomm/s25/full_tflite_gpu_profile/README.md) has a newly measured 100-sample p50/p95 **0.839/1.157 s** for 2 s of audio, with mixed CPU/GPU placement and component median RTF **0.420**. Its 4.24% same-fixture waveform relative L2, unknown exact checkpoint revision and diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/README.md b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/README.md new file mode 100644 index 000000000..0bf302df6 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/README.md @@ -0,0 +1,42 @@ +# Spark native Windows restart after in-flight cancellation + +The pinned Spark-X2.5-1.7B Q4_K_M GGUF and llama.cpp b11124 binary from the +[native Windows Omni text profile](../../../e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md) +were reused without changing model or precision. The model SHA-256 is +`902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4`; +the server SHA-256 is +`9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52`. +The host was Windows 11 build 26200 on Ryzen AI 9 HX 370, with Radeon 890M +driver 32.0.22018.6001 and the Performance power scheme. Native Python used +vLLM 0.29.0 and this Omni checkout. The RTX 5090 Laptop GPU was present but +not selected in the Radeon route. The current OS, CPU, GPU driver and power +scheme are retained in the [Windows hardware record](hardware.json). + +The [probe](../../../../experiments/probe_omni_llamacpp_spark.py) ran one +complete `Paris` request, one 120-item inventory request (`1511`), and one +serial inventory measurement before submitting a second inventory request +for cancellation. It waited until the owned llama.cpp server log recorded a +fourth task start, checked no terminal output had arrived, then cancelled. +The first StageRuntime released its 4 GiB host-RAM reservation and terminated +the worker. A fresh StageRuntime loaded the same hashed artifact and returned +`Paris` with a different worker generation, no late output and no remaining +reservation or quarantine. The [auditor](../../../../experiments/audit_spark_windows_restart.py) +binds the [CPU](cpu_v4_report.json) and [Radeon](radeon_v4_report.json) reports +to their separate [CPU first](cpu_v4_server.log), +[CPU restart](cpu_v4_server_restart.log), +[Radeon first](radeon_v4_server.log) and +[Radeon restart](radeon_v4_server_restart.log) server logs. Both routes +[passed](audit_report.json); CPU offloaded 0/29 layers with CPU model buffers, +while Radeon 890M Vulkan1 offloaded 29/29 layers before and after restart. +This is **one fresh-stage restart per placement**, not automatic same-session +state migration or incremental token streaming. + +The initial [CPU](cpu_report.json) and [Radeon](radeon_report.json) probes, +and the [second Radeon run](radeon_v2_report.json), cancelled after admission +but before the server had started the abort task; they were insufficient for +the in-flight claim. A [third CPU attempt](cpu_v3_report.json) waited for a +server task start but used a short counting prompt that finished before +cancellation, and failed explicitly. Those raw failures are retained. The +final inventory-triggered runs are separate from the earlier 20-request +latency profiles; they do not establish a new p50/p95, broader text quality, +concurrency, load peak, power or thermal behavior. diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/audit_report.json b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/audit_report.json new file mode 100644 index 000000000..427102ba8 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/audit_report.json @@ -0,0 +1,63 @@ +{ + "scope": "one in-flight cancel and fresh-stage restart per native Windows Spark placement", + "checkpoint": "Spark-X2.5-1.7B Q4_K_M", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "llama_server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "routes": { + "cpu": { + "status": "passed", + "device": "cpu", + "old_generation": "2ff1298272904e81afa406b3c4361654", + "new_generation": "464ec54f6eac420b86f5ee5bb27db2cf", + "first_server_task_starts": 4, + "second_server_task_starts": 1, + "started_before_cancel": true, + "fresh_answer": "Paris", + "ledger_cleared": true, + "files": { + "cpu_v4_report.json": { + "bytes": 6906, + "sha256": "3b47fdf72040d9c6d62d8dc46c1080d1332874c5904eaa83af6e5815b4681fa6" + }, + "cpu_v4_server.log": { + "bytes": 26687, + "sha256": "e03d4468206c171a3ecac2ce8af6910c76286b2207e3cf6b8ef4cf89cbf0be3d" + }, + "cpu_v4_server_restart.log": { + "bytes": 17235, + "sha256": "1d6661b9512ccb197c012003e64702e7dee0f8b47b310b9909fbecb7ce3f2341" + } + } + }, + "radeon": { + "status": "passed", + "device": "Vulkan1", + "old_generation": "e96bf7f656d24b39bcf5744e577cc073", + "new_generation": "171ce688813848b38ffa8f66ec236618", + "first_server_task_starts": 4, + "second_server_task_starts": 1, + "started_before_cancel": true, + "fresh_answer": "Paris", + "ledger_cleared": true, + "files": { + "radeon_v4_report.json": { + "bytes": 7024, + "sha256": "85232352c3a988ea89cf703485feb6c58a9e1b4124d2ab3f6efcd7d406f78fee" + }, + "radeon_v4_server.log": { + "bytes": 26466, + "sha256": "a14f34b001aa62a1860a3519866a7dc127d35b6cae6a18059781e52ebe0d2918" + }, + "radeon_v4_server_restart.log": { + "bytes": 17502, + "sha256": "7bc88261a85c320dea2c0bedf84ad849c4f8bca9385ba99451d2ab0e3cec6b66" + } + } + } + }, + "limits": [ + "A fresh StageRuntime is created after cancellation; no same-session state migration is claimed.", + "The backend returns complete requests rather than incremental token events.", + "This is one restart per placement, not concurrent or sustained recovery qualification." + ] +} diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_report.json b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_report.json new file mode 100644 index 000000000..0b5337522 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_report.json @@ -0,0 +1,247 @@ +{ + "scope": "complete Spark GGUF text requests through Omni StageRuntime/StagePool and external llama.cpp backend", + "model_file": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "device": "cpu", + "memory_budget": { + "capacities": { + "host_ram": 12884901888 + }, + "demands": { + "host_ram": 4294967296 + } + }, + "warmup_count": 0, + "measured_count": 1, + "startup_s": 2.7457405999302864, + "execution_plan": { + "backend": "external.llamacpp.text.v1", + "stage_id": 0, + "worker_generation": "57cc5c72a7c5471aaf6a94224f16568a", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "requested_device": "cpu", + "expected_device_name": null, + "layer_assignments": 0, + "offloaded_layers": [ + "0", + "29" + ], + "model_buffer_devices": [], + "all_model_buffers": [ + "CPU_Mapped", + "CPU_REPACK" + ], + "cpu_only_layout": false, + "worker_pid": 15580, + "loaded_rss_bytes": 1970765824, + "reserved_bytes": { + "host_ram": 4294967296 + }, + "memory_pool": "host_ram", + "memory_overhead_bytes": 2147483648, + "context_tokens": 4096, + "max_new_tokens": 96, + "max_io_bytes": 1048576, + "request_capacity": 1, + "stateful_session": "llama.cpp-owned; reset for each complete request", + "evidence": "B" + }, + "checks": [ + { + "case": "short", + "answer": "Paris", + "expected": "Paris", + "correct": true, + "wall_s": 0.18201659992337227, + "stage_event": { + "request_id": "short-688448640000000", + "stage_id": 0, + "epoch": 1, + "seq": 1, + "kind": "text", + "worker_generation": "57cc5c72a7c5471aaf6a94224f16568a", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 0.17684849991928786, + "usage": { + "completion_tokens": 2, + "prompt_tokens": 28, + "total_tokens": 30, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + }, + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 7.185723599977791, + "stage_event": { + "request_id": "inventory_120-688448812000000", + "stage_id": 0, + "epoch": 2, + "seq": 1, + "kind": "text", + "worker_generation": "57cc5c72a7c5471aaf6a94224f16568a", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 7.166244800086133, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "warmups": [], + "measured": [ + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 7.143662300077267, + "stage_event": { + "request_id": "inventory_120-688456000000000", + "stage_id": 0, + "epoch": 3, + "seq": 1, + "kind": "text", + "worker_generation": "57cc5c72a7c5471aaf6a94224f16568a", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 7.125400600023568, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "nearest_rank_p50_wall_s": 7.143662300077267, + "nearest_rank_p95_wall_s": 7.143662300077267, + "abort_check": { + "stale_output": false, + "worker_exited": true, + "ledger_after_abort": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + } + }, + "restart_check": { + "first_runtime_ledger_after_shutdown": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + }, + "fresh_startup_s": 2.300211500027217, + "fresh_server_log": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\vllm-omni-edge\\benchmarks\\edge_harness\\results\\e2e_expansion_20260924\\evidence\\spark_omni_windows_restart\\cpu_server_restart.log", + "fresh_execution_plan": { + "backend": "external.llamacpp.text.v1", + "stage_id": 0, + "worker_generation": "97b21d33af3e41649e4949bb2918cd47", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "requested_device": "cpu", + "expected_device_name": null, + "layer_assignments": 0, + "offloaded_layers": [ + "0", + "29" + ], + "model_buffer_devices": [], + "all_model_buffers": [ + "CPU_Mapped", + "CPU_REPACK" + ], + "cpu_only_layout": false, + "worker_pid": 48280, + "loaded_rss_bytes": 1970434048, + "reserved_bytes": { + "host_ram": 4294967296 + }, + "memory_pool": "host_ram", + "memory_overhead_bytes": 2147483648, + "context_tokens": 4096, + "max_new_tokens": 96, + "max_io_bytes": 1048576, + "request_capacity": 1, + "stateful_session": "llama.cpp-owned; reset for each complete request", + "evidence": "B" + }, + "fresh_request": { + "case": "short", + "answer": "Paris", + "expected": "Paris", + "correct": true, + "wall_s": 0.17802889994345605, + "stage_event": { + "request_id": "short-688465703000000", + "stage_id": 0, + "epoch": 1, + "seq": 1, + "kind": "text", + "worker_generation": "97b21d33af3e41649e4949bb2918cd47", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 0.175863099982962, + "usage": { + "completion_tokens": 2, + "prompt_tokens": 28, + "total_tokens": 30, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + }, + "late_output": false + }, + "status": "passed", + "ledger_after_shutdown": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + } +} diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_server.log b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_server.log new file mode 100644 index 000000000..f23d0d006 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_server.log @@ -0,0 +1,270 @@ +0.00.000.940 I srv llama_server: initializing ... +0.00.217.371 I cmn common_param: common_params_print_info: build 11124 (1a679828f) with Clang 20.1.8 for Windows x86_64 +0.00.217.378 I cmn common_param: common_params_print_info: verbosity = 4 (adjust with the `-lv N` CLI arg) +0.00.217.380 I cmn common_param: device_info: +0.00.221.703 I cmn common_param: - Vulkan0 : NVIDIA GeForce RTX 5090 Laptop GPU (24137 MiB, 23369 MiB free) +0.00.225.749 I cmn common_param: - Vulkan1 : AMD Radeon(TM) 890M Graphics (48505 MiB, 46080 MiB free) +0.00.225.763 I cmn common_param: - CPU : AMD Ryzen AI 9 HX 370 w/ Radeon 890M (64633 MiB, 25780 MiB free) +0.00.225.851 I cmn common_param: system_info: n_threads = 12 (n_threads_batch = 12) / 24 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | +0.00.225.972 I srv init_listene: running without SSL +0.00.226.050 I srv init_listene: The UI is disabled +0.00.226.051 I srv init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable +0.00.226.057 I srv init: using 23 threads for HTTP server +0.00.226.143 W srv llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655) +0.00.232.206 I srv load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.232.207 I srv load_model: local path 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.256.388 I llama_model_loader: loaded meta data with 37 key-value pairs and 226 tensors from C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf (version GGUF V3 (latest)) +0.00.256.401 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. +0.00.256.407 I llama_model_loader: - kv 0: general.architecture str = spark2_5 +0.00.256.408 I llama_model_loader: - kv 1: general.type str = model +0.00.256.412 I llama_model_loader: - kv 2: general.sampling.top_k i32 = 0 +0.00.256.419 I llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000 +0.00.256.420 I llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000 +0.00.256.420 I llama_model_loader: - kv 5: general.name str = Hf_Format +0.00.256.421 I llama_model_loader: - kv 6: general.size_label str = 1.7B +0.00.256.424 I llama_model_loader: - kv 7: spark2_5.block_count u32 = 28 +0.00.256.424 I llama_model_loader: - kv 8: spark2_5.context_length u32 = 1048576 +0.00.256.425 I llama_model_loader: - kv 9: spark2_5.embedding_length u32 = 2048 +0.00.256.425 I llama_model_loader: - kv 10: spark2_5.feed_forward_length u32 = 6656 +0.00.256.426 I llama_model_loader: - kv 11: spark2_5.attention.head_count u32 = 8 +0.00.256.426 I llama_model_loader: - kv 12: spark2_5.attention.head_count_kv u32 = 2 +0.00.256.427 I llama_model_loader: - kv 13: spark2_5.rope.freq_base f32 = 5000000.000000 +0.00.256.428 I llama_model_loader: - kv 14: spark2_5.rope.freq_base_swa f32 = 10000.000000 +0.00.256.430 I llama_model_loader: - kv 15: spark2_5.attention.layer_norm_rms_epsilon f32 = 0.000001 +0.00.256.430 I llama_model_loader: - kv 16: spark2_5.attention.key_length u32 = 256 +0.00.256.431 I llama_model_loader: - kv 17: spark2_5.attention.value_length u32 = 256 +0.00.256.431 I llama_model_loader: - kv 18: spark2_5.vocab_size u32 = 131072 +0.00.256.432 I llama_model_loader: - kv 19: spark2_5.attention.sliding_window u32 = 512 +0.00.256.453 I llama_model_loader: - kv 20: spark2_5.attention.sliding_window_pattern arr[bool,28] = [true, true, true, false, true, true,... +0.00.256.453 I llama_model_loader: - kv 21: spark2_5.rope.dimension_count u32 = 64 +0.00.256.454 I llama_model_loader: - kv 22: spark2_5.rope.dimension_count_swa u32 = 256 +0.00.256.454 I llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 +0.00.256.454 I llama_model_loader: - kv 24: tokenizer.ggml.pre str = spark2_5 +0.00.273.720 I llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,131072] = ["<｜start▁of▁sentence｜>", "<... +0.00.278.575 I llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,131072] = [3, 3, 3, 4, 4, 3, 3, 3, 4, 4, 4, 3, ... +0.00.294.911 I llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,130716] = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "Ġ... +0.00.294.919 I llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 0 +0.00.294.921 I llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 1 +0.00.294.922 I llama_model_loader: - kv 30: tokenizer.ggml.unknown_token_id u32 = 5 +0.00.294.923 I llama_model_loader: - kv 31: tokenizer.ggml.padding_token_id u32 = 2 +0.00.294.924 I llama_model_loader: - kv 32: tokenizer.ggml.add_bos_token bool = false +0.00.294.924 I llama_model_loader: - kv 33: tokenizer.ggml.add_eos_token bool = false +0.00.294.930 I llama_model_loader: - kv 34: tokenizer.chat_template str = {%- if not messages %}\n {{- raise_... +0.00.294.930 I llama_model_loader: - kv 35: general.quantization_version u32 = 2 +0.00.294.931 I llama_model_loader: - kv 36: general.file_type u32 = 15 +0.00.294.940 I llama_model_loader: - type f32: 57 tensors +0.00.294.941 I llama_model_loader: - type q4_K: 140 tensors +0.00.294.941 I llama_model_loader: - type q6_K: 29 tensors +0.00.294.946 I print_info: file format = GGUF V3 (latest) +0.00.294.947 I print_info: file type = Q4_K - Medium +0.00.294.951 I print_info: file size = 1.03 GiB (5.16 BPW) +0.00.355.087 I load: 86 unused tokens +0.00.384.520 I load: printing all EOG tokens: +0.00.384.534 I load: - 1 ('<｜end▁of▁sentence｜>') +0.00.384.783 I load: special tokens cache size = 113 +0.00.401.597 I load: token to piece cache size = 0.7975 MB +0.00.401.627 I print_info: arch = spark2_5 +0.00.401.628 I print_info: vocab_only = 0 +0.00.401.629 I print_info: no_alloc = 0 +0.00.401.629 I print_info: n_ctx_train = 1048576 +0.00.401.630 I print_info: n_embd_inp = 2048 +0.00.401.630 I print_info: n_embd = 2048 +0.00.401.631 I print_info: n_embd_out = 2048 +0.00.401.631 I print_info: n_layer = 28 +0.00.401.632 I print_info: n_layer_all = 28 +0.00.401.649 I print_info: n_head = 8 +0.00.401.651 I print_info: n_head_kv = 2 +0.00.401.652 I print_info: n_rot = 64 +0.00.401.652 I print_info: n_swa = 512 +0.00.401.652 I print_info: is_swa_any = 1 +0.00.401.653 I print_info: non_causal_type = 0 +0.00.401.653 I print_info: n_embd_head_k = 256 +0.00.401.653 I print_info: n_embd_head_v = 256 +0.00.401.655 I print_info: n_gqa = 4 +0.00.401.656 I print_info: n_embd_k_gqa = 512 +0.00.401.658 I print_info: n_embd_v_gqa = 512 +0.00.401.661 I print_info: f_norm_eps = 0.0e+00 +0.00.401.662 I print_info: f_norm_rms_eps = 1.0e-06 +0.00.401.662 I print_info: f_clamp_kqv = 0.0e+00 +0.00.401.663 I print_info: f_max_alibi_bias = 0.0e+00 +0.00.401.663 I print_info: f_logit_scale = 0.0e+00 +0.00.401.663 I print_info: f_attn_scale = 0.0e+00 +0.00.401.664 I print_info: f_attn_value_scale = 0.0000 +0.00.401.665 I print_info: n_ff = 6656 +0.00.401.665 I print_info: n_expert = 0 +0.00.401.666 I print_info: n_expert_used = 0 +0.00.401.668 I print_info: n_expert_groups = 0 +0.00.401.668 I print_info: n_group_used = 0 +0.00.401.669 I print_info: causal attn = 1 +0.00.401.669 I print_info: pooling type = -1 +0.00.401.669 I print_info: rope type = 2 +0.00.401.670 I print_info: rope scaling = linear +0.00.401.671 I print_info: freq_base_train = 5000000.0 +0.00.401.672 I print_info: freq_scale_train = 1 +0.00.401.673 I print_info: freq_base_swa = 10000.0 +0.00.401.674 I print_info: freq_scale_swa = 1 +0.00.401.674 I print_info: n_embd_head_k_swa = 256 +0.00.401.674 I print_info: n_embd_head_v_swa = 256 +0.00.401.674 I print_info: n_rot_swa = 256 +0.00.401.675 I print_info: n_ctx_orig_yarn = 1048576 +0.00.401.675 I print_info: rope_yarn_log_mul = 0.0000 +0.00.401.675 I print_info: rope_finetuned = unknown +0.00.401.676 I print_info: model type = 1.7B +0.00.401.677 I print_info: model params = 1.71 B +0.00.401.677 I print_info: general.name = Hf_Format +0.00.401.678 I print_info: vocab type = BPE +0.00.401.679 I print_info: n_vocab = 131072 +0.00.401.679 I print_info: n_merges = 130716 +0.00.401.680 I print_info: BOS token = 0 '<｜start▁of▁sentence｜>' +0.00.401.680 I print_info: EOS token = 1 '<｜end▁of▁sentence｜>' +0.00.401.680 I print_info: EOT token = 1 '<｜end▁of▁sentence｜>' +0.00.401.681 I print_info: UNK token = 5 '' +0.00.401.681 I print_info: PAD token = 2 '<｜▁pad▁｜>' +0.00.401.682 I print_info: LF token = 198 'Ċ' +0.00.401.682 I print_info: FIM PRE token = 11 '<｜fim▁begin｜>' +0.00.401.683 I print_info: FIM SUF token = 12 '<｜fim▁hole｜>' +0.00.401.683 I print_info: FIM MID token = 13 '<｜fim▁end｜>' +0.00.401.683 I print_info: EOG token = 1 '<｜end▁of▁sentence｜>' +0.00.401.684 I print_info: max token length = 210 +0.00.401.685 I load_tensors: loading model tensors, this can take a while... (load_mode = mmap) +0.00.450.111 I load_tensors: offloading 0 repeating layers to GPU +0.00.450.113 I load_tensors: offloaded 0/29 layers to GPU +0.00.450.117 I load_tensors: CPU_Mapped model buffer size = 1043.71 MiB +0.00.450.118 I load_tensors: CPU_REPACK model buffer size = 622.37 MiB +0.00.824.248 I cmn common_init_: added <｜end▁of▁sentence｜> logit bias = -inf +0.00.824.673 I llama_context: constructing llama_context +0.00.824.678 I llama_context: n_seq_max = 1 +0.00.824.679 I llama_context: n_ctx = 4096 +0.00.824.679 I llama_context: n_ctx_seq = 4096 +0.00.824.679 I llama_context: n_batch = 2048 +0.00.824.680 I llama_context: n_ubatch = 512 +0.00.824.680 I llama_context: causal_attn = 1 +0.00.824.681 I llama_context: flash_attn = auto +0.00.824.681 I llama_context: kv_unified = false +0.00.824.684 I llama_context: freq_base = 5000000.0 +0.00.824.686 I llama_context: freq_scale = 1 +0.00.824.687 I llama_context: n_rs_seq = 0 +0.00.824.687 I llama_context: n_outputs_max = 1 +0.00.824.687 I llama_context: n_outputs_max_per_seq = 1 +0.00.824.688 I llama_context: n_ctx_seq (4096) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized +0.00.824.779 I llama_context: CPU output buffer size = 0.50 MiB +0.00.824.783 I llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells +0.00.824.885 I llama_kv_cache: CPU KV buffer size = 56.00 MiB +0.00.828.699 I llama_kv_cache: size = 56.00 MiB ( 4096 cells, 7 layers, 1/1 seqs), K (f16): 28.00 MiB, V (f16): 28.00 MiB +0.00.828.707 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.828.708 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.828.713 I llama_kv_cache_iswa: creating SWA KV cache, size = 1024 cells +0.00.828.799 I llama_kv_cache: CPU KV buffer size = 42.00 MiB +0.00.832.448 I llama_kv_cache: size = 42.00 MiB ( 1024 cells, 21 layers, 1/1 seqs), K (f16): 21.00 MiB, V (f16): 21.00 MiB +0.00.832.455 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.832.455 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.832.462 I sched_reserve: reserving ... +0.00.833.513 I resolve_fused_ops: Flash Attention enabled +0.00.833.514 I resolve_fused_ops: resolving fused DeepSeek V4 HC support: +0.00.833.875 I resolve_fused_ops: fused DeepSeek V4 HC pre enabled +0.00.834.196 I resolve_fused_ops: fused DeepSeek V4 HC comb enabled +0.00.834.523 I resolve_fused_ops: fused DeepSeek V4 HC post enabled +0.00.835.654 I sched_reserve: CPU compute buffer size = 60.02 MiB +0.00.835.656 I sched_reserve: graph: nodes = 986, splits = 1, input objects = 4, input tensors = 9 +0.00.835.657 I sched_reserve: reserve took 3.19 ms, sched copies = 1 +0.00.835.751 I cmn init: llama threadpool init, n_threads = 12 +0.00.835.764 I cmn common_init_: warming up the model with an empty run - please wait ... (--no-warmup to disable) +0.00.975.856 I srv load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false' +0.00.975.927 I spec common_specu: no implementations specified for speculative decoding +0.00.975.929 I slot load_model: id 0 | task -1 | new slot, n_ctx = 4096 +0.00.976.056 I srv load_model: prompt cache is disabled - use `--cache-ram N` to enable it +0.00.976.056 I srv load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391 +0.00.976.057 I srv load_model: context checkpoints enabled, max = 32, min spacing = 8192 +0.00.976.090 W srv init: --cache-idle-slots requires --cache-ram, disabling +0.00.991.948 I srv init: init: chat template, example_format: '<｜start▁of▁sentence｜><|System|> +you are a helpful assistant. + +You are a helpful assistant<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>Hello<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>Hi there<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>How are you?<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>' +0.01.004.124 I srv init: init: chat template, thinking = 0 +0.01.004.136 I srv init: preserve_reasoning kwarg: true +0.01.004.137 W srv init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve) +0.01.004.164 I srv llama_server: model loaded +0.01.004.167 I srv llama_server: listening on http://127.0.0.1:59618 +0.01.004.268 I srv update_slots: all slots are idle +0.01.165.436 I srv server_strea: conv_id= (empty=1) +0.01.165.713 I srv operator(): chat format: peg-native +0.01.165.873 I slot get_availabl: id 0 | task -1 | - skipping, slot is empty +0.01.165.876 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = -1 +0.01.166.197 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.166.219 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.166.223 I slot launch_slot_: id 0 | task 0 | processing task, is_child = 0 +0.01.166.243 I slot operator(): id 0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 28 +0.01.166.252 I slot operator(): id 0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.212.242 I slot operator(): id 0 | task 0 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.212.444 I slot create_check: id 0 | task 0 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.01.274.979 I slot operator(): id 0 | task 0 | cached n_tokens = 24, memory_seq_rm [24, end) +0.01.275.048 I slot init_sampler: id 0 | task 0 | init sampler, took 0.01 ms, tokens: text = 28, total = 28 +0.01.275.316 I slot create_check: id 0 | task 0 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 23, n_tokens = 24, size = 0.985 MiB) +0.01.326.780 I slot print_timing: id 0 | task 0 | prompt eval time = 136.89 ms / 28 tokens ( 4.89 ms per token, 204.54 tokens per second) +0.01.326.794 I slot print_timing: id 0 | task 0 | eval time = 23.61 ms / 2 tokens ( 23.61 ms per token, 42.35 tokens per second) +0.01.326.796 I slot print_timing: id 0 | task 0 | total time = 160.50 ms / 30 tokens +0.01.326.798 I slot print_timing: id 0 | task 0 | graphs reused = 1 +0.01.326.836 I slot release: id 0 | task 0 | stop processing: n_tokens = 29, truncated = 0 +0.01.326.848 I srv update_slots: all slots are idle +0.01.374.860 I srv server_strea: conv_id= (empty=1) +0.01.377.098 I srv operator(): chat format: peg-native +0.01.377.280 I slot get_availabl: id 0 | task -1 | - checking sim = 0.005 (12/2463) > 0.100 +0.01.377.282 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = 1111656 +0.01.377.667 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.377.681 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.377.684 I slot launch_slot_: id 0 | task 5 | processing task, is_child = 0 +0.01.377.701 I slot operator(): id 0 | task 5 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.01.377.706 I slot operator(): id 0 | task 5 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.01.377.710 I slot operator(): id 0 | task 5 | erased invalidated context checkpoint (pos_min = 0, pos_max = 23, n_tokens = 24, n_swa = 512, pos_next = 0, size = 0.985 MiB) +0.01.377.712 I slot operator(): id 0 | task 5 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.420.613 I slot operator(): id 0 | task 5 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.420.999 I slot create_check: id 0 | task 5 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.06.791.097 I slot print_timing: id 0 | task 5 | prompt processing, n_tokens = 1947, progress = 0.79, t = 5.41 s / 359.67 tokens per second +0.06.791.105 I slot operator(): id 0 | task 5 | cached n_tokens = 1947, memory_seq_rm [1947, end) +0.06.796.043 I slot create_check: id 0 | task 5 | created context checkpoint 2 of 32 (pos_min = 923, pos_max = 1946, n_tokens = 1947, size = 21.006 MiB) +0.08.349.078 I slot print_timing: id 0 | task 5 | prompt processing, n_tokens = 2459, progress = 1.00, t = 6.97 s / 352.73 tokens per second +0.08.349.086 I slot operator(): id 0 | task 5 | cached n_tokens = 2459, memory_seq_rm [2459, end) +0.08.349.348 I slot init_sampler: id 0 | task 5 | init sampler, took 0.24 ms, tokens: text = 2463, total = 2463 +0.08.354.366 I slot create_check: id 0 | task 5 | created context checkpoint 3 of 32 (pos_min = 1435, pos_max = 2458, n_tokens = 2459, size = 21.006 MiB) +0.08.512.684 I slot print_timing: id 0 | task 5 | prompt eval time = 7017.96 ms / 2463 tokens ( 2.85 ms per token, 350.96 tokens per second) +0.08.512.693 I slot print_timing: id 0 | task 5 | eval time = 117.00 ms / 5 tokens ( 29.25 ms per token, 34.19 tokens per second) +0.08.512.694 I slot print_timing: id 0 | task 5 | total time = 7134.97 ms / 2468 tokens +0.08.512.695 I slot print_timing: id 0 | task 5 | graphs reused = 4 +0.08.512.750 I slot release: id 0 | task 5 | stop processing: n_tokens = 2467, truncated = 0 +0.08.512.775 I srv update_slots: all slots are idle +0.08.557.239 I srv server_strea: conv_id= (empty=1) +0.08.559.706 I srv operator(): chat format: peg-native +0.08.559.938 I slot get_availabl: id 0 | task -1 | - checking sim = 1.000 (2463/2463) > 0.100 +0.08.559.941 I slot get_availabl: id 0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998 +0.08.560.193 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.08.560.205 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.08.560.207 I slot launch_slot_: id 0 | task 14 | processing task, is_child = 0 +0.08.560.215 I slot operator(): id 0 | task 14 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.08.560.218 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.08.560.221 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 923, pos_max = 1946, n_tokens = 1947, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.08.561.110 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 1435, pos_max = 2458, n_tokens = 2459, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.08.561.967 I slot operator(): id 0 | task 14 | cached n_tokens = 0, memory_seq_rm [0, end) +0.08.607.867 I slot operator(): id 0 | task 14 | cached n_tokens = 10, memory_seq_rm [10, end) +0.08.608.004 I slot create_check: id 0 | task 14 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.13.986.838 I slot print_timing: id 0 | task 14 | prompt processing, n_tokens = 1947, progress = 0.79, t = 5.43 s / 358.79 tokens per second +0.13.986.850 I slot operator(): id 0 | task 14 | cached n_tokens = 1947, memory_seq_rm [1947, end) +0.13.991.868 I slot create_check: id 0 | task 14 | created context checkpoint 2 of 32 (pos_min = 923, pos_max = 1946, n_tokens = 1947, size = 21.006 MiB) +0.15.500.187 I slot print_timing: id 0 | task 14 | prompt processing, n_tokens = 2459, progress = 1.00, t = 6.94 s / 354.33 tokens per second +0.15.500.194 I slot operator(): id 0 | task 14 | cached n_tokens = 2459, memory_seq_rm [2459, end) +0.15.500.558 I slot init_sampler: id 0 | task 14 | init sampler, took 0.34 ms, tokens: text = 2463, total = 2463 +0.15.504.464 I slot create_check: id 0 | task 14 | created context checkpoint 3 of 32 (pos_min = 1435, pos_max = 2458, n_tokens = 2459, size = 21.006 MiB) +0.15.656.370 I slot print_timing: id 0 | task 14 | prompt eval time = 6981.49 ms / 2463 tokens ( 2.83 ms per token, 352.79 tokens per second) +0.15.656.384 I slot print_timing: id 0 | task 14 | eval time = 114.65 ms / 5 tokens ( 28.66 ms per token, 34.89 tokens per second) +0.15.656.386 I \ No newline at end of file diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_server_restart.log b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_server_restart.log new file mode 100644 index 000000000..74e328470 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_server_restart.log @@ -0,0 +1,208 @@ +0.00.000.957 I srv llama_server: initializing ... +0.00.227.026 I cmn common_param: common_params_print_info: build 11124 (1a679828f) with Clang 20.1.8 for Windows x86_64 +0.00.227.034 I cmn common_param: common_params_print_info: verbosity = 4 (adjust with the `-lv N` CLI arg) +0.00.227.035 I cmn common_param: device_info: +0.00.232.036 I cmn common_param: - Vulkan0 : NVIDIA GeForce RTX 5090 Laptop GPU (24137 MiB, 23369 MiB free) +0.00.235.833 I cmn common_param: - Vulkan1 : AMD Radeon(TM) 890M Graphics (48505 MiB, 46080 MiB free) +0.00.235.842 I cmn common_param: - CPU : AMD Ryzen AI 9 HX 370 w/ Radeon 890M (64633 MiB, 25775 MiB free) +0.00.235.897 I cmn common_param: system_info: n_threads = 12 (n_threads_batch = 12) / 24 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | +0.00.235.978 I srv init_listene: running without SSL +0.00.236.037 I srv init_listene: The UI is disabled +0.00.236.038 I srv init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable +0.00.236.042 I srv init: using 23 threads for HTTP server +0.00.236.105 W srv llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655) +0.00.242.834 I srv load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.242.835 I srv load_model: local path 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.266.405 I llama_model_loader: loaded meta data with 37 key-value pairs and 226 tensors from C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf (version GGUF V3 (latest)) +0.00.266.420 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. +0.00.266.427 I llama_model_loader: - kv 0: general.architecture str = spark2_5 +0.00.266.428 I llama_model_loader: - kv 1: general.type str = model +0.00.266.432 I llama_model_loader: - kv 2: general.sampling.top_k i32 = 0 +0.00.266.440 I llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000 +0.00.266.441 I llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000 +0.00.266.441 I llama_model_loader: - kv 5: general.name str = Hf_Format +0.00.266.442 I llama_model_loader: - kv 6: general.size_label str = 1.7B +0.00.266.444 I llama_model_loader: - kv 7: spark2_5.block_count u32 = 28 +0.00.266.445 I llama_model_loader: - kv 8: spark2_5.context_length u32 = 1048576 +0.00.266.446 I llama_model_loader: - kv 9: spark2_5.embedding_length u32 = 2048 +0.00.266.446 I llama_model_loader: - kv 10: spark2_5.feed_forward_length u32 = 6656 +0.00.266.447 I llama_model_loader: - kv 11: spark2_5.attention.head_count u32 = 8 +0.00.266.447 I llama_model_loader: - kv 12: spark2_5.attention.head_count_kv u32 = 2 +0.00.266.454 I llama_model_loader: - kv 13: spark2_5.rope.freq_base f32 = 5000000.000000 +0.00.266.455 I llama_model_loader: - kv 14: spark2_5.rope.freq_base_swa f32 = 10000.000000 +0.00.266.457 I llama_model_loader: - kv 15: spark2_5.attention.layer_norm_rms_epsilon f32 = 0.000001 +0.00.266.457 I llama_model_loader: - kv 16: spark2_5.attention.key_length u32 = 256 +0.00.266.458 I llama_model_loader: - kv 17: spark2_5.attention.value_length u32 = 256 +0.00.266.458 I llama_model_loader: - kv 18: spark2_5.vocab_size u32 = 131072 +0.00.266.458 I llama_model_loader: - kv 19: spark2_5.attention.sliding_window u32 = 512 +0.00.266.480 I llama_model_loader: - kv 20: spark2_5.attention.sliding_window_pattern arr[bool,28] = [true, true, true, false, true, true,... +0.00.266.481 I llama_model_loader: - kv 21: spark2_5.rope.dimension_count u32 = 64 +0.00.266.482 I llama_model_loader: - kv 22: spark2_5.rope.dimension_count_swa u32 = 256 +0.00.266.482 I llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 +0.00.266.482 I llama_model_loader: - kv 24: tokenizer.ggml.pre str = spark2_5 +0.00.283.849 I llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,131072] = ["<｜start▁of▁sentence｜>", "<... +0.00.289.023 I llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,131072] = [3, 3, 3, 4, 4, 3, 3, 3, 4, 4, 4, 3, ... +0.00.307.306 I llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,130716] = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "Ġ... +0.00.307.316 I llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 