Qualify scoped WSL CPU Qwen3-TTS Omni complete requests
This commit is contained in:
tzhouam committed
1 parent
10ae2818c7
commit
e6ea35ca12
28 files changed
+15097
-9
No files matched your search
@@ -8,6 +8,7 @@ import asyncio
|
||||
import hashlib
|
||||
import json
|
||||
import math
|
||||
import platform
|
||||
import time
|
||||
import wave
|
||||
from pathlib import Path
|
||||
@@ -41,6 +42,9 @@ async def main() -> None:
|
||||
parser.add_argument("--repeats", type=int, default=20)
|
||||
parser.add_argument("--reserve-gib", type=int, default=10)
|
||||
parser.add_argument("--abort-check", action="store_true")
|
||||
parser.add_argument("--allow-platform-variation", action="store_true")
|
||||
parser.add_argument("--overlay-marker-sha256")
|
||||
parser.add_argument("--expect-admission-refusal", action="store_true")
|
||||
args = parser.parse_args()
|
||||
if args.warmups < 0 or args.repeats <= 0 or not 0 < args.reserve_gib <= 16:
|
||||
parser.error("invalid profile or budget")
|
||||
@@ -55,6 +59,7 @@ async def main() -> None:
|
||||
"python_sha256": args.python_sha256,
|
||||
"model_dir": args.model_dir,
|
||||
"overlay_dir": args.overlay_dir,
|
||||
"overlay_marker_sha256": args.overlay_marker_sha256,
|
||||
"talker_sha256": args.talker_sha256,
|
||||
"tokenizer_sha256": args.tokenizer_sha256,
|
||||
"log_file": args.server_log,
|
||||
@@ -81,7 +86,9 @@ async def main() -> None:
|
||||
configs = [stage.to_omegaconf() for stage in merge_pipeline_deploy(pipeline, deploy)]
|
||||
runtime = StageRuntime(configs, "local-qwen-tts-cpu", "", stage_init_timeout=180, async_chunk=False)
|
||||
report = {
|
||||
"scope": "complete resident native Windows CPU Qwen3-TTS text-to-WAV through Omni",
|
||||
"scope": "complete resident local CPU Qwen3-TTS text-to-WAV through Omni",
|
||||
"host_os": platform.platform(),
|
||||
"allow_platform_variation": args.allow_platform_variation,
|
||||
"warmup_count": args.warmups,
|
||||
"measured_count": args.repeats,
|
||||
"memory_budget": deploy.stages[0].resource_budget,
|
||||
@@ -90,6 +97,16 @@ async def main() -> None:
|
||||
stop_sampling = asyncio.Event()
|
||||
try:
|
||||
started = time.perf_counter()
|
||||
if args.expect_admission_refusal:
|
||||
try:
|
||||
runtime.initialize()
|
||||
except Exception as exc:
|
||||
if "exceed reservation" not in str(exc):
|
||||
raise
|
||||
report["admission_refusal"] = f"{type(exc).__name__}: {exc}"
|
||||
report["status"] = "passed"
|
||||
return
|
||||
raise AssertionError("Qwen CPU TTS accepted an insufficient reservation")
|
||||
runtime.initialize()
|
||||
report["startup_s"] = time.perf_counter() - started
|
||||
pool = runtime.stage_pools[0]
|
||||
@@ -168,7 +185,7 @@ async def main() -> None:
|
||||
report["standalone_pcm_parity"] = [
|
||||
row["pcm_sha256"] == digest for row, digest in zip(report["checks"], expected)
|
||||
]
|
||||
if report["standalone_pcm_parity"] != [True, True]:
|
||||
if report["standalone_pcm_parity"] != [True, True] and not args.allow_platform_variation:
|
||||
raise RuntimeError("CPU Omni PCM differs from the same checkpoint's standalone output")
|
||||