0 +0.00.307.317 I llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 1 +0.00.307.318 I llama_model_loader: - kv 30: tokenizer.ggml.unknown_token_id u32 = 5 +0.00.307.319 I llama_model_loader: - kv 31: tokenizer.ggml.padding_token_id u32 = 2 +0.00.307.320 I llama_model_loader: - kv 32: tokenizer.ggml.add_bos_token bool = false +0.00.307.321 I llama_model_loader: - kv 33: tokenizer.ggml.add_eos_token bool = false +0.00.307.326 I llama_model_loader: - kv 34: tokenizer.chat_template str = {%- if not messages %}\n {{- raise_... +0.00.307.327 I llama_model_loader: - kv 35: general.quantization_version u32 = 2 +0.00.307.328 I llama_model_loader: - kv 36: general.file_type u32 = 15 +0.00.307.340 I llama_model_loader: - type f32: 57 tensors +0.00.307.341 I llama_model_loader: - type q4_K: 140 tensors +0.00.307.341 I llama_model_loader: - type q6_K: 29 tensors +0.00.307.349 I print_info: file format = GGUF V3 (latest) +0.00.307.349 I print_info: file type = Q4_K - Medium +0.00.307.353 I print_info: file size = 1.03 GiB (5.16 BPW) +0.00.375.357 I load: 86 unused tokens +0.00.402.789 I load: printing all EOG tokens: +0.00.402.796 I load: - 1 ('<｜end▁of▁sentence｜>') +0.00.402.944 I load: special tokens cache size = 113 +0.00.418.917 I load: token to piece cache size = 0.7975 MB +0.00.418.945 I print_info: arch = spark2_5 +0.00.418.946 I print_info: vocab_only = 0 +0.00.418.946 I print_info: no_alloc = 0 +0.00.418.947 I print_info: n_ctx_train = 1048576 +0.00.418.947 I print_info: n_embd_inp = 2048 +0.00.418.948 I print_info: n_embd = 2048 +0.00.418.948 I print_info: n_embd_out = 2048 +0.00.418.948 I print_info: n_layer = 28 +0.00.418.949 I print_info: n_layer_all = 28 +0.00.418.967 I print_info: n_head = 8 +0.00.418.968 I print_info: n_head_kv = 2 +0.00.418.968 I print_info: n_rot = 64 +0.00.418.969 I print_info: n_swa = 512 +0.00.418.969 I print_info: is_swa_any = 1 +0.00.418.970 I print_info: non_causal_type = 0 +0.00.418.970 I print_info: n_embd_head_k = 256 +0.00.418.970 I print_info: n_embd_head_v = 256 +0.00.418.971 I print_info: n_gqa = 4 +0.00.418.973 I print_info: n_embd_k_gqa = 512 +0.00.418.974 I print_info: n_embd_v_gqa = 512 +0.00.418.977 I print_info: f_norm_eps = 0.0e+00 +0.00.418.978 I print_info: f_norm_rms_eps = 1.0e-06 +0.00.418.978 I print_info: f_clamp_kqv = 0.0e+00 +0.00.418.979 I print_info: f_max_alibi_bias = 0.0e+00 +0.00.418.979 I print_info: f_logit_scale = 0.0e+00 +0.00.418.979 I print_info: f_attn_scale = 0.0e+00 +0.00.418.980 I print_info: f_attn_value_scale = 0.0000 +0.00.418.981 I print_info: n_ff = 6656 +0.00.418.982 I print_info: n_expert = 0 +0.00.418.982 I print_info: n_expert_used = 0 +0.00.418.984 I print_info: n_expert_groups = 0 +0.00.418.984 I print_info: n_group_used = 0 +0.00.418.984 I print_info: causal attn = 1 +0.00.418.985 I print_info: pooling type = -1 +0.00.418.985 I print_info: rope type = 2 +0.00.418.986 I print_info: rope scaling = linear +0.00.418.987 I print_info: freq_base_train = 5000000.0 +0.00.418.988 I print_info: freq_scale_train = 1 +0.00.418.989 I print_info: freq_base_swa = 10000.0 +0.00.418.989 I print_info: freq_scale_swa = 1 +0.00.418.990 I print_info: n_embd_head_k_swa = 256 +0.00.418.990 I print_info: n_embd_head_v_swa = 256 +0.00.418.990 I print_info: n_rot_swa = 256 +0.00.418.990 I print_info: n_ctx_orig_yarn = 1048576 +0.00.418.991 I print_info: rope_yarn_log_mul = 0.0000 +0.00.418.991 I print_info: rope_finetuned = unknown +0.00.418.992 I print_info: model type = 1.7B +0.00.418.993 I print_info: model params = 1.71 B +0.00.418.993 I print_info: general.name = Hf_Format +0.00.418.994 I print_info: vocab type = BPE +0.00.418.995 I print_info: n_vocab = 131072 +0.00.418.995 I print_info: n_merges = 130716 +0.00.418.996 I print_info: BOS token = 0 '<｜start▁of▁sentence｜>' +0.00.418.996 I print_info: EOS token = 1 '<｜end▁of▁sentence｜>' +0.00.418.997 I print_info: EOT token = 1 '<｜end▁of▁sentence｜>' +0.00.418.997 I print_info: UNK token = 5 '' +0.00.418.998 I print_info: PAD token = 2 '<｜▁pad▁｜>' +0.00.418.998 I print_info: LF token = 198 'Ċ' +0.00.418.999 I print_info: FIM PRE token = 11 '<｜fim▁begin｜>' +0.00.418.999 I print_info: FIM SUF token = 12 '<｜fim▁hole｜>' +0.00.419.000 I print_info: FIM MID token = 13 '<｜fim▁end｜>' +0.00.419.000 I print_info: EOG token = 1 '<｜end▁of▁sentence｜>' +0.00.419.000 I print_info: max token length = 210 +0.00.419.002 I load_tensors: loading model tensors, this can take a while... (load_mode = mmap) +0.00.467.339 I load_tensors: offloading 0 repeating layers to GPU +0.00.467.340 I load_tensors: offloaded 0/29 layers to GPU +0.00.467.344 I load_tensors: CPU_Mapped model buffer size = 1043.71 MiB +0.00.467.345 I load_tensors: CPU_REPACK model buffer size = 622.37 MiB +0.00.843.877 I cmn common_init_: added <｜end▁of▁sentence｜> logit bias = -inf +0.00.844.376 I llama_context: constructing llama_context +0.00.844.382 I llama_context: n_seq_max = 1 +0.00.844.382 I llama_context: n_ctx = 4096 +0.00.844.383 I llama_context: n_ctx_seq = 4096 +0.00.844.383 I llama_context: n_batch = 2048 +0.00.844.383 I llama_context: n_ubatch = 512 +0.00.844.384 I llama_context: causal_attn = 1 +0.00.844.385 I llama_context: flash_attn = auto +0.00.844.385 I llama_context: kv_unified = false +0.00.844.388 I llama_context: freq_base = 5000000.0 +0.00.844.391 I llama_context: freq_scale = 1 +0.00.844.391 I llama_context: n_rs_seq = 0 +0.00.844.392 I llama_context: n_outputs_max = 1 +0.00.844.392 I llama_context: n_outputs_max_per_seq = 1 +0.00.844.393 I llama_context: n_ctx_seq (4096) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized +0.00.844.502 I llama_context: CPU output buffer size = 0.50 MiB +0.00.844.507 I llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells +0.00.844.666 I llama_kv_cache: CPU KV buffer size = 56.00 MiB +0.00.849.110 I llama_kv_cache: size = 56.00 MiB ( 4096 cells, 7 layers, 1/1 seqs), K (f16): 28.00 MiB, V (f16): 28.00 MiB +0.00.849.117 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.849.118 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.849.122 I llama_kv_cache_iswa: creating SWA KV cache, size = 1024 cells +0.00.849.187 I llama_kv_cache: CPU KV buffer size = 42.00 MiB +0.00.851.886 I llama_kv_cache: size = 42.00 MiB ( 1024 cells, 21 layers, 1/1 seqs), K (f16): 21.00 MiB, V (f16): 21.00 MiB +0.00.851.890 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.851.890 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.851.896 I sched_reserve: reserving ... +0.00.852.677 I resolve_fused_ops: Flash Attention enabled +0.00.852.678 I resolve_fused_ops: resolving fused DeepSeek V4 HC support: +0.00.852.978 I resolve_fused_ops: fused DeepSeek V4 HC pre enabled +0.00.853.255 I resolve_fused_ops: fused DeepSeek V4 HC comb enabled +0.00.853.525 I resolve_fused_ops: fused DeepSeek V4 HC post enabled +0.00.854.677 I sched_reserve: CPU compute buffer size = 60.02 MiB +0.00.854.679 I sched_reserve: graph: nodes = 986, splits = 1, input objects = 4, input tensors = 9 +0.00.854.680 I sched_reserve: reserve took 2.78 ms, sched copies = 1 +0.00.854.770 I cmn init: llama threadpool init, n_threads = 12 +0.00.854.784 I cmn common_init_: warming up the model with an empty run - please wait ... (--no-warmup to disable) +0.00.987.296 I srv load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false' +0.00.987.359 I spec common_specu: no implementations specified for speculative decoding +0.00.987.361 I slot load_model: id 0 | task -1 | new slot, n_ctx = 4096 +0.00.987.580 I srv load_model: prompt cache is disabled - use `--cache-ram N` to enable it +0.00.987.581 I srv load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391 +0.00.987.583 I srv load_model: context checkpoints enabled, max = 32, min spacing = 8192 +0.00.987.630 W srv init: --cache-idle-slots requires --cache-ram, disabling +0.01.003.420 I srv init: init: chat template, example_format: '<｜start▁of▁sentence｜><|System|> +you are a helpful assistant. + +You are a helpful assistant<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>Hello<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>Hi there<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>How are you?<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>' +0.01.014.566 I srv init: init: chat template, thinking = 0 +0.01.014.576 I srv init: preserve_reasoning kwarg: true +0.01.014.577 W srv init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve) +0.01.014.600 I srv llama_server: model loaded +0.01.014.603 I srv llama_server: listening on http://127.0.0.1:62987 +0.01.014.702 I srv update_slots: all slots are idle +0.01.188.894 I srv server_strea: conv_id= (empty=1) +0.01.189.119 I srv operator(): chat format: peg-native +0.01.189.260 I slot get_availabl: id 0 | task -1 | - skipping, slot is empty +0.01.189.263 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = -1 +0.01.189.342 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.189.371 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.189.375 I slot launch_slot_: id 0 | task 0 | processing task, is_child = 0 +0.01.189.395 I slot operator(): id 0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 28 +0.01.189.407 I slot operator(): id 0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.234.192 I slot operator(): id 0 | task 0 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.234.397 I slot create_check: id 0 | task 0 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.01.284.959 I slot operator(): id 0 | task 0 | cached n_tokens = 24, memory_seq_rm [24, end) +0.01.285.029 I slot init_sampler: id 0 | task 0 | init sampler, took 0.01 ms, tokens: text = 28, total = 28 +0.01.285.301 I slot create_check: id 0 | task 0 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 23, n_tokens = 24, size = 0.985 MiB) +0.01.337.651 I slot print_timing: id 0 | task 0 | prompt eval time = 124.18 ms / 28 tokens ( 4.44 ms per token, 225.48 tokens per second) +0.01.337.666 I \ No newline at end of file diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v3_report.json b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v3_report.json new file mode 100644 index 000000000..49f2b1018 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v3_report.json @@ -0,0 +1,156 @@ +{ + "scope": "complete Spark GGUF text requests through Omni StageRuntime/StagePool and external llama.cpp backend", + "model_file": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "device": "cpu", + "memory_budget": { + "capacities": { + "host_ram": 12884901888 + }, + "demands": { + "host_ram": 4294967296 + } + }, + "warmup_count": 0, + "measured_count": 1, + "startup_s": 2.8147969000274315, + "execution_plan": { + "backend": "external.llamacpp.text.v1", + "stage_id": 0, + "worker_generation": "d758f2b3af094853a8db4844e92b5066", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "requested_device": "cpu", + "expected_device_name": null, + "layer_assignments": 0, + "offloaded_layers": [ + "0", + "29" + ], + "model_buffer_devices": [], + "all_model_buffers": [ + "CPU_Mapped", + "CPU_REPACK" + ], + "cpu_only_layout": false, + "worker_pid": 11764, + "loaded_rss_bytes": 1970462720, + "reserved_bytes": { + "host_ram": 4294967296 + }, + "memory_pool": "host_ram", + "memory_overhead_bytes": 2147483648, + "context_tokens": 4096, + "max_new_tokens": 96, + "max_io_bytes": 1048576, + "request_capacity": 1, + "stateful_session": "llama.cpp-owned; reset for each complete request", + "evidence": "B" + }, + "checks": [ + { + "case": "short", + "answer": "Paris", + "expected": "Paris", + "correct": true, + "wall_s": 0.1861239999998361, + "stage_event": { + "request_id": "short-688621078000000", + "stage_id": 0, + "epoch": 1, + "seq": 1, + "kind": "text", + "worker_generation": "d758f2b3af094853a8db4844e92b5066", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 0.18202489998657256, + "usage": { + "completion_tokens": 2, + "prompt_tokens": 28, + "total_tokens": 30, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + }, + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 7.082891099969856, + "stage_event": { + "request_id": "inventory_120-688621265000000", + "stage_id": 0, + "epoch": 2, + "seq": 1, + "kind": "text", + "worker_generation": "d758f2b3af094853a8db4844e92b5066", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 7.060311499983072, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "warmups": [], + "measured": [ + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 7.12928680004552, + "stage_event": { + "request_id": "inventory_120-688628359000000", + "stage_id": 0, + "epoch": 3, + "seq": 1, + "kind": "text", + "worker_generation": "d758f2b3af094853a8db4844e92b5066", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 7.123366800020449, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "nearest_rank_p50_wall_s": 7.12928680004552, + "nearest_rank_p95_wall_s": 7.12928680004552, + "status": "failed", + "error": "RuntimeError: abort request completed before in-flight cancellation", + "ledger_after_shutdown": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + } +} diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v3_server.log b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v3_server.log new file mode 100644 index 000000000..32c07288e --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v3_server.log @@ -0,0 +1,295 @@ +0.00.000.817 I srv llama_server: initializing ... +0.00.210.936 I cmn common_param: common_params_print_info: build 11124 (1a679828f) with Clang 20.1.8 for Windows x86_64 +0.00.210.944 I cmn common_param: common_params_print_info: verbosity = 4 (adjust with the `-lv N` CLI arg) +0.00.210.946 I cmn common_param: device_info: +0.00.215.074 I cmn common_param: - Vulkan0 : NVIDIA GeForce RTX 5090 Laptop GPU (24137 MiB, 23369 MiB free) +0.00.218.719 I cmn common_param: - Vulkan1 : AMD Radeon(TM) 890M Graphics (48505 MiB, 46080 MiB free) +0.00.218.732 I cmn common_param: - CPU : AMD Ryzen AI 9 HX 370 w/ Radeon 890M (64633 MiB, 26576 MiB free) +0.00.218.788 I cmn common_param: system_info: n_threads = 12 (n_threads_batch = 12) / 24 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | +0.00.218.862 I srv init_listene: running without SSL +0.00.218.923 I srv init_listene: The UI is disabled +0.00.218.924 I srv init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable +0.00.218.928 I srv init: using 23 threads for HTTP server +0.00.218.990 W srv llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655) +0.00.233.790 I srv load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.233.795 I srv load_model: local path 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.256.902 I llama_model_loader: loaded meta data with 37 key-value pairs and 226 tensors from C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf (version GGUF V3 (latest)) +0.00.256.916 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. +0.00.256.923 I llama_model_loader: - kv 0: general.architecture str = spark2_5 +0.00.256.928 I llama_model_loader: - kv 1: general.type str = model +0.00.256.932 I llama_model_loader: - kv 2: general.sampling.top_k i32 = 0 +0.00.256.939 I llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000 +0.00.256.940 I llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000 +0.00.256.940 I llama_model_loader: - kv 5: general.name str = Hf_Format +0.00.256.941 I llama_model_loader: - kv 6: general.size_label str = 1.7B +0.00.256.943 I llama_model_loader: - kv 7: spark2_5.block_count u32 = 28 +0.00.256.944 I llama_model_loader: - kv 8: spark2_5.context_length u32 = 1048576 +0.00.256.944 I llama_model_loader: - kv 9: spark2_5.embedding_length u32 = 2048 +0.00.256.945 I llama_model_loader: - kv 10: spark2_5.feed_forward_length u32 = 6656 +0.00.256.945 I llama_model_loader: - kv 11: spark2_5.attention.head_count u32 = 8 +0.00.256.945 I llama_model_loader: - kv 12: spark2_5.attention.head_count_kv u32 = 2 +0.00.256.947 I llama_model_loader: - kv 13: spark2_5.rope.freq_base f32 = 5000000.000000 +0.00.256.948 I llama_model_loader: - kv 14: spark2_5.rope.freq_base_swa f32 = 10000.000000 +0.00.256.950 I llama_model_loader: - kv 15: spark2_5.attention.layer_norm_rms_epsilon f32 = 0.000001 +0.00.256.950 I llama_model_loader: - kv 16: spark2_5.attention.key_length u32 = 256 +0.00.256.950 I llama_model_loader: - kv 17: spark2_5.attention.value_length u32 = 256 +0.00.256.951 I llama_model_loader: - kv 18: spark2_5.vocab_size u32 = 131072 +0.00.256.951 I llama_model_loader: - kv 19: spark2_5.attention.sliding_window u32 = 512 +0.00.256.971 I llama_model_loader: - kv 20: spark2_5.attention.sliding_window_pattern arr[bool,28] = [true, true, true, false, true, true,... +0.00.256.971 I llama_model_loader: - kv 21: spark2_5.rope.dimension_count u32 = 64 +0.00.256.971 I llama_model_loader: - kv 22: spark2_5.rope.dimension_count_swa u32 = 256 +0.00.256.972 I llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 +0.00.256.972 I llama_model_loader: - kv 24: tokenizer.ggml.pre str = spark2_5 +0.00.273.377 I llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,131072] = ["<｜start▁of▁sentence｜>", "<... +0.00.278.359 I llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,131072] = [3, 3, 3, 4, 4, 3, 3, 3, 4, 4, 4, 3, ... +0.00.295.035 I llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,130716] = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "Ġ... +0.00.295.045 I llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 0 +0.00.295.046 I llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 1 +0.00.295.047 I llama_model_loader: - kv 30: tokenizer.ggml.unknown_token_id u32 = 5 +0.00.295.047 I llama_model_loader: - kv 31: tokenizer.ggml.padding_token_id u32 = 2 +0.00.295.049 I llama_model_loader: - kv 32: tokenizer.ggml.add_bos_token bool = false +0.00.295.050 I llama_model_loader: - kv 33: tokenizer.ggml.add_eos_token bool = false +0.00.295.055 I llama_model_loader: - kv 34: tokenizer.chat_template str = {%- if not messages %}\n {{- raise_... +0.00.295.056 I llama_model_loader: - kv 35: general.quantization_version u32 = 2 +0.00.295.056 I llama_model_loader: - kv 36: general.file_type u32 = 15 +0.00.295.063 I llama_model_loader: - type f32: 57 tensors +0.00.295.064 I llama_model_loader: - type q4_K: 140 tensors +0.00.295.064 I llama_model_loader: - type q6_K: 29 tensors +0.00.295.069 I print_info: file format = GGUF V3 (latest) +0.00.295.070 I print_info: file type = Q4_K - Medium +0.00.295.073 I print_info: file size = 1.03 GiB (5.16 BPW) +0.00.358.012 I load: 86 unused tokens +0.00.379.854 I load: printing all EOG tokens: +0.00.379.862 I load: - 1 ('<｜end▁of▁sentence｜>') +0.00.379.997 I load: special tokens cache size = 113 +0.00.396.304 I load: token to piece cache size = 0.7975 MB +0.00.396.349 I print_info: arch = spark2_5 +0.00.396.351 I print_info: vocab_only = 0 +0.00.396.352 I print_info: no_alloc = 0 +0.00.396.352 I print_info: n_ctx_train = 1048576 +0.00.396.353 I print_info: n_embd_inp = 2048 +0.00.396.354 I print_info: n_embd = 2048 +0.00.396.354 I print_info: n_embd_out = 2048 +0.00.396.354 I print_info: n_layer = 28 +0.00.396.355 I print_info: n_layer_all = 28 +0.00.396.380 I print_info: n_head = 8 +0.00.396.382 I print_info: n_head_kv = 2 +0.00.396.383 I print_info: n_rot = 64 +0.00.396.383 I print_info: n_swa = 512 +0.00.396.383 I print_info: is_swa_any = 1 +0.00.396.384 I print_info: non_causal_type = 0 +0.00.396.384 I print_info: n_embd_head_k = 256 +0.00.396.384 I print_info: n_embd_head_v = 256 +0.00.396.386 I print_info: n_gqa = 4 +0.00.396.387 I print_info: n_embd_k_gqa = 512 +0.00.396.389 I print_info: n_embd_v_gqa = 512 +0.00.396.394 I print_info: f_norm_eps = 0.0e+00 +0.00.396.395 I print_info: f_norm_rms_eps = 1.0e-06 +0.00.396.396 I print_info: f_clamp_kqv = 0.0e+00 +0.00.396.397 I print_info: f_max_alibi_bias = 0.0e+00 +0.00.396.397 I print_info: f_logit_scale = 0.0e+00 +0.00.396.398 I print_info: f_attn_scale = 0.0e+00 +0.00.396.399 I print_info: f_attn_value_scale = 0.0000 +0.00.396.401 I print_info: n_ff = 6656 +0.00.396.401 I print_info: n_expert = 0 +0.00.396.401 I print_info: n_expert_used = 0 +0.00.396.404 I print_info: n_expert_groups = 0 +0.00.396.404 I print_info: n_group_used = 0 +0.00.396.405 I print_info: causal attn = 1 +0.00.396.406 I print_info: pooling type = -1 +0.00.396.406 I print_info: rope type = 2 +0.00.396.407 I print_info: rope scaling = linear +0.00.396.408 I print_info: freq_base_train = 5000000.0 +0.00.396.410 I print_info: freq_scale_train = 1 +0.00.396.411 I print_info: freq_base_swa = 10000.0 +0.00.396.412 I print_info: freq_scale_swa = 1 +0.00.396.412 I print_info: n_embd_head_k_swa = 256 +0.00.396.413 I print_info: n_embd_head_v_swa = 256 +0.00.396.413 I print_info: n_rot_swa = 256 +0.00.396.413 I print_info: n_ctx_orig_yarn = 1048576 +0.00.396.414 I print_info: rope_yarn_log_mul = 0.0000 +0.00.396.414 I print_info: rope_finetuned = unknown +0.00.396.416 I print_info: model type = 1.7B +0.00.396.417 I print_info: model params = 1.71 B +0.00.396.417 I print_info: general.name = Hf_Format +0.00.396.418 I print_info: vocab type = BPE +0.00.396.419 I print_info: n_vocab = 131072 +0.00.396.419 I print_info: n_merges = 130716 +0.00.396.420 I print_info: BOS token = 0 '<｜start▁of▁sentence｜>' +0.00.396.421 I print_info: EOS token = 1 '<｜end▁of▁sentence｜>' +0.00.396.421 I print_info: EOT token = 1 '<｜end▁of▁sentence｜>' +0.00.396.422 I print_info: UNK token = 5 '' +0.00.396.423 I print_info: PAD token = 2 '<｜▁pad▁｜>' +0.00.396.423 I print_info: LF token = 198 'Ċ' +0.00.396.424 I print_info: FIM PRE token = 11 '<｜fim▁begin｜>' +0.00.396.424 I print_info: FIM SUF token = 12 '<｜fim▁hole｜>' +0.00.396.425 I print_info: FIM MID token = 13 '<｜fim▁end｜>' +0.00.396.425 I print_info: EOG token = 1 '<｜end▁of▁sentence｜>' +0.00.396.426 I print_info: max token length = 210 +0.00.396.428 I load_tensors: loading model tensors, this can take a while... (load_mode = mmap) +0.00.444.894 I load_tensors: offloading 0 repeating layers to GPU +0.00.444.896 I load_tensors: offloaded 0/29 layers to GPU +0.00.444.899 I load_tensors: CPU_Mapped model buffer size = 1043.71 MiB +0.00.444.900 I load_tensors: CPU_REPACK model buffer size = 622.37 MiB +0.00.827.736 I cmn common_init_: added <｜end▁of▁sentence｜> logit bias = -inf +0.00.828.144 I llama_context: constructing llama_context +0.00.828.150 I llama_context: n_seq_max = 1 +0.00.828.150 I llama_context: n_ctx = 4096 +0.00.828.151 I llama_context: n_ctx_seq = 4096 +0.00.828.151 I llama_context: n_batch = 2048 +0.00.828.151 I llama_context: n_ubatch = 512 +0.00.828.152 I llama_context: causal_attn = 1 +0.00.828.152 I llama_context: flash_attn = auto +0.00.828.153 I llama_context: kv_unified = false +0.00.828.156 I llama_context: freq_base = 5000000.0 +0.00.828.158 I llama_context: freq_scale = 1 +0.00.828.158 I llama_context: n_rs_seq = 0 +0.00.828.159 I llama_context: n_outputs_max = 1 +0.00.828.159 I llama_context: n_outputs_max_per_seq = 1 +0.00.828.160 I llama_context: n_ctx_seq (4096) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized +0.00.828.244 I llama_context: CPU output buffer size = 0.50 MiB +0.00.828.249 I llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells +0.00.828.370 I llama_kv_cache: CPU KV buffer size = 56.00 MiB +0.00.832.272 I llama_kv_cache: size = 56.00 MiB ( 4096 cells, 7 layers, 1/1 seqs), K (f16): 28.00 MiB, V (f16): 28.00 MiB +0.00.832.281 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.832.281 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.832.288 I llama_kv_cache_iswa: creating SWA KV cache, size = 1024 cells +0.00.832.408 I llama_kv_cache: CPU KV buffer size = 42.00 MiB +0.00.836.524 I llama_kv_cache: size = 42.00 MiB ( 1024 cells, 21 layers, 1/1 seqs), K (f16): 21.00 MiB, V (f16): 21.00 MiB +0.00.836.531 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.836.531 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.836.539 I sched_reserve: reserving ... +0.00.837.496 I resolve_fused_ops: Flash Attention enabled +0.00.837.497 I resolve_fused_ops: resolving fused DeepSeek V4 HC support: +0.00.837.795 I resolve_fused_ops: fused DeepSeek V4 HC pre enabled +0.00.838.073 I resolve_fused_ops: fused DeepSeek V4 HC comb enabled +0.00.838.346 I resolve_fused_ops: fused DeepSeek V4 HC post enabled +0.00.839.438 I sched_reserve: CPU compute buffer size = 60.02 MiB +0.00.839.440 I sched_reserve: graph: nodes = 986, splits = 1, input objects = 4, input tensors = 9 +0.00.839.441 I sched_reserve: reserve took 2.90 ms, sched copies = 1 +0.00.839.535 I cmn init: llama threadpool init, n_threads = 12 +0.00.839.550 I cmn common_init_: warming up the model with an empty run - please wait ... (--no-warmup to disable) +0.00.988.064 I srv load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false' +0.00.988.134 I spec common_specu: no implementations specified for speculative decoding +0.00.988.136 I slot load_model: id 0 | task -1 | new slot, n_ctx = 4096 +0.00.988.303 I srv load_model: prompt cache is disabled - use `--cache-ram N` to enable it +0.00.988.304 I srv load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391 +0.00.988.305 I srv load_model: context checkpoints enabled, max = 32, min spacing = 8192 +0.00.988.416 W srv init: --cache-idle-slots requires --cache-ram, disabling +0.01.006.854 I srv init: init: chat template, example_format: '<｜start▁of▁sentence｜><|System|> +you are a helpful assistant. + +You are a helpful assistant<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>Hello<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>Hi there<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>How are you?<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>' +0.01.019.492 I srv init: init: chat template, thinking = 0 +0.01.019.509 I srv init: preserve_reasoning kwarg: true +0.01.019.510 W srv init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve) +0.01.019.546 I srv llama_server: model loaded +0.01.019.549 I srv llama_server: listening on http://127.0.0.1:57245 +0.01.019.676 I srv update_slots: all slots are idle +0.01.192.962 I srv server_strea: conv_id= (empty=1) +0.01.193.171 I srv operator(): chat format: peg-native +0.01.193.281 I slot get_availabl: id 0 | task -1 | - skipping, slot is empty +0.01.193.283 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = -1 +0.01.193.322 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.193.335 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.193.337 I slot launch_slot_: id 0 | task 0 | processing task, is_child = 0 +0.01.193.345 I slot operator(): id 0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 28 +0.01.193.351 I slot operator(): id 0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.236.711 I slot operator(): id 0 | task 0 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.236.906 I slot create_check: id 0 | task 0 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.01.286.385 I slot operator(): id 0 | task 0 | cached n_tokens = 24, memory_seq_rm [24, end) +0.01.286.452 I slot init_sampler: id 0 | task 0 | init sampler, took 0.01 ms, tokens: text = 28, total = 28 +0.01.286.713 I slot create_check: id 0 | task 0 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 23, n_tokens = 24, size = 0.985 MiB) +0.01.345.118 I slot print_timing: id 0 | task 0 | prompt eval time = 128.12 ms / 28 tokens ( 4.58 ms per token, 218.55 tokens per second) +0.01.345.130 I slot print_timing: id 0 | task 0 | eval time = 23.62 ms / 2 tokens ( 23.62 ms per token, 42.34 tokens per second) +0.01.345.132 I slot print_timing: id 0 | task 0 | total time = 151.74 ms / 30 tokens +0.01.345.133 I slot print_timing: id 0 | task 0 | graphs reused = 1 +0.01.345.176 I slot release: id 0 | task 0 | stop processing: n_tokens = 29, truncated = 0 +0.01.345.190 I srv update_slots: all slots are idle +0.01.385.144 I srv server_strea: conv_id= (empty=1) +0.01.387.517 I srv operator(): chat format: peg-native +0.01.387.718 I slot get_availabl: id 0 | task -1 | - checking sim = 0.005 (12/2463) > 0.100 +0.01.387.720 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = 1136446 +0.01.387.966 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.387.979 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.387.981 I slot launch_slot_: id 0 | task 5 | processing task, is_child = 0 +0.01.387.994 I slot operator(): id 0 | task 5 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.01.388.000 I slot operator(): id 0 | task 5 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.01.388.004 I slot operator(): id 0 | task 5 | erased invalidated context checkpoint (pos_min = 0, pos_max = 23, n_tokens = 24, n_swa = 512, pos_next = 0, size = 0.985 MiB) +0.01.388.006 I slot operator(): id 0 | task 5 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.432.897 I slot operator(): id 0 | task 5 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.433.057 I slot create_check: id 0 | task 5 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.06.795.843 I slot print_timing: id 0 | task 5 | prompt processing, n_tokens = 1947, progress = 0.79, t = 5.41 s / 360.03 tokens per second +0.06.795.855 I slot operator(): id 0 | task 5 | cached n_tokens = 1947, memory_seq_rm [1947, end) +0.06.801.020 I slot create_check: id 0 | task 5 | created context checkpoint 2 of 32 (pos_min = 923, pos_max = 1946, n_tokens = 1947, size = 21.006 MiB) +0.08.274.591 I slot print_timing: id 0 | task 5 | prompt processing, n_tokens = 2459, progress = 1.00, t = 6.89 s / 357.07 tokens per second +0.08.274.601 I slot operator(): id 0 | task 5 | cached n_tokens = 2459, memory_seq_rm [2459, end) +0.08.274.968 I slot init_sampler: id 0 | task 5 | init sampler, took 0.34 ms, tokens: text = 2463, total = 2463 +0.08.279.242 I slot create_check: id 0 | task 5 | created context checkpoint 3 of 32 (pos_min = 1435, pos_max = 2458, n_tokens = 2459, size = 21.006 MiB) +0.08.428.008 I slot print_timing: id 0 | task 5 | prompt eval time = 6928.08 ms / 2463 tokens ( 2.81 ms per token, 355.51 tokens per second) +0.08.428.023 I slot print_timing: id 0 | task 5 | eval time = 111.91 ms / 5 tokens ( 27.98 ms per token, 35.74 tokens per second) +0.08.428.024 I slot print_timing: id 0 | task 5 | total time = 7039.99 ms / 2468 tokens +0.08.428.026 I slot print_timing: id 0 | task 5 | graphs reused = 4 +0.08.428.088 I slot release: id 0 | task 5 | stop processing: n_tokens = 2467, truncated = 0 +0.08.428.110 I srv update_slots: all slots are idle +0.08.470.962 I srv server_strea: conv_id= (empty=1) +0.08.474.562 I srv operator(): chat format: peg-native +0.08.474.805 I slot get_availabl: id 0 | task -1 | - checking sim = 1.000 (2463/2463) > 0.100 +0.08.474.808 I slot get_availabl: id 0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998 +0.08.475.137 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.08.475.158 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.08.475.162 I slot launch_slot_: id 0 | task 14 | processing task, is_child = 0 +0.08.475.172 I slot operator(): id 0 | task 14 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.08.475.176 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.08.475.179 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 923, pos_max = 1946, n_tokens = 1947, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.08.476.070 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 1435, pos_max = 2458, n_tokens = 2459, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.08.477.029 I slot operator(): id 0 | task 14 | cached n_tokens = 0, memory_seq_rm [0, end) +0.08.520.049 I slot operator(): id 0 | task 14 | cached n_tokens = 10, memory_seq_rm [10, end) +0.08.520.183 I slot create_check: id 0 | task 14 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.13.849.530 I slot print_timing: id 0 | task 14 | prompt processing, n_tokens = 1947, progress = 0.79, t = 5.37 s / 362.28 tokens per second +0.13.849.542 I slot operator(): id 0 | task 14 | cached n_tokens = 1947, memory_seq_rm [1947, end) +0.13.854.277 I slot create_check: id 0 | task 14 | created context checkpoint 2 of 32 (pos_min = 923, pos_max = 1946, n_tokens = 1947, size = 21.006 MiB) +0.15.394.775 I slot print_timing: id 0 | task 14 | prompt processing, n_tokens = 2459, progress = 1.00, t = 6.92 s / 355.37 tokens per second +0.15.394.783 I slot operator(): id 0 | task 14 | cached n_tokens = 2459, memory_seq_rm [2459, end) +0.15.395.163 I slot init_sampler: id 0 | task 14 | init sampler, took 0.35 ms, tokens: text = 2463, total = 2463 +0.15.400.272 I slot create_check: id 0 | task 14 | created context checkpoint 3 of 32 (pos_min = 1435, pos_max = 2458, n_tokens = 2459, size = 21.006 MiB) +0.15.557.579 I slot print_timing: id 0 | task 14 | prompt eval time = 6966.53 ms / 2463 tokens ( 2.83 ms per token, 353.55 tokens per second) +0.15.557.593 I slot print_timing: id 0 | task 14 | eval time = 115.85 ms / 5 tokens ( 28.96 ms per token, 34.53 tokens per second) +0.15.557.594 I slot print_timing: id 0 | task 14 | total time = 7082.39 ms / 2468 tokens +0.15.557.596 I slot print_timing: id 0 | task 14 | graphs reused = 7 +0.15.557.673 I slot release: id 0 | task 14 | stop processing: n_tokens = 2467, truncated = 0 +0.15.557.694 I srv update_slots: all slots are idle +0.15.642.630 I srv server_strea: conv_id= (empty=1) +0.15.642.938 I srv operator(): chat format: peg-native +0.15.643.085 I slot get_availabl: id 0 | task -1 | - checking sim = 0.444 (12/27) > 0.100 +0.15.643.089 I slot get_availabl: id 0 | task -1 | selected slot by LCP similarity, f_sim_best = 0.444 (> 0.100 thold), f_keep = 0.005 +0.15.643.389 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.15.643.402 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.15.643.405 I slot launch_slot_: id 0 | task 23 | processing task, is_child = 0 +0.15.643.415 I slot operator(): id 0 | task 23 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 27 +0.15.643.419 I slot operator(): id 0 | task 23 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.15.643.422 I slot operator(): id 0 | task 23 | erased invalidated context checkpoint (pos_min = 923, pos_max = 1946, n_tokens = 1947, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.15.644.402 I slot operator(): id 0 | task 23 | erased invalidated context checkpoint (pos_min = 1435, pos_max = 2458, n_tokens = 2459, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.15.645.387 I slot operator(): id 0 | task 23 | cached n_tokens = 0, memory_seq_rm [0, end) +0.15.687.945 I slot operator(): id 0 | task 23 | cached n_tokens = 10, memory_seq_rm [10, end) +0.15.688.067 I slot create_check: id 0 | task 23 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.15.738.337 I slot operator(): id 0 | task 23 | cached n_tokens = 23, memory_seq_rm [23, end) +0.15.738.376 I slot init_sampler: id 0 | task 23 | init sampler, took 0.01 ms, tokens: text = 27, total = 27 +0.15.738.642 I slot create_check: id 0 | task 23 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 22, n_tokens = 23, size = 0.944 MiB) +0.15.799.028 I slot print_timing: id 0 | task 23 | prompt eval time = 125.53 ms / 27 tokens ( 4.65 ms per token, 215.09 tokens per second) diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v4_report.json b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v4_report.json new file mode 100644 index 000000000..86cb688b4 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v4_report.json @@ -0,0 +1,250 @@ +{ + "scope": "complete Spark GGUF text requests through Omni StageRuntime/StagePool and external llama.cpp backend", + "model_file": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "device": "cpu", + "memory_budget": { + "capacities": { + "host_ram": 12884901888 + }, + "demands": { + "host_ram": 4294967296 + } + }, + "warmup_count": 0, + "measured_count": 1, + "startup_s": 3.164883299963549, + "execution_plan": { + "backend": "external.llamacpp.text.v1", + "stage_id": 0, + "worker_generation": "2ff1298272904e81afa406b3c4361654", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "requested_device": "cpu", + "expected_device_name": null, + "layer_assignments": 0, + "offloaded_layers": [ + "0", + "29" + ], + "model_buffer_devices": [], + "all_model_buffers": [ + "CPU_Mapped", + "CPU_REPACK" + ], + "cpu_only_layout": false, + "worker_pid": 67568, + "loaded_rss_bytes": 1970728960, + "reserved_bytes": { + "host_ram": 4294967296 + }, + "memory_pool": "host_ram", + "memory_overhead_bytes": 2147483648, + "context_tokens": 4096, + "max_new_tokens": 96, + "max_io_bytes": 1048576, + "request_capacity": 1, + "stateful_session": "llama.cpp-owned; reset for each complete request", + "evidence": "B" + }, + "checks": [ + { + "case": "short", + "answer": "Paris", + "expected": "Paris", + "correct": true, + "wall_s": 0.1657647000392899, + "stage_event": { + "request_id": "short-688698843000000", + "stage_id": 0, + "epoch": 1, + "seq": 1, + "kind": "text", + "worker_generation": "2ff1298272904e81afa406b3c4361654", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 0.1493499999633059, + "usage": { + "completion_tokens": 2, + "prompt_tokens": 28, + "total_tokens": 30, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + }, + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 6.759047100087628, + "stage_event": { + "request_id": "inventory_120-688699000000000", + "stage_id": 0, + "epoch": 2, + "seq": 1, + "kind": "text", + "worker_generation": "2ff1298272904e81afa406b3c4361654", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 6.754019699990749, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "warmups": [], + "measured": [ + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 7.075117200030945, + "stage_event": { + "request_id": "inventory_120-688705765000000", + "stage_id": 0, + "epoch": 3, + "seq": 1, + "kind": "text", + "worker_generation": "2ff1298272904e81afa406b3c4361654", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 7.069126900052652, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "nearest_rank_p50_wall_s": 7.075117200030945, + "nearest_rank_p95_wall_s": 7.075117200030945, + "abort_check": { + "abort_case": "inventory_120", + "server_tasks_before_abort_request": 3, + "server_tasks_after_abort_request_start": 4, + "stale_output": false, + "worker_exited": true, + "ledger_after_abort": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + } + }, + "restart_check": { + "first_runtime_ledger_after_shutdown": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + }, + "fresh_startup_s": 2.27150899998378, + "fresh_server_log": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\vllm-omni-edge\\benchmarks\\edge_harness\\results\\e2e_expansion_20260924\\evidence\\spark_omni_windows_restart\\cpu_v4_server_restart.log", + "fresh_execution_plan": { + "backend": "external.llamacpp.text.v1", + "stage_id": 0, + "worker_generation": "464ec54f6eac420b86f5ee5bb27db2cf", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "requested_device": "cpu", + "expected_device_name": null, + "layer_assignments": 0, + "offloaded_layers": [ + "0", + "29" + ], + "model_buffer_devices": [], + "all_model_buffers": [ + "CPU_Mapped", + "CPU_REPACK" + ], + "cpu_only_layout": false, + "worker_pid": 57384, + "loaded_rss_bytes": 1970438144, + "reserved_bytes": { + "host_ram": 4294967296 + }, + "memory_pool": "host_ram", + "memory_overhead_bytes": 2147483648, + "context_tokens": 4096, + "max_new_tokens": 96, + "max_io_bytes": 1048576, + "request_capacity": 1, + "stateful_session": "llama.cpp-owned; reset for each complete request", + "evidence": "B" + }, + "fresh_request": { + "case": "short", + "answer": "Paris", + "expected": "Paris", + "correct": true, + "wall_s": 0.14341949997469783, + "stage_event": { + "request_id": "short-688715500000000", + "stage_id": 0, + "epoch": 1, + "seq": 1, + "kind": "text", + "worker_generation": "464ec54f6eac420b86f5ee5bb27db2cf", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 0.141190700000152, + "usage": { + "completion_tokens": 2, + "prompt_tokens": 28, + "total_tokens": 30, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + }, + "late_output": false + }, + "status": "passed", + "ledger_after_shutdown": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + } +} diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v4_server.log b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v4_server.log new file mode 100644 index 000000000..2216c0578 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v4_server.log @@ -0,0 +1,289 @@ +0.00.000.752 I srv llama_server: initializing ... +0.00.207.629 I cmn common_param: common_params_print_info: build 11124 (1a679828f) with Clang 20.1.8 for Windows x86_64 +0.00.207.635 I cmn common_param: common_params_print_info: verbosity = 4 (adjust with the `-lv N` CLI arg) +0.00.207.637 I cmn common_param: device_info: +0.00.212.464 I cmn common_param: - Vulkan0 : NVIDIA GeForce RTX 5090 Laptop GPU (24137 MiB, 23369 MiB free) +0.00.216.159 I cmn common_param: - Vulkan1 : AMD Radeon(TM) 890M Graphics (48505 MiB, 46080 MiB free) +0.00.216.167 I cmn common_param: - CPU : AMD Ryzen AI 9 HX 370 w/ Radeon 890M (64633 MiB, 26292 MiB free) +0.00.216.220 I cmn common_param: system_info: n_threads = 12 (n_threads_batch = 12) / 24 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | +0.00.216.293 I srv init_listene: running without SSL +0.00.216.349 I srv init_listene: The UI is disabled +0.00.216.350 I srv init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable +0.00.216.354 I srv init: using 23 threads for HTTP server +0.00.216.416 W srv llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655) +0.00.224.562 I srv load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.224.566 I srv load_model: local path 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.248.307 I llama_model_loader: loaded meta data with 37 key-value pairs and 226 tensors from C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf (version GGUF V3 (latest)) +0.00.248.319 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. +0.00.248.326 I llama_model_loader: - kv 0: general.architecture str = spark2_5 +0.00.248.327 I llama_model_loader: - kv 1: general.type str = model +0.00.248.330 I llama_model_loader: - kv 2: general.sampling.top_k i32 = 0 +0.00.248.340 I llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000 +0.00.248.341 I llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000 +0.00.248.341 I llama_model_loader: - kv 5: general.name str = Hf_Format +0.00.248.342 I llama_model_loader: - kv 6: general.size_label str = 1.7B +0.00.248.346 I llama_model_loader: - kv 7: spark2_5.block_count u32 = 28 +0.00.248.347 I llama_model_loader: - kv 8: spark2_5.context_length u32 = 1048576 +0.00.248.348 I llama_model_loader: - kv 9: spark2_5.embedding_length u32 = 2048 +0.00.248.348 I llama_model_loader: - kv 10: spark2_5.feed_forward_length u32 = 6656 +0.00.248.349 I llama_model_loader: - kv 11: spark2_5.attention.head_count u32 = 8 +0.00.248.349 I llama_model_loader: - kv 12: spark2_5.attention.head_count_kv u32 = 2 +0.00.248.351 I llama_model_loader: - kv 13: spark2_5.rope.freq_base f32 = 5000000.000000 +0.00.248.357 I llama_model_loader: - kv 14: spark2_5.rope.freq_base_swa f32 = 10000.000000 +0.00.248.359 I llama_model_loader: - kv 15: spark2_5.attention.layer_norm_rms_epsilon f32 = 0.000001 +0.00.248.359 I llama_model_loader: - kv 16: spark2_5.attention.key_length u32 = 256 +0.00.248.360 I llama_model_loader: - kv 17: spark2_5.attention.value_length u32 = 256 +0.00.248.360 I llama_model_loader: - kv 18: spark2_5.vocab_size u32 = 131072 +0.00.248.361 I llama_model_loader: - kv 19: spark2_5.attention.sliding_window u32 = 512 +0.00.248.383 I llama_model_loader: - kv 20: spark2_5.attention.sliding_window_pattern arr[bool,28] = [true, true, true, false, true, true,... +0.00.248.384 I llama_model_loader: - kv 21: spark2_5.rope.dimension_count u32 = 64 +0.00.248.384 I llama_model_loader: - kv 22: spark2_5.rope.dimension_count_swa u32 = 256 +0.00.248.385 I llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 +0.00.248.385 I llama_model_loader: - kv 24: tokenizer.ggml.pre str = spark2_5 +0.00.264.353 I llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,131072] = ["<｜start▁of▁sentence｜>", "<... +0.00.269.486 I llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,131072] = [3, 3, 3, 4, 4, 3, 3, 3, 4, 4, 4, 3, ... +0.00.286.102 I llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,130716] = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "Ġ... +0.00.286.111 I llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 0 +0.00.286.112 I llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 1 +0.00.286.113 I llama_model_loader: - kv 30: tokenizer.ggml.unknown_token_id u32 = 5 +0.00.286.114 I llama_model_loader: - kv 31: tokenizer.ggml.padding_token_id u32 = 2 +0.00.286.115 I llama_model_loader: - kv 32: tokenizer.ggml.add_bos_token bool = false +0.00.286.115 I llama_model_loader: - kv 33: tokenizer.ggml.add_eos_token bool = false +0.00.286.120 I llama_model_loader: - kv 34: tokenizer.chat_template str = {%- if not messages %}\n {{- raise_... +0.00.286.120 I llama_model_loader: - kv 35: general.quantization_version u32 = 2 +0.00.286.121 I llama_model_loader: - kv 36: general.file_type u32 = 15 +0.00.286.130 I llama_model_loader: - type f32: 57 tensors +0.00.286.131 I llama_model_loader: - type q4_K: 140 tensors +0.00.286.132 I llama_model_loader: - type q6_K: 29 tensors +0.00.286.137 I print_info: file format = GGUF V3 (latest) +0.00.286.137 I print_info: file type = Q4_K - Medium +0.00.286.142 I print_info: file size = 1.03 GiB (5.16 BPW) +0.00.341.608 I load: 86 unused tokens +0.00.358.878 I load: printing all EOG tokens: +0.00.358.886 I load: - 1 ('<｜end▁of▁sentence｜>') +0.00.359.016 I load: special tokens cache size = 113 +0.00.375.025 I load: token to piece cache size = 0.7975 MB +0.00.375.056 I print_info: arch = spark2_5 +0.00.375.057 I print_info: vocab_only = 0 +0.00.375.057 I print_info: no_alloc = 0 +0.00.375.058 I print_info: n_ctx_train = 1048576 +0.00.375.058 I print_info: n_embd_inp = 2048 +0.00.375.059 I print_info: n_embd = 2048 +0.00.375.059 I print_info: n_embd_out = 2048 +0.00.375.060 I print_info: n_layer = 28 +0.00.375.060 I print_info: n_layer_all = 28 +0.00.375.076 I print_info: n_head = 8 +0.00.375.078 I print_info: n_head_kv = 2 +0.00.375.078 I print_info: n_rot = 64 +0.00.375.078 I print_info: n_swa = 512 +0.00.375.079 I print_info: is_swa_any = 1 +0.00.375.079 I print_info: non_causal_type = 0 +0.00.375.079 I print_info: n_embd_head_k = 256 +0.00.375.080 I print_info: n_embd_head_v = 256 +0.00.375.081 I print_info: n_gqa = 4 +0.00.375.082 I print_info: n_embd_k_gqa = 512 +0.00.375.084 I print_info: n_embd_v_gqa = 512 +0.00.375.087 I print_info: f_norm_eps = 0.0e+00 +0.00.375.088 I print_info: f_norm_rms_eps = 1.0e-06 +0.00.375.088 I print_info: f_clamp_kqv = 0.0e+00 +0.00.375.089 I print_info: f_max_alibi_bias = 0.0e+00 +0.00.375.089 I print_info: f_logit_scale = 0.0e+00 +0.00.375.089 I print_info: f_attn_scale = 0.0e+00 +0.00.375.090 I print_info: f_attn_value_scale = 0.0000 +0.00.375.091 I print_info: n_ff = 6656 +0.00.375.092 I print_info: n_expert = 0 +0.00.375.092 I print_info: n_expert_used = 0 +0.00.375.094 I print_info: n_expert_groups = 0 +0.00.375.094 I print_info: n_group_used = 0 +0.00.375.094 I print_info: causal attn = 1 +0.00.375.095 I print_info: pooling type = -1 +0.00.375.095 I print_info: rope type = 2 +0.00.375.095 I print_info: rope scaling = linear +0.00.375.097 I print_info: freq_base_train = 5000000.0 +0.00.375.098 I print_info: freq_scale_train = 1 +0.00.375.099 I print_info: freq_base_swa = 10000.0 +0.00.375.099 I print_info: freq_scale_swa = 1 +0.00.375.100 I print_info: n_embd_head_k_swa = 256 +0.00.375.100 I print_info: n_embd_head_v_swa = 256 +0.00.375.100 I print_info: n_rot_swa = 256 +0.00.375.100 I print_info: n_ctx_orig_yarn = 1048576 +0.00.375.101 I print_info: rope_yarn_log_mul = 0.0000 +0.00.375.101 I print_info: rope_finetuned = unknown +0.00.375.102 I print_info: model type = 1.7B +0.00.375.103 I print_info: model params = 1.71 B +0.00.375.103 I print_info: general.name = Hf_Format +0.00.375.104 I print_info: vocab type = BPE +0.00.375.104 I print_info: n_vocab = 131072 +0.00.375.105 I print_info: n_merges = 130716 +0.00.375.106 I print_info: BOS token = 0 '<｜start▁of▁sentence｜>' +0.00.375.106 I print_info: EOS token = 1 '<｜end▁of▁sentence｜>' +0.00.375.107 I print_info: EOT token = 1 '<｜end▁of▁sentence｜>' +0.00.375.107 I print_info: UNK token = 5 '' +0.00.375.108 I print_info: PAD token = 2 '<｜▁pad▁｜>' +0.00.375.108 I print_info: LF token = 198 'Ċ' +0.00.375.109 I print_info: FIM PRE token = 11 '<｜fim▁begin｜>' +0.00.375.109 I print_info: FIM SUF token = 12 '<｜fim▁hole｜>' +0.00.375.110 I print_info: FIM MID token = 13 '<｜fim▁end｜>' +0.00.375.110 I print_info: EOG token = 1 '<｜end▁of▁sentence｜>' +0.00.375.110 I print_info: max token length = 210 +0.00.375.112 I load_tensors: loading model tensors, this can take a while... (load_mode = mmap) +0.00.422.680 I load_tensors: offloading 0 repeating layers to GPU +0.00.422.682 I load_tensors: offloaded 0/29 layers to GPU +0.00.422.686 I load_tensors: CPU_Mapped model buffer size = 1043.71 MiB +0.00.422.687 I load_tensors: CPU_REPACK model buffer size = 622.37 MiB +0.00.777.403 I cmn common_init_: added <｜end▁of▁sentence｜> logit bias = -inf +0.00.777.812 I llama_context: constructing llama_context +0.00.777.817 I llama_context: n_seq_max = 1 +0.00.777.818 I llama_context: n_ctx = 4096 +0.00.777.818 I llama_context: n_ctx_seq = 4096 +0.00.777.818 I llama_context: n_batch = 2048 +0.00.777.819 I llama_context: n_ubatch = 512 +0.00.777.819 I llama_context: causal_attn = 1 +0.00.777.820 I llama_context: flash_attn = auto +0.00.777.820 I llama_context: kv_unified = false +0.00.777.823 I llama_context: freq_base = 5000000.0 +0.00.777.825 I llama_context: freq_scale = 1 +0.00.777.825 I llama_context: n_rs_seq = 0 +0.00.777.826 I llama_context: n_outputs_max = 1 +0.00.777.826 I llama_context: n_outputs_max_per_seq = 1 +0.00.777.827 I llama_context: n_ctx_seq (4096) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized +0.00.777.916 I llama_context: CPU output buffer size = 0.50 MiB +0.00.777.921 I llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells +0.00.778.026 I llama_kv_cache: CPU KV buffer size = 56.00 MiB +0.00.781.782 I llama_kv_cache: size = 56.00 MiB ( 4096 cells, 7 layers, 1/1 seqs), K (f16): 28.00 MiB, V (f16): 28.00 MiB +0.00.781.788 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.781.788 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.781.794 I llama_kv_cache_iswa: creating SWA KV cache, size = 1024 cells +0.00.781.884 I llama_kv_cache: CPU KV buffer size = 42.00 MiB +0.00.784.873 I llama_kv_cache: size = 42.00 MiB ( 1024 cells, 21 layers, 1/1 seqs), K (f16): 21.00 MiB, V (f16): 21.00 MiB +0.00.784.878 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.784.878 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.784.883 I sched_reserve: reserving ... +0.00.785.728 I resolve_fused_ops: Flash Attention enabled +0.00.785.729 I resolve_fused_ops: resolving fused DeepSeek V4 HC support: +0.00.786.017 I resolve_fused_ops: fused DeepSeek V4 HC pre enabled +0.00.786.285 I resolve_fused_ops: fused DeepSeek V4 HC comb enabled +0.00.786.540 I resolve_fused_ops: fused DeepSeek V4 HC post enabled +0.00.787.644 I sched_reserve: CPU compute buffer size = 60.02 MiB +0.00.787.646 I sched_reserve: graph: nodes = 986, splits = 1, input objects = 4, input tensors = 9 +0.00.787.647 I sched_reserve: reserve took 2.76 ms, sched copies = 1 +0.00.787.754 I cmn init: llama threadpool init, n_threads = 12 +0.00.787.767 I cmn common_init_: warming up the model with an empty run - please wait ... (--no-warmup to disable) +0.00.915.934 I srv load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false' +0.00.915.970 I spec common_specu: no implementations specified for speculative decoding +0.00.915.973 I slot load_model: id 0 | task -1 | new slot, n_ctx = 4096 +0.00.916.131 I srv load_model: prompt cache is disabled - use `--cache-ram N` to enable it +0.00.916.132 I srv load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391 +0.00.916.132 I srv load_model: context checkpoints enabled, max = 32, min spacing = 8192 +0.00.916.168 W srv init: --cache-idle-slots requires --cache-ram, disabling +0.00.932.716 I srv init: init: chat template, example_format: '<｜start▁of▁sentence｜><|System|> +you are a helpful assistant. + +You are a helpful assistant<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>Hello<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>Hi there<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>How are you?<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>' +0.00.944.622 I srv init: init: chat template, thinking = 0 +0.00.944.642 I srv init: preserve_reasoning kwarg: true +0.00.944.643 W srv init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve) +0.00.944.691 I srv llama_server: model loaded +0.00.944.694 I srv llama_server: listening on http://127.0.0.1:57355 +0.00.944.810 I srv update_slots: all slots are idle +0.01.182.565 I srv server_strea: conv_id= (empty=1) +0.01.182.747 I srv operator(): chat format: peg-native +0.01.182.884 I slot get_availabl: id 0 | task -1 | - skipping, slot is empty +0.01.182.888 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = -1 +0.01.182.954 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.182.984 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.182.987 I slot launch_slot_: id 0 | task 0 | processing task, is_child = 0 +0.01.183.005 I slot operator(): id 0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 28 +0.01.183.015 I slot operator(): id 0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.221.294 I slot operator(): id 0 | task 0 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.221.526 I slot create_check: id 0 | task 0 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.01.273.703 I slot operator(): id 0 | task 0 | cached n_tokens = 24, memory_seq_rm [24, end) +0.01.273.759 I slot init_sampler: id 0 | task 0 | init sampler, took 0.01 ms, tokens: text = 28, total = 28 +0.01.274.001 I slot create_check: id 0 | task 0 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 23, n_tokens = 24, size = 0.985 MiB) +0.01.321.379 I slot print_timing: id 0 | task 0 | prompt eval time = 117.97 ms / 28 tokens ( 4.21 ms per token, 237.34 tokens per second) +0.01.321.394 I slot print_timing: id 0 | task 0 | eval time = 20.36 ms / 2 tokens ( 20.36 ms per token, 49.10 tokens per second) +0.01.321.396 I slot print_timing: id 0 | task 0 | total time = 138.34 ms / 30 tokens +0.01.321.397 I slot print_timing: id 0 | task 0 | graphs reused = 1 +0.01.321.440 I slot release: id 0 | task 0 | stop processing: n_tokens = 29, truncated = 0 +0.01.321.452 I srv update_slots: all slots are idle +0.01.341.088 I srv server_strea: conv_id= (empty=1) +0.01.343.098 I srv operator(): chat format: peg-native +0.01.343.232 I slot get_availabl: id 0 | task -1 | - checking sim = 0.005 (12/2463) > 0.100 +0.01.343.233 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = 1115888 +0.01.343.449 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.343.459 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.343.460 I slot launch_slot_: id 0 | task 5 | processing task, is_child = 0 +0.01.343.467 I slot operator(): id 0 | task 5 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.01.343.469 I slot operator(): id 0 | task 5 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.01.343.472 I slot operator(): id 0 | task 5 | erased invalidated context checkpoint (pos_min = 0, pos_max = 23, n_tokens = 24, n_swa = 512, pos_next = 0, size = 0.985 MiB) +0.01.343.473 I slot operator(): id 0 | task 5 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.378.816 I slot operator(): id 0 | task 5 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.378.965 I slot create_check: id 0 | task 5 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.06.494.619 I slot print_timing: id 0 | task 5 | prompt processing, n_tokens = 1947, progress = 0.79, t = 5.15 s / 377.98 tokens per second +0.06.494.625 I slot operator(): id 0 | task 5 | cached n_tokens = 1947, memory_seq_rm [1947, end) +0.06.498.014 I slot create_check: id 0 | task 5 | created context checkpoint 2 of 32 (pos_min = 923, pos_max = 1946, n_tokens = 1947, size = 21.006 MiB) +0.07.952.543 I slot print_timing: id 0 | task 5 | prompt processing, n_tokens = 2459, progress = 1.00, t = 6.61 s / 372.07 tokens per second +0.07.952.551 I slot operator(): id 0 | task 5 | cached n_tokens = 2459, memory_seq_rm [2459, end) +0.07.952.838 I slot init_sampler: id 0 | task 5 | init sampler, took 0.26 ms, tokens: text = 2463, total = 2463 +0.07.956.714 I slot create_check: id 0 | task 5 | created context checkpoint 3 of 32 (pos_min = 1435, pos_max = 2458, n_tokens = 2459, size = 21.006 MiB) +0.08.080.808 I slot print_timing: id 0 | task 5 | prompt eval time = 6647.30 ms / 2463 tokens ( 2.70 ms per token, 370.53 tokens per second) +0.08.080.824 I slot print_timing: id 0 | task 5 | eval time = 90.02 ms / 5 tokens ( 22.50 ms per token, 44.44 tokens per second) +0.08.080.826 I slot print_timing: id 0 | task 5 | total time = 6737.32 ms / 2468 tokens +0.08.080.827 I slot print_timing: id 0 | task 5 | graphs reused = 4 +0.08.080.886 I slot release: id 0 | task 5 | stop processing: n_tokens = 2467, truncated = 0 +0.08.080.907 I srv update_slots: all slots are idle +0.08.127.645 I srv server_strea: conv_id= (empty=1) +0.08.130.227 I srv operator(): chat format: peg-native +0.08.130.470 I slot get_availabl: id 0 | task -1 | - checking sim = 1.000 (2463/2463) > 0.100 +0.08.130.473 I slot get_availabl: id 0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998 +0.08.130.726 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.08.130.741 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.08.130.744 I slot launch_slot_: id 0 | task 14 | processing task, is_child = 0 +0.08.130.755 I slot operator(): id 0 | task 14 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.08.130.758 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.08.130.761 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 923, pos_max = 1946, n_tokens = 1947, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.08.131.641 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 1435, pos_max = 2458, n_tokens = 2459, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.08.132.467 I slot operator(): id 0 | task 14 | cached n_tokens = 0, memory_seq_rm [0, end) +0.08.173.058 I slot operator(): id 0 | task 14 | cached n_tokens = 10, memory_seq_rm [10, end) +0.08.173.189 I slot create_check: id 0 | task 14 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.13.531.539 I slot print_timing: id 0 | task 14 | prompt processing, n_tokens = 1947, progress = 0.79, t = 5.40 s / 360.51 tokens per second +0.13.531.547 I slot operator(): id 0 | task 14 | cached n_tokens = 1947, memory_seq_rm [1947, end) +0.13.535.776 I slot create_check: id 0 | task 14 | created context checkpoint 2 of 32 (pos_min = 923, pos_max = 1946, n_tokens = 1947, size = 21.006 MiB) +0.15.017.275 I slot print_timing: id 0 | task 14 | prompt processing, n_tokens = 2459, progress = 1.00, t = 6.89 s / 357.08 tokens per second +0.15.017.286 I slot operator(): id 0 | task 14 | cached n_tokens = 2459, memory_seq_rm [2459, end) +0.15.017.649 I slot init_sampler: id 0 | task 14 | init sampler, took 0.34 ms, tokens: text = 2463, total = 2463 +0.15.021.913 I slot create_check: id 0 | task 14 | created context checkpoint 3 of 32 (pos_min = 1435, pos_max = 2458, n_tokens = 2459, size = 21.006 MiB) +0.15.155.952 I slot print_timing: id 0 | task 14 | prompt eval time = 6925.35 ms / 2463 tokens ( 2.81 ms per token, 355.65 tokens per second) +0.15.155.962 I slot print_timing: id 0 | task 14 | eval time = 99.82 ms / 5 tokens ( 24.96 ms per token, 40.07 tokens per second) +0.15.155.964 I slot print_timing: id 0 | task 14 | total time = 7025.18 ms / 2468 tokens +0.15.155.965 I slot print_timing: id 0 | task 14 | graphs reused = 7 +0.15.156.037 I slot release: id 0 | task 14 | stop processing: n_tokens = 2467, truncated = 0 +0.15.156.057 I srv update_slots: all slots are idle +0.15.252.941 I srv server_strea: conv_id= (empty=1) +0.15.255.963 I srv operator(): chat format: peg-native +0.15.256.177 I slot get_availabl: id 0 | task -1 | - checking sim = 1.000 (2463/2463) > 0.100 +0.15.256.179 I slot get_availabl: id 0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998 +0.15.256.397 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.15.256.407 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.15.256.410 I slot launch_slot_: id 0 | task 23 | processing task, is_child = 0 +0.15.256.419 I slot operator(): id 0 | task 23 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.15.256.422 I slot operator(): id 0 | task 23 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.15.256.425 I slot operator(): id 0 | task 23 | erased invalidated context checkpoint (pos_min = 923, pos_max = 1946, n_tokens = 1947, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.15.257.136 I slot operator(): id 0 | task 23 | erased invalidated context checkpoint (pos_min = 1435, pos_max = 2458, n_tokens = 2459, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.15.257.828 I slot operator(): id 0 | task 23 | cached n_tokens = 0, memory_seq_rm [0, end) diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v4_server_restart.log b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v4_server_restart.log new file mode 100644 index 000000000..8376b0e00 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/cpu_v4_server_restart.log @@ -0,0 +1,207 @@ +0.00.001.212 I srv llama_server: initializing ... +0.00.208.106 I cmn common_param: common_params_print_info: build 11124 (1a679828f) with Clang 20.1.8 for Windows x86_64 +0.00.208.113 I cmn common_param: common_params_print_info: verbosity = 4 (adjust with the `-lv N` CLI arg) +0.00.208.114 I cmn common_param: device_info: +0.00.212.753 I cmn common_param: - Vulkan0 : NVIDIA GeForce RTX 5090 Laptop GPU (24137 MiB, 23369 MiB free) +0.00.216.704 I cmn common_param: - Vulkan1 : AMD Radeon(TM) 890M Graphics (48505 MiB, 46080 MiB free) +0.00.216.714 I cmn common_param: - CPU : AMD Ryzen AI 9 HX 370 w/ Radeon 890M (64633 MiB, 26140 MiB free) +0.00.216.778 I cmn common_param: system_info: n_threads = 12 (n_threads_batch = 12) / 24 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | +0.00.216.858 I srv init_listene: running without SSL +0.00.216.916 I srv init_listene: The UI is disabled +0.00.216.917 I srv init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable +0.00.216.921 I srv init: using 23 threads for HTTP server +0.00.216.992 W srv llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655) +0.00.220.516 I srv load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.220.517 I srv load_model: local path 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.244.323 I llama_model_loader: loaded meta data with 37 key-value pairs and 226 tensors from C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf (version GGUF V3 (latest)) +0.00.244.335 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. +0.00.244.341 I llama_model_loader: - kv 0: general.architecture str = spark2_5 +0.00.244.342 I llama_model_loader: - kv 1: general.type str = model +0.00.244.346 I llama_model_loader: - kv 2: general.sampling.top_k i32 = 0 +0.00.244.353 I llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000 +0.00.244.354 I llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000 +0.00.244.354 I llama_model_loader: - kv 5: general.name str = Hf_Format +0.00.244.355 I llama_model_loader: - kv 6: general.size_label str = 1.7B +0.00.244.358 I llama_model_loader: - kv 7: spark2_5.block_count u32 = 28 +0.00.244.358 I llama_model_loader: - kv 8: spark2_5.context_length u32 = 1048576 +0.00.244.359 I llama_model_loader: - kv 9: spark2_5.embedding_length u32 = 2048 +0.00.244.359 I llama_model_loader: - kv 10: spark2_5.feed_forward_length u32 = 6656 +0.00.244.360 I llama_model_loader: - kv 11: spark2_5.attention.head_count u32 = 8 +0.00.244.360 I llama_model_loader: - kv 12: spark2_5.attention.head_count_kv u32 = 2 +0.00.244.361 I llama_model_loader: - kv 13: spark2_5.rope.freq_base f32 = 5000000.000000 +0.00.244.362 I llama_model_loader: - kv 14: spark2_5.rope.freq_base_swa f32 = 10000.000000 +0.00.244.364 I llama_model_loader: - kv 15: spark2_5.attention.layer_norm_rms_epsilon f32 = 0.000001 +0.00.244.364 I llama_model_loader: - kv 16: spark2_5.attention.key_length u32 = 256 +0.00.244.365 I llama_model_loader: - kv 17: spark2_5.attention.value_length u32 = 256 +0.00.244.365 I llama_model_loader: - kv 18: spark2_5.vocab_size u32 = 131072 +0.00.244.366 I llama_model_loader: - kv 19: spark2_5.attention.sliding_window u32 = 512 +0.00.244.386 I llama_model_loader: - kv 20: spark2_5.attention.sliding_window_pattern arr[bool,28] = [true, true, true, false, true, true,... +0.00.244.386 I llama_model_loader: - kv 21: spark2_5.rope.dimension_count u32 = 64 +0.00.244.387 I llama_model_loader: - kv 22: spark2_5.rope.dimension_count_swa u32 = 256 +0.00.244.387 I llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 +0.00.244.387 I llama_model_loader: - kv 24: tokenizer.ggml.pre str = spark2_5 +0.00.261.437 I llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,131072] = ["<｜start▁of▁sentence｜>", "<... +0.00.266.746 I llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,131072] = [3, 3, 3, 4, 4, 3, 3, 3, 4, 4, 4, 3, ... +0.00.284.473 I llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,130716] = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "Ġ... +0.00.284.496 I llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 0 +0.00.284.499 I llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 1 +0.00.284.501 I llama_model_loader: - kv 30: tokenizer.ggml.unknown_token_id u32 = 5 +0.00.284.502 I llama_model_loader: - kv 31: tokenizer.ggml.padding_token_id u32 = 2 +0.00.284.503 I llama_model_loader: - kv 32: tokenizer.ggml.add_bos_token bool = false +0.00.284.504 I llama_model_loader: - kv 33: tokenizer.ggml.add_eos_token bool = false +0.00.284.513 I llama_model_loader: - kv 34: tokenizer.chat_template str = {%- if not messages %}\n {{- raise_... +0.00.284.514 I llama_model_loader: - kv 35: general.quantization_version u32 = 2 +0.00.284.515 I llama_model_loader: - kv 36: general.file_type u32 = 15 +0.00.284.526 I llama_model_loader: - type f32: 57 tensors +0.00.284.527 I llama_model_loader: - type q4_K: 140 tensors +0.00.284.528 I llama_model_loader: - type q6_K: 29 tensors +0.00.284.555 I print_info: file format = GGUF V3 (latest) +0.00.284.556 I print_info: file type = Q4_K - Medium +0.00.284.563 I print_info: file size = 1.03 GiB (5.16 BPW) +0.00.339.620 I load: 86 unused tokens +0.00.359.220 I load: printing all EOG tokens: +0.00.359.229 I load: - 1 ('<｜end▁of▁sentence｜>') +0.00.359.382 I load: special tokens cache size = 113 +0.00.375.450 I load: token to piece cache size = 0.7975 MB +0.00.375.479 I print_info: arch = spark2_5 +0.00.375.480 I print_info: vocab_only = 0 +0.00.375.481 I print_info: no_alloc = 0 +0.00.375.481 I print_info: n_ctx_train = 1048576 +0.00.375.482 I print_info: n_embd_inp = 2048 +0.00.375.482 I print_info: n_embd = 2048 +0.00.375.482 I print_info: n_embd_out = 2048 +0.00.375.483 I print_info: n_layer = 28 +0.00.375.483 I print_info: n_layer_all = 28 +0.00.375.501 I print_info: n_head = 8 +0.00.375.502 I print_info: n_head_kv = 2 +0.00.375.503 I print_info: n_rot = 64 +0.00.375.503 I print_info: n_swa = 512 +0.00.375.503 I print_info: is_swa_any = 1 +0.00.375.504 I print_info: non_causal_type = 0 +0.00.375.504 I print_info: n_embd_head_k = 256 +0.00.375.505 I print_info: n_embd_head_v = 256 +0.00.375.506 I print_info: n_gqa = 4 +0.00.375.508 I print_info: n_embd_k_gqa = 512 +0.00.375.509 I print_info: n_embd_v_gqa = 512 +0.00.375.512 I print_info: f_norm_eps = 0.0e+00 +0.00.375.513 I print_info: f_norm_rms_eps = 1.0e-06 +0.00.375.514 I print_info: f_clamp_kqv = 0.0e+00 +0.00.375.514 I print_info: f_max_alibi_bias = 0.0e+00 +0.00.375.515 I print_info: f_logit_scale = 0.0e+00 +0.00.375.515 I print_info: f_attn_scale = 0.0e+00 +0.00.375.515 I print_info: f_attn_value_scale = 0.0000 +0.00.375.517 I print_info: n_ff = 6656 +0.00.375.517 I print_info: n_expert = 0 +0.00.375.517 I print_info: n_expert_used = 0 +0.00.375.519 I print_info: n_expert_groups = 0 +0.00.375.520 I print_info: n_group_used = 0 +0.00.375.520 I print_info: causal attn = 1 +0.00.375.520 I print_info: pooling type = -1 +0.00.375.521 I print_info: rope type = 2 +0.00.375.521 I print_info: rope scaling = linear +0.00.375.522 I print_info: freq_base_train = 5000000.0 +0.00.375.524 I print_info: freq_scale_train = 1 +0.00.375.524 I print_info: freq_base_swa = 10000.0 +0.00.375.525 I print_info: freq_scale_swa = 1 +0.00.375.525 I print_info: n_embd_head_k_swa = 256 +0.00.375.525 I print_info: n_embd_head_v_swa = 256 +0.00.375.526 I print_info: n_rot_swa = 256 +0.00.375.526 I print_info: n_ctx_orig_yarn = 1048576 +0.00.375.526 I print_info: rope_yarn_log_mul = 0.0000 +0.00.375.527 I print_info: rope_finetuned = unknown +0.00.375.528 I print_info: model type = 1.7B +0.00.375.528 I print_info: model params = 1.71 B +0.00.375.529 I print_info: general.name = Hf_Format +0.00.375.530 I print_info: vocab type = BPE +0.00.375.530 I print_info: n_vocab = 131072 +0.00.375.530 I print_info: n_merges = 130716 +0.00.375.531 I print_info: BOS token = 0 '<｜start▁of▁sentence｜>' +0.00.375.532 I print_info: EOS token = 1 '<｜end▁of▁sentence｜>' +0.00.375.532 I print_info: EOT token = 1 '<｜end▁of▁sentence｜>' +0.00.375.533 I print_info: UNK token = 5 '' +0.00.375.533 I print_info: PAD token = 2 '<｜▁pad▁｜>' +0.00.375.533 I print_info: LF token = 198 'Ċ' +0.00.375.534 I print_info: FIM PRE token = 11 '<｜fim▁begin｜>' +0.00.375.535 I print_info: FIM SUF token = 12 '<｜fim▁hole｜>' +0.00.375.535 I print_info: FIM MID token = 13 '<｜fim▁end｜>' +0.00.375.535 I print_info: EOG token = 1 '<｜end▁of▁sentence｜>' +0.00.375.536 I print_info: max token length = 210 +0.00.375.538 I load_tensors: loading model tensors, this can take a while... (load_mode = mmap) +0.00.424.840 I load_tensors: offloading 0 repeating layers to GPU +0.00.424.842 I load_tensors: offloaded 0/29 layers to GPU +0.00.424.846 I load_tensors: CPU_Mapped model buffer size = 1043.71 MiB +0.00.424.847 I load_tensors: CPU_REPACK model buffer size = 622.37 MiB +0.00.786.917 I cmn common_init_: added <｜end▁of▁sentence｜> logit bias = -inf +0.00.787.344 I llama_context: constructing llama_context +0.00.787.349 I llama_context: n_seq_max = 1 +0.00.787.349 I llama_context: n_ctx = 4096 +0.00.787.349 I llama_context: n_ctx_seq = 4096 +0.00.787.350 I llama_context: n_batch = 2048 +0.00.787.350 I llama_context: n_ubatch = 512 +0.00.787.351 I llama_context: causal_attn = 1 +0.00.787.352 I llama_context: flash_attn = auto +0.00.787.352 I llama_context: kv_unified = false +0.00.787.355 I llama_context: freq_base = 5000000.0 +0.00.787.357 I llama_context: freq_scale = 1 +0.00.787.358 I llama_context: n_rs_seq = 0 +0.00.787.358 I llama_context: n_outputs_max = 1 +0.00.787.358 I llama_context: n_outputs_max_per_seq = 1 +0.00.787.359 I llama_context: n_ctx_seq (4096) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized +0.00.787.478 I llama_context: CPU output buffer size = 0.50 MiB +0.00.787.483 I llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells +0.00.787.626 I llama_kv_cache: CPU KV buffer size = 56.00 MiB +0.00.791.296 I llama_kv_cache: size = 56.00 MiB ( 4096 cells, 7 layers, 1/1 seqs), K (f16): 28.00 MiB, V (f16): 28.00 MiB +0.00.791.299 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.791.300 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.791.304 I llama_kv_cache_iswa: creating SWA KV cache, size = 1024 cells +0.00.791.340 I llama_kv_cache: CPU KV buffer size = 42.00 MiB +0.00.794.057 I llama_kv_cache: size = 42.00 MiB ( 1024 cells, 21 layers, 1/1 seqs), K (f16): 21.00 MiB, V (f16): 21.00 MiB +0.00.794.059 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.794.060 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.794.064 I sched_reserve: reserving ... +0.00.794.863 I resolve_fused_ops: Flash Attention enabled +0.00.794.864 I resolve_fused_ops: resolving fused DeepSeek V4 HC support: +0.00.795.139 I resolve_fused_ops: fused DeepSeek V4 HC pre enabled +0.00.795.399 I resolve_fused_ops: fused DeepSeek V4 HC comb enabled +0.00.795.655 I resolve_fused_ops: fused DeepSeek V4 HC post enabled +0.00.796.677 I sched_reserve: CPU compute buffer size = 60.02 MiB +0.00.796.678 I sched_reserve: graph: nodes = 986, splits = 1, input objects = 4, input tensors = 9 +0.00.796.679 I sched_reserve: reserve took 2.61 ms, sched copies = 1 +0.00.796.760 I cmn init: llama threadpool init, n_threads = 12 +0.00.796.772 I cmn common_init_: warming up the model with an empty run - please wait ... (--no-warmup to disable) +0.00.921.184 I srv load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false' +0.00.921.230 I spec common_specu: no implementations specified for speculative decoding +0.00.921.232 I slot load_model: id 0 | task -1 | new slot, n_ctx = 4096 +0.00.921.349 I srv load_model: prompt cache is disabled - use `--cache-ram N` to enable it +0.00.921.349 I srv load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391 +0.00.921.350 I srv load_model: context checkpoints enabled, max = 32, min spacing = 8192 +0.00.921.374 W srv init: --cache-idle-slots requires --cache-ram, disabling +0.00.937.363 I srv init: init: chat template, example_format: '<｜start▁of▁sentence｜><|System|> +you are a helpful assistant. + +You are a helpful assistant<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>Hello<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>Hi there<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>How are you?