report["warmups"] = [await request_one(first) for _ in range(args.warmups)]
|
||||
report["measured"] = [await request_one(first, save=i == 0) for i in range(args.repeats)]
|
||||
|
||||
@@ -22,6 +22,12 @@ async def main() -> None:
|
||||
"tokenizer_sha256", "server_log", "output_report",
|
||||
):
|
||||
parser.add_argument("--" + name.replace("_", "-"), required=True)
|
||||
parser.add_argument("--overlay-marker-sha256")
|
||||
parser.add_argument("--expected-frames", type=int, default=109440)
|
||||
parser.add_argument(
|
||||
"--expected-pcm-sha256",
|
||||
default="c8a49b5c73efd642a1eb04be973d6aefdc8ed9cfd18b34ff67690b8464906bed",
|
||||
)
|
||||
args = parser.parse_args()
|
||||
|
||||
from vllm_omni.entrypoints.async_omni import AsyncOmni
|
||||
@@ -33,6 +39,7 @@ async def main() -> None:
|
||||
"python_sha256": args.python_sha256,
|
||||
"model_dir": args.model_dir,
|
||||
"overlay_dir": args.overlay_dir,
|
||||
"overlay_marker_sha256": args.overlay_marker_sha256,
|
||||
"talker_sha256": args.talker_sha256,
|
||||
"tokenizer_sha256": args.tokenizer_sha256,
|
||||
"log_file": args.server_log,
|
||||
@@ -87,8 +94,8 @@ async def main() -> None:
|
||||
report["request_wall_s"] = time.perf_counter() - started
|
||||
report["outputs"] = outputs
|
||||
assert len(outputs) == 1 and outputs[0]["request_id"] == "public-cpu-tts-1"
|
||||
assert outputs[0]["sample_rate"] == 24000 and outputs[0]["frames"] == 109440
|
||||
assert outputs[0]["pcm_sha256"] == "c8a49b5c73efd642a1eb04be973d6aefdc8ed9cfd18b34ff67690b8464906bed"
|
||||
assert outputs[0]["sample_rate"] == 24000 and outputs[0]["frames"] == args.expected_frames
|
||||
assert outputs[0]["pcm_sha256"] == args.expected_pcm_sha256
|
||||
report["status"] = "passed"
|
||||
except BaseException as exc:
|
||||
report["status"] = "failed"
|
||||
|
||||
@@ -4,6 +4,7 @@ This extends the [60-cell device × model matrix](../e2e_profiling_20260922/evid
|
||||
|
||||
| Device / model | Verified depth | Observed output and time | Remaining gate |
|
||||
|---|---|---|---|
|
||||
| WSL HX370 CPU / Qwen3-TTS 0.6B CustomVoice BF16 | Scoped complete-request Omni resident CPU whole-session stage (P for two named texts; not playable streaming) | One warmup + 20 serial requests returned identical 3.44 s WAVs; nearest-rank complete-request wall p50/p95 8.27/8.47 s (RTF 2.40/2.46), excluding 19.82 s startup. One pinned Whisper tiny.en transcription matched the reference (WER 0); public `AsyncOmni` audio, 10 GiB admission, 4 GiB refusal and cancellation drain passed. The WSL PCM differs from the native Windows output for the same checkpoint/settings. | Cross-OS numerical/quality investigation, playable streaming, post-cancel restart, loading peak, concurrency and sustained power; [raw evidence](evidence/qwen_tts_omni_wsl_cpu/README.md). |
|
||||
| WSL HX370 CPU / MiniCPM-o 4.5 Q4_K_M/F16 GGUF | Scoped complete-request Omni C++ whole-session graph stage (P for one synthetic combined spoken-question+image input; CPU-only model execution) | All 20 measured requests after one warmup answered the red-square question correctly with complete nonzero 24 kHz speech. Nearest-rank complete-request wall p50/p95 22.67/24.83 s; audio duration 1.60–2.92 s, maximum sampled worker RSS 13.70 GB. One WAV independently transcribed exactly (Whisper tiny.en WER 0). 24 GiB admission, 8 GiB refusal and cancellation with zero remaining ledger passed. | Real multimodal suites, playable streaming, speech alignment, post-cancel restart, loading peak, concurrency and sustained power; [raw evidence](evidence/minicpmo_omni_wsl_cpu/README.md). |
|
||||
| Native Windows HX370 CPU / MiniCPM-o 4.5 Q4_K_M/F16 GGUF | Scoped complete-request Omni C++ whole-session graph stage (P for one synthetic combined audio+image input) | Correct red-square answer, 2.24 s 24 kHz WAV; one cold request 24.34 s, sampled RSS 13.11 GB. 24 GiB admission, 8 GiB refusal and in-flight cancellation with zero remaining reservation passed. | Real multimodal suites, playable streaming, repeated latency, loading peak, post-cancel restart and sustained power; [raw evidence](evidence/minicpmo_omni_cpp/README.md). |
|
||||
| Native Windows HX370 CPU+Radeon 890M / same MiniCPM-o GGUF | Scoped complete-request Omni hybrid graph stage (P for one synthetic combined audio+image input) | Correct red-square answer, 2.60 s 24 kHz WAV; one cold request 18.27 s, sampled RSS 15.57 GB. Log verifies Radeon language+vision and CPU TTS/Token2Wav; cancellation drained without stale output or reservation. | Real multimodal suites, playable streaming, repeated/concurrent latency, loading peak and sustained power; [raw evidence](evidence/minicpmo_omni_cpp/README.md). |
|
||||
|
||||
+76
@@ -0,0 +1,76 @@
|
||||
# WSL CPU Qwen3-TTS complete requests through Omni
|
||||
|
||||
This is a scoped text-to-complete-WAV pass on Ryzen AI 9 HX 370 under Ubuntu
|
||||
26.04/WSL2 kernel `6.18.33.2-microsoft-standard-WSL2`. The public
|
||||
`AsyncOmni` API and `StageRuntime` used the existing
|
||||
`external.qwen_tts.cpu.v1` whole-session stage, with a resident isolated
|
||||
PyTorch worker. The worker's model parameters and output are CPU-only BF16/SDPA,
|
||||
eight threads, `Ryan`/English, seed 42 and `max_new_tokens=64`. CUDA is masked;
|
||||
no Radeon, RTX or NPU execution is claimed. The Omni controller uses installed
|
||||
vLLM 0.29.0; the worker uses PyTorch 2.13.0+cu130 on CPU, `qwen-tts` 0.1.1 and
|
||||
Transformers 4.57.3. Its exact [overlay manifest](overlay_manifest.json) and
|
||||
[optional-kernels CPU shim](kernels/__init__.py) are retained. The local SoX
|
||||
Python package emitted a missing SoX CLI warning; the named text-to-WAV path
|
||||
completed, while SoX-dependent audio-input paths were not tested.
|
||||
|
||||
The checkpoint is Qwen3-TTS-12Hz-0.6B-CustomVoice revision
|
||||
`85e237c12c027371202489a0ec509ded67b5e4b5`, with talker SHA256
|
||||
`bc3c7e785eb961179c25450d1acff03f839e0002f2f3a5aeb67b5735c0fa2adb`
|
||||
and speech-tokenizer SHA256
|
||||
`836b7b357f5ea43e889936a3709af68dfe3751881acefe4ecf0dbd30ba571258`.
|
||||
The stage checked both, the declared interpreter hash and the optional shim
|
||||
hash before accepting requests. The [profile report](profile20.json) records
|
||||
the worker's package versions, actual CPU placement and loaded memory.