<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>' +0.00.949.520 I srv init: init: chat template, thinking = 0 +0.00.949.536 I srv init: preserve_reasoning kwarg: true +0.00.949.537 W srv init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve) +0.00.949.576 I srv llama_server: model loaded +0.00.949.579 I srv llama_server: listening on http://127.0.0.1:61580 +0.00.949.685 I srv update_slots: all slots are idle +0.01.175.423 I srv server_strea: conv_id= (empty=1) +0.01.175.600 I srv operator(): chat format: peg-native +0.01.175.725 I slot get_availabl: id 0 | task -1 | - skipping, slot is empty +0.01.175.729 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = -1 +0.01.175.803 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.175.832 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.175.835 I slot launch_slot_: id 0 | task 0 | processing task, is_child = 0 +0.01.175.853 I slot operator(): id 0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 28 +0.01.175.863 I slot operator(): id 0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.212.793 I slot operator(): id 0 | task 0 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.213.000 I slot create_check: id 0 | task 0 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.01.260.176 I slot operator(): id 0 | task 0 | cached n_tokens = 24, memory_seq_rm [24, end) +0.01.260.230 I slot init_sampler: id 0 | task 0 | init sampler, took 0.01 ms, tokens: text = 28, total = 28 +0.01.260.490 I slot create_check: id 0 | task 0 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 23, n_tokens = 24, size = 0.985 MiB) +0.01.305.468 I slot print_timing: id 0 | task 0 | prompt eval time = 109.87 ms / 28 tokens ( 3.92 ms per token, 254.85 tokens per second) diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/hardware.json b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/hardware.json new file mode 100644 index 000000000..436f92058 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/hardware.json @@ -0,0 +1,10 @@ +{ + "TimestampUtc": "2026-09-24T07:17:39.2184813Z", + "OSCaption": "Microsoft Windows 11 家庭版 中文版", + "OSVersion": "10.0.26200", + "BuildNumber": "26200", + "CPU": "AMD Ryzen AI 9 HX 370 w/ Radeon 890M", + "GPU": "AMD Radeon(TM) 890M Graphics", + "GPUDriverVersion": "32.0.22018.6001", + "PowerScheme": "电源方案 GUID: 27fa6203-3987-4dcc-918d-748559d549ec (Performance)" +} diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_report.json b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_report.json new file mode 100644 index 000000000..17ab1f3cc --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_report.json @@ -0,0 +1,250 @@ +{ + "scope": "complete Spark GGUF text requests through Omni StageRuntime/StagePool and external llama.cpp backend", + "model_file": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "device": "Vulkan1", + "memory_budget": { + "capacities": { + "host_ram": 12884901888 + }, + "demands": { + "host_ram": 4294967296 + } + }, + "warmup_count": 0, + "measured_count": 1, + "startup_s": 3.6517130999127403, + "execution_plan": { + "backend": "external.llamacpp.text.v1", + "stage_id": 0, + "worker_generation": "134aca8adaf7404fb16427c7ac8be89d", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "requested_device": "Vulkan1", + "expected_device_name": "AMD Radeon(TM) 890M Graphics", + "layer_assignments": 0, + "offloaded_layers": [ + "29", + "29" + ], + "model_buffer_devices": [ + "Vulkan1" + ], + "all_model_buffers": [ + "Vulkan1", + "Vulkan_Host" + ], + "cpu_only_layout": false, + "worker_pid": 3604, + "loaded_rss_bytes": 1644249088, + "reserved_bytes": { + "host_ram": 4294967296 + }, + "memory_pool": "host_ram", + "memory_overhead_bytes": 2147483648, + "context_tokens": 4096, + "max_new_tokens": 96, + "max_io_bytes": 1048576, + "request_capacity": 1, + "stateful_session": "llama.cpp-owned; reset for each complete request", + "evidence": "B" + }, + "checks": [ + { + "case": "short", + "answer": "Paris", + "expected": "Paris", + "correct": true, + "wall_s": 0.15901159995701164, + "stage_event": { + "request_id": "short-688313250000000", + "stage_id": 0, + "epoch": 1, + "seq": 1, + "kind": "text", + "worker_generation": "134aca8adaf7404fb16427c7ac8be89d", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 0.15466610004659742, + "usage": { + "completion_tokens": 2, + "prompt_tokens": 28, + "total_tokens": 30, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + }, + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 2.2468628000933677, + "stage_event": { + "request_id": "inventory_120-688313406000000", + "stage_id": 0, + "epoch": 2, + "seq": 1, + "kind": "text", + "worker_generation": "134aca8adaf7404fb16427c7ac8be89d", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 2.218514999956824, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "warmups": [], + "measured": [ + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 2.229528099996969, + "stage_event": { + "request_id": "inventory_120-688315656000000", + "stage_id": 0, + "epoch": 3, + "seq": 1, + "kind": "text", + "worker_generation": "134aca8adaf7404fb16427c7ac8be89d", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 2.226126999943517, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "nearest_rank_p50_wall_s": 2.229528099996969, + "nearest_rank_p95_wall_s": 2.229528099996969, + "abort_check": { + "stale_output": false, + "worker_exited": true, + "ledger_after_abort": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + } + }, + "restart_check": { + "first_runtime_ledger_after_shutdown": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + }, + "fresh_startup_s": 1.918302800040692, + "fresh_execution_plan": { + "backend": "external.llamacpp.text.v1", + "stage_id": 0, + "worker_generation": "59dde2e3fbe74524912225d40849d4ac", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "requested_device": "Vulkan1", + "expected_device_name": "AMD Radeon(TM) 890M Graphics", + "layer_assignments": 0, + "offloaded_layers": [ + "29", + "29" + ], + "model_buffer_devices": [ + "Vulkan1" + ], + "all_model_buffers": [ + "Vulkan1", + "Vulkan_Host" + ], + "cpu_only_layout": false, + "worker_pid": 39192, + "loaded_rss_bytes": 1644154880, + "reserved_bytes": { + "host_ram": 4294967296 + }, + "memory_pool": "host_ram", + "memory_overhead_bytes": 2147483648, + "context_tokens": 4096, + "max_new_tokens": 96, + "max_io_bytes": 1048576, + "request_capacity": 1, + "stateful_session": "llama.cpp-owned; reset for each complete request", + "evidence": "B" + }, + "fresh_request": { + "case": "short", + "answer": "Paris", + "expected": "Paris", + "correct": true, + "wall_s": 0.17051299999002367, + "stage_event": { + "request_id": "short-688320109000000", + "stage_id": 0, + "epoch": 1, + "seq": 1, + "kind": "text", + "worker_generation": "59dde2e3fbe74524912225d40849d4ac", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 0.15348049998283386, + "usage": { + "completion_tokens": 2, + "prompt_tokens": 28, + "total_tokens": 30, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + }, + "late_output": false + }, + "status": "passed", + "ledger_after_shutdown": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + } +} diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_server.log b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_server.log new file mode 100644 index 000000000..e7bb8abf8 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_server.log @@ -0,0 +1,211 @@ +0.00.000.848 I srv llama_server: initializing ... +0.00.221.600 I cmn common_param: common_params_print_info: build 11124 (1a679828f) with Clang 20.1.8 for Windows x86_64 +0.00.221.608 I cmn common_param: common_params_print_info: verbosity = 4 (adjust with the `-lv N` CLI arg) +0.00.221.609 I cmn common_param: device_info: +0.00.226.371 I cmn common_param: - Vulkan0 : NVIDIA GeForce RTX 5090 Laptop GPU (24137 MiB, 23369 MiB free) +0.00.230.342 I cmn common_param: - Vulkan1 : AMD Radeon(TM) 890M Graphics (48505 MiB, 46080 MiB free) +0.00.230.353 I cmn common_param: - CPU : AMD Ryzen AI 9 HX 370 w/ Radeon 890M (64633 MiB, 25929 MiB free) +0.00.230.424 I cmn common_param: system_info: n_threads = 12 (n_threads_batch = 12) / 24 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | +0.00.230.509 I srv init_listene: running without SSL +0.00.230.577 I srv init_listene: The UI is disabled +0.00.230.578 I srv init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable +0.00.230.584 I srv init: using 23 threads for HTTP server +0.00.230.655 W srv llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655) +0.00.239.335 I srv load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.239.342 I srv load_model: local path 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.262.993 I llama_model_loader: loaded meta data with 37 key-value pairs and 226 tensors from C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf (version GGUF V3 (latest)) +0.00.263.006 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. +0.00.263.013 I llama_model_loader: - kv 0: general.architecture str = spark2_5 +0.00.263.014 I llama_model_loader: - kv 1: general.type str = model +0.00.263.018 I llama_model_loader: - kv 2: general.sampling.top_k i32 = 0 +0.00.263.025 I llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000 +0.00.263.026 I llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000 +0.00.263.027 I llama_model_loader: - kv 5: general.name str = Hf_Format +0.00.263.027 I llama_model_loader: - kv 6: general.size_label str = 1.7B +0.00.263.030 I llama_model_loader: - kv 7: spark2_5.block_count u32 = 28 +0.00.263.031 I llama_model_loader: - kv 8: spark2_5.context_length u32 = 1048576 +0.00.263.031 I llama_model_loader: - kv 9: spark2_5.embedding_length u32 = 2048 +0.00.263.032 I llama_model_loader: - kv 10: spark2_5.feed_forward_length u32 = 6656 +0.00.263.032 I llama_model_loader: - kv 11: spark2_5.attention.head_count u32 = 8 +0.00.263.033 I llama_model_loader: - kv 12: spark2_5.attention.head_count_kv u32 = 2 +0.00.263.034 I llama_model_loader: - kv 13: spark2_5.rope.freq_base f32 = 5000000.000000 +0.00.263.035 I llama_model_loader: - kv 14: spark2_5.rope.freq_base_swa f32 = 10000.000000 +0.00.263.037 I llama_model_loader: - kv 15: spark2_5.attention.layer_norm_rms_epsilon f32 = 0.000001 +0.00.263.038 I llama_model_loader: - kv 16: spark2_5.attention.key_length u32 = 256 +0.00.263.038 I llama_model_loader: - kv 17: spark2_5.attention.value_length u32 = 256 +0.00.263.038 I llama_model_loader: - kv 18: spark2_5.vocab_size u32 = 131072 +0.00.263.039 I llama_model_loader: - kv 19: spark2_5.attention.sliding_window u32 = 512 +0.00.263.060 I llama_model_loader: - kv 20: spark2_5.attention.sliding_window_pattern arr[bool,28] = [true, true, true, false, true, true,... +0.00.263.060 I llama_model_loader: - kv 21: spark2_5.rope.dimension_count u32 = 64 +0.00.263.061 I llama_model_loader: - kv 22: spark2_5.rope.dimension_count_swa u32 = 256 +0.00.263.061 I llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 +0.00.263.062 I llama_model_loader: - kv 24: tokenizer.ggml.pre str = spark2_5 +0.00.279.493 I llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,131072] = ["<｜start▁of▁sentence｜>", "<... +0.00.284.599 I llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,131072] = [3, 3, 3, 4, 4, 3, 3, 3, 4, 4, 4, 3, ... +0.00.300.759 I llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,130716] = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "Ġ... +0.00.300.769 I llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 0 +0.00.300.769 I llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 1 +0.00.300.770 I llama_model_loader: - kv 30: tokenizer.ggml.unknown_token_id u32 = 5 +0.00.300.771 I llama_model_loader: - kv 31: tokenizer.ggml.padding_token_id u32 = 2 +0.00.300.772 I llama_model_loader: - kv 32: tokenizer.ggml.add_bos_token bool = false +0.00.300.772 I llama_model_loader: - kv 33: tokenizer.ggml.add_eos_token bool = false +0.00.300.777 I llama_model_loader: - kv 34: tokenizer.chat_template str = {%- if not messages %}\n {{- raise_... +0.00.300.777 I llama_model_loader: - kv 35: general.quantization_version u32 = 2 +0.00.300.778 I llama_model_loader: - kv 36: general.file_type u32 = 15 +0.00.300.783 I llama_model_loader: - type f32: 57 tensors +0.00.300.784 I llama_model_loader: - type q4_K: 140 tensors +0.00.300.784 I llama_model_loader: - type q6_K: 29 tensors +0.00.300.789 I print_info: file format = GGUF V3 (latest) +0.00.300.789 I print_info: file type = Q4_K - Medium +0.00.300.795 I print_info: file size = 1.03 GiB (5.16 BPW) +0.00.307.410 I llama_prepare_model_devices: using device Vulkan1 (AMD Radeon(TM) 890M Graphics) (unknown id) - 46080 MiB free +0.00.361.717 I load: 86 unused tokens +0.00.388.375 I load: printing all EOG tokens: +0.00.388.384 I load: - 1 ('<｜end▁of▁sentence｜>') +0.00.388.528 I load: special tokens cache size = 113 +0.00.404.007 I load: token to piece cache size = 0.7975 MB +0.00.404.036 I print_info: arch = spark2_5 +0.00.404.036 I print_info: vocab_only = 0 +0.00.404.037 I print_info: no_alloc = 0 +0.00.404.038 I print_info: n_ctx_train = 1048576 +0.00.404.038 I print_info: n_embd_inp = 2048 +0.00.404.039 I print_info: n_embd = 2048 +0.00.404.039 I print_info: n_embd_out = 2048 +0.00.404.039 I print_info: n_layer = 28 +0.00.404.039 I print_info: n_layer_all = 28 +0.00.404.056 I print_info: n_head = 8 +0.00.404.058 I print_info: n_head_kv = 2 +0.00.404.058 I print_info: n_rot = 64 +0.00.404.059 I print_info: n_swa = 512 +0.00.404.059 I print_info: is_swa_any = 1 +0.00.404.059 I print_info: non_causal_type = 0 +0.00.404.059 I print_info: n_embd_head_k = 256 +0.00.404.060 I print_info: n_embd_head_v = 256 +0.00.404.061 I print_info: n_gqa = 4 +0.00.404.063 I print_info: n_embd_k_gqa = 512 +0.00.404.064 I print_info: n_embd_v_gqa = 512 +0.00.404.065 I print_info: f_norm_eps = 0.0e+00 +0.00.404.067 I print_info: f_norm_rms_eps = 1.0e-06 +0.00.404.067 I print_info: f_clamp_kqv = 0.0e+00 +0.00.404.067 I print_info: f_max_alibi_bias = 0.0e+00 +0.00.404.068 I print_info: f_logit_scale = 0.0e+00 +0.00.404.068 I print_info: f_attn_scale = 0.0e+00 +0.00.404.068 I print_info: f_attn_value_scale = 0.0000 +0.00.404.070 I print_info: n_ff = 6656 +0.00.404.070 I print_info: n_expert = 0 +0.00.404.071 I print_info: n_expert_used = 0 +0.00.404.073 I print_info: n_expert_groups = 0 +0.00.404.073 I print_info: n_group_used = 0 +0.00.404.073 I print_info: causal attn = 1 +0.00.404.074 I print_info: pooling type = -1 +0.00.404.074 I print_info: rope type = 2 +0.00.404.074 I print_info: rope scaling = linear +0.00.404.075 I print_info: freq_base_train = 5000000.0 +0.00.404.077 I print_info: freq_scale_train = 1 +0.00.404.077 I print_info: freq_base_swa = 10000.0 +0.00.404.078 I print_info: freq_scale_swa = 1 +0.00.404.078 I print_info: n_embd_head_k_swa = 256 +0.00.404.078 I print_info: n_embd_head_v_swa = 256 +0.00.404.079 I print_info: n_rot_swa = 256 +0.00.404.079 I print_info: n_ctx_orig_yarn = 1048576 +0.00.404.079 I print_info: rope_yarn_log_mul = 0.0000 +0.00.404.080 I print_info: rope_finetuned = unknown +0.00.404.081 I print_info: model type = 1.7B +0.00.404.081 I print_info: model params = 1.71 B +0.00.404.082 I print_info: general.name = Hf_Format +0.00.404.083 I print_info: vocab type = BPE +0.00.404.083 I print_info: n_vocab = 131072 +0.00.404.083 I print_info: n_merges = 130716 +0.00.404.084 I print_info: BOS token = 0 '<｜start▁of▁sentence｜>' +0.00.404.085 I print_info: EOS token = 1 '<｜end▁of▁sentence｜>' +0.00.404.085 I print_info: EOT token = 1 '<｜end▁of▁sentence｜>' +0.00.404.085 I print_info: UNK token = 5 '' +0.00.404.086 I print_info: PAD token = 2 '<｜▁pad▁｜>' +0.00.404.086 I print_info: LF token = 198 'Ċ' +0.00.404.087 I print_info: FIM PRE token = 11 '<｜fim▁begin｜>' +0.00.404.087 I print_info: FIM SUF token = 12 '<｜fim▁hole｜>' +0.00.404.088 I print_info: FIM MID token = 13 '<｜fim▁end｜>' +0.00.404.088 I print_info: EOG token = 1 '<｜end▁of▁sentence｜>' +0.00.404.089 I print_info: max token length = 210 +0.00.412.965 I load_tensors: loading model tensors, this can take a while... (load_mode = none) +0.00.604.239 I load_tensors: offloading output layer to GPU +0.00.604.245 I load_tensors: offloading 27 repeating layers to GPU +0.00.604.246 I load_tensors: offloaded 29/29 layers to GPU +0.00.604.251 I load_tensors: Vulkan1 model buffer size = 1051.02 MiB +0.00.604.252 I load_tensors: Vulkan_Host model buffer size = 210.00 MiB +0.00.929.665 I cmn common_init_: added <｜end▁of▁sentence｜> logit bias = -inf +0.00.930.472 I llama_context: constructing llama_context +0.00.930.477 I llama_context: n_seq_max = 1 +0.00.930.477 I llama_context: n_ctx = 4096 +0.00.930.478 I llama_context: n_ctx_seq = 4096 +0.00.930.479 I llama_context: n_batch = 2048 +0.00.930.479 I llama_context: n_ubatch = 512 +0.00.930.479 I llama_context: causal_attn = 1 +0.00.930.480 I llama_context: flash_attn = auto +0.00.930.480 I llama_context: kv_unified = false +0.00.930.484 I llama_context: freq_base = 5000000.0 +0.00.930.486 I llama_context: freq_scale = 1 +0.00.930.487 I llama_context: n_rs_seq = 0 +0.00.930.487 I llama_context: n_outputs_max = 1 +0.00.930.487 I llama_context: n_outputs_max_per_seq = 1 +0.00.930.488 I llama_context: n_ctx_seq (4096) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized +0.00.930.816 I llama_context: Vulkan_Host output buffer size = 0.50 MiB +0.00.930.824 I llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells +0.00.935.985 I llama_kv_cache: Vulkan1 KV buffer size = 56.00 MiB +0.00.936.626 I llama_kv_cache: size = 56.00 MiB ( 4096 cells, 7 layers, 1/1 seqs), K (f16): 28.00 MiB, V (f16): 28.00 MiB +0.00.936.633 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.936.633 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.936.638 I llama_kv_cache_iswa: creating SWA KV cache, size = 1024 cells +0.00.940.395 I llama_kv_cache: Vulkan1 KV buffer size = 42.00 MiB +0.00.940.868 I llama_kv_cache: size = 42.00 MiB ( 1024 cells, 21 layers, 1/1 seqs), K (f16): 21.00 MiB, V (f16): 21.00 MiB +0.00.940.874 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.940.874 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.940.884 I sched_reserve: reserving ... +0.00.942.633 I resolve_fused_ops: Flash Attention enabled +0.00.942.634 I resolve_fused_ops: resolving fused DeepSeek V4 HC support: +0.00.944.138 I resolve_fused_ops: fused DeepSeek V4 HC pre enabled +0.00.945.074 I resolve_fused_ops: fused DeepSeek V4 HC comb enabled +0.00.945.993 I resolve_fused_ops: fused DeepSeek V4 HC post enabled +0.00.954.689 I sched_reserve: Vulkan1 compute buffer size = 56.02 MiB +0.00.954.691 I sched_reserve: Vulkan_Host compute buffer size = 13.02 MiB +0.00.954.692 I sched_reserve: graph: nodes = 986, splits = 2, input objects = 4, input tensors = 9 +0.00.954.694 I sched_reserve: reserve took 13.80 ms, sched copies = 1 +0.00.954.801 I cmn init: llama threadpool init, n_threads = 12 +0.00.954.815 I cmn common_init_: warming up the model with an empty run - please wait ... (--no-warmup to disable) +0.01.005.055 I srv load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false' +0.01.005.077 I spec common_specu: no implementations specified for speculative decoding +0.01.005.078 I slot load_model: id 0 | task -1 | new slot, n_ctx = 4096 +0.01.005.167 I srv load_model: prompt cache is disabled - use `--cache-ram N` to enable it +0.01.005.167 I srv load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391 +0.01.005.168 I srv load_model: context checkpoints enabled, max = 32, min spacing = 8192 +0.01.005.191 W srv init: --cache-idle-slots requires --cache-ram, disabling +0.01.015.875 I srv init: init: chat template, example_format: '<｜start▁of▁sentence｜><|System|> +you are a helpful assistant. + +You are a helpful assistant<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>Hello<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>Hi there<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>How are you?<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>' +0.01.023.563 I srv init: init: chat template, thinking = 0 +0.01.023.575 I srv init: preserve_reasoning kwarg: true +0.01.023.575 W srv init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve) +0.01.023.598 I srv llama_server: model loaded +0.01.023.601 I srv llama_server: listening on http://127.0.0.1:64909 +0.01.023.681 I srv update_slots: all slots are idle +0.01.185.796 I srv server_strea: conv_id= (empty=1) +0.01.186.015 I srv operator(): chat format: peg-native +0.01.186.127 I slot get_availabl: id 0 | task -1 | - skipping, slot is empty +0.01.186.129 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = -1 +0.01.186.165 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.186.180 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.186.181 I slot launch_slot_: id 0 | task 0 | processing task, is_child = 0 +0.01.186.192 I slot operator(): id 0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 28 +0.01.186.197 I slot operator(): id 0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.196.664 I slot operator(): id 0 | task 0 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.231.698 I slot create_check: id 0 | task 0 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.01.234.871 I slot operator(): id 0 | task 0 | cached n_tokens = 24, memory_seq_rm [24, end) +0.01.234.906 I slot init_sampler: id 0 | task 0 | init sampler, took 0.00 ms, tokens: text = 28, total = 28 +0.01.273.005 I slot create_check: id 0 | task 0 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 23, n_tokens = 24, size = 0.985 MiB) +0.01.314.039 I slot print_timing: id 0 | task 0 | prompt eval time = 110.00 ms / 28 tokens ( 3.93 ms per token, 254.54 tokens per second) +0.01.314.046 \ No newline at end of file diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v2_report.json b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v2_report.json new file mode 100644 index 000000000..88b9bec6c --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v2_report.json @@ -0,0 +1,251 @@ +{ + "scope": "complete Spark GGUF text requests through Omni StageRuntime/StagePool and external llama.cpp backend", + "model_file": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "device": "Vulkan1", + "memory_budget": { + "capacities": { + "host_ram": 12884901888 + }, + "demands": { + "host_ram": 4294967296 + } + }, + "warmup_count": 0, + "measured_count": 1, + "startup_s": 3.1089909001020715, + "execution_plan": { + "backend": "external.llamacpp.text.v1", + "stage_id": 0, + "worker_generation": "bad1ece82058464bb849c4d4d8233330", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "requested_device": "Vulkan1", + "expected_device_name": "AMD Radeon(TM) 890M Graphics", + "layer_assignments": 0, + "offloaded_layers": [ + "29", + "29" + ], + "model_buffer_devices": [ + "Vulkan1" + ], + "all_model_buffers": [ + "Vulkan1", + "Vulkan_Host" + ], + "cpu_only_layout": false, + "worker_pid": 50344, + "loaded_rss_bytes": 1644072960, + "reserved_bytes": { + "host_ram": 4294967296 + }, + "memory_pool": "host_ram", + "memory_overhead_bytes": 2147483648, + "context_tokens": 4096, + "max_new_tokens": 96, + "max_io_bytes": 1048576, + "request_capacity": 1, + "stateful_session": "llama.cpp-owned; reset for each complete request", + "evidence": "B" + }, + "checks": [ + { + "case": "short", + "answer": "Paris", + "expected": "Paris", + "correct": true, + "wall_s": 0.2012867999728769, + "stage_event": { + "request_id": "short-688393843000000", + "stage_id": 0, + "epoch": 1, + "seq": 1, + "kind": "text", + "worker_generation": "bad1ece82058464bb849c4d4d8233330", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 0.18573420005850494, + "usage": { + "completion_tokens": 2, + "prompt_tokens": 28, + "total_tokens": 30, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + }, + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 2.721963300020434, + "stage_event": { + "request_id": "inventory_120-688394031000000", + "stage_id": 0, + "epoch": 2, + "seq": 1, + "kind": "text", + "worker_generation": "bad1ece82058464bb849c4d4d8233330", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 2.7165678000310436, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "warmups": [], + "measured": [ + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 2.7230667000403628, + "stage_event": { + "request_id": "inventory_120-688396765000000", + "stage_id": 0, + "epoch": 3, + "seq": 1, + "kind": "text", + "worker_generation": "bad1ece82058464bb849c4d4d8233330", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 2.692699500010349, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "nearest_rank_p50_wall_s": 2.7230667000403628, + "nearest_rank_p95_wall_s": 2.7230667000403628, + "abort_check": { + "stale_output": false, + "worker_exited": true, + "ledger_after_abort": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + } + }, + "restart_check": { + "first_runtime_ledger_after_shutdown": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + }, + "fresh_startup_s": 2.520516899996437, + "fresh_server_log": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\vllm-omni-edge\\benchmarks\\edge_harness\\results\\e2e_expansion_20260924\\evidence\\spark_omni_windows_restart\\radeon_v2_server_restart.log", + "fresh_execution_plan": { + "backend": "external.llamacpp.text.v1", + "stage_id": 0, + "worker_generation": "c8387a3ba9f047c7b282dfea553a7668", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "requested_device": "Vulkan1", + "expected_device_name": "AMD Radeon(TM) 890M Graphics", + "layer_assignments": 0, + "offloaded_layers": [ + "29", + "29" + ], + "model_buffer_devices": [ + "Vulkan1" + ], + "all_model_buffers": [ + "Vulkan1", + "Vulkan_Host" + ], + "cpu_only_layout": false, + "worker_pid": 16564, + "loaded_rss_bytes": 1643921408, + "reserved_bytes": { + "host_ram": 4294967296 + }, + "memory_pool": "host_ram", + "memory_overhead_bytes": 2147483648, + "context_tokens": 4096, + "max_new_tokens": 96, + "max_io_bytes": 1048576, + "request_capacity": 1, + "stateful_session": "llama.cpp-owned; reset for each complete request", + "evidence": "B" + }, + "fresh_request": { + "case": "short", + "answer": "Paris", + "expected": "Paris", + "correct": true, + "wall_s": 0.19192340003792197, + "stage_event": { + "request_id": "short-688402375000000", + "stage_id": 0, + "epoch": 1, + "seq": 1, + "kind": "text", + "worker_generation": "c8387a3ba9f047c7b282dfea553a7668", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 0.17687359999399632, + "usage": { + "completion_tokens": 2, + "prompt_tokens": 28, + "total_tokens": 30, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + }, + "late_output": false + }, + "status": "passed", + "ledger_after_shutdown": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + } +} diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v2_server.log b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v2_server.log new file mode 100644 index 000000000..b0d4153e9 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v2_server.log @@ -0,0 +1,272 @@ +0.00.001.182 I srv llama_server: initializing ... +0.00.226.616 I cmn common_param: common_params_print_info: build 11124 (1a679828f) with Clang 20.1.8 for Windows x86_64 +0.00.226.623 I cmn common_param: common_params_print_info: verbosity = 4 (adjust with the `-lv N` CLI arg) +0.00.226.624 I cmn common_param: device_info: +0.00.230.847 I cmn common_param: - Vulkan0 : NVIDIA GeForce RTX 5090 Laptop GPU (24137 MiB, 23369 MiB free) +0.00.234.558 I cmn common_param: - Vulkan1 : AMD Radeon(TM) 890M Graphics (48505 MiB, 46080 MiB free) +0.00.234.566 I cmn common_param: - CPU : AMD Ryzen AI 9 HX 370 w/ Radeon 890M (64633 MiB, 25628 MiB free) +0.00.234.627 I cmn common_param: system_info: n_threads = 12 (n_threads_batch = 12) / 24 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | +0.00.234.705 I srv init_listene: running without SSL +0.00.234.768 I srv init_listene: The UI is disabled +0.00.234.768 I srv init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable +0.00.234.772 I srv init: using 23 threads for HTTP server +0.00.234.843 W srv llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655) +0.00.252.485 I srv load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.252.494 I srv load_model: local path 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.276.251 I llama_model_loader: loaded meta data with 37 key-value pairs and 226 tensors from C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf (version GGUF V3 (latest)) +0.00.276.263 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. +0.00.276.270 I llama_model_loader: - kv 0: general.architecture str = spark2_5 +0.00.276.271 I llama_model_loader: - kv 1: general.type str = model +0.00.276.274 I llama_model_loader: - kv 2: general.sampling.top_k i32 = 0 +0.00.276.282 I llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000 +0.00.276.283 I llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000 +0.00.276.283 I llama_model_loader: - kv 5: general.name str = Hf_Format +0.00.276.284 I llama_model_loader: - kv 6: general.size_label str = 1.7B +0.00.276.287 I llama_model_loader: - kv 7: spark2_5.block_count u32 = 28 +0.00.276.287 I llama_model_loader: - kv 8: spark2_5.context_length u32 = 1048576 +0.00.276.288 I llama_model_loader: - kv 9: spark2_5.embedding_length u32 = 2048 +0.00.276.288 I llama_model_loader: - kv 10: spark2_5.feed_forward_length u32 = 6656 +0.00.276.289 I llama_model_loader: - kv 11: spark2_5.attention.head_count u32 = 8 +0.00.276.289 I llama_model_loader: - kv 12: spark2_5.attention.head_count_kv u32 = 2 +0.00.276.290 I llama_model_loader: - kv 13: spark2_5.rope.freq_base f32 = 5000000.000000 +0.00.276.291 I llama_model_loader: - kv 14: spark2_5.rope.freq_base_swa f32 = 10000.000000 +0.00.276.293 I llama_model_loader: - kv 15: spark2_5.attention.layer_norm_rms_epsilon f32 = 0.000001 +0.00.276.294 I llama_model_loader: - kv 16: spark2_5.attention.key_length u32 = 256 +0.00.276.294 I llama_model_loader: - kv 17: spark2_5.attention.value_length u32 = 256 +0.00.276.294 I llama_model_loader: - kv 18: spark2_5.vocab_size u32 = 131072 +0.00.276.295 I llama_model_loader: - kv 19: spark2_5.attention.sliding_window u32 = 512 +0.00.276.317 I llama_model_loader: - kv 