|
||||
|
||||
| Check | WSL CPU observation | Evidence |
|
||||
|---|---|---|
|
||||
| Two named complete requests | “Hello from the local computer.” returned 82,560 frames (3.44 s); “The blue car is parked beside the library.” returned 80,640 frames (3.36 s), both 24 kHz mono WAVs. | [report](profile20.json), [worker log](profile20_worker.log) |
|
||||
| Serial profile | One warmup excluded; **20/20** measured first-sentence requests returned the same PCM SHA256. Nearest-rank complete-request wall p50/p95 **8.27/8.47 s** for 3.44 s audio (RTF 2.40/2.46). These are resident-worker times after 19.82 s stage startup, not first-audio latency. | [all timings and memory samples](profile20.json), [first measured WAV](profile20.wav) |
|
||||
| Independent speech proxy | Pinned Whisper tiny.en transcribed the first measured WAV as “Hello from the local computer!”, WER 0 against the stated sentence. This covers one output, not perceptual or speaker quality. | [ASR report](profile20_asr.json) |
|
||||
| Public API | `AsyncOmni.generate` returned one terminal audio event with the same 82,560-frame PCM hash as the measured stage run. Startup was 20.29 s; the one public request took 12.10 s. | [public report](public_report.json), [worker log](public_worker.log) |
|
||||
| Admission and cancellation | A 10 GiB host-RAM reservation from a declared 16 GiB pool passed; 4 GiB was refused before worker launch. In-flight cancellation left no stale output, exited the worker and returned the ledger to zero. | [refusal report](low_budget_report.json), [profile report](profile20.json) |
|
||||
|
||||
The two WSL WAV hashes and durations **differ from the same-checkpoint native
|
||||
Windows CPU outputs** recorded [here](../qwen_tts_omni_cpu/README.md), despite
|
||||
matching declared BF16 settings and seed. The WSL profile's own 20 repeated
|
||||
outputs are identical. The cause of the cross-OS difference is not established,
|
||||
so Windows PCM parity is not claimed. The first WSL output passed the named ASR
|
||||
proxy, but broader speech quality and numerical agreement remain open.
|
||||
|
||||
A [post-portability-change native Windows regression](windows_regression_report.json)
|
||||
repeated the two named requests and one measured request with the original
|
||||
Windows pins: both standalone PCM parity checks still passed, in-flight
|
||||
cancellation drained the worker, and the ledger returned to zero. Its
|
||||
[worker log](windows_regression_worker.log) and [WAV](windows_regression.wav)
|
||||
are retained. This verifies the launch-path change did not alter the already
|
||||
qualified Windows fixture; it does not explain the WSL waveform difference.
|
||||
|
||||
The 0.25 s profile samples observed at most 3.39 GB summed process-tree RSS;
|
||||
Linux's `psutil` private-memory fallback used RSS, so that column is not a
|
||||
separate private-memory measurement. The samples do not prove loading peak,
|
||||
unique physical RAM, pagefile peak or concurrent admission. No power mode or
|
||||
thermal history was captured. The worker returns complete WAVs only: playable
|
||||
streaming, post-cancel restart, larger prompts and sustained real-time use are
|
||||
unqualified; measured RTF remains above one.
|
||||
|
||||
The [first failed attempt](attempt1_report.json) resolved the Linux virtualenv
|
||||
`python` symlink to a bare interpreter, which lost the venv packages. The
|
||||
adapter now preserves the declared launch path while hashing its target. The
|
||||
[second failed attempt](attempt2_report.json) produced both complete WAVs but
|
||||
hit the probe's Windows-only PCM assertion; `--allow-platform-variation` now
|
||||
records that difference explicitly. The [original logs](attempt1_driver.log),
|
||||
[worker error](attempt1_worker.log), [second driver](attempt2_driver.log) and
|
||||
[second worker](attempt2_worker.log) remain available.
|
||||
|
||||
Reproduce with
|
||||
[`probe_omni_qwen_tts_cpu.py`](../../../../experiments/probe_omni_qwen_tts_cpu.py)
|
||||
and [`probe_omni_qwen_tts_cpu_entrypoint.py`](../../../../experiments/probe_omni_qwen_tts_cpu_entrypoint.py)
|
||||
using the checkpoint and hashes above, the recorded Python interpreter hash
|
||||
`021044895e95be79dc2f110367607e684119afbc8ce75f6f0eec94844e0acec7`,
|
||||
the pinned overlay packages in the manifest, and a copy of the shim at
|
||||
`<overlay>/kernels/__init__.py`. Pass its SHA256
|
||||
`e155388b1045b573082534132778828a71203522f95a22680f8245362f1e06ee`
|
||||
as `--overlay-marker-sha256`. The serial profile used
|
||||
`--warmups 1 --repeats 20 --abort-check --allow-platform-variation`; the
|
||||
public check used `--expected-frames 82560` and PCM SHA256
|
||||
`66ca76a468807381c208b21ba458e5005510c752efa89323ee909749047ab1db`.