20: spark2_5.attention.sliding_window_pattern arr[bool,28] = [true, true, true, false, true, true,... +0.00.276.318 I llama_model_loader: - kv 21: spark2_5.rope.dimension_count u32 = 64 +0.00.276.318 I llama_model_loader: - kv 22: spark2_5.rope.dimension_count_swa u32 = 256 +0.00.276.318 I llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 +0.00.276.319 I llama_model_loader: - kv 24: tokenizer.ggml.pre str = spark2_5 +0.00.292.563 I llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,131072] = ["<｜start▁of▁sentence｜>", "<... +0.00.297.497 I llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,131072] = [3, 3, 3, 4, 4, 3, 3, 3, 4, 4, 4, 3, ... +0.00.313.613 I llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,130716] = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "Ġ... +0.00.313.623 I llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 0 +0.00.313.624 I llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 1 +0.00.313.625 I llama_model_loader: - kv 30: tokenizer.ggml.unknown_token_id u32 = 5 +0.00.313.626 I llama_model_loader: - kv 31: tokenizer.ggml.padding_token_id u32 = 2 +0.00.313.626 I llama_model_loader: - kv 32: tokenizer.ggml.add_bos_token bool = false +0.00.313.627 I llama_model_loader: - kv 33: tokenizer.ggml.add_eos_token bool = false +0.00.313.632 I llama_model_loader: - kv 34: tokenizer.chat_template str = {%- if not messages %}\n {{- raise_... +0.00.313.632 I llama_model_loader: - kv 35: general.quantization_version u32 = 2 +0.00.313.633 I llama_model_loader: - kv 36: general.file_type u32 = 15 +0.00.313.639 I llama_model_loader: - type f32: 57 tensors +0.00.313.640 I llama_model_loader: - type q4_K: 140 tensors +0.00.313.640 I llama_model_loader: - type q6_K: 29 tensors +0.00.313.644 I print_info: file format = GGUF V3 (latest) +0.00.313.645 I print_info: file type = Q4_K - Medium +0.00.313.650 I print_info: file size = 1.03 GiB (5.16 BPW) +0.00.318.241 I llama_prepare_model_devices: using device Vulkan1 (AMD Radeon(TM) 890M Graphics) (unknown id) - 46080 MiB free +0.00.371.726 I load: 86 unused tokens +0.00.393.628 I load: printing all EOG tokens: +0.00.393.636 I load: - 1 ('<｜end▁of▁sentence｜>') +0.00.393.762 I load: special tokens cache size = 113 +0.00.409.950 I load: token to piece cache size = 0.7975 MB +0.00.409.975 I print_info: arch = spark2_5 +0.00.409.976 I print_info: vocab_only = 0 +0.00.409.977 I print_info: no_alloc = 0 +0.00.409.977 I print_info: n_ctx_train = 1048576 +0.00.409.978 I print_info: n_embd_inp = 2048 +0.00.409.978 I print_info: n_embd = 2048 +0.00.409.978 I print_info: n_embd_out = 2048 +0.00.409.979 I print_info: n_layer = 28 +0.00.409.979 I print_info: n_layer_all = 28 +0.00.409.998 I print_info: n_head = 8 +0.00.409.999 I print_info: n_head_kv = 2 +0.00.410.000 I print_info: n_rot = 64 +0.00.410.000 I print_info: n_swa = 512 +0.00.410.001 I print_info: is_swa_any = 1 +0.00.410.001 I print_info: non_causal_type = 0 +0.00.410.001 I print_info: n_embd_head_k = 256 +0.00.410.002 I print_info: n_embd_head_v = 256 +0.00.410.003 I print_info: n_gqa = 4 +0.00.410.004 I print_info: n_embd_k_gqa = 512 +0.00.410.006 I print_info: n_embd_v_gqa = 512 +0.00.410.007 I print_info: f_norm_eps = 0.0e+00 +0.00.410.008 I print_info: f_norm_rms_eps = 1.0e-06 +0.00.410.009 I print_info: f_clamp_kqv = 0.0e+00 +0.00.410.009 I print_info: f_max_alibi_bias = 0.0e+00 +0.00.410.010 I print_info: f_logit_scale = 0.0e+00 +0.00.410.010 I print_info: f_attn_scale = 0.0e+00 +0.00.410.010 I print_info: f_attn_value_scale = 0.0000 +0.00.410.012 I print_info: n_ff = 6656 +0.00.410.012 I print_info: n_expert = 0 +0.00.410.012 I print_info: n_expert_used = 0 +0.00.410.014 I print_info: n_expert_groups = 0 +0.00.410.015 I print_info: n_group_used = 0 +0.00.410.015 I print_info: causal attn = 1 +0.00.410.015 I print_info: pooling type = -1 +0.00.410.016 I print_info: rope type = 2 +0.00.410.016 I print_info: rope scaling = linear +0.00.410.017 I print_info: freq_base_train = 5000000.0 +0.00.410.019 I print_info: freq_scale_train = 1 +0.00.410.020 I print_info: freq_base_swa = 10000.0 +0.00.410.021 I print_info: freq_scale_swa = 1 +0.00.410.021 I print_info: n_embd_head_k_swa = 256 +0.00.410.022 I print_info: n_embd_head_v_swa = 256 +0.00.410.022 I print_info: n_rot_swa = 256 +0.00.410.022 I print_info: n_ctx_orig_yarn = 1048576 +0.00.410.023 I print_info: rope_yarn_log_mul = 0.0000 +0.00.410.023 I print_info: rope_finetuned = unknown +0.00.410.024 I print_info: model type = 1.7B +0.00.410.025 I print_info: model params = 1.71 B +0.00.410.026 I print_info: general.name = Hf_Format +0.00.410.027 I print_info: vocab type = BPE +0.00.410.027 I print_info: n_vocab = 131072 +0.00.410.028 I print_info: n_merges = 130716 +0.00.410.029 I print_info: BOS token = 0 '<｜start▁of▁sentence｜>' +0.00.410.029 I print_info: EOS token = 1 '<｜end▁of▁sentence｜>' +0.00.410.030 I print_info: EOT token = 1 '<｜end▁of▁sentence｜>' +0.00.410.030 I print_info: UNK token = 5 '' +0.00.410.031 I print_info: PAD token = 2 '<｜▁pad▁｜>' +0.00.410.031 I print_info: LF token = 198 'Ċ' +0.00.410.032 I print_info: FIM PRE token = 11 '<｜fim▁begin｜>' +0.00.410.032 I print_info: FIM SUF token = 12 '<｜fim▁hole｜>' +0.00.410.033 I print_info: FIM MID token = 13 '<｜fim▁end｜>' +0.00.410.034 I print_info: EOG token = 1 '<｜end▁of▁sentence｜>' +0.00.410.034 I print_info: max token length = 210 +0.00.418.552 I load_tensors: loading model tensors, this can take a while... (load_mode = none) +0.00.609.154 I load_tensors: offloading output layer to GPU +0.00.609.160 I load_tensors: offloading 27 repeating layers to GPU +0.00.609.160 I load_tensors: offloaded 29/29 layers to GPU +0.00.609.165 I load_tensors: Vulkan1 model buffer size = 1051.02 MiB +0.00.609.166 I load_tensors: Vulkan_Host model buffer size = 210.00 MiB +0.00.981.646 I cmn common_init_: added <｜end▁of▁sentence｜> logit bias = -inf +0.00.982.053 I llama_context: constructing llama_context +0.00.982.059 I llama_context: n_seq_max = 1 +0.00.982.059 I llama_context: n_ctx = 4096 +0.00.982.059 I llama_context: n_ctx_seq = 4096 +0.00.982.060 I llama_context: n_batch = 2048 +0.00.982.060 I llama_context: n_ubatch = 512 +0.00.982.061 I llama_context: causal_attn = 1 +0.00.982.062 I llama_context: flash_attn = auto +0.00.982.062 I llama_context: kv_unified = false +0.00.982.065 I llama_context: freq_base = 5000000.0 +0.00.982.067 I llama_context: freq_scale = 1 +0.00.982.068 I llama_context: n_rs_seq = 0 +0.00.982.068 I llama_context: n_outputs_max = 1 +0.00.982.068 I llama_context: n_outputs_max_per_seq = 1 +0.00.982.069 I llama_context: n_ctx_seq (4096) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized +0.00.982.361 I llama_context: Vulkan_Host output buffer size = 0.50 MiB +0.00.982.368 I llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells +0.00.988.150 I llama_kv_cache: Vulkan1 KV buffer size = 56.00 MiB +0.00.988.908 I llama_kv_cache: size = 56.00 MiB ( 4096 cells, 7 layers, 1/1 seqs), K (f16): 28.00 MiB, V (f16): 28.00 MiB +0.00.988.918 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.988.919 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.988.928 I llama_kv_cache_iswa: creating SWA KV cache, size = 1024 cells +0.00.992.901 I llama_kv_cache: Vulkan1 KV buffer size = 42.00 MiB +0.00.993.374 I llama_kv_cache: size = 42.00 MiB ( 1024 cells, 21 layers, 1/1 seqs), K (f16): 21.00 MiB, V (f16): 21.00 MiB +0.00.993.383 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.993.384 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.993.397 I sched_reserve: reserving ... +0.00.995.290 I resolve_fused_ops: Flash Attention enabled +0.00.995.291 I resolve_fused_ops: resolving fused DeepSeek V4 HC support: +0.00.996.226 I resolve_fused_ops: fused DeepSeek V4 HC pre enabled +0.00.997.082 I resolve_fused_ops: fused DeepSeek V4 HC comb enabled +0.00.997.927 I resolve_fused_ops: fused DeepSeek V4 HC post enabled +0.01.009.137 I sched_reserve: Vulkan1 compute buffer size = 56.02 MiB +0.01.009.140 I sched_reserve: Vulkan_Host compute buffer size = 13.02 MiB +0.01.009.144 I sched_reserve: graph: nodes = 986, splits = 2, input objects = 4, input tensors = 9 +0.01.009.146 I sched_reserve: reserve took 15.74 ms, sched copies = 1 +0.01.009.298 I cmn init: llama threadpool init, n_threads = 12 +0.01.009.321 I cmn common_init_: warming up the model with an empty run - please wait ... (--no-warmup to disable) +0.01.058.096 I srv load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false' +0.01.058.119 I spec common_specu: no implementations specified for speculative decoding +0.01.058.120 I slot load_model: id 0 | task -1 | new slot, n_ctx = 4096 +0.01.058.202 I srv load_model: prompt cache is disabled - use `--cache-ram N` to enable it +0.01.058.203 I srv load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391 +0.01.058.203 I srv load_model: context checkpoints enabled, max = 32, min spacing = 8192 +0.01.058.225 W srv init: --cache-idle-slots requires --cache-ram, disabling +0.01.068.143 I srv init: init: chat template, example_format: '<｜start▁of▁sentence｜><|System|> +you are a helpful assistant. + +You are a helpful assistant<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>Hello<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>Hi there<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>How are you?<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>' +0.01.077.118 I srv init: init: chat template, thinking = 0 +0.01.077.131 I srv init: preserve_reasoning kwarg: true +0.01.077.132 W srv init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve) +0.01.077.158 I srv llama_server: model loaded +0.01.077.161 I srv llama_server: listening on http://127.0.0.1:61776 +0.01.077.232 I srv update_slots: all slots are idle +0.01.191.134 I srv server_strea: conv_id= (empty=1) +0.01.191.357 I srv operator(): chat format: peg-native +0.01.191.464 I slot get_availabl: id 0 | task -1 | - skipping, slot is empty +0.01.191.465 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = -1 +0.01.191.503 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.191.516 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.191.518 I slot launch_slot_: id 0 | task 0 | processing task, is_child = 0 +0.01.191.527 I slot operator(): id 0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 28 +0.01.191.533 I slot operator(): id 0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.205.010 I slot operator(): id 0 | task 0 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.254.007 I slot create_check: id 0 | task 0 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.01.257.566 I slot operator(): id 0 | task 0 | cached n_tokens = 24, memory_seq_rm [24, end) +0.01.257.603 I slot init_sampler: id 0 | task 0 | init sampler, took 0.01 ms, tokens: text = 28, total = 28 +0.01.312.893 I slot create_check: id 0 | task 0 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 23, n_tokens = 24, size = 0.985 MiB) +0.01.365.917 I slot print_timing: id 0 | task 0 | prompt eval time = 150.22 ms / 28 tokens ( 5.37 ms per token, 186.39 tokens per second) +0.01.365.923 I slot print_timing: id 0 | task 0 | eval time = 24.15 ms / 2 tokens ( 24.15 ms per token, 41.41 tokens per second) +0.01.365.925 I slot print_timing: id 0 | task 0 | total time = 174.37 ms / 30 tokens +0.01.365.925 I slot print_timing: id 0 | task 0 | graphs reused = 1 +0.01.365.955 I slot release: id 0 | task 0 | stop processing: n_tokens = 29, truncated = 0 +0.01.365.967 I srv update_slots: all slots are idle +0.01.393.741 I srv server_strea: conv_id= (empty=1) +0.01.395.413 I srv operator(): chat format: peg-native +0.01.395.558 I slot get_availabl: id 0 | task -1 | - checking sim = 0.005 (12/2463) > 0.100 +0.01.395.561 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = 1157284 +0.01.395.819 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.395.830 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.395.832 I slot launch_slot_: id 0 | task 5 | processing task, is_child = 0 +0.01.395.843 I slot operator(): id 0 | task 5 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.01.395.846 I slot operator(): id 0 | task 5 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.01.395.849 I slot operator(): id 0 | task 5 | erased invalidated context checkpoint (pos_min = 0, pos_max = 23, n_tokens = 24, n_swa = 512, pos_next = 0, size = 0.985 MiB) +0.01.395.852 I slot operator(): id 0 | task 5 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.399.894 I slot operator(): id 0 | task 5 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.439.045 I slot create_check: id 0 | task 5 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.02.819.131 I slot operator(): id 0 | task 5 | cached n_tokens = 1947, memory_seq_rm [1947, end) +0.03.333.521 I slot create_check: id 0 | task 5 | created context checkpoint 2 of 32 (pos_min = 923, pos_max = 1946, n_tokens = 1947, size = 21.006 MiB) +0.03.341.790 I slot operator(): id 0 | task 5 | cached n_tokens = 2459, memory_seq_rm [2459, end) +0.03.342.398 I slot init_sampler: id 0 | task 5 | init sampler, took 0.57 ms, tokens: text = 2463, total = 2463 +0.03.943.029 I slot create_check: id 0 | task 5 | created context checkpoint 3 of 32 (pos_min = 1435, pos_max = 2458, n_tokens = 2459, size = 21.006 MiB) +0.04.087.475 I slot print_timing: id 0 | task 5 | prompt eval time = 2578.84 ms / 2463 tokens ( 1.05 ms per token, 955.08 tokens per second) +0.04.087.490 I slot print_timing: id 0 | task 5 | eval time = 112.77 ms / 5 tokens ( 28.19 ms per token, 35.47 tokens per second) +0.04.087.492 I slot print_timing: id 0 | task 5 | total time = 2691.61 ms / 2468 tokens +0.04.087.493 I slot print_timing: id 0 | task 5 | graphs reused = 4 +0.04.087.589 I slot release: id 0 | task 5 | stop processing: n_tokens = 2467, truncated = 0 +0.04.087.643 I srv update_slots: all slots are idle +0.04.136.115 I srv server_strea: conv_id= (empty=1) +0.04.139.673 I srv operator(): chat format: peg-native +0.04.139.950 I slot get_availabl: id 0 | task -1 | - checking sim = 1.000 (2463/2463) > 0.100 +0.04.139.953 I slot get_availabl: id 0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998 +0.04.140.289 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.04.140.304 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.04.140.306 I slot launch_slot_: id 0 | task 14 | processing task, is_child = 0 +0.04.140.323 I slot operator(): id 0 | task 14 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.04.140.327 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.04.140.332 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 923, pos_max = 1946, n_tokens = 1947, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.04.141.509 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 1435, pos_max = 2458, n_tokens = 2459, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.04.142.492 I slot operator(): id 0 | task 14 | cached n_tokens = 0, memory_seq_rm [0, end) +0.04.147.920 I slot operator(): id 0 | task 14 | cached n_tokens = 10, memory_seq_rm [10, end) +0.04.191.486 I slot create_check: id 0 | task 14 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.05.569.203 I slot operator(): id 0 | task 14 | cached n_tokens = 1947, memory_seq_rm [1947, end) +0.06.074.567 I slot create_check: id 0 | task 14 | created context checkpoint 2 of 32 (pos_min = 923, pos_max = 1946, n_tokens = 1947, size = 21.006 MiB) +0.06.082.773 I slot operator(): id 0 | task 14 | cached n_tokens = 2459, memory_seq_rm [2459, end) +0.06.083.225 I slot init_sampler: id 0 | task 14 | init sampler, took 0.42 ms, tokens: text = 2463, total = 2463 +0.06.695.574 I slot create_check: id 0 | task 14 | created context checkpoint 3 of 32 (pos_min = 1435, pos_max = 2458, n_tokens = 2459, size = 21.006 MiB) +0.06.811.155 I slot print_timing: id 0 | task 14 | prompt eval time = 2586.49 ms / 2463 tokens ( 1.05 ms per token, 952.25 tokens per second) +0.06.811.168 I slot print_timing: id 0 | task 14 | eval time = 84.31 ms / 5 tokens ( 21.08 ms per token, 47.44 tokens per second) +0.06.811.170 I slot print_timing: id 0 | task 14 | total time = 2670.80 ms / 2468 tokens +0.06.811.173 I slot print_timing: id 0 | task 14 | graphs reused = 7 +0.06.811.287 I slot release: id 0 | task 14 | stop processing: n_tokens = 2467, truncated = 0 +0.06.811.322 I srv update_slots: all slots are idle diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v2_server_restart.log b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v2_server_restart.log new file mode 100644 index 000000000..fecfafcd3 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v2_server_restart.log @@ -0,0 +1,211 @@ +0.00.000.978 I srv llama_server: initializing ... +0.00.237.982 I cmn common_param: common_params_print_info: build 11124 (1a679828f) with Clang 20.1.8 for Windows x86_64 +0.00.237.989 I cmn common_param: common_params_print_info: verbosity = 4 (adjust with the `-lv N` CLI arg) +0.00.237.991 I cmn common_param: device_info: +0.00.242.510 I cmn common_param: - Vulkan0 : NVIDIA GeForce RTX 5090 Laptop GPU (24137 MiB, 23369 MiB free) +0.00.247.157 I cmn common_param: - Vulkan1 : AMD Radeon(TM) 890M Graphics (48505 MiB, 46080 MiB free) +0.00.247.170 I cmn common_param: - CPU : AMD Ryzen AI 9 HX 370 w/ Radeon 890M (64633 MiB, 25783 MiB free) +0.00.247.239 I cmn common_param: system_info: n_threads = 12 (n_threads_batch = 12) / 24 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | +0.00.247.324 I srv init_listene: running without SSL +0.00.247.401 I srv init_listene: The UI is disabled +0.00.247.402 I srv init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable +0.00.247.406 I srv init: using 23 threads for HTTP server +0.00.247.481 W srv llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655) +0.00.260.742 I srv load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.260.748 I srv load_model: local path 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.289.823 I llama_model_loader: loaded meta data with 37 key-value pairs and 226 tensors from C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf (version GGUF V3 (latest)) +0.00.289.837 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. +0.00.289.843 I llama_model_loader: - kv 0: general.architecture str = spark2_5 +0.00.289.844 I llama_model_loader: - kv 1: general.type str = model +0.00.289.848 I llama_model_loader: - kv 2: general.sampling.top_k i32 = 0 +0.00.289.855 I llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000 +0.00.289.856 I llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000 +0.00.289.857 I llama_model_loader: - kv 5: general.name str = Hf_Format +0.00.289.857 I llama_model_loader: - kv 6: general.size_label str = 1.7B +0.00.289.860 I llama_model_loader: - kv 7: spark2_5.block_count u32 = 28 +0.00.289.861 I llama_model_loader: - kv 8: spark2_5.context_length u32 = 1048576 +0.00.289.861 I llama_model_loader: - kv 9: spark2_5.embedding_length u32 = 2048 +0.00.289.862 I llama_model_loader: - kv 10: spark2_5.feed_forward_length u32 = 6656 +0.00.289.862 I llama_model_loader: - kv 11: spark2_5.attention.head_count u32 = 8 +0.00.289.863 I llama_model_loader: - kv 12: spark2_5.attention.head_count_kv u32 = 2 +0.00.289.864 I llama_model_loader: - kv 13: spark2_5.rope.freq_base f32 = 5000000.000000 +0.00.289.865 I llama_model_loader: - kv 14: spark2_5.rope.freq_base_swa f32 = 10000.000000 +0.00.289.867 I llama_model_loader: - kv 15: spark2_5.attention.layer_norm_rms_epsilon f32 = 0.000001 +0.00.289.867 I llama_model_loader: - kv 16: spark2_5.attention.key_length u32 = 256 +0.00.289.868 I llama_model_loader: - kv 17: spark2_5.attention.value_length u32 = 256 +0.00.289.868 I llama_model_loader: - kv 18: spark2_5.vocab_size u32 = 131072 +0.00.289.869 I llama_model_loader: - kv 19: spark2_5.attention.sliding_window u32 = 512 +0.00.289.890 I llama_model_loader: - kv 20: spark2_5.attention.sliding_window_pattern arr[bool,28] = [true, true, true, false, true, true,... +0.00.289.890 I llama_model_loader: - kv 21: spark2_5.rope.dimension_count u32 = 64 +0.00.289.891 I llama_model_loader: - kv 22: spark2_5.rope.dimension_count_swa u32 = 256 +0.00.289.891 I llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 +0.00.289.892 I llama_model_loader: - kv 24: tokenizer.ggml.pre str = spark2_5 +0.00.306.236 I llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,131072] = ["<｜start▁of▁sentence｜>", "<... +0.00.311.192 I llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,131072] = [3, 3, 3, 4, 4, 3, 3, 3, 4, 4, 4, 3, ... +0.00.327.489 I llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,130716] = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "Ġ... +0.00.327.498 I llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 0 +0.00.327.499 I llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 1 +0.00.327.501 I llama_model_loader: - kv 30: tokenizer.ggml.unknown_token_id u32 = 5 +0.00.327.501 I llama_model_loader: - kv 31: tokenizer.ggml.padding_token_id u32 = 2 +0.00.327.502 I llama_model_loader: - kv 32: tokenizer.ggml.add_bos_token bool = false +0.00.327.503 I llama_model_loader: - kv 33: tokenizer.ggml.add_eos_token bool = false +0.00.327.507 I llama_model_loader: - kv 34: tokenizer.chat_template str = {%- if not messages %}\n {{- raise_... +0.00.327.508 I llama_model_loader: - kv 35: general.quantization_version u32 = 2 +0.00.327.509 I llama_model_loader: - kv 36: general.file_type u32 = 15 +0.00.327.517 I llama_model_loader: - type f32: 57 tensors +0.00.327.517 I llama_model_loader: - type q4_K: 140 tensors +0.00.327.518 I llama_model_loader: - type q6_K: 29 tensors +0.00.327.523 I print_info: file format = GGUF V3 (latest) +0.00.327.523 I print_info: file type = Q4_K - Medium +0.00.327.528 I print_info: file size = 1.03 GiB (5.16 BPW) +0.00.332.405 I llama_prepare_model_devices: using device Vulkan1 (AMD Radeon(TM) 890M Graphics) (unknown id) - 46080 MiB free +0.00.416.321 I load: 86 unused tokens +0.00.439.243 I load: printing all EOG tokens: +0.00.439.252 I load: - 1 ('<｜end▁of▁sentence｜>') +0.00.439.549 I load: special tokens cache size = 113 +0.00.455.836 I load: token to piece cache size = 0.7975 MB +0.00.455.869 I print_info: arch = spark2_5 +0.00.455.870 I print_info: vocab_only = 0 +0.00.455.870 I print_info: no_alloc = 0 +0.00.455.871 I print_info: n_ctx_train = 1048576 +0.00.455.871 I print_info: n_embd_inp = 2048 +0.00.455.872 I print_info: n_embd = 2048 +0.00.455.872 I print_info: n_embd_out = 2048 +0.00.455.872 I print_info: n_layer = 28 +0.00.455.873 I print_info: n_layer_all = 28 +0.00.455.890 I print_info: n_head = 8 +0.00.455.891 I print_info: n_head_kv = 2 +0.00.455.892 I print_info: n_rot = 64 +0.00.455.892 I print_info: n_swa = 512 +0.00.455.893 I print_info: is_swa_any = 1 +0.00.455.893 I print_info: non_causal_type = 0 +0.00.455.893 I print_info: n_embd_head_k = 256 +0.00.455.893 I print_info: n_embd_head_v = 256 +0.00.455.895 I print_info: n_gqa = 4 +0.00.455.896 I print_info: n_embd_k_gqa = 512 +0.00.455.897 I print_info: n_embd_v_gqa = 512 +0.00.455.899 I print_info: f_norm_eps = 0.0e+00 +0.00.455.900 I print_info: f_norm_rms_eps = 1.0e-06 +0.00.455.901 I print_info: f_clamp_kqv = 0.0e+00 +0.00.455.901 I print_info: f_max_alibi_bias = 0.0e+00 +0.00.455.901 I print_info: f_logit_scale = 0.0e+00 +0.00.455.902 I print_info: f_attn_scale = 0.0e+00 +0.00.455.902 I print_info: f_attn_value_scale = 0.0000 +0.00.455.904 I print_info: n_ff = 6656 +0.00.455.904 I print_info: n_expert = 0 +0.00.455.904 I print_info: n_expert_used = 0 +0.00.455.906 I print_info: n_expert_groups = 0 +0.00.455.906 I print_info: n_group_used = 0 +0.00.455.907 I print_info: causal attn = 1 +0.00.455.907 I print_info: pooling type = -1 +0.00.455.907 I print_info: rope type = 2 +0.00.455.908 I print_info: rope scaling = linear +0.00.455.909 I print_info: freq_base_train = 5000000.0 +0.00.455.910 I print_info: freq_scale_train = 1 +0.00.455.911 I print_info: freq_base_swa = 10000.0 +0.00.455.911 I print_info: freq_scale_swa = 1 +0.00.455.912 I print_info: n_embd_head_k_swa = 256 +0.00.455.912 I print_info: n_embd_head_v_swa = 256 +0.00.455.912 I print_info: n_rot_swa = 256 +0.00.455.913 I print_info: n_ctx_orig_yarn = 1048576 +0.00.455.913 I print_info: rope_yarn_log_mul = 0.0000 +0.00.455.913 I print_info: rope_finetuned = unknown +0.00.455.914 I print_info: model type = 1.7B +0.00.455.915 I print_info: model params = 1.71 B +0.00.455.916 I print_info: general.name = Hf_Format +0.00.455.917 I print_info: vocab type = BPE +0.00.455.917 I print_info: n_vocab = 131072 +0.00.455.917 I print_info: n_merges = 130716 +0.00.455.918 I print_info: BOS token = 0 '<｜start▁of▁sentence｜>' +0.00.455.918 I print_info: EOS token = 1 '<｜end▁of▁sentence｜>' +0.00.455.919 I print_info: EOT token = 1 '<｜end▁of▁sentence｜>' +0.00.455.919 I print_info: UNK token = 5 '' +0.00.455.920 I print_info: PAD token = 2 '<｜▁pad▁｜>' +0.00.455.920 I print_info: LF token = 198 'Ċ' +0.00.455.921 I print_info: FIM PRE token = 11 '<｜fim▁begin｜>' +0.00.455.921 I print_info: FIM SUF token = 12 '<｜fim▁hole｜>' +0.00.455.922 I print_info: FIM MID token = 13 '<｜fim▁end｜>' +0.00.455.922 I print_info: EOG token = 1 '<｜end▁of▁sentence｜>' +0.00.455.923 I print_info: max token length = 210 +0.00.465.828 I load_tensors: loading model tensors, this can take a while... (load_mode = none) +0.00.650.868 I load_tensors: offloading output layer to GPU +0.00.650.874 I load_tensors: offloading 27 repeating layers to GPU +0.00.650.874 I load_tensors: offloaded 29/29 layers to GPU +0.00.650.879 I load_tensors: Vulkan1 model buffer size = 1051.02 MiB +0.00.650.880 I load_tensors: Vulkan_Host model buffer size = 210.00 MiB +0.01.020.180 I cmn common_init_: added <｜end▁of▁sentence｜> logit bias = -inf +0.01.020.594 I llama_context: constructing llama_context +0.01.020.600 I llama_context: n_seq_max = 1 +0.01.020.600 I llama_context: n_ctx = 4096 +0.01.020.601 I llama_context: n_ctx_seq = 4096 +0.01.020.602 I llama_context: n_batch = 2048 +0.01.020.602 I llama_context: n_ubatch = 512 +0.01.020.602 I llama_context: causal_attn = 1 +0.01.020.603 I llama_context: flash_attn = auto +0.01.020.603 I llama_context: kv_unified = false +0.01.020.606 I llama_context: freq_base = 5000000.0 +0.01.020.608 I llama_context: freq_scale = 1 +0.01.020.609 I llama_context: n_rs_seq = 0 +0.01.020.609 I llama_context: n_outputs_max = 1 +0.01.020.610 I llama_context: n_outputs_max_per_seq = 1 +0.01.020.610 I llama_context: n_ctx_seq (4096) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized +0.01.020.934 I llama_context: Vulkan_Host output buffer size = 0.50 MiB +0.01.020.940 I llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells +0.01.027.328 I llama_kv_cache: Vulkan1 KV buffer size = 56.00 MiB +0.01.027.942 I llama_kv_cache: size = 56.00 MiB ( 4096 cells, 7 layers, 1/1 seqs), K (f16): 28.00 MiB, V (f16): 28.00 MiB +0.01.027.950 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.01.027.951 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.01.027.956 I llama_kv_cache_iswa: creating SWA KV cache, size = 1024 cells +0.01.031.624 I llama_kv_cache: Vulkan1 KV buffer size = 42.00 MiB +0.01.032.147 I llama_kv_cache: size = 42.00 MiB ( 1024 cells, 21 layers, 1/1 seqs), K (f16): 21.00 MiB, V (f16): 21.00 MiB +0.01.032.154 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.01.032.155 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.01.032.164 I sched_reserve: reserving ... +0.01.033.957 I resolve_fused_ops: Flash Attention enabled +0.01.033.958 I resolve_fused_ops: resolving fused DeepSeek V4 HC support: +0.01.035.026 I resolve_fused_ops: fused DeepSeek V4 HC pre enabled +0.01.035.946 I resolve_fused_ops: fused DeepSeek V4 HC comb enabled +0.01.036.821 I resolve_fused_ops: fused DeepSeek V4 HC post enabled +0.01.047.016 I sched_reserve: Vulkan1 compute buffer size = 56.02 MiB +0.01.047.018 I sched_reserve: Vulkan_Host compute buffer size = 13.02 MiB +0.01.047.020 I sched_reserve: graph: nodes = 986, splits = 2, input objects = 4, input tensors = 9 +0.01.047.021 I sched_reserve: reserve took 14.85 ms, sched copies = 1 +0.01.047.121 I cmn init: llama threadpool init, n_threads = 12 +0.01.047.137 I cmn common_init_: warming up the model with an empty run - please wait ... (--no-warmup to disable) +0.01.111.473 I srv load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false' +0.01.111.497 I spec common_specu: no implementations specified for speculative decoding +0.01.111.498 I slot load_model: id 0 | task -1 | new slot, n_ctx = 4096 +0.01.111.607 I srv load_model: prompt cache is disabled - use `--cache-ram N` to enable it +0.01.111.608 I srv load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391 +0.01.111.609 I srv load_model: context checkpoints enabled, max = 32, min spacing = 8192 +0.01.111.630 W srv init: --cache-idle-slots requires --cache-ram, disabling +0.01.121.515 I srv init: init: chat template, example_format: '<｜start▁of▁sentence｜><|System|> +you are a helpful assistant. + +You are a helpful assistant<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>Hello<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>Hi there<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>How are you?<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>' +0.01.128.881 I srv init: init: chat template, thinking = 0 +0.01.128.892 I srv init: preserve_reasoning kwarg: true +0.01.128.893 W srv init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve) +0.01.128.919 I srv llama_server: model loaded +0.01.128.921 I srv llama_server: listening on http://127.0.0.1:61807 +0.01.128.994 I srv update_slots: all slots are idle +0.01.400.210 I srv server_strea: conv_id= (empty=1) +0.01.400.456 I srv operator(): chat format: peg-native +0.01.400.593 I slot get_availabl: id 0 | task -1 | - skipping, slot is empty +0.01.400.595 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = -1 +0.01.400.651 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.400.669 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.400.671 I slot launch_slot_: id 0 | task 0 | processing task, is_child = 0 +0.01.400.685 I slot operator(): id 0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 28 +0.01.400.693 I slot operator(): id 0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.417.204 I slot operator(): id 0 | task 0 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.460.637 I slot create_check: id 0 | task 0 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.01.464.017 I slot operator(): id 0 | task 0 | cached n_tokens = 24, memory_seq_rm [24, end) +0.01.464.046 I slot init_sampler: id 0 | task 0 | init sampler, took 0.00 ms, tokens: text = 28, total = 28 +0.01.505.554 I slot create_check: id 0 | task 0 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 23, n_tokens = 24, size = 0.985 MiB) +0.01.552.281 I slot print_timing: id 0 | task 0 | prompt eval time = 130.98 ms / 28 tokens ( 4.68 ms per token, 213.77 tokens per second) +0.01.552.287 I \ No newline at end of file diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v3_report.json b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v3_report.json new file mode 100644 index 000000000..602a0f9f6 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v3_report.json @@ -0,0 +1,253 @@ +{ + "scope": "complete Spark GGUF text requests through Omni StageRuntime/StagePool and external llama.cpp backend", + "model_file": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "device": "Vulkan1", + "memory_budget": { + "capacities": { + "host_ram": 12884901888 + }, + "demands": { + "host_ram": 4294967296 + } + }, + "warmup_count": 0, + "measured_count": 1, + "startup_s": 2.678464900003746, + "execution_plan": { + "backend": "external.llamacpp.text.v1", + "stage_id": 0, + "worker_generation": "efa4702441ff40c3bc62bfa4833f2db2", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "requested_device": "Vulkan1", + "expected_device_name": "AMD Radeon(TM) 890M Graphics", + "layer_assignments": 0, + "offloaded_layers": [ + "29", + "29" + ], + "model_buffer_devices": [ + "Vulkan1" + ], + "all_model_buffers": [ + "Vulkan1", + "Vulkan_Host" + ], + "cpu_only_layout": false, + "worker_pid": 66304, + "loaded_rss_bytes": 1643982848, + "reserved_bytes": { + "host_ram": 4294967296 + }, + "memory_pool": "host_ram", + "memory_overhead_bytes": 2147483648, + "context_tokens": 4096, + "max_new_tokens": 96, + "max_io_bytes": 1048576, + "request_capacity": 1, + "stateful_session": "llama.cpp-owned; reset for each complete request", + "evidence": "B" + }, + "checks": [ + { + "case": "short", + "answer": "Paris", + "expected": "Paris", + "correct": true, + "wall_s": 0.18514120008330792, + "stage_event": { + "request_id": "short-688568093000000", + "stage_id": 0, + "epoch": 1, + "seq": 1, + "kind": "text", + "worker_generation": "efa4702441ff40c3bc62bfa4833f2db2", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 0.16857679991517216, + "usage": { + "completion_tokens": 2, + "prompt_tokens": 28, + "total_tokens": 30, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + }, + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 2.5507894000038505, + "stage_event": { + "request_id": "inventory_120-688568281000000", + "stage_id": 0, + "epoch": 2, + "seq": 1, + "kind": "text", + "worker_generation": "efa4702441ff40c3bc62bfa4833f2db2", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 2.5293748999247327, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "warmups": [], + "measured": [ + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 2.599552399944514, + "stage_event": { + "request_id": "inventory_120-688570828000000", + "stage_id": 0, + "epoch": 3, + "seq": 1, + "kind": "text", + "worker_generation": "efa4702441ff40c3bc62bfa4833f2db2", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 2.572201499948278, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "nearest_rank_p50_wall_s": 2.599552399944514, + "nearest_rank_p95_wall_s": 2.599552399944514, + "abort_check": { + "server_tasks_before_abort_request": 3, + "server_tasks_after_abort_request_start": 4, + "stale_output": false, + "worker_exited": true, + "ledger_after_abort": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + } + }, + "restart_check": { + "first_runtime_ledger_after_shutdown": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + }, + "fresh_startup_s": 2.2718601999804378, + "fresh_server_log": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\vllm-omni-edge\\benchmarks\\edge_harness\\results\\e2e_expansion_20260924\\evidence\\spark_omni_windows_restart\\radeon_v3_server_restart.log", + "fresh_execution_plan": { + "backend": "external.llamacpp.text.v1", + "stage_id": 0, + "worker_generation": "715af02878c74202aa5c3916abee6c79", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "requested_device": "Vulkan1", + "expected_device_name": "AMD