|
||||
+82
@@ -0,0 +1,82 @@
|
||||
INFO 09-24 00:25:57 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
|
||||
INFO 09-24 00:25:57 [patch.py:510] inductor factorable-divisibility patch: installed.
|
||||
INFO 09-24 00:25:57 [patch.py:589] [cumem-cuda] CuMemAllocator._python_free_callback patched: asleep guard extended to all platforms.
|
||||
INFO 09-24 00:25:58 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn
|
||||
WARNING 09-24 00:25:58 [stage_init_utils.py:1887] [stage_init] Failed to load transfer config: [Errno 21] Is a directory: '/home/zhout/project/edge_infer/vllm-omni-edge'
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] [StageRuntime] Stage initialization failed; shutting down 0 initialized client(s)
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] Traceback (most recent call last):
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/stage_runtime.py", line 252, in initialize
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] initialized_clients = self._initialize_stage_replicas(stage_plans, self._stage_init_timeout)
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/stage_runtime.py", line 589, in _initialize_stage_replicas
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] raise primary_exc
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/stage_runtime.py", line 550, in _init_group
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] client = self._initialize_replica(
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] ^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/stage_runtime.py", line 817, in _initialize_replica
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] return create_graph_client(
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] ^^^^^^^^^^^^^^^^^^^^
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/backends/__init__.py", line 23, in create_graph_client
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] return QwenTTSCPUStageClient(metadata, backend, ledger, reservation)
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/backends/qwen_tts_cpu.py", line 107, in __init__
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] raise RuntimeError(f"CPU TTS worker exited during load; see {self._log_path}")
|
||||
ERROR 09-24 00:26:11 [stage_runtime.py:265] RuntimeError: CPU TTS worker exited during load; see /home/zhout/project/edge_infer/vllm-omni-edge/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_wsl_cpu/worker.log
|
||||
{
|
||||
"scope": "complete resident native Windows CPU Qwen3-TTS text-to-WAV through Omni",
|
||||
"warmup_count": 0,
|
||||
"measured_count": 1,
|
||||
"memory_budget": {
|
||||
"capacities": {
|
||||
"host_ram": 17179869184
|
||||
},
|
||||
"demands": {
|
||||
"host_ram": 10737418240
|
||||
}
|
||||
},
|
||||
"status": "failed",
|
||||
"error": "RuntimeError: CPU TTS worker exited during load; see /home/zhout/project/edge_infer/vllm-omni-edge/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_wsl_cpu/worker.log",
|
||||
"ledger_after_shutdown": {
|
||||
"capacities": {
|
||||
"host_ram": 17179869184
|
||||
},
|
||||
"reserved": {
|
||||
"host_ram": 0
|
||||
},
|
||||
"owners": [],
|
||||
"quarantined": []
|
||||
}
|
||||
}
|
||||
Traceback (most recent call last):
|
||||
File "/home/zhout/project/edge_infer/vllm-omni-edge/benchmarks/edge_harness/experiments/probe_omni_qwen_tts_cpu.py", line 216, in <module>
|
||||
asyncio.run(main())
|
||||
File "/home/zhout/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/asyncio/runners.py", line 195, in run
|
||||
return runner.run(main)
|
||||
^^^^^^^^^^^^^^^^
|
||||
File "/home/zhout/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/asyncio/runners.py", line 118, in run
|
||||
return self._loop.run_until_complete(task)
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
File "/home/zhout/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/asyncio/base_events.py", line 691, in run_until_complete
|
||||
return future.result()
|
||||
^^^^^^^^^^^^^^^
|
||||
File "/home/zhout/project/edge_infer/vllm-omni-edge/benchmarks/edge_harness/experiments/probe_omni_qwen_tts_cpu.py", line 93, in main
|
||||
runtime.initialize()
|
||||
File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/stage_runtime.py", line 278, in initialize
|
||||
raise exc
|
||||
File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/stage_runtime.py", line 252, in initialize
|
||||
initialized_clients = self._initialize_stage_replicas(stage_plans, self._stage_init_timeout)
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/stage_runtime.py", line 589, in _initialize_stage_replicas
|
||||
raise primary_exc
|
||||
File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/stage_runtime.py", line 550, in _init_group
|
||||
client = self._initialize_replica(
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/stage_runtime.py", line 817, in _initialize_replica
|
||||
return create_graph_client(
|
||||
^^^^^^^^^^^^^^^^^^^^
|
||||
File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/backends/__init__.py", line 23, in create_graph_client
|
||||
return QwenTTSCPUStageClient(metadata, backend, ledger, reservation)
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/backends/qwen_tts_cpu.py", line 107, in __init__
|
||||
raise RuntimeError(f"CPU TTS worker exited during load; see {self._log_path}")
|
||||
RuntimeError: CPU TTS worker exited during load; see /home/zhout/project/edge_infer/vllm-omni-edge/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_wsl_cpu/worker.log
|
||||
+25
@@ -0,0 +1,25 @@
|
||||
{
|
||||
"scope": "complete resident native Windows CPU Qwen3-TTS text-to-WAV through Omni",
|
||||
"warmup_count": 0,
|
||||
"measured_count": 1,
|
||||
"memory_budget": {
|
||||
"capacities": {
|
||||
"host_ram": 17179869184
|
||||
},
|
||||
"demands": {
|
||||
"host_ram": 10737418240
|
||||
}
|
||||
},
|
||||
"status": "failed",
|
||||
"error": "RuntimeError: CPU TTS worker exited during load; see /home/zhout/project/edge_infer/vllm-omni-edge/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_wsl_cpu/worker.log",
|
||||
"ledger_after_shutdown": {
|
||||
"capacities": {
|
||||
"host_ram": 17179869184
|
||||
},
|
||||
"reserved": {
|
||||
"host_ram": 0
|
||||
},
|
||||
"owners": [],
|
||||
"quarantined": []
|
||||
}
|
||||
}
|
||||
+6
@@ -0,0 +1,6 @@
|
||||
Traceback (most recent call last):
|
||||
File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/backends/qwen_tts_cpu_worker.py", line 138, in <module>
|
||||
main()
|
||||
File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/backends/qwen_tts_cpu_worker.py", line 36, in main
|
||||
import numpy as np
|
||||
ModuleNotFoundError: No module named 'numpy'
|
||||
+137
@@ -0,0 +1,137 @@
|
||||
INFO 09-24 00:27:04 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
|
||||
INFO 09-24 00:27:04 [patch.py:510] inductor factorable-divisibility patch: installed.
|
||||
INFO 09-24 00:27:04 [patch.py:589] [cumem-cuda] CuMemAllocator._python_free_callback patched: asleep guard extended to all platforms.