Radeon(TM) 890M Graphics", + "layer_assignments": 0, + "offloaded_layers": [ + "29", + "29" + ], + "model_buffer_devices": [ + "Vulkan1" + ], + "all_model_buffers": [ + "Vulkan1", + "Vulkan_Host" + ], + "cpu_only_layout": false, + "worker_pid": 40188, + "loaded_rss_bytes": 1644060672, + "reserved_bytes": { + "host_ram": 4294967296 + }, + "memory_pool": "host_ram", + "memory_overhead_bytes": 2147483648, + "context_tokens": 4096, + "max_new_tokens": 96, + "max_io_bytes": 1048576, + "request_capacity": 1, + "stateful_session": "llama.cpp-owned; reset for each complete request", + "evidence": "B" + }, + "fresh_request": { + "case": "short", + "answer": "Paris", + "expected": "Paris", + "correct": true, + "wall_s": 0.18440090003423393, + "stage_event": { + "request_id": "short-688576265000000", + "stage_id": 0, + "epoch": 1, + "seq": 1, + "kind": "text", + "worker_generation": "715af02878c74202aa5c3916abee6c79", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 0.18165389995556325, + "usage": { + "completion_tokens": 2, + "prompt_tokens": 28, + "total_tokens": 30, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + }, + "late_output": false + }, + "status": "passed", + "ledger_after_shutdown": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + } +} diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v3_server.log b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v3_server.log new file mode 100644 index 000000000..ffa79a739 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v3_server.log @@ -0,0 +1,293 @@ +0.00.000.797 I srv llama_server: initializing ... +0.00.231.626 I cmn common_param: common_params_print_info: build 11124 (1a679828f) with Clang 20.1.8 for Windows x86_64 +0.00.231.633 I cmn common_param: common_params_print_info: verbosity = 4 (adjust with the `-lv N` CLI arg) +0.00.231.634 I cmn common_param: device_info: +0.00.235.643 I cmn common_param: - Vulkan0 : NVIDIA GeForce RTX 5090 Laptop GPU (24137 MiB, 23369 MiB free) +0.00.239.457 I cmn common_param: - Vulkan1 : AMD Radeon(TM) 890M Graphics (48505 MiB, 46080 MiB free) +0.00.239.466 I cmn common_param: - CPU : AMD Ryzen AI 9 HX 370 w/ Radeon 890M (64633 MiB, 26657 MiB free) +0.00.239.527 I cmn common_param: system_info: n_threads = 12 (n_threads_batch = 12) / 24 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | +0.00.239.609 I srv init_listene: running without SSL +0.00.239.669 I srv init_listene: The UI is disabled +0.00.239.669 I srv init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable +0.00.239.673 I srv init: using 23 threads for HTTP server +0.00.239.746 W srv llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655) +0.00.257.205 I srv load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.257.226 I srv load_model: local path 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.283.315 I llama_model_loader: loaded meta data with 37 key-value pairs and 226 tensors from C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf (version GGUF V3 (latest)) +0.00.283.332 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. +0.00.283.341 I llama_model_loader: - kv 0: general.architecture str = spark2_5 +0.00.283.342 I llama_model_loader: - kv 1: general.type str = model +0.00.283.348 I llama_model_loader: - kv 2: general.sampling.top_k i32 = 0 +0.00.283.354 I llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000 +0.00.283.355 I llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000 +0.00.283.356 I llama_model_loader: - kv 5: general.name str = Hf_Format +0.00.283.356 I llama_model_loader: - kv 6: general.size_label str = 1.7B +0.00.283.361 I llama_model_loader: - kv 7: spark2_5.block_count u32 = 28 +0.00.283.362 I llama_model_loader: - kv 8: spark2_5.context_length u32 = 1048576 +0.00.283.363 I llama_model_loader: - kv 9: spark2_5.embedding_length u32 = 2048 +0.00.283.363 I llama_model_loader: - kv 10: spark2_5.feed_forward_length u32 = 6656 +0.00.283.364 I llama_model_loader: - kv 11: spark2_5.attention.head_count u32 = 8 +0.00.283.364 I llama_model_loader: - kv 12: spark2_5.attention.head_count_kv u32 = 2 +0.00.283.365 I llama_model_loader: - kv 13: spark2_5.rope.freq_base f32 = 5000000.000000 +0.00.283.366 I llama_model_loader: - kv 14: spark2_5.rope.freq_base_swa f32 = 10000.000000 +0.00.283.368 I llama_model_loader: - kv 15: spark2_5.attention.layer_norm_rms_epsilon f32 = 0.000001 +0.00.283.369 I llama_model_loader: - kv 16: spark2_5.attention.key_length u32 = 256 +0.00.283.369 I llama_model_loader: - kv 17: spark2_5.attention.value_length u32 = 256 +0.00.283.370 I llama_model_loader: - kv 18: spark2_5.vocab_size u32 = 131072 +0.00.283.370 I llama_model_loader: - kv 19: spark2_5.attention.sliding_window u32 = 512 +0.00.283.436 I llama_model_loader: - kv 20: spark2_5.attention.sliding_window_pattern arr[bool,28] = [true, true, true, false, true, true,... +0.00.283.438 I llama_model_loader: - kv 21: spark2_5.rope.dimension_count u32 = 64 +0.00.283.439 I llama_model_loader: - kv 22: spark2_5.rope.dimension_count_swa u32 = 256 +0.00.283.439 I llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 +0.00.283.440 I llama_model_loader: - kv 24: tokenizer.ggml.pre str = spark2_5 +0.00.300.142 I llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,131072] = ["<｜start▁of▁sentence｜>", "<... +0.00.305.534 I llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,131072] = [3, 3, 3, 4, 4, 3, 3, 3, 4, 4, 4, 3, ... +0.00.322.155 I llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,130716] = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "Ġ... +0.00.322.164 I llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 0 +0.00.322.164 I llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 1 +0.00.322.165 I llama_model_loader: - kv 30: tokenizer.ggml.unknown_token_id u32 = 5 +0.00.322.166 I llama_model_loader: - kv 31: tokenizer.ggml.padding_token_id u32 = 2 +0.00.322.167 I llama_model_loader: - kv 32: tokenizer.ggml.add_bos_token bool = false +0.00.322.167 I llama_model_loader: - kv 33: tokenizer.ggml.add_eos_token bool = false +0.00.322.175 I llama_model_loader: - kv 34: tokenizer.chat_template str = {%- if not messages %}\n {{- raise_... +0.00.322.176 I llama_model_loader: - kv 35: general.quantization_version u32 = 2 +0.00.322.176 I llama_model_loader: - kv 36: general.file_type u32 = 15 +0.00.322.184 I llama_model_loader: - type f32: 57 tensors +0.00.322.185 I llama_model_loader: - type q4_K: 140 tensors +0.00.322.185 I llama_model_loader: - type q6_K: 29 tensors +0.00.322.190 I print_info: file format = GGUF V3 (latest) +0.00.322.190 I print_info: file type = Q4_K - Medium +0.00.322.194 I print_info: file size = 1.03 GiB (5.16 BPW) +0.00.326.661 I llama_prepare_model_devices: using device Vulkan1 (AMD Radeon(TM) 890M Graphics) (unknown id) - 46080 MiB free +0.00.386.252 I load: 86 unused tokens +0.00.408.788 I load: printing all EOG tokens: +0.00.408.796 I load: - 1 ('<｜end▁of▁sentence｜>') +0.00.408.942 I load: special tokens cache size = 113 +0.00.424.454 I load: token to piece cache size = 0.7975 MB +0.00.424.482 I print_info: arch = spark2_5 +0.00.424.483 I print_info: vocab_only = 0 +0.00.424.483 I print_info: no_alloc = 0 +0.00.424.484 I print_info: n_ctx_train = 1048576 +0.00.424.484 I print_info: n_embd_inp = 2048 +0.00.424.485 I print_info: n_embd = 2048 +0.00.424.485 I print_info: n_embd_out = 2048 +0.00.424.485 I print_info: n_layer = 28 +0.00.424.486 I print_info: n_layer_all = 28 +0.00.424.501 I print_info: n_head = 8 +0.00.424.502 I print_info: n_head_kv = 2 +0.00.424.502 I print_info: n_rot = 64 +0.00.424.503 I print_info: n_swa = 512 +0.00.424.503 I print_info: is_swa_any = 1 +0.00.424.504 I print_info: non_causal_type = 0 +0.00.424.504 I print_info: n_embd_head_k = 256 +0.00.424.504 I print_info: n_embd_head_v = 256 +0.00.424.505 I print_info: n_gqa = 4 +0.00.424.507 I print_info: n_embd_k_gqa = 512 +0.00.424.508 I print_info: n_embd_v_gqa = 512 +0.00.424.509 I print_info: f_norm_eps = 0.0e+00 +0.00.424.511 I print_info: f_norm_rms_eps = 1.0e-06 +0.00.424.511 I print_info: f_clamp_kqv = 0.0e+00 +0.00.424.512 I print_info: f_max_alibi_bias = 0.0e+00 +0.00.424.512 I print_info: f_logit_scale = 0.0e+00 +0.00.424.512 I print_info: f_attn_scale = 0.0e+00 +0.00.424.513 I print_info: f_attn_value_scale = 0.0000 +0.00.424.514 I print_info: n_ff = 6656 +0.00.424.514 I print_info: n_expert = 0 +0.00.424.515 I print_info: n_expert_used = 0 +0.00.424.517 I print_info: n_expert_groups = 0 +0.00.424.517 I print_info: n_group_used = 0 +0.00.424.517 I print_info: causal attn = 1 +0.00.424.518 I print_info: pooling type = -1 +0.00.424.518 I print_info: rope type = 2 +0.00.424.518 I print_info: rope scaling = linear +0.00.424.519 I print_info: freq_base_train = 5000000.0 +0.00.424.521 I print_info: freq_scale_train = 1 +0.00.424.521 I print_info: freq_base_swa = 10000.0 +0.00.424.521 I print_info: freq_scale_swa = 1 +0.00.424.522 I print_info: n_embd_head_k_swa = 256 +0.00.424.522 I print_info: n_embd_head_v_swa = 256 +0.00.424.522 I print_info: n_rot_swa = 256 +0.00.424.523 I print_info: n_ctx_orig_yarn = 1048576 +0.00.424.523 I print_info: rope_yarn_log_mul = 0.0000 +0.00.424.523 I print_info: rope_finetuned = unknown +0.00.424.524 I print_info: model type = 1.7B +0.00.424.525 I print_info: model params = 1.71 B +0.00.424.525 I print_info: general.name = Hf_Format +0.00.424.526 I print_info: vocab type = BPE +0.00.424.527 I print_info: n_vocab = 131072 +0.00.424.527 I print_info: n_merges = 130716 +0.00.424.528 I print_info: BOS token = 0 '<｜start▁of▁sentence｜>' +0.00.424.528 I print_info: EOS token = 1 '<｜end▁of▁sentence｜>' +0.00.424.529 I print_info: EOT token = 1 '<｜end▁of▁sentence｜>' +0.00.424.529 I print_info: UNK token = 5 '' +0.00.424.530 I print_info: PAD token = 2 '<｜▁pad▁｜>' +0.00.424.530 I print_info: LF token = 198 'Ċ' +0.00.424.531 I print_info: FIM PRE token = 11 '<｜fim▁begin｜>' +0.00.424.531 I print_info: FIM SUF token = 12 '<｜fim▁hole｜>' +0.00.424.531 I print_info: FIM MID token = 13 '<｜fim▁end｜>' +0.00.424.532 I print_info: EOG token = 1 '<｜end▁of▁sentence｜>' +0.00.424.532 I print_info: max token length = 210 +0.00.434.986 I load_tensors: loading model tensors, this can take a while... (load_mode = none) +0.00.639.411 I load_tensors: offloading output layer to GPU +0.00.639.418 I load_tensors: offloading 27 repeating layers to GPU +0.00.639.419 I load_tensors: offloaded 29/29 layers to GPU +0.00.639.423 I load_tensors: Vulkan1 model buffer size = 1051.02 MiB +0.00.639.424 I load_tensors: Vulkan_Host model buffer size = 210.00 MiB +0.00.989.278 I cmn common_init_: added <｜end▁of▁sentence｜> logit bias = -inf +0.00.989.689 I llama_context: constructing llama_context +0.00.989.695 I llama_context: n_seq_max = 1 +0.00.989.695 I llama_context: n_ctx = 4096 +0.00.989.695 I llama_context: n_ctx_seq = 4096 +0.00.989.696 I llama_context: n_batch = 2048 +0.00.989.696 I llama_context: n_ubatch = 512 +0.00.989.697 I llama_context: causal_attn = 1 +0.00.989.697 I llama_context: flash_attn = auto +0.00.989.698 I llama_context: kv_unified = false +0.00.989.700 I llama_context: freq_base = 5000000.0 +0.00.989.703 I llama_context: freq_scale = 1 +0.00.989.703 I llama_context: n_rs_seq = 0 +0.00.989.703 I llama_context: n_outputs_max = 1 +0.00.989.704 I llama_context: n_outputs_max_per_seq = 1 +0.00.989.704 I llama_context: n_ctx_seq (4096) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized +0.00.990.015 I llama_context: Vulkan_Host output buffer size = 0.50 MiB +0.00.990.023 I llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells +0.00.995.091 I llama_kv_cache: Vulkan1 KV buffer size = 56.00 MiB +0.00.995.755 I llama_kv_cache: size = 56.00 MiB ( 4096 cells, 7 layers, 1/1 seqs), K (f16): 28.00 MiB, V (f16): 28.00 MiB +0.00.995.763 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.995.763 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.995.769 I llama_kv_cache_iswa: creating SWA KV cache, size = 1024 cells +0.01.001.477 I llama_kv_cache: Vulkan1 KV buffer size = 42.00 MiB +0.01.001.944 I llama_kv_cache: size = 42.00 MiB ( 1024 cells, 21 layers, 1/1 seqs), K (f16): 21.00 MiB, V (f16): 21.00 MiB +0.01.001.950 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.01.001.951 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.01.001.961 I sched_reserve: reserving ... +0.01.003.501 I resolve_fused_ops: Flash Attention enabled +0.01.003.502 I resolve_fused_ops: resolving fused DeepSeek V4 HC support: +0.01.004.390 I resolve_fused_ops: fused DeepSeek V4 HC pre enabled +0.01.005.264 I resolve_fused_ops: fused DeepSeek V4 HC comb enabled +0.01.006.101 I resolve_fused_ops: fused DeepSeek V4 HC post enabled +0.01.016.530 I sched_reserve: Vulkan1 compute buffer size = 56.02 MiB +0.01.016.532 I sched_reserve: Vulkan_Host compute buffer size = 13.02 MiB +0.01.016.534 I sched_reserve: graph: nodes = 986, splits = 2, input objects = 4, input tensors = 9 +0.01.016.535 I sched_reserve: reserve took 14.57 ms, sched copies = 1 +0.01.016.629 I cmn init: llama threadpool init, n_threads = 12 +0.01.016.643 I cmn common_init_: warming up the model with an empty run - please wait ... (--no-warmup to disable) +0.01.069.493 I srv load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false' +0.01.069.517 I spec common_specu: no implementations specified for speculative decoding +0.01.069.518 I slot load_model: id 0 | task -1 | new slot, n_ctx = 4096 +0.01.069.616 I srv load_model: prompt cache is disabled - use `--cache-ram N` to enable it +0.01.069.617 I srv load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391 +0.01.069.618 I srv load_model: context checkpoints enabled, max = 32, min spacing = 8192 +0.01.069.640 W srv init: --cache-idle-slots requires --cache-ram, disabling +0.01.080.061 I srv init: init: chat template, example_format: '<｜start▁of▁sentence｜><|System|> +you are a helpful assistant. + +You are a helpful assistant<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>Hello<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>Hi there<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>How are you?<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>' +0.01.087.240 I srv init: init: chat template, thinking = 0 +0.01.087.251 I srv init: preserve_reasoning kwarg: true +0.01.087.251 W srv init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve) +0.01.087.274 I srv llama_server: model loaded +0.01.087.277 I srv llama_server: listening on http://127.0.0.1:51689 +0.01.087.355 I srv update_slots: all slots are idle +0.01.172.029 I srv server_strea: conv_id= (empty=1) +0.01.172.249 I srv operator(): chat format: peg-native +0.01.172.386 I slot get_availabl: id 0 | task -1 | - skipping, slot is empty +0.01.172.390 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = -1 +0.01.172.464 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.172.491 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.172.494 I slot launch_slot_: id 0 | task 0 | processing task, is_child = 0 +0.01.172.513 I slot operator(): id 0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 28 +0.01.172.522 I slot operator(): id 0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.184.969 I slot operator(): id 0 | task 0 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.223.896 I slot create_check: id 0 | task 0 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.01.227.729 I slot operator(): id 0 | task 0 | cached n_tokens = 24, memory_seq_rm [24, end) +0.01.227.770 I slot init_sampler: id 0 | task 0 | init sampler, took 0.01 ms, tokens: text = 28, total = 28 +0.01.271.380 I slot create_check: id 0 | task 0 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 23, n_tokens = 24, size = 0.985 MiB) +0.01.317.432 I slot print_timing: id 0 | task 0 | prompt eval time = 123.10 ms / 28 tokens ( 4.40 ms per token, 227.46 tokens per second) +0.01.317.440 I slot print_timing: id 0 | task 0 | eval time = 21.77 ms / 2 tokens ( 21.77 ms per token, 45.92 tokens per second) +0.01.317.442 I slot print_timing: id 0 | task 0 | total time = 144.88 ms / 30 tokens +0.01.317.443 I slot print_timing: id 0 | task 0 | graphs reused = 1 +0.01.317.485 I slot release: id 0 | task 0 | stop processing: n_tokens = 29, truncated = 0 +0.01.317.504 I srv update_slots: all slots are idle +0.01.349.316 I srv server_strea: conv_id= (empty=1) +0.01.350.891 I srv operator(): chat format: peg-native +0.01.350.994 I slot get_availabl: id 0 | task -1 | - checking sim = 0.005 (12/2463) > 0.100 +0.01.350.995 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = 1103594 +0.01.351.175 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.351.185 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.351.187 I slot launch_slot_: id 0 | task 5 | processing task, is_child = 0 +0.01.351.194 I slot operator(): id 0 | task 5 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.01.351.197 I slot operator(): id 0 | task 5 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.01.351.200 I slot operator(): id 0 | task 5 | erased invalidated context checkpoint (pos_min = 0, pos_max = 23, n_tokens = 24, n_swa = 512, pos_next = 0, size = 0.985 MiB) +0.01.351.202 I slot operator(): id 0 | task 5 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.354.280 I slot operator(): id 0 | task 5 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.394.572 I slot create_check: id 0 | task 5 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.02.659.565 I slot operator(): id 0 | task 5 | cached n_tokens = 1947, memory_seq_rm [1947, end) +0.03.152.589 I slot create_check: id 0 | task 5 | created context checkpoint 2 of 32 (pos_min = 923, pos_max = 1946, n_tokens = 1947, size = 21.006 MiB) +0.03.161.470 I slot operator(): id 0 | task 5 | cached n_tokens = 2459, memory_seq_rm [2459, end) +0.03.162.022 I slot init_sampler: id 0 | task 5 | init sampler, took 0.50 ms, tokens: text = 2463, total = 2463 +0.03.758.090 I slot create_check: id 0 | task 5 | created context checkpoint 3 of 32 (pos_min = 1435, pos_max = 2458, n_tokens = 2459, size = 21.006 MiB) +0.03.868.732 I slot print_timing: id 0 | task 5 | prompt eval time = 2436.26 ms / 2463 tokens ( 0.99 ms per token, 1010.98 tokens per second) +0.03.868.742 I slot print_timing: id 0 | task 5 | eval time = 81.26 ms / 5 tokens ( 20.32 ms per token, 49.22 tokens per second) +0.03.868.744 I slot print_timing: id 0 | task 5 | total time = 2517.52 ms / 2468 tokens +0.03.868.745 I slot print_timing: id 0 | task 5 | graphs reused = 4 +0.03.868.833 I slot release: id 0 | task 5 | stop processing: n_tokens = 2467, truncated = 0 +0.03.868.853 I srv update_slots: all slots are idle +0.03.922.265 I srv server_strea: conv_id= (empty=1) +0.03.925.875 I srv operator(): chat format: peg-native +0.03.926.114 I slot get_availabl: id 0 | task -1 | - checking sim = 1.000 (2463/2463) > 0.100 +0.03.926.118 I slot get_availabl: id 0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998 +0.03.926.415 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.03.926.432 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.03.926.435 I slot launch_slot_: id 0 | task 14 | processing task, is_child = 0 +0.03.926.446 I slot operator(): id 0 | task 14 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.03.926.451 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.03.926.455 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 923, pos_max = 1946, n_tokens = 1947, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.03.927.796 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 1435, pos_max = 2458, n_tokens = 2459, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.03.928.983 I slot operator(): id 0 | task 14 | cached n_tokens = 0, memory_seq_rm [0, end) +0.03.933.487 I slot operator(): id 0 | task 14 | cached n_tokens = 10, memory_seq_rm [10, end) +0.03.972.810 I slot create_check: id 0 | task 14 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.05.238.221 I slot operator(): id 0 | task 14 | cached n_tokens = 1947, memory_seq_rm [1947, end) +0.05.734.261 I slot create_check: id 0 | task 14 | created context checkpoint 2 of 32 (pos_min = 923, pos_max = 1946, n_tokens = 1947, size = 21.006 MiB) +0.05.745.360 I slot operator(): id 0 | task 14 | cached n_tokens = 2459, memory_seq_rm [2459, end) +0.05.745.780 I slot init_sampler: id 0 | task 14 | init sampler, took 0.37 ms, tokens: text = 2463, total = 2463 +0.06.346.715 I slot create_check: id 0 | task 14 | created context checkpoint 3 of 32 (pos_min = 1435, pos_max = 2458, n_tokens = 2459, size = 21.006 MiB) +0.06.467.153 I slot print_timing: id 0 | task 14 | prompt eval time = 2453.52 ms / 2463 tokens ( 1.00 ms per token, 1003.87 tokens per second) +0.06.467.163 I slot print_timing: id 0 | task 14 | eval time = 87.17 ms / 5 tokens ( 21.79 ms per token, 45.89 tokens per second) +0.06.467.165 I slot print_timing: id 0 | task 14 | total time = 2540.68 ms / 2468 tokens +0.06.467.167 I slot print_timing: id 0 | task 14 | graphs reused = 7 +0.06.467.277 I slot release: id 0 | task 14 | stop processing: n_tokens = 2467, truncated = 0 +0.06.467.330 I srv update_slots: all slots are idle +0.06.551.820 I srv server_strea: conv_id= (empty=1) +0.06.552.033 I srv operator(): chat format: peg-native +0.06.552.197 I slot get_availabl: id 0 | task -1 | - checking sim = 0.444 (12/27) > 0.100 +0.06.552.203 I slot get_availabl: id 0 | task -1 | selected slot by LCP similarity, f_sim_best = 0.444 (> 0.100 thold), f_keep = 0.005 +0.06.552.628 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.06.552.651 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.06.552.654 I slot launch_slot_: id 0 | task 23 | processing task, is_child = 0 +0.06.552.667 I slot operator(): id 0 | task 23 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 27 +0.06.552.671 I slot operator(): id 0 | task 23 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.06.552.675 I slot operator(): id 0 | task 23 | erased invalidated context checkpoint (pos_min = 923, pos_max = 1946, n_tokens = 1947, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.06.553.857 I slot operator(): id 0 | task 23 | erased invalidated context checkpoint (pos_min = 1435, pos_max = 2458, n_tokens = 2459, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.06.554.905 I slot operator(): id 0 | task 23 | cached n_tokens = 0, memory_seq_rm [0, end) +0.06.560.851 I slot operator(): id 0 | task 23 | cached n_tokens = 10, memory_seq_rm [10, end) +0.06.606.029 I slot create_check: id 0 | task 23 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.06.610.057 I slot operator(): id 0 | task 23 | cached n_tokens = 23, memory_seq_rm [23, end) +0.06.610.122 I slot init_sampler: id 0 | task 23 | init sampler, took 0.01 ms, tokens: text = 27, total = 27 +0.06.656.097 I slot create_check: id 0 | task 23 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 22, n_tokens = 23, size = 0.944 MiB) diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v3_server_restart.log b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v3_server_restart.log new file mode 100644 index 000000000..8de8a7722 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v3_server_restart.log @@ -0,0 +1,211 @@ +0.00.001.390 I srv llama_server: initializing ... +0.00.238.298 I cmn common_param: common_params_print_info: build 11124 (1a679828f) with Clang 20.1.8 for Windows x86_64 +0.00.238.305 I cmn common_param: common_params_print_info: verbosity = 4 (adjust with the `-lv N` CLI arg) +0.00.238.306 I cmn common_param: device_info: +0.00.242.958 I cmn common_param: - Vulkan0 : NVIDIA GeForce RTX 5090 Laptop GPU (24137 MiB, 23369 MiB free) +0.00.247.610 I cmn common_param: - Vulkan1 : AMD Radeon(TM) 890M Graphics (48505 MiB, 46080 MiB free) +0.00.247.620 I cmn common_param: - CPU : AMD Ryzen AI 9 HX 370 w/ Radeon 890M (64633 MiB, 26641 MiB free) +0.00.247.680 I cmn common_param: system_info: n_threads = 12 (n_threads_batch = 12) / 24 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | +0.00.247.761 I srv init_listene: running without SSL +0.00.247.834 I srv init_listene: The UI is disabled +0.00.247.834 I srv init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable +0.00.247.839 I srv init: using 23 threads for HTTP server +0.00.247.912 W srv llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655) +0.00.260.525 I srv load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.260.531 I srv load_model: local path 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.285.906 I llama_model_loader: loaded meta data with 37 key-value pairs and 226 tensors from C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf (version GGUF V3 (latest)) +0.00.285.919 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. +0.00.285.926 I llama_model_loader: - kv 0: general.architecture str = spark2_5 +0.00.285.927 I llama_model_loader: - kv 1: general.type str = model +0.00.285.930 I llama_model_loader: - kv 2: general.sampling.top_k i32 = 0 +0.00.285.935 I llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000 +0.00.285.936 I llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000 +0.00.285.937 I llama_model_loader: - kv 5: general.name str = Hf_Format +0.00.285.937 I llama_model_loader: - kv 6: general.size_label str = 1.7B +0.00.285.940 I llama_model_loader: - kv 7: spark2_5.block_count u32 = 28 +0.00.285.941 I llama_model_loader: - kv 8: spark2_5.context_length u32 = 1048576 +0.00.285.941 I llama_model_loader: - kv 9: spark2_5.embedding_length u32 = 2048 +0.00.285.942 I llama_model_loader: - kv 10: spark2_5.feed_forward_length u32 = 6656 +0.00.285.942 I llama_model_loader: - kv 11: spark2_5.attention.head_count u32 = 8 +0.00.285.943 I llama_model_loader: - kv 12: spark2_5.attention.head_count_kv u32 = 2 +0.00.285.944 I llama_model_loader: - kv 13: spark2_5.rope.freq_base f32 = 5000000.000000 +0.00.285.945 I llama_model_loader: - kv 14: spark2_5.rope.freq_base_swa f32 = 10000.000000 +0.00.285.947 I llama_model_loader: - kv 15: spark2_5.attention.layer_norm_rms_epsilon f32 = 0.000001 +0.00.285.947 I llama_model_loader: - kv 16: spark2_5.attention.key_length u32 = 256 +0.00.285.947 I llama_model_loader: - kv 17: spark2_5.attention.value_length u32 = 256 +0.00.285.948 I llama_model_loader: - kv 18: spark2_5.vocab_size u32 = 131072 +0.00.285.948 I llama_model_loader: - kv 19: spark2_5.attention.sliding_window u32 = 512 +0.00.285.968 I llama_model_loader: - kv 20: spark2_5.attention.sliding_window_pattern arr[bool,28] = [true, true, true, false, true, true,... +0.00.285.969 I llama_model_loader: - kv 21: spark2_5.rope.dimension_count u32 = 64 +0.00.285.970 I llama_model_loader: - kv 22: spark2_5.rope.dimension_count_swa u32 = 256 +0.00.285.970 I llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 +0.00.285.970 I llama_model_loader: - kv 24: tokenizer.ggml.pre str = spark2_5 +0.00.302.958 I llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,131072] = ["<｜start▁of▁sentence｜>", "<... +0.00.307.973 I llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,131072] = [3, 3, 3, 4, 4, 3, 3, 3, 4, 4, 4, 3, ... +0.00.324.596 I llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,130716] = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "Ġ... +0.00.324.607 I llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 0 +0.00.324.608 I llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 1 +0.00.324.609 I llama_model_loader: - kv 30: tokenizer.ggml.unknown_token_id u32 = 5 +0.00.324.609 I llama_model_loader: - kv 31: tokenizer.ggml.padding_token_id u32 = 2 +0.00.324.611 I llama_model_loader: - kv 32: tokenizer.ggml.add_bos_token bool = false +0.00.324.611 I llama_model_loader: - kv 33: tokenizer.ggml.add_eos_token bool = false +0.00.324.616 I llama_model_loader: - kv 34: tokenizer.chat_template str = {%- if not messages %}\n {{- raise_... +0.00.324.617 I llama_model_loader: - kv 35: general.quantization_version u32 = 2 +0.00.324.617 I llama_model_loader: - kv 36: general.file_type u32 = 15 +0.00.324.626 I llama_model_loader: - type f32: 57 tensors +0.00.324.627 I llama_model_loader: - type q4_K: 140 tensors +0.00.324.627 I llama_model_loader: - type q6_K: 29 tensors +0.00.324.632 I print_info: file format = GGUF V3 (latest) +0.00.324.633 I print_info: file type = Q4_K - Medium +0.00.324.636 I print_info: file size = 1.03 GiB (5.16 BPW) +0.00.329.166 I llama_prepare_model_devices: using device Vulkan1 (AMD Radeon(TM) 890M Graphics) (unknown id) - 46080 MiB free +0.00.400.441 I load: 86 unused tokens +0.00.428.364 I load: printing all EOG tokens: +0.00.428.372 I load: - 1 ('<｜end▁of▁sentence｜>') +0.00.428.517 I load: special tokens cache size = 113 +0.00.444.657 I load: token to piece cache size = 0.7975 MB +0.00.444.690 I print_info: arch = spark2_5 +0.00.444.692 I print_info: vocab_only = 0 +0.00.444.692 I print_info: no_alloc = 0 +0.00.444.692 I print_info: n_ctx_train = 1048576 +0.00.444.693 I print_info: n_embd_inp = 2048 +0.00.444.693 I print_info: n_embd = 2048 +0.00.444.694 I print_info: n_embd_out = 2048 +0.00.444.694 I print_info: n_layer = 28 +0.00.444.694 I print_info: n_layer_all = 28 +0.00.444.712 I print_info: n_head = 8 +0.00.444.713 I print_info: n_head_kv = 2 +0.00.444.713 I print_info: n_rot = 64 +0.00.444.714 I print_info: n_swa = 512 +0.00.444.714 I print_info: is_swa_any = 1 +0.00.444.714 I print_info: non_causal_type = 0 +0.00.444.715 I print_info: n_embd_head_k = 256 +0.00.444.715 I print_info: n_embd_head_v = 256 +0.00.444.716 I print_info: n_gqa = 4 +0.00.444.718 I print_info: n_embd_k_gqa = 512 +0.00.444.719 I print_info: n_embd_v_gqa = 512 +0.00.444.721 I print_info: f_norm_eps = 0.0e+00 +0.00.444.722 I print_info: f_norm_rms_eps = 1.0e-06 +0.00.444.722 I print_info: f_clamp_kqv = 0.0e+00 +0.00.444.723 I print_info: f_max_alibi_bias = 0.0e+00 +0.00.444.723 I print_info: f_logit_scale = 0.0e+00 +0.00.444.723 I print_info: f_attn_scale = 0.0e+00 +0.00.444.724 I print_info: f_attn_value_scale = 0.0000 +0.00.444.725 I print_info: n_ff = 6656 +0.00.444.726 I print_info: n_expert = 0 +0.00.444.726 I print_info: n_expert_used = 0 +0.00.444.728 I print_info: n_expert_groups = 0 +0.00.444.728 I print_info: n_group_used = 0 +0.00.444.728 I print_info: causal attn = 1 +0.00.444.729 I print_info: pooling type = -1 +0.00.444.729 I print_info: rope type = 2 +0.00.444.730 I print_info: rope scaling = linear +0.00.444.730 I print_info: freq_base_train = 5000000.0 +0.00.444.732 I print_info: freq_scale_train = 1 +0.00.444.733 I print_info: freq_base_swa = 10000.0 +0.00.444.733 I print_info: freq_scale_swa = 1 +0.00.444.734 I print_info: n_embd_head_k_swa = 256 +0.00.444.734 I print_info: n_embd_head_v_swa = 256 +0.00.444.734 I print_info: n_rot_swa = 256 +0.00.444.734 I print_info: n_ctx_orig_yarn = 1048576 +0.00.444.735 I print_info: rope_yarn_log_mul = 0.0000 +0.00.444.735 I print_info: rope_finetuned = unknown +0.00.444.736 I print_info: model type = 1.7B +0.00.444.737 I print_info: model params = 1.71 B +0.00.444.737 I print_info: general.name = Hf_Format +0.00.444.738 I print_info: vocab type = BPE +0.00.444.739 I print_info: n_vocab = 131072 +0.00.444.739 I print_info: n_merges = 130716 +0.00.444.740 I print_info: BOS token = 0 '<｜start▁of▁sentence｜>' +0.00.444.740 I print_info: EOS token = 1 '<｜end▁of▁sentence｜>' +0.00.444.740 I print_info: EOT token = 1 '<｜end▁of▁sentence｜>' +0.00.444.741 I print_info: UNK token = 5 '' +0.00.444.741 I print_info: PAD token = 2 '<｜▁pad▁｜>' +0.00.444.742 I print_info: LF token = 198 'Ċ' +0.00.444.742 I print_info: FIM PRE token = 11 '<｜fim▁begin｜>' +0.00.444.743 I print_info: FIM SUF token = 12 '<｜fim▁hole｜>' +0.00.444.743 I print_info: FIM MID token = 13 '<｜fim▁end｜>' +0.00.444.743 I print_info: EOG token = 1 '<｜end▁of▁sentence｜>' +0.00.444.744 I print_info: max token length = 210 +0.00.453.742 I load_tensors: loading model tensors, this can take a while... (load_mode = none) +0.00.651.955 I load_tensors: offloading output layer to GPU +0.00.651.962 I load_tensors: offloading 27 repeating layers to GPU +0.00.651.962 I load_tensors: offloaded 29/29 layers to GPU +0.00.651.965 I load_tensors: Vulkan1 model buffer size = 1051.02 MiB +0.00.651.967 I load_tensors: Vulkan_Host model buffer size = 210.00 MiB +0.01.024.719 I cmn common_init_: added <｜end▁of▁sentence｜> logit bias = -inf +0.01.025.167 I llama_context: constructing llama_context +0.01.025.173 I llama_context: n_seq_max = 1 +0.01.025.173 I llama_context: n_ctx = 4096 +0.01.025.173 I llama_context: n_ctx_seq = 4096 +0.01.025.174 I llama_context: n_batch = 2048 +0.01.025.174 I llama_context: n_ubatch = 512 +0.01.025.175 I llama_context: causal_attn = 1 +0.01.025.176 I llama_context: flash_attn = auto +0.01.025.176 I llama_context: kv_unified = false +0.01.025.179 I llama_context: freq_base = 5000000.0 +0.01.025.181 I llama_context: freq_scale = 1 +0.01.025.181 I llama_context: n_rs_seq = 0 +0.01.025.182 I llama_context: n_outputs_max = 1 +0.01.025.182 I llama_context: n_outputs_max_per_seq = 1 +0.01.025.183 I llama_context: n_ctx_seq (4096) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized +0.01.025.496 I llama_context: Vulkan_Host output buffer size = 0.50 MiB +0.01.025.505 I llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells +0.01.030.718 I llama_kv_cache: Vulkan1 KV buffer size = 56.00 MiB +0.01.031.489 I llama_kv_cache: size = 56.00 MiB ( 4096 cells, 7 layers, 1/1 seqs), K (f16): 28.00 MiB, V (f16): 28.00 MiB +0.01.031.500 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.01.031.500 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.01.031.509 I llama_kv_cache_iswa: creating SWA KV cache, size = 1024 cells +0.01.035.802 I llama_kv_cache: Vulkan1 KV buffer size = 42.00 MiB +0.01.036.295 I llama_kv_cache: size = 42.00 MiB ( 1024 cells, 21 layers, 1/1 seqs), K (f16): 21.00 MiB, V (f16): 21.00 MiB +0.01.036.304 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.01.036.304 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.01.036.318 I sched_reserve: reserving ... +0.01.037.998 I resolve_fused_ops: Flash Attention enabled +0.01.037.999 I resolve_fused_ops: resolving fused DeepSeek V4 HC support: +0.01.038.974 I resolve_fused_ops: fused DeepSeek V4 HC pre enabled +0.01.039.873 I resolve_fused_ops: fused DeepSeek V4 HC comb enabled +0.01.041.007 I resolve_fused_ops: fused DeepSeek V4 HC post enabled +0.01.050.055 I sched_reserve: Vulkan1 compute buffer size = 56.02 MiB +0.01.050.057 I sched_reserve: Vulkan_Host compute buffer size = 13.02 MiB +0.01.050.059 I sched_reserve: graph: nodes = 986, splits = 2, input objects = 4, input tensors = 9 +0.01.050.060 I sched_reserve: reserve took 13.73 ms, sched copies = 1 +0.01.050.171 I cmn init: llama threadpool init, n_threads = 12 +0.01.050.187 I cmn common_init_: warming up the model with an empty run - please wait ... (--no-warmup to disable) +0.01.104.347 I srv load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false' +0.01.104.373 I spec common_specu: no implementations specified for speculative decoding +0.01.104.375 I slot load_model: id 0 | task -1 | new slot, n_ctx = 4096 +0.01.104.480 I srv load_model: prompt cache is disabled - use `--cache-ram N` to enable it +0.01.104.481 I srv load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391 +0.01.104.481 I srv load_model: context checkpoints enabled, max = 32, min spacing = 8192 +0.01.104.520 W srv init: --cache-idle-slots requires --cache-ram, disabling +0.01.114.947 I srv init: init: chat template, example_format: '<｜start▁of▁sentence｜><|System|> +you are a helpful assistant. + +You are a helpful assistant<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>Hello<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>Hi there<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>How are you?<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>' +0.01.124.138 I srv init: init: chat template, thinking = 0 +0.01.124.149 I srv init: preserve_reasoning kwarg: true +0.01.124.150 W srv init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve) +0.01.124.176 I srv llama_server: model loaded +0.01.124.179 I srv llama_server: listening on http://127.0.0.1:51717 +0.01.124.254 I srv update_slots: all slots are idle +0.01.188.984 I srv server_strea: conv_id= (empty=1) +0.01.189.206 I srv operator(): chat format: peg-native +0.01.189.307 I slot get_availabl: id 0 | task -1 | - skipping, slot is empty +0.01.189.309 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = -1 +0.01.189.354 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.189.366 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.189.367 I slot launch_slot_: id 0 | task 0 | processing task, is_child = 0 +0.01.189.377 I slot operator(): id 0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 28 +0.01.189.382 I slot operator(): id 0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.198.472 I slot operator(): id 0 | task 0 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.239.897 I slot create_check: id 0 | task 0 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.01.243.864 I slot operator(): id 0 | task 0 | cached n_tokens = 24, memory_seq_rm [24, end) +0.01.243.895 I slot init_sampler: id 0 | task 0 | init sampler, took 0.00 ms, tokens: text = 28, total = 28 +0.01.289.828 I slot create_check: id 0 | task 0 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 23, n_tokens = 24, size = 0.985 MiB) +0.01.335.955 I slot print_timing: id 0 | task 0 | prompt eval time = 126.17 ms / 28 tokens ( 4.51 ms per token, 221.92 tokens per second) +0.01.335.962 I \ No newline at end of file diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v4_report.json b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v4_report.json new file mode 100644 index 000000000..97111cf92 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v4_report.json @@ -0,0 +1,254 @@ +{ + "scope": "complete Spark GGUF text requests through Omni StageRuntime/StagePool and external llama.cpp backend", + "model_file": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "device": "Vulkan1", + "memory_budget": { + "capacities": { + "host_ram": 12884901888 + }, + "demands": { + "host_ram": 4294967296 + } + }, + "warmup_count": 0, + "measured_count": 1, + "startup_s": 2.8167914000805467, + "execution_plan": { + "backend": "external.llamacpp.text.v1", + "stage_id": 0, + "worker_generation": "e96bf7f656d24b39bcf5744e577cc073", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "requested_device": "Vulkan1", + "expected_device_name": "AMD Radeon(TM) 890M Graphics", + "layer_assignments": 0, + "offloaded_layers": [ + "29", + "29" + ], + "model_buffer_devices": [ + "Vulkan1" + ], + "all_model_buffers": [ + "Vulkan1", + "Vulkan_Host" + ], + "cpu_only_layout": false, + "worker_pid": 62780, + "loaded_rss_bytes": 1643917312, + "reserved_bytes": { + "host_ram": 4294967296 + }, + "memory_pool": "host_ram", + "memory_overhead_bytes": 2147483648, + "context_tokens": 4096, + "max_new_tokens": 96, + "max_io_bytes": 1048576, + "request_capacity": 1, + "stateful_session": "llama.cpp-owned; reset for each complete request", + "evidence": "B" + }, + "checks": [ + { + "case": "short", + "answer": "Paris", + "expected": "Paris", + "correct": true, + "wall_s": 0.16570489993318915, + "stage_event": { + "request_id": "short-688754296000000", + "stage_id": 0, + "epoch": 1, + "seq": 1, + "kind": "text", + "worker_generation": "e96bf7f656d24b39bcf5744e577cc073", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 0.14932229998521507, + "usage": { + "completion_tokens": 2, + "prompt_tokens": 28, + "total_tokens": 30, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + }, + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 2.198886900092475, + "stage_event": { + "request_id": "inventory_120-688754468000000", + "stage_id": 0, + "epoch": 2, + "seq": 1, + "kind": "text", + "worker_generation": "e96bf7f656d24b39bcf5744e577cc073", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 2.174214500002563, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "warmups": [], + "measured": [ + { + "case": "inventory_120", + "answer": "1511", + "expected": "1511", + "correct": true, + "wall_s": 2.210317499935627, + "stage_event": { + "request_id": "inventory_120-688756656000000", + "stage_id": 0, + "epoch": 3, + "seq": 1, + "kind": "text", + "worker_generation": "e96bf7f656d24b39bcf5744e577cc073", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 2.1807818999513984, + "usage": { + "completion_tokens": 5, + "prompt_tokens": 2463, + "total_tokens": 2468, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + } + ], + "nearest_rank_p50_wall_s": 2.210317499935627, + "nearest_rank_p95_wall_s": 2.210317499935627, + "abort_check": { + "abort_case": "inventory_120", + "server_tasks_before_abort_request": 3, + "server_tasks_after_abort_request_start": 4, + "stale_output": false, + "worker_exited": true, + "ledger_after_abort": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + } + }, + "restart_check": { + "first_runtime_ledger_after_shutdown": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + }, + "fresh_startup_s": 2.283025399898179, + "fresh_server_log": "\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\vllm-omni-edge\\benchmarks\\edge_harness\\results\\e2e_expansion_20260924\\evidence\\spark_omni_windows_restart\\radeon_v4_server_restart.log", + "fresh_execution_plan": { + "backend": "external.llamacpp.text.v1", + "stage_id": 0, + "worker_generation": "171ce688813848b38ffa8f66ec236618", + "model_sha256": "902bde2522394954ac17821b3e5fd0df02defbc6944f122253f2580acf0503f4", + "server_sha256": "9ffc5919acb4cb43c7be5f3053b59014cce70a87d3a801fcad49c4f459984f52", + "model_alias": "C:\\Users\\zhout\\w2\\models\\Spark-X2.5-1.7B-Q4_K_M.gguf", + "requested_device": "Vulkan1", + "expected_device_name": "AMD Radeon(TM) 890M Graphics", + "layer_assignments": 0, + "offloaded_layers": [ + "29", + "29" + ], + "model_buffer_devices": [ + "Vulkan1" + ], + "all_model_buffers": [ + "Vulkan1", + "Vulkan_Host" + ], + "cpu_only_layout": false, + "worker_pid": 45788, + "loaded_rss_bytes": 1644441600, + "reserved_bytes": { + "host_ram": 4294967296 + }, + "memory_pool": "host_ram", + "memory_overhead_bytes": 2147483648, + "context_tokens": 4096, + "max_new_tokens": 96, + "max_io_bytes": 1048576, + "request_capacity": 1, + "stateful_session": "llama.cpp-owned; reset for each complete request", + "evidence": "B" + }, + "fresh_request": { + "case": "short", + "answer": "Paris", + "expected": "Paris", + "correct": true, + "wall_s": 0.14121619996149093, + "stage_event": { + "request_id": "short-688761640000000", + "stage_id": 0, + "epoch": 1, + "seq": 1, + "kind": "text", + "worker_generation": "171ce688813848b38ffa8f66ec236618", + "buffers": [], + "terminal": true, + "error": null + }, + "metrics": { + "llamacpp_wall_s": 0.13696709997020662, + "usage": { + "completion_tokens": 2, + "prompt_tokens": 28, + "total_tokens": 30, + "prompt_tokens_details": { + "cached_tokens": 0 + } + } + } + }, + "late_output": false + }, + "status": "passed", + "ledger_after_shutdown": { + "capacities": { + "host_ram": 12884901888 + }, + "reserved": { + "host_ram": 0 + }, + "owners": [], + "quarantined": [] + } +} diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v4_server.log b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v4_server.log new file mode 100644 index 000000000..8549d159e --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v4_server.log @@ -0,0 +1,289 @@ +0.00.000.910 I srv llama_server: initializing ... +0.00.217.038 I cmn common_param: common_params_print_info: build 11124 (1a679828f) with Clang 20.1.8 for Windows x86_64 +0.00.217.045 I cmn common_param: common_params_print_info: verbosity = 4 (adjust with the `-lv N` CLI arg) +0.00.217.046 I cmn common_param: device_info: +0.00.221.208 I cmn common_param: - Vulkan0 : NVIDIA GeForce RTX 5090 Laptop GPU (24137 MiB, 23369 MiB free) +0.00.226.686 I cmn common_param: - Vulkan1 : AMD Radeon(TM) 890M Graphics (48505 MiB, 46080 MiB free) +0.00.226.696 I cmn common_param: - CPU : AMD Ryzen AI 9 HX 370 w/ Radeon 890M (64633 MiB, 26120 MiB free) +0.00.226.763 I cmn common_param: system_info: n_threads = 12 (n_threads_batch = 12) / 24 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | +0.00.226.850 I srv init_listene: running without SSL +0.00.226.923 I srv init_listene: The UI is disabled +0.00.226.923 I srv init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable +0.00.226.928 I srv init: using 23 threads for HTTP server +0.00.226.998 W srv llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655) +0.00.239.786 I srv load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.239.794 I srv load_model: local path 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.262.609 I llama_model_loader: loaded meta data with 37 key-value pairs and 226 tensors from C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf (version GGUF V3 (latest)) +0.00.262.622 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. +0.00.262.628 I llama_model_loader: - kv 0: general.architecture str = spark2_5 +0.00.262.629 I llama_model_loader: - kv 1: general.type str = model +0.00.262.633 I llama_model_loader: - kv 2: general.sampling.top_k i32 = 0 +0.00.262.640 I llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000 +0.00.262.641 I llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000 +0.00.262.641 I llama_model_loader: - kv 5: general.name str = Hf_Format +0.00.262.642 I llama_model_loader: - kv 6: general.size_label str = 1.7B +0.00.262.645 I llama_model_loader: - kv 7: spark2_5.block_count u32 = 28 +0.00.262.645 I llama_model_loader: - kv 8: spark2_5.context_length u32 = 1048576 +0.00.262.646 I llama_model_loader: - kv 9: spark2_5.embedding_length u32 = 2048 +0.00.262.646 I llama_model_loader: - kv 10: spark2_5.feed_forward_length u32 = 6656 +0.00.262.646 I llama_model_loader: - kv 11: spark2_5.attention.head_count u32 = 8 +0.00.262.647 I llama_model_loader: - kv 12: spark2_5.attention.head_count_kv u32 = 2 +0.00.262.648 I llama_model_loader: - kv 13: spark2_5.rope.freq_base f32 = 5000000.000000 +0.00.262.649 I llama_model_loader: - kv 14: spark2_5.rope.freq_base_swa f32 = 10000.000000 +0.00.262.651 I llama_model_loader: - kv 15: spark2_5.attention.layer_norm_rms_epsilon f32 = 0.000001 +0.00.262.651 I llama_model_loader: - kv 16: spark2_5.attention.key_length u32 = 256 +0.00.262.651 I llama_model_loader: - kv 17: spark2_5.attention.value_length u32 = 256 +0.00.262.652 I llama_model_loader: - kv 18: spark2_5.vocab_size u32 = 131072 +0.00.262.652 I llama_model_loader: - kv 19: spark2_5.attention.sliding_window u32 = 512 +0.00.262.672 I llama_model_loader: - kv 20: spark2_5.attention.sliding_window_pattern arr[bool,28] = [true, true, true, false, true, true,... +0.00.262.673 I llama_model_loader: - kv 21: spark2_5.rope.dimension_count u32 = 64 +0.00.262.673 I llama_model_loader: - kv 22: spark2_5.rope.dimension_count_swa u32 = 256 +0.00.262.674 I llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 +0.00.262.674 I llama_model_loader: - kv 24: tokenizer.ggml.pre str = spark2_5 +0.00.278.603 I llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,131072] = ["<｜start▁of▁sentence｜>", "<... +0.00.283.726 I llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,131072] = [3, 3, 3, 4, 4, 3, 3, 3, 4, 4, 4, 3, ... +0.00.299.936 I llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,130716] = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "Ġ... +0.00.299.945 I llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 0 +0.00.299.946 I llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 1 +0.00.299.946 I llama_model_loader: - kv 30: tokenizer.ggml.unknown_token_id u32 = 5 +0.00.299.947 I llama_model_loader: - kv 31: tokenizer.ggml.padding_token_id u32 = 2 +0.00.299.947 I llama_model_loader: - kv 32: tokenizer.ggml.add_bos_token bool = false +0.00.299.948 I llama_model_loader: - kv 33: tokenizer.ggml.add_eos_token bool = false +0.00.299.953 I llama_model_loader: - kv 34: tokenizer.chat_template str = {%- if not messages %}\n {{- raise_... +0.00.299.953 I llama_model_loader: - kv 35: general.quantization_version u32 = 2 +0.00.299.954 I llama_model_loader: - kv 36: general.file_type u32 = 15 +0.00.299.960 I llama_model_loader: - type f32: 57 tensors +0.00.299.960 I llama_model_loader: - type q4_K: 140 tensors +0.00.299.961 I llama_model_loader: - type q6_K: 29 tensors +0.00.299.965 I print_info: file format = GGUF V3 (latest) +0.00.299.966 I print_info: file type = Q4_K - Medium +0.00.299.971 I print_info: file size = 1.03 GiB (5.16 BPW) +0.00.305.556 I llama_prepare_model_devices: using device Vulkan1 (AMD Radeon(TM) 890M Graphics) (unknown id) - 46080 MiB free +0.00.366.815 I load: 86 unused tokens +0.00.387.616 I load: printing all EOG tokens: +0.00.387.624 I load: - 1 ('<｜end▁of▁sentence｜>') +0.00.387.762 I load: special tokens cache size = 113 +0.00.403.326 I load: token to piece cache size = 0.7975 MB +0.00.403.355 I print_info: arch = spark2_5 +0.00.403.356 I print_info: vocab_only = 0 +0.00.403.356 I print_info: no_alloc = 0 +0.00.403.357 I print_info: n_ctx_train = 1048576 +0.00.403.357 I print_info: n_embd_inp = 2048 +0.00.403.358 I print_info: n_embd = 2048 +0.00.403.358 I print_info: n_embd_out = 2048 +0.00.403.358 I print_info: n_layer = 28 +0.00.403.358 I print_info: n_layer_all = 28 +0.00.403.375 I print_info: n_head = 8 +0.00.403.377 I print_info: n_head_kv = 2 +0.00.403.377 I print_info: n_rot = 64 +0.00.403.378 I print_info: n_swa = 512 +0.00.403.378 I print_info: is_swa_any = 1 +0.00.403.378 I print_info: non_causal_type = 0 +0.00.403.379 I print_info: n_embd_head_k = 256 +0.00.403.379 I print_info: n_embd_head_v = 256 +0.00.403.380 I print_info: n_gqa = 4 +0.00.403.382 I print_info: n_embd_k_gqa = 512 +0.00.403.383 I print_info: n_embd_v_gqa = 512 +0.00.403.384 I print_info: f_norm_eps = 0.0e+00 +0.00.403.386 I print_info: f_norm_rms_eps = 1.0e-06 +0.00.403.386 I print_info: f_clamp_kqv = 0.0e+00 +0.00.403.387 I print_info: f_max_alibi_bias = 0.0e+00 +0.00.403.387 I print_info: f_logit_scale = 0.0e+00 +0.00.403.387 I print_info: f_attn_scale = 0.0e+00 +0.00.403.388 I print_info: f_attn_value_scale = 0.0000 +0.00.403.389 I print_info: n_ff = 6656 +0.00.403.390 I print_info: n_expert = 0 +0.00.403.390 I print_info: n_expert_used = 0 +0.00.403.392 I print_info: n_expert_groups = 0 +0.00.403.392 I print_info: n_group_used = 0 +0.00.403.393 I print_info: causal attn = 1 +0.00.403.393 I print_info: pooling type = -1 +0.00.403.393 I print_info: rope type = 2 +0.00.403.394 I print_info: rope scaling = linear +0.00.403.395 I print_info: freq_base_train = 5000000.0 +0.00.403.396 I print_info: freq_scale_train = 1 +0.00.403.397 I print_info: freq_base_swa = 10000.0 +0.00.403.397 I print_info: freq_scale_swa = 1 +0.00.403.397 I print_info: n_embd_head_k_swa = 256 +0.00.403.398 I print_info: n_embd_head_v_swa = 256 +0.00.403.398 I print_info: n_rot_swa = 256 +0.00.403.398 I print_info: n_ctx_orig_yarn = 1048576 +0.00.403.399 I print_info: rope_yarn_log_mul = 0.0000 +0.00.403.399 I print_info: rope_finetuned = unknown +0.00.403.400 I print_info: model type = 1.7B +0.00.403.401 I print_info: model params = 1.71 B +0.00.403.401 I print_info: general.name = Hf_Format +0.00.403.402 I print_info: vocab type = BPE +0.00.403.402 I print_info: n_vocab = 131072 +0.00.403.402 I print_info: n_merges = 130716 +0.00.403.403 I print_info: BOS token = 0 '<｜start▁of▁sentence｜>' +0.00.403.404 I print_info: EOS token = 1 '<｜end▁of▁sentence｜>' +0.00.403.404 I print_info: EOT token = 1 '<｜end▁of▁sentence｜>' +0.00.403.404 I print_info: UNK token = 5 '' +0.00.403.405 I print_info: PAD token = 2 '<｜▁pad▁｜>' +0.00.403.405 I print_info: LF token = 198 'Ċ' +0.00.403.406 I print_info: FIM PRE token = 11 '<｜fim▁begin｜>' +0.00.403.406 I print_info: FIM SUF token = 12 '<｜fim▁hole｜>' +0.00.403.406 I print_info: FIM MID token = 13 '<｜fim▁end｜>' +0.00.403.407 I print_info: EOG token = 1 '<｜end▁of▁sentence｜>' +0.00.403.407 I print_info: max token length = 210 +0.00.411.383 I load_tensors: loading model tensors, this can take a while... (load_mode = none) +0.00.602.932 I load_tensors: offloading output layer to GPU +0.00.602.938 I load_tensors: offloading 27 repeating layers to GPU +0.00.602.938 I load_tensors: offloaded 29/29 layers to GPU +0.00.602.944 I load_tensors: Vulkan1 model buffer size = 1051.02 MiB +0.00.602.945 I load_tensors: Vulkan_Host model buffer size = 210.00 MiB +0.00.939.849 I cmn common_init_: added <｜end▁of▁sentence｜> logit bias = -inf +0.00.940.264 I llama_context: constructing llama_context +0.00.940.269 I llama_context: n_seq_max = 1 +0.00.940.269 I llama_context: n_ctx = 4096 +0.00.940.270 I llama_context: n_ctx_seq = 4096 +0.00.940.270 I llama_context: n_batch = 2048 +0.00.940.271 I llama_context: n_ubatch = 512 +0.00.940.271 I llama_context: causal_attn = 1 +0.00.940.272 I llama_context: flash_attn = auto +0.00.940.272 I llama_context: kv_unified = false +0.00.940.275 I llama_context: freq_base = 5000000.0 +0.00.940.278 I llama_context: freq_scale = 1 +0.00.940.278 I llama_context: n_rs_seq = 0 +0.00.940.279 I llama_context: n_outputs_max = 1 +0.00.940.279 I llama_context: n_outputs_max_per_seq = 1 +0.00.940.280 I llama_context: n_ctx_seq (4096) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized +0.00.940.570 I llama_context: Vulkan_Host output buffer size = 0.50 MiB +0.00.940.580 I llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells +0.00.945.402 I llama_kv_cache: Vulkan1 KV buffer size = 56.00 MiB +0.00.945.993 I llama_kv_cache: size = 56.00 MiB ( 4096 cells, 7 layers, 1/1 seqs), K (f16): 28.00 MiB, V (f16): 28.00 MiB +0.00.946.000 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.946.001 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.946.005 I llama_kv_cache_iswa: creating SWA KV cache, size = 1024 cells +0.00.949.595 I llama_kv_cache: Vulkan1 KV buffer size = 42.00 MiB +0.00.950.019 I llama_kv_cache: size = 42.00 MiB ( 1024 cells, 21 layers, 1/1 seqs), K (f16): 21.00 MiB, V (f16): 21.00 MiB +0.00.950.025 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.950.026 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.950.035 I sched_reserve: reserving ... +0.00.951.578 I resolve_fused_ops: Flash Attention enabled +0.00.951.579 I resolve_fused_ops: resolving fused DeepSeek V4 HC support: +0.00.952.491 I resolve_fused_ops: fused DeepSeek V4 HC pre enabled +0.00.953.348 I resolve_fused_ops: fused DeepSeek V4 HC comb enabled +0.00.954.199 I resolve_fused_ops: fused DeepSeek V4 HC post enabled +0.00.962.653 I sched_reserve: Vulkan1 compute buffer size = 56.02 MiB +0.00.962.655 I sched_reserve: Vulkan_Host compute buffer size = 13.02 MiB +0.00.962.657 I sched_reserve: graph: nodes = 986, splits = 2, input objects = 4, input tensors = 9 +0.00.962.659 I sched_reserve: reserve took 12.62 ms, sched copies = 1 +0.00.962.762 I cmn init: llama threadpool init, n_threads = 12 +0.00.962.778 I cmn common_init_: warming up the model with an empty run - please wait ... (--no-warmup to disable) +0.01.017.598 I srv load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false' +0.01.017.617 I spec common_specu: no implementations specified for speculative decoding +0.01.017.618 I slot load_model: id 0 | task -1 | new slot, n_ctx = 4096 +0.01.017.680 I srv load_model: prompt cache is disabled - use `--cache-ram N` to enable it +0.01.017.681 I srv load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391 +0.01.017.681 I srv load_model: context checkpoints enabled, max = 32, min spacing = 8192 +0.01.017.703 W srv init: --cache-idle-slots requires --cache-ram, disabling +0.01.028.655 I srv init: init: chat template, example_format: '<｜start▁of▁sentence｜><|System|> +you are a helpful assistant. + +You are a helpful assistant<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>Hello<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>Hi there<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>How are you?<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>' +0.01.035.639 I srv init: init: chat template, thinking = 0 +0.01.035.650 I srv init: preserve_reasoning kwarg: true +0.01.035.651 W srv init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve) +0.01.035.673 I srv llama_server: model loaded +0.01.035.675 I srv llama_server: listening on http://127.0.0.1:50164 +0.01.035.753 I srv update_slots: all slots are idle +0.01.238.593 I srv server_strea: conv_id= (empty=1) +0.01.238.807 I srv operator(): chat format: peg-native +0.01.238.925 I slot get_availabl: id 0 | task -1 | - skipping, slot is empty +0.01.238.927 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = -1 +0.01.238.965 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.238.980 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.238.981 I slot launch_slot_: id 0 | task 0 | processing task, is_child = 0 +0.01.238.990 I slot operator(): id 0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 28 +0.01.238.995 I slot operator(): id 0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.249.031 I slot operator(): id 0 | task 0 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.283.124 I slot create_check: id 0 | task 0 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.01.285.972 I slot operator(): id 0 | task 0 | cached n_tokens = 24, memory_seq_rm [24, end) +0.01.286.002 I slot init_sampler: id 0 | task 0 | init sampler, took 0.01 ms, tokens: text = 28, total = 28 +0.01.323.803 I slot create_check: id 0 | task 0 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 23, n_tokens = 24, size = 0.985 MiB) +0.01.364.502 I slot print_timing: id 0 | task 0 | prompt eval time = 106.53 ms / 28 tokens ( 3.80 ms per token, 262.84 tokens per second) +0.01.364.510 I slot print_timing: id 0 | task 0 | eval time = 18.97 ms / 2 tokens ( 18.97 ms per token, 52.72 tokens per second) +0.01.364.511 I slot print_timing: id 0 | task 0 | total time = 125.49 ms / 30 tokens +0.01.364.512 I slot print_timing: id 0 | task 0 | graphs reused = 1 +0.01.364.536 I slot release: id 0 | task 0 | stop processing: n_tokens = 29, truncated = 0 +0.01.364.543 I srv update_slots: all slots are idle +0.01.398.737 I srv server_strea: conv_id= (empty=1) +0.01.400.284 I srv operator(): chat format: peg-native +0.01.400.418 I slot get_availabl: id 0 | task -1 | - checking sim = 0.005 (12/2463) > 0.100 +0.01.400.421 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = 1165006 +0.01.400.762 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.400.784 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.400.787 I slot launch_slot_: id 0 | task 5 | processing task, is_child = 0 +0.01.400.798 I slot operator(): id 0 | task 5 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.01.400.803 I slot operator(): id 0 | task 5 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.01.400.807 I slot operator(): id 0 | task 5 | erased invalidated context checkpoint (pos_min = 0, pos_max = 23, n_tokens = 24, n_swa = 512, pos_next = 0, size = 0.985 MiB) +0.01.400.810 I slot operator(): id 0 | task 5 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.404.785 I slot operator(): id 0 | task 5 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.441.884 I slot create_check: id 0 | task 5 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.02.538.388 I slot operator(): id 0 | task 5 | cached n_tokens = 1947, memory_seq_rm [1947, end) +0.02.955.955 I slot create_check: id 0 | task 5 | created context checkpoint 2 of 32 (pos_min = 923, pos_max = 1946, n_tokens = 1947, size = 21.006 MiB) +0.02.960.624 I slot operator(): id 0 | task 5 | cached n_tokens = 2459, memory_seq_rm [2459, end) +0.02.960.873 I slot init_sampler: id 0 | task 5 | init sampler, took 0.23 ms, tokens: text = 2463, total = 2463 +0.03.456.981 I slot create_check: id 0 | task 5 | created context checkpoint 3 of 32 (pos_min = 1435, pos_max = 2458, n_tokens = 2459, size = 21.006 MiB) +0.03.563.279 I slot print_timing: id 0 | task 5 | prompt eval time = 2084.48 ms / 2463 tokens ( 0.85 ms per token, 1181.59 tokens per second) +0.03.563.290 I slot print_timing: id 0 | task 5 | eval time = 77.98 ms / 5 tokens ( 19.50 ms per token, 51.29 tokens per second) +0.03.563.292 I slot print_timing: id 0 | task 5 | total time = 2162.46 ms / 2468 tokens +0.03.563.293 I slot print_timing: id 0 | task 5 | graphs reused = 4 +0.03.563.381 I slot release: id 0 | task 5 | stop processing: n_tokens = 2467, truncated = 0 +0.03.563.402 I srv update_slots: all slots are idle +0.03.615.421 I srv server_strea: conv_id= (empty=1) +0.03.617.052 I srv operator(): chat format: peg-native +0.03.617.293 I slot get_availabl: id 0 | task -1 | - checking sim = 1.000 (2463/2463) > 0.100 +0.03.617.296 I slot get_availabl: id 0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998 +0.03.617.572 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.03.617.588 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.03.617.591 I slot launch_slot_: id 0 | task 14 | processing task, is_child = 0 +0.03.617.601 I slot operator(): id 0 | task 14 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.03.617.605 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.03.617.609 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 923, pos_max = 1946, n_tokens = 1947, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.03.618.467 I slot operator(): id 0 | task 14 | erased invalidated context checkpoint (pos_min = 1435, pos_max = 2458, n_tokens = 2459, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.03.619.439 I slot operator(): id 0 | task 14 | cached n_tokens = 0, memory_seq_rm [0, end) +0.03.626.247 I slot operator(): id 0 | task 14 | cached n_tokens = 10, memory_seq_rm [10, end) +0.03.664.643 I slot create_check: id 0 | task 14 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.04.731.305 I slot operator(): id 0 | task 14 | cached n_tokens = 1947, memory_seq_rm [1947, end) +0.05.157.167 I slot create_check: id 0 | task 14 | created context checkpoint 2 of 32 (pos_min = 923, pos_max = 1946, n_tokens = 1947, size = 21.006 MiB) +0.05.164.767 I slot operator(): id 0 | task 14 | cached n_tokens = 2459, memory_seq_rm [2459, end) +0.05.165.180 I slot init_sampler: id 0 | task 14 | init sampler, took 0.38 ms, tokens: text = 2463, total = 2463 +0.05.668.053 I slot create_check: id 0 | task 14 | created context checkpoint 3 of 32 (pos_min = 1435, pos_max = 2458, n_tokens = 2459, size = 21.006 MiB) +0.05.773.969 I slot print_timing: id 0 | task 14 | prompt eval time = 2078.27 ms / 2463 tokens ( 0.84 ms per token, 1185.12 tokens per second) +0.05.773.984 I slot print_timing: id 0 | task 14 | eval time = 78.06 ms / 5 tokens ( 19.52 ms per token, 51.24 tokens per second) +0.05.773.986 I slot print_timing: id 0 | task 14 | total time = 2156.34 ms / 2468 tokens +0.05.773.988 I slot print_timing: id 0 | task 14 | graphs reused = 7 +0.05.774.099 I slot release: id 0 | task 14 | stop processing: n_tokens = 2467, truncated = 0 +0.05.774.138 I srv update_slots: all slots are idle +0.05.859.535 I srv server_strea: conv_id= (empty=1) +0.05.862.169 I srv operator(): chat format: peg-native +0.05.862.356 I slot get_availabl: id 0 | task -1 | - checking sim = 1.000 (2463/2463) > 0.100 +0.05.862.359 I slot get_availabl: id 0 | task -1 | selected slot by LCP similarity, f_sim_best = 1.000 (> 0.100 thold), f_keep = 0.998 +0.05.862.781 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.05.862.805 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.05.862.809 I slot launch_slot_: id 0 | task 23 | processing task, is_child = 0 +0.05.862.832 I slot operator(): id 0 | task 23 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2463 +0.05.862.838 I slot operator(): id 0 | task 23 | erased invalidated context checkpoint (pos_min = 0, pos_max = 9, n_tokens = 10, n_swa = 512, pos_next = 0, size = 0.411 MiB) +0.05.862.842 I slot operator(): id 0 | task 23 | erased invalidated context checkpoint (pos_min = 923, pos_max = 1946, n_tokens = 1947, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.05.863.832 I slot operator(): id 0 | task 23 | erased invalidated context checkpoint (pos_min = 1435, pos_max = 2458, n_tokens = 2459, n_swa = 512, pos_next = 0, size = 21.006 MiB) +0.05.864.588 I slot operator(): id 0 | task 23 | cached n_tokens = 0, memory_seq_rm [0, end) +0.05.870.017 I slot operator(): id 0 | task 23 | cached n_tokens = 10, memory_seq_rm [10, end) diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v4_server_restart.log b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v4_server_restart.log new file mode 100644 index 000000000..4d4c174ab --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/radeon_v4_server_restart.log @@ -0,0 +1,210 @@ +0.00.000.840 I srv llama_server: initializing ... +0.00.219.431 I cmn common_param: common_params_print_info: build 11124 (1a679828f) with Clang 20.1.8 for Windows x86_64 +0.00.219.437 I cmn common_param: common_params_print_info: verbosity = 4 (adjust with the `-lv N` CLI arg) +0.00.219.439 I cmn common_param: device_info: +0.00.224.079 I cmn common_param: - Vulkan0 : NVIDIA GeForce RTX 5090 Laptop GPU (24137 MiB, 23369 MiB free) +0.00.227.919 I cmn common_param: - Vulkan1 : AMD Radeon(TM) 890M Graphics (48505 MiB, 46080 MiB free) +0.00.227.929 I cmn common_param: - CPU : AMD Ryzen AI 9 HX 370 w/ Radeon 890M (64633 MiB, 26160 MiB free) +0.00.227.991 I cmn common_param: system_info: n_threads = 12 (n_threads_batch = 12) / 24 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | +0.00.228.076 I srv init_listene: running without SSL +0.00.228.144 I srv init_listene: The UI is disabled +0.00.228.145 I srv init_listene: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable +0.00.228.149 I srv init: using 23 threads for HTTP server +0.00.228.227 W srv llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655) +0.00.236.133 I srv load_model: loading model 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.236.139 I srv load_model: local path 'C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf' +0.00.262.661 I llama_model_loader: loaded meta data with 37 key-value pairs and 226 tensors from C:\Users\zhout\w2\models\Spark-X2.5-1.7B-Q4_K_M.gguf (version GGUF V3 (latest)) +0.00.262.676 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. +0.00.262.682 I llama_model_loader: - kv 0: general.architecture str = spark2_5 +0.00.262.683 I llama_model_loader: - kv 1: general.type str = model +0.00.262.687 I llama_model_loader: - kv 2: general.sampling.top_k i32 = 0 +0.00.262.694 I llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000 +0.00.262.695 I llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000 +0.00.262.696 I llama_model_loader: - kv 5: general.name str = Hf_Format +0.00.262.696 I llama_model_loader: - kv 6: general.size_label str = 1.7B +0.00.262.699 I llama_model_loader: - kv 7: spark2_5.block_count u32 = 28 +0.00.262.700 I llama_model_loader: - kv 8: spark2_5.context_length u32 = 1048576 +0.00.262.700 I llama_model_loader: - kv 9: spark2_5.embedding_length u32 = 2048 +0.00.262.701 I llama_model_loader: - kv 10: spark2_5.feed_forward_length u32 = 6656 +0.00.262.701 I llama_model_loader: - kv 11: spark2_5.attention.head_count u32 = 8 +0.00.262.701 I llama_model_loader: - kv 12: spark2_5.attention.head_count_kv u32 = 2 +0.00.262.703 I llama_model_loader: - kv 13: spark2_5.rope.freq_base f32 = 5000000.000000 +0.00.262.704 I llama_model_loader: - kv 14: spark2_5.rope.freq_base_swa f32 = 10000.000000 +0.00.262.706 I llama_model_loader: - kv 15: spark2_5.attention.layer_norm_rms_epsilon f32 = 0.000001 +0.00.262.706 I llama_model_loader: - kv 16: spark2_5.attention.key_length u32 = 256 +0.00.262.706 I llama_model_loader: - kv 17: spark2_5.attention.value_length u32 = 256 +0.00.262.707 I llama_model_loader: - kv 18: spark2_5.vocab_size u32 = 131072 +0.00.262.707 I llama_model_loader: - kv 19: spark2_5.attention.sliding_window u32 = 512 +0.00.262.728 I llama_model_loader: - kv 20: spark2_5.attention.sliding_window_pattern arr[bool,28] = [true, true, true, false, true, true,... +0.00.262.729 I llama_model_loader: - kv 21: spark2_5.rope.dimension_count u32 = 64 +0.00.262.729 I llama_model_loader: - kv 22: spark2_5.rope.dimension_count_swa u32 = 256 +0.00.262.729 I llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 +0.00.262.730 I llama_model_loader: - kv 24: tokenizer.ggml.pre str = spark2_5 +0.00.279.153 I llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,131072] = ["<｜start▁of▁sentence｜>", "<... +0.00.284.342 I llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,131072] = [3, 3, 3, 4, 4, 3, 3, 3, 4, 4, 4, 3, ... +0.00.300.733 I llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,130716] = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "Ġ... +0.00.300.742 I llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 0 +0.00.300.743 I llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 1 +0.00.300.744 I llama_model_loader: - kv 30: tokenizer.ggml.unknown_token_id u32 = 5 +0.00.300.745 I llama_model_loader: - kv 31: tokenizer.ggml.padding_token_id u32 = 2 +0.00.300.747 I llama_model_loader: - kv 32: tokenizer.ggml.add_bos_token bool = false +0.00.300.747 I llama_model_loader: - kv 33: tokenizer.ggml.add_eos_token bool = false +0.00.300.751 I llama_model_loader: - kv 34: tokenizer.chat_template str = {%- if not messages %}\n {{- raise_... +0.00.300.752 I llama_model_loader: - kv 35: general.quantization_version u32 = 2 +0.00.300.753 I llama_model_loader: - kv 36: general.file_type u32 = 15 +0.00.300.762 I llama_model_loader: - type f32: 57 tensors +0.00.300.762 I llama_model_loader: - type q4_K: 140 tensors +0.00.300.763 I llama_model_loader: - type q6_K: 29 tensors +0.00.300.768 I print_info: file format = GGUF V3 (latest) +0.00.300.768 I print_info: file type = Q4_K - Medium +0.00.300.773 I print_info: file size = 1.03 GiB (5.16 BPW) +0.00.306.049 I llama_prepare_model_devices: using device Vulkan1 (AMD Radeon(TM) 890M Graphics) (unknown id) - 46080 MiB free +0.00.365.248 I load: 86 unused tokens +0.00.394.284 I load: printing all EOG tokens: +0.00.394.291 I load: - 1 ('<｜end▁of▁sentence｜>') +0.00.394.431 I load: special tokens cache size = 113 +0.00.410.435 I load: token to piece cache size = 0.7975 MB +0.00.410.462 I print_info: arch = spark2_5 +0.00.410.463 I print_info: vocab_only = 0 +0.00.410.464 I print_info: no_alloc = 0 +0.00.410.464 I print_info: n_ctx_train = 1048576 +0.00.410.465 I print_info: n_embd_inp = 2048 +0.00.410.465 I print_info: n_embd = 2048 +0.00.410.465 I print_info: n_embd_out = 2048 +0.00.410.466 I print_info: n_layer = 28 +0.00.410.466 I print_info: n_layer_all = 28 +0.00.410.483 I print_info: n_head = 8 +0.00.410.485 I print_info: n_head_kv = 2 +0.00.410.485 I print_info: n_rot = 64 +0.00.410.486 I print_info: n_swa = 512 +0.00.410.486 I print_info: is_swa_any = 1 +0.00.410.486 I print_info: non_causal_type = 0 +0.00.410.487 I print_info: n_embd_head_k = 256 +0.00.410.487 I print_info: n_embd_head_v = 256 +0.00.410.488 I print_info: n_gqa = 4 +0.00.410.490 I print_info: n_embd_k_gqa = 512 +0.00.410.491 I print_info: n_embd_v_gqa = 512 +0.00.410.492 I print_info: f_norm_eps = 0.0e+00 +0.00.410.494 I print_info: f_norm_rms_eps = 1.0e-06 +0.00.410.494 I print_info: f_clamp_kqv = 0.0e+00 +0.00.410.495 I print_info: f_max_alibi_bias = 0.0e+00 +0.00.410.495 I print_info: f_logit_scale = 0.0e+00 +0.00.410.495 I print_info: f_attn_scale = 0.0e+00 +0.00.410.496 I print_info: f_attn_value_scale = 0.0000 +0.00.410.497 I print_info: n_ff = 6656 +0.00.410.497 I print_info: n_expert = 0 +0.00.410.498 I print_info: n_expert_used = 0 +0.00.410.500 I print_info: n_expert_groups = 0 +0.00.410.500 I print_info: n_group_used = 0 +0.00.410.500 I print_info: causal attn = 1 +0.00.410.501 I print_info: pooling type = -1 +0.00.410.501 I print_info: rope type = 2 +0.00.410.501 I print_info: rope scaling = linear +0.00.410.502 I print_info: freq_base_train = 5000000.0 +0.00.410.504 I print_info: freq_scale_train = 1 +0.00.410.505 I print_info: freq_base_swa = 10000.0 +0.00.410.505 I print_info: freq_scale_swa = 1 +0.00.410.506 I print_info: n_embd_head_k_swa = 256 +0.00.410.506 I print_info: n_embd_head_v_swa = 256 +0.00.410.506 I print_info: n_rot_swa = 256 +0.00.410.507 I print_info: n_ctx_orig_yarn = 1048576 +0.00.410.507 I print_info: rope_yarn_log_mul = 0.0000 +0.00.410.507 I print_info: rope_finetuned = unknown +0.00.410.508 I print_info: model type = 1.7B +0.00.410.509 I print_info: model params = 1.71 B +0.00.410.509 I print_info: general.name = Hf_Format +0.00.410.510 I print_info: vocab type = BPE +0.00.410.511 I print_info: n_vocab = 131072 +0.00.410.511 I print_info: n_merges = 130716 +0.00.410.512 I print_info: BOS token = 0 '<｜start▁of▁sentence｜>' +0.00.410.512 I print_info: EOS token = 1 '<｜end▁of▁sentence｜>' +0.00.410.513 I print_info: EOT token = 1 '<｜end▁of▁sentence｜>' +0.00.410.513 I print_info: UNK token = 5 '' +0.00.410.514 I print_info: PAD token = 2 '<｜▁pad▁｜>' +0.00.410.514 I print_info: LF token = 198 'Ċ' +0.00.410.515 I print_info: FIM PRE token = 11 '<｜fim▁begin｜>' +0.00.410.515 I print_info: FIM SUF token = 12 '<｜fim▁hole｜>' +0.00.410.516 I print_info: FIM MID token = 13 '<｜fim▁end｜>' +0.00.410.516 I print_info: EOG token = 1 '<｜end▁of▁sentence｜>' +0.00.410.517 I print_info: max token length = 210 +0.00.418.609 I load_tensors: loading model tensors, this can take a while... (load_mode = none) +0.00.615.432 I load_tensors: offloading output layer to GPU +0.00.615.438 I load_tensors: offloading 27 repeating layers to GPU +0.00.615.439 I load_tensors: offloaded 29/29 layers to GPU +0.00.615.444 I load_tensors: Vulkan1 model buffer size = 1051.02 MiB +0.00.615.447 I load_tensors: Vulkan_Host model buffer size = 210.00 MiB +0.00.954.470 I cmn common_init_: added <｜end▁of▁sentence｜> logit bias = -inf +0.00.954.877 I llama_context: constructing llama_context +0.00.954.883 I llama_context: n_seq_max = 1 +0.00.954.883 I llama_context: n_ctx = 4096 +0.00.954.884 I llama_context: n_ctx_seq = 4096 +0.00.954.884 I llama_context: n_batch = 2048 +0.00.954.885 I llama_context: n_ubatch = 512 +0.00.954.885 I llama_context: causal_attn = 1 +0.00.954.886 I llama_context: flash_attn = auto +0.00.954.886 I llama_context: kv_unified = false +0.00.954.889 I llama_context: freq_base = 5000000.0 +0.00.954.892 I llama_context: freq_scale = 1 +0.00.954.892 I llama_context: n_rs_seq = 0 +0.00.954.892 I llama_context: n_outputs_max = 1 +0.00.954.893 I llama_context: n_outputs_max_per_seq = 1 +0.00.954.893 I llama_context: n_ctx_seq (4096) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized +0.00.955.204 I llama_context: Vulkan_Host output buffer size = 0.50 MiB +0.00.955.212 I llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells +0.00.960.200 I llama_kv_cache: Vulkan1 KV buffer size = 56.00 MiB +0.00.960.791 I llama_kv_cache: size = 56.00 MiB ( 4096 cells, 7 layers, 1/1 seqs), K (f16): 28.00 MiB, V (f16): 28.00 MiB +0.00.960.799 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.960.799 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.960.804 I llama_kv_cache_iswa: creating SWA KV cache, size = 1024 cells +0.00.964.690 I llama_kv_cache: Vulkan1 KV buffer size = 42.00 MiB +0.00.965.182 I llama_kv_cache: size = 42.00 MiB ( 1024 cells, 21 layers, 1/1 seqs), K (f16): 21.00 MiB, V (f16): 21.00 MiB +0.00.965.187 I llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256 +0.00.965.188 I llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256 +0.00.965.198 I sched_reserve: reserving ... +0.00.966.658 I resolve_fused_ops: Flash Attention enabled +0.00.966.658 I resolve_fused_ops: resolving fused DeepSeek V4 HC support: +0.00.967.604 I resolve_fused_ops: fused DeepSeek V4 HC pre enabled +0.00.968.503 I resolve_fused_ops: fused DeepSeek V4 HC comb enabled +0.00.969.360 I resolve_fused_ops: fused DeepSeek V4 HC post enabled +0.00.977.527 I sched_reserve: Vulkan1 compute buffer size = 56.02 MiB +0.00.977.529 I sched_reserve: Vulkan_Host compute buffer size = 13.02 MiB +0.00.977.531 I sched_reserve: graph: nodes = 986, splits = 2, input objects = 4, input tensors = 9 +0.00.977.532 I sched_reserve: reserve took 12.33 ms, sched copies = 1 +0.00.977.623 I cmn init: llama threadpool init, n_threads = 12 +0.00.977.638 I cmn common_init_: warming up the model with an empty run - please wait ... (--no-warmup to disable) +0.01.029.750 I srv load_model: initializing, n_slots = 1, n_ctx_slot = 4096, kv_unified = 'false' +0.01.029.771 I spec common_specu: no implementations specified for speculative decoding +0.01.029.772 I slot load_model: id 0 | task -1 | new slot, n_ctx = 4096 +0.01.029.845 I srv load_model: prompt cache is disabled - use `--cache-ram N` to enable it +0.01.029.845 I srv load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391 +0.01.029.846 I srv load_model: context checkpoints enabled, max = 32, min spacing = 8192 +0.01.029.867 W srv init: --cache-idle-slots requires --cache-ram, disabling +0.01.041.131 I srv init: init: chat template, example_format: '<｜start▁of▁sentence｜><|System|> +you are a helpful assistant. + +You are a helpful assistant<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>Hello<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>Hi there<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|User|>How are you?<｜end▁of▁sentence｜><｜start▁of▁sentence｜><|Bot|>' +0.01.048.367 I srv init: init: chat template, thinking = 0 +0.01.048.377 I srv init: preserve_reasoning kwarg: true +0.01.048.378 W srv init: chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve) +0.01.048.399 I srv llama_server: model loaded +0.01.048.402 I srv llama_server: listening on http://127.0.0.1:50197 +0.01.048.474 I srv update_slots: all slots are idle +0.01.144.095 I srv server_strea: conv_id= (empty=1) +0.01.144.316 I srv operator(): chat format: peg-native +0.01.144.419 I slot get_availabl: id 0 | task -1 | - skipping, slot is empty +0.01.144.421 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = -1 +0.01.144.468 I slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> ?top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist +0.01.144.481 I slot launch_slot_: id 0 | task -1 | sampler params: + repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 + dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64 + top_k = 0, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000 + mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900 +0.01.144.483 I slot launch_slot_: id 0 | task 0 | processing task, is_child = 0 +0.01.144.492 I slot operator(): id 0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 28 +0.01.144.497 I slot operator(): id 0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end) +0.01.155.986 I slot operator(): id 0 | task 0 | cached n_tokens = 10, memory_seq_rm [10, end) +0.01.190.148 I slot create_check: id 0 | task 0 | created context checkpoint 1 of 32 (pos_min = 0, pos_max = 9, n_tokens = 10, size = 0.411 MiB) +0.01.193.063 I slot operator(): id 0 | task 0 | cached n_tokens = 24, memory_seq_rm [24, end) +0.01.193.096 I slot init_sampler: id 0 | task 0 | init sampler, took 0.01 ms, tokens: text = 28, total = 28 +0.01.230.815 I slot create_check: id 0 | task 0 | created context checkpoint 2 of 32 (pos_min = 0, pos_max = 23, n_tokens = 24, size = 0.985 MiB) +0.01.270.903 I slot print_timing: id 0 | task 0 | prompt eval time = 108.55 ms / 28 tokens ( 3.88 ms per token, 257.95 tokens per second) diff --git a/benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md b/benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md index 2096b838a..11042d6ce 100644 --- a/benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md +++ b/benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md @@ -43,10 +43,10 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i | Device | Spark-X2.5 | Qwen3-TTS 0.6B CustomVoice | Qwen3.8-27B | MiniCPM-o 4.5 | InternVLA-A1 | |---|---|---|---|---|---| | PC HX370: CPU / WSL | scoped E2E: Omni GGUF or BF16 Spark text | scoped E2E: Omni CPU text → complete WAV | NOT E2E | scoped E2E: Omni GGUF combined image+audio → text+speech | synthetic policy pass: Omni CPU | -| PC HX370: CPU / native Windows | scoped E2E: Omni Spark text | scoped E2E: Omni CPU text → complete WAV | NOT E2E | scoped E2E: Omni GGUF image+audio → text+speech (native-storage 20-run profile) | synthetic policy pass: Omni CPU | +| PC HX370: CPU / native Windows | scoped E2E: Omni Spark text; fresh restart after in-flight cancel | scoped E2E: Omni CPU text → complete WAV | NOT E2E | scoped E2E: Omni GGUF image+audio → text+speech (native-storage 20-run profile) | synthetic policy pass: Omni CPU | | PC HX370 + RTX 5090 Laptop / WSL | profile completed | profile completed | scoped E2E: text + image | scoped E2E: text/image/audio → speech | synthetic policy pass: admitted Omni CUDA | | PC HX370 + RTX 5090 Laptop / Windows | profile completed | profile completed; 30-min repeat no collapse, brief playback jitter | NOT E2E | scoped E2E: text/image/audio → speech | synthetic policy pass: admitted Omni CUDA | -| 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 + Radeon 890M / Windows worker | scoped E2E: Omni Spark text; fresh restart after in-flight cancel | 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 (one Spark attention layer on NPU) | NOT E2E (CPU talker/vocoder steps; two-frame QNN vocoder on NPU, component RTF 22.89) | NOT E2E | NOT E2E (NPU speech-head step profiled) | NOT E2E (Cosmos NPU encoder profiled) | @@ -64,7 +64,7 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i | PC HX370: CPU / WSL | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No compatible CPU artifact/deployment for this model was found in the checked model roots. Available Qwen 27B FP8/NVFP4 CUDA artifacts do not establish a CPU route. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. | | PC HX370: CPU / WSL | MiniCPM-o 4.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST | The BF16/vLLM 0.28 three-stage path previously passed separate text, synthetic red-square image and synthetic 440 Hz audio inputs to text+speech; its 20-request serial **text-only** profile was p50/p95 21.86/22.97 s with swap and shutdown warnings. A separate pinned Q4_K_M/F16 GGUF worker now runs a **combined spoken question+image → text+speech** request through a bounded Omni StageRuntime/StagePool graph stage on WSL CPU, with installed vLLM 0.29 for the controller and CPU-only C++ model execution. All 20 measured requests after one warmup answered “The square in the image is red.”, emitted complete nonzero speech and a terminal audio event; nearest-rank complete-request wall p50/p95 was 22.67/24.83 s, speech duration 1.60–2.92 s, and maximum sampled worker RSS 13.70 GB. The first measured WAV independently transcribed exactly with pinned Whisper tiny.en (WER 0, one-output intelligibility proxy). Ten weight hashes and CPU placement were verified; 24 GiB shared RAM was reserved, 8 GiB was refused before launch, and in-flight cancellation left no stale output or reservation. Host RAM/swap samples covered only part of the profile. One synthetic fixture and variable speech PCM do not establish general quality, playable streaming, loading peak or sustained performance. | Validate real image/audio suites, speech alignment and quality, video input, concurrent admission, post-cancel restart, loading peak and sustained thermal behavior; [GGUF Omni raw evidence](../../../e2e_expansion_20260923/evidence/minicpmo_omni_wsl_cpu/README.md), [BF16/vLLM CPU evidence](../../../e2e_expansion_20260923/README.md), [earlier standalone GGUF run](../../../e2e_expansion_20260923/evidence/minicpmo_cpp_cpu/README.md). | | PC HX370: CPU / WSL | InternVLA-A1 | SYNTHETIC_POLICY_PASS_OMNI; NOT_TASK_E2E | The real Place_Markpen checkpoint ran through an admitted Omni StageRuntime whole-policy stage with a CPU-only isolated worker. One warmup plus 20 serial patterned synthetic camera/state/noise requests returned identical finite `[1,50,32]` action chunks with terminal action events; nearest-rank complete-request wall p50/p95 was 3.175/3.219 s, excluding 40.03 s startup. The stage verified checkpoint/interpreter hashes and CPU placement, reserved 16 GiB of a declared 30 GiB WSL host-RAM pool, refused 8 GiB before launch, and released the reservation on shutdown. The earlier direct zero-input policy/CUDA comparison was a different fixture and remains separate. No real-observation task-quality tolerance exists. | Validate real observations/reference actions, physical units/order/step time, action deadlines, loading peak, concurrency and sustained power/thermal behavior; [Omni raw evidence](../../../e2e_expansion_20260923/evidence/internvla_omni_policy/README.md), [earlier direct policy evidence](../../../e2e_expansion_20260923/README.md). | -| PC HX370: CPU / native Windows | Spark-X2.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST | The pinned Spark-X2.5-1.7B Q4_K_M GGUF ran in native Windows llama.cpp CPU (`-dev none -ngl 0`) as a whole-session Omni StageRuntime/StagePool backend. `Paris` and the correct `1511` inventory answer passed; one warmup plus 20 serial complete requests were 20/20 correct, nearest-rank wall p50/p95 6.93/7.05 s. The public AsyncOmni short request passed. The stage verifies artifact hashes and CPU placement (0/29 layers offloaded), reserves 4 GiB host RAM, emits a terminal text event and releases reservation after in-flight cancellation. The separate prior standalone profile was 6.65/6.86 s. No vLLM CPU attention repair, incremental stream, restart-after-cancel, loading peak or sustained power/thermal gate is claimed. | Validate broader text/token quality, streaming and cancellation recovery, memory peak and sustained load; [Omni raw evidence](../../../e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md), [standalone evidence](../../../e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/README.md). | +| PC HX370: CPU / native Windows | Spark-X2.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST; RESTART_AFTER_INFLIGHT_CANCEL_PASS | The pinned Spark-X2.5-1.7B Q4_K_M GGUF ran in native Windows llama.cpp CPU (`-dev none -ngl 0`) as a whole-session Omni StageRuntime/StagePool backend. `Paris` and the correct `1511` inventory answer passed; one warmup plus 20 serial complete requests were 20/20 correct, nearest-rank wall p50/p95 6.93/7.05 s. The public AsyncOmni short request passed. The stage verifies artifact hashes and CPU placement (0/29 layers offloaded), reserves 4 GiB host RAM, emits a terminal text event and releases reservation after in-flight cancellation. The separate prior standalone profile was 6.65/6.86 s. A separate confirmed server-start cancellation of the 120-item request terminated the worker, left no stale output or reservation, and a fresh StageRuntime with the same pinned artifact returned Paris under a new worker generation. This is fresh-stage recovery, not same-session continuation. No vLLM CPU attention repair, incremental stream, loading peak or sustained power/thermal gate is claimed. | Validate broader text/token quality, streaming, concurrent recovery, memory peak and sustained load; [restart evidence](../../../e2e_expansion_20260924/evidence/spark_omni_windows_restart/README.md), [Omni raw evidence](../../../e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md), [standalone evidence](../../../e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/README.md). | | PC HX370: CPU / native Windows | Qwen3-TTS 0.6B CustomVoice | SCOPED_E2E_OMNI_COMPLETE_REQUEST | The pinned 0.6B CustomVoice BF16/SDPA checkpoint ran wholly on CPU through an isolated Qwen wrapper worker owned by one bounded Omni StageRuntime stage. Two named 24 kHz WAVs matched the earlier standalone PCM SHA256 values exactly. One warmup + 20 serial complete Omni requests for 4.56 s of audio had 20/20 identical PCM and nearest-rank wall p50/p95 15.32/15.60 s (RTF 3.36/3.42); independent Whisper tiny.en ASR gave WER 0.20 on the first Omni WAV, an intelligibility proxy only. A public AsyncOmni audio request passed. The stage verified pinned weights/interpreter and CPU placement, reserved 10 GiB host RAM, rejected an insufficient 4 GiB demand before worker launch, and after in-flight cancellation emitted no stale output and returned the ledger to zero. Sampled process-tree private memory peaked at 5.39 GB, not a loading peak. Native vLLM CPU attention operators remain absent; playable streaming, post-cancel restart, broader speech quality, loading peak and sustained power/thermal behavior remain unverified. | Validate playable streaming and restart/recovery, broader voice quality, peak memory, concurrency and sustained behavior; [Omni raw evidence](../../../e2e_expansion_20260923/evidence/qwen_tts_omni_cpu/README.md), [standalone baseline](../../../e2e_expansion_20260923/evidence/qwen_tts_native_cpu/README.md). | | PC HX370: CPU / native Windows | Qwen3.8-27B | NOT_E2E_PROFILEABLE | The installed native Windows vLLM extension lacks cpu_attn_reshape_and_cache and cpu_attn_get_scheduler_metadata. The separate ORT CPU graph worker is not a whole-model text/TTS adapter; a qualified CPU build/binding is missing. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. | | PC HX370: CPU / native Windows | MiniCPM-o 4.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST; NATIVE_STORAGE_PROFILED | The pinned Q4_K_M/F16 GGUF combined synthetic JPEG+spoken-question → text+speech route passed one warmup and 20/20 serial complete requests with weights on native NTFS: nearest-rank p50/p95 23.62/23.88 s, excluding 9.23 s controller startup; every request returned the expected red-square sentence and nonzero 1.60–2.88 s 24 kHz speech. The first measured WAV exactly matched pinned Whisper tiny.en ASR (WER 0, one-output intelligibility proxy). CPU placement and ten weight hashes were verified; a 21 GiB shared-RAM demand was admitted, sampled worker RSS peaked at 13.10 GB, no worker logged a prefetch warning, and in-flight cancellation left no stale output or reservation. A same-artifact one-request control from WSL UNC weights took 79.20 s, versus 21.16 s from native weights, with 40.90 vs 12.75 s worker load and two vs zero prefetch warnings. That sequential comparison associates UNC storage with the older slowdown but does not isolate causality or prove stable cross-placement speedup. The older UNC CPU result remains valid for that path. This is a whole-request C++ worker under Omni control, not native vLLM execution or playable streaming; the synthetic fixture is not general quality qualification. | Validate real image/audio suites, playable streaming, concurrent admission, post-cancel restart, loading peak and sustained thermal/power behavior; [native/UNC raw evidence](../../../e2e_expansion_20260923/evidence/minicpmo_omni_native_weights_io/README.md), [earlier UNC CPU control](../../../e2e_expansion_20260923/evidence/minicpmo_omni_radeon_profile20/README.md), [earlier Omni evidence](../../../e2e_expansion_20260923/evidence/minicpmo_omni_cpp/README.md). | @@ -79,7 +79,7 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i | PC HX370 + RTX 5090 Laptop / Windows | Qwen3.8-27B | NOT_E2E_PROFILEABLE | The Omni pipeline binding now exists and passed scoped text/image cases in WSL, but no native Windows Omni run has verified the NVFP4 checkpoint, memory admission or modality behavior. | Run the same checkpoint and explicit context on native Windows, then broaden quality and profiling; [WSL-only evidence](../../../e2e_recovery_20260923/README.md). | | PC HX370 + RTX 5090 Laptop / Windows | MiniCPM-o 4.5 | SCOPED_E2E | The verified checkpoint completed native Windows three-stage BF16 text, one synthetic image, and one synthetic audio input to text+speech in separate constrained plans with 5 GiB thinker host offload and explicit 2,048/512-token limits. Text and all 32 thinker token IDs matched WSL CUDA; a 5.2 s WAV was nonzero. The red-square image produced the same description and 27 thinker token IDs as WSL plus a nonzero 5.2 s WAV in 17.15 s. The 2 s/440 Hz tone produced the same whistling description and nine IDs as WSL plus a nonzero 2.16 s WAV in 8.26 s. The initial text run needed 900/600 s startup limits after a 300 s timeout; a separate 1-warmup/20-measured serial text run had request wall p50 18.38 s/p95 18.73 s, with sampled host available RAM as low as 0.78 GB and pagefile use up to 11.17 GB. These synthetic inputs do not establish general image/audio or speech quality. | Check real image/audio suites, video input, speech intelligibility/alignment, concurrency, interruption, loading-peak admission and sustained thermal/power behavior before broader qualification; [native evidence](../../../e2e_expansion_20260923/README.md). | | PC HX370 + RTX 5090 Laptop / Windows | InternVLA-A1 | SYNTHETIC_POLICY_PASS_OMNI_CUDA; NOT_TASK_E2E | The real Place_Markpen checkpoint ran through an admitted native Windows Omni whole-policy graph stage with BF16 policy and Cosmos encoder on the RTX 5090 Laptop GPU. After one warmup, 20/20 serial patterned synthetic requests returned identical finite float32 [1,50,32] actions with terminal events; nearest-rank complete-request wall p50/p95 was 0.746/0.806 s, excluding 147.77 s startup. The 16 GiB host-RAM and 16 GiB CUDA demands were within explicit 18/21 GiB ceilings checked against available memory. Actual CUDA placement was verified, PyTorch peak reserved VRAM reached 7.86 GB, and an 8 GiB VRAM plan was refused before worker launch. In-flight cancellation and shutdown released both ledgers with no stale output; a separate public AsyncOmni.generate request returned the same action hash. Relative to the separate native CPU patterned output, action relative L2 difference was 2.340% with no task tolerance; native-vs-WSL CUDA relative L2 was 1.555%. The earlier one-forward zero-input direct CUDA run was a different experiment. This is synthetic action-policy integration, not robot-task E2E; units/order/step time remain unverified and control_ready=false. | Validate real observations/reference actions and physical action semantics, post-cancel restart, concurrent admission, loading peak, deadlines and sustained latency/power/thermal behavior; [Omni Windows CUDA evidence](../../../e2e_expansion_20260923/evidence/internvla_omni_windows_cuda/README.md), [earlier direct result](../../../e2e_expansion_20260923/README.md). | -| PC HX370 + Radeon 890M / Windows worker | Spark-X2.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST | The same pinned Spark-X2.5-1.7B Q4_K_M GGUF completed text prefill/decode through a bounded Omni whole-session stage with llama.cpp Vulkan1. The log identifies Radeon 890M and 29/29 layers offloaded there, rather than RTX 5090; an intentionally wrong device name was refused. `Paris` and `1511` passed; one warmup plus 20 serial complete requests were 20/20 correct, nearest-rank wall p50/p95 2.10/2.14 s. A public AsyncOmni request passed; an oversized context was rejected with a later correct request; in-flight cancellation produced no stale output and released the ledger. The prior standalone run was 2.18/2.20 s. No streaming, restart-after-cancel, GPU-memory peak or sustained power/thermal check is claimed. | Qualify public configuration, broader text/token quality, streaming and cancellation recovery, shared-RAM/GPU peak and sustained use; [Omni raw evidence](../../../e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md), [standalone evidence](../../../e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/README.md). | +| PC HX370 + Radeon 890M / Windows worker | Spark-X2.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST; RESTART_AFTER_INFLIGHT_CANCEL_PASS | The same pinned Spark-X2.5-1.7B Q4_K_M GGUF completed text prefill/decode through a bounded Omni whole-session stage with llama.cpp Vulkan1. The log identifies Radeon 890M and 29/29 layers offloaded there, rather than RTX 5090; an intentionally wrong device name was refused. `Paris` and `1511` passed; one warmup plus 20 serial complete requests were 20/20 correct, nearest-rank wall p50/p95 2.10/2.14 s. A public AsyncOmni request passed; an oversized context was rejected with a later correct request; in-flight cancellation produced no stale output and released the ledger. The prior standalone run was 2.18/2.20 s. A separate confirmed server-start cancellation of the 120-item request terminated the Vulkan1 worker, left no stale output or reservation, and a fresh StageRuntime returned Paris under a new worker generation with Radeon 890M placement again verified. This is fresh-stage recovery, not same-session continuation. No streaming, GPU-memory peak or sustained power/thermal check is claimed. | Qualify public configuration, broader text/token quality, streaming, concurrent recovery, shared-RAM/GPU peak and sustained use; [restart evidence](../../../e2e_expansion_20260924/evidence/spark_omni_windows_restart/README.md), [Omni raw evidence](../../../e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md), [standalone evidence](../../../e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/README.md). | | PC HX370 + Radeon 890M / Windows worker | Qwen3-TTS 0.6B CustomVoice | SCOPED_E2E_OMNI_COMPLETE_REQUEST_HYBRID | The pinned Q8_0 CustomVoice talker, F16 codec and punctuation GGUFs completed two named text-to-WAV requests through an Omni whole-session CrispASR stage on native Windows. The runtime log identifies Radeon 890M Vulkan0 for talker/codec and the 432 MB code predictor pinned to CPU FP32; this is not all-Radeon or NPU execution. Both PCM hashes exactly match the earlier standalone outputs. Whisper tiny.en transcribed the first Omni WAV and both standalone WAVs exactly (WER 0 as an intelligibility proxy). One warmup plus 20 serial complete Omni requests for 2.56 s of audio had identical PCM and nearest-rank wall p50/p95 3.298/3.336 s (RTF 1.29/1.30). A public AsyncOmni audio request passed; cancellation drained the worker without stale output or remaining reservation. The earlier standalone profile was 3.310/3.345 s. The default Vulkan configuration previously fell back to CPU, all-Vulkan ran away to a 360-frame cap, and DirectML failed waveform quality. No playable stream, post-cancel restart, loading peak, broad speech quality or sustained power/thermal run is claimed. | Validate playable streaming, restart/recovery, broader voice quality, shared-RAM/GPU peak, concurrency and sustained real-time behavior before wider qualification; [Omni raw evidence](../../../e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/README.md), [standalone evidence](../../../e2e_expansion_20260923/evidence/qwen_tts_radeon890m_joint/README.md), [failed DirectML route](../../../e2e_expansion_20260923/evidence/qwen_tts_radeon890m_directml/README.md). | | PC HX370 + Radeon 890M / Windows worker | Qwen3.8-27B | NOT_E2E_PROFILEABLE | FP32 448x448 vision tower: DirectML 85/113 nodes; normalized max error 9.095e-06 versus torch CPU. Historical component path, not full 27B or current graph-stage image-to-text integration. | Recover the hashed tower bundle and qualify its current Omni graph handoff with the language stage. | | PC HX370 + Radeon 890M / Windows worker | MiniCPM-o 4.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST_HYBRID; NATIVE_STORAGE_PROFILED | The pinned Q4_K_M/F16 GGUF combined synthetic spoken-question+JPEG → text+speech route passed one warmup and 20/20 serial complete requests through a bounded Omni whole-session stage with weights on native NTFS. Nearest-rank wall p50/p95 was **16.35/17.20 s**, range 14.08–17.30 s, excluding 8.75 s startup; all requests returned the red-square sentence and nonzero 2.08–2.88 s 24 kHz speech with terminal audio events. The first WAV exactly matched pinned Whisper tiny.en ASR (WER 0, one-output intelligibility proxy). All 21 successful worker logs identified Radeon 890M Vulkan0 language (37/37 layers) and vision plus CPU Token2Wav, with zero prefetch warnings; sampled RSS peaked at 15.30 GB. The 21 GiB shared-RAM reservation admitted after file-cache reclaim; an initial identical attempt was explicitly refused before model load when available RAM had fallen below reservation. Cancellation left no stale output or reservation. A sequential same-artifact one-request UNC control took 90.63 s versus 16.89 s native and logged two prefetch warnings; the previous UNC 20-request profile remains p50/p95 87.27/91.45 s for that path. Storage is strongly associated with the slowdown, but OS cache/run order are uncontrolled and the separate CPU/Radeon profiles are not a qualified iGPU speedup comparison. This is CPU+iGPU joint execution, not all-Radeon, NPU, general quality, playable streaming or sustained thermal/power qualification. | Validate real multimodal quality, playable streaming, post-cancel restart, concurrency, loading peak and sustained power/thermal behavior; capture matched OS/shared-iGPU memory traces. [Native/UNC raw evidence](../../../e2e_expansion_20260923/evidence/minicpmo_omni_native_weights_io/README.md), [earlier UNC 20-request evidence](../../../e2e_expansion_20260923/evidence/minicpmo_omni_radeon_profile20/README.md). | diff --git a/docs/design/omni_edge_e2e_roadmap.md b/docs/design/omni_edge_e2e_roadmap.md index 4bba3fa66..7e5b18e09 100644 --- a/docs/design/omni_edge_e2e_roadmap.md +++ b/docs/design/omni_edge_e2e_roadmap.md @@ -325,7 +325,7 @@ precision and workload. New concurrency, state-boundary or quality evidence is required before broadening that acceptance. Current batch-dependent greedy outputs need reference analysis; timing alone cannot identify their cause. -The 2026-09-23 [Spark GGUF Omni follow-up](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md) adds a separate native Windows CPU/Radeon 890M whole-session option. Its pinned 1.7B Q4_K_M artifact passed two named complete text requests and 20 serial inventory requests per device through `StageRuntime`/`StagePool`; public `AsyncOmni` short requests passed on both devices. It verifies CPU or Vulkan1 placement, explicit 4 GiB host reservation, terminal event/acknowledgement and cancellation drain. This does not change the vLLM M0 acceptance or establish M1 mobile generation. Its next gates are incremental output, post-cancel restart, broader token/quality parity, loading/device-memory peaks, concurrency and sustained power/thermal behavior. +The 2026-09-23 [Spark GGUF Omni follow-up](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md) adds a separate native Windows CPU/Radeon 890M whole-session option. Its pinned 1.7B Q4_K_M artifact passed two named complete text requests and 20 serial inventory requests per device through `StageRuntime`/`StagePool`; public `AsyncOmni` short requests passed on both devices. It verifies CPU or Vulkan1 placement, explicit 4 GiB host reservation, terminal event/acknowledgement and cancellation drain. A [2026-09-24 restart audit](../../benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/spark_omni_windows_restart/README.md) confirmed the server had begun the long request before cancellation on each placement, then a fresh StageRuntime with the same artifact returned Paris under a new worker generation with no stale output or reservation. This does not change the vLLM M0 acceptance or establish M1 mobile generation. Its next gates are incremental output, broader token/quality parity, loading/device-memory peaks, concurrency and sustained power/thermal behavior. The [Qwen3-TTS Omni hybrid follow-up](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/README.md) adds one complete-request stage for the native Windows HX370 CPU+Radeon 890M route. A pinned Q8_0 talker and F16 codec produced two named PCM outputs exactly matching the prior standalone backend; one warmup plus 20 serial requests measured 3.298/3.336 s p50/p95 for 2.56 s of audio. Public `AsyncOmni` audio, cancellation drain and explicit 8 GiB host admission passed. This is progress toward M2 integration, not M2 acceptance: RTF remains above 1, and playable chunk streaming, post-cancel restart, loading/iGPU peaks, broader voice quality and 30-minute thermal behavior remain open. No NPU participated.