|
||||
INFO 09-24 00:27:05 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn
|
||||
WARNING 09-24 00:27:05 [stage_init_utils.py:1887] [stage_init] Failed to load transfer config: [Errno 21] Is a directory: '/home/zhout/project/edge_infer/vllm-omni-edge'
|
||||
{
|
||||
"scope": "complete resident native Windows CPU Qwen3-TTS text-to-WAV through Omni",
|
||||
"warmup_count": 0,
|
||||
"measured_count": 1,
|
||||
"memory_budget": {
|
||||
"capacities": {
|
||||
"host_ram": 17179869184
|
||||
},
|
||||
"demands": {
|
||||
"host_ram": 10737418240
|
||||
}
|
||||
},
|
||||
"startup_s": 20.292010181990918,
|
||||
"execution_plan": {
|
||||
"backend": "external.qwen_tts.cpu.v1",
|
||||
"stage_id": 0,
|
||||
"worker_generation": "11b36820eb34430f80bebc322a38ab74",
|
||||
"worker_pid": 3642346,
|
||||
"worker_script_sha256": "862177bec7221114ec93f496602434dbefd36ca903d009ea35ffdca37b529205",
|
||||
"artifact_sha256": {
|
||||
"python_sha256": "021044895e95be79dc2f110367607e684119afbc8ce75f6f0eec94844e0acec7",
|
||||
"talker_sha256": "bc3c7e785eb961179c25450d1acff03f839e0002f2f3a5aeb67b5735c0fa2adb",
|
||||
"tokenizer_sha256": "836b7b357f5ea43e889936a3709af68dfe3751881acefe4ecf0dbd30ba571258"
|
||||
},
|
||||
"model_dir": "/mnt/c/Users/zhout/w2/models/Qwen3-TTS-0.6B-85e237c",
|
||||
"placement": "CPU BF16 SDPA, 8 threads; isolated Transformers 4.57.3",
|
||||
"worker_props": {
|
||||
"model_dir": "/mnt/c/Users/zhout/w2/models/Qwen3-TTS-0.6B-85e237c",
|
||||
"placement": "cpu",
|
||||
"torch": "2.13.0+cu130",
|
||||
"transformers": "4.57.3",
|
||||
"qwen_tts": "0.1.1",
|
||||
"dtype": "bfloat16",
|
||||
"attention": "sdpa",
|
||||
"threads": 8,
|
||||
"sample_rate": 24000,
|
||||
"speaker": "Ryan",
|
||||
"max_new_tokens": 64,
|
||||
"load_s": 3.402708667970728
|
||||
},
|
||||
"loaded_rss_bytes": 2125488128,
|
||||
"loaded_private_bytes": 2125488128,
|
||||
"loaded_process_tree": [
|
||||
{
|
||||
"pid": 3642346,
|
||||
"rss_bytes": 2125488128,
|
||||
"private_bytes": 2125488128
|
||||
}
|
||||
],
|
||||
"reserved_bytes": {
|
||||
"host_ram": 10737418240
|
||||
},
|
||||
"memory_overhead_bytes": 6442450944,
|
||||
"max_text_bytes": 4096,
|
||||
"max_wav_bytes": 4194304,
|
||||
"max_audio_s": 10.0,
|
||||
"voice": "Ryan",
|
||||
"seed": 42,
|
||||
"request_capacity": 1,
|
||||
"stateful_session": "Qwen PyTorch worker-owned; complete WAV per request",
|
||||
"evidence": "B"
|
||||
},
|
||||
"checks": [
|
||||
{
|
||||
"text": "Hello from the local computer.",
|
||||
"wall_s": 12.656105556990951,
|
||||
"frames": 82560,
|
||||
"duration_s": 3.44,
|
||||
"pcm_sha256": "66ca76a468807381c208b21ba458e5005510c752efa89323ee909749047ab1db",
|
||||
"stage_event": {
|
||||
"request_id": "cpu-tts-434866642434220",
|
||||
"stage_id": 0,
|
||||
"epoch": 1,
|
||||
"seq": 1,
|
||||
"kind": "audio",
|
||||
"worker_generation": "11b36820eb34430f80bebc322a38ab74",
|
||||
"buffers": [],
|
||||
"terminal": true,
|
||||
"error": null
|
||||
}
|
||||
},
|
||||
{
|
||||
"text": "The blue car is parked beside the library.",
|
||||
"wall_s": 8.73547542700544,
|
||||
"frames": 80640,
|
||||
"duration_s": 3.36,
|
||||
"pcm_sha256": "c661953310ee1b6b39806156f4180a0793d4372425d13f58bc1440b1d027eeb0",
|
||||
"stage_event": {
|
||||
"request_id": "cpu-tts-434879299154073",
|
||||
"stage_id": 0,
|
||||
"epoch": 2,
|
||||
"seq": 1,
|
||||
"kind": "audio",
|
||||
"worker_generation": "11b36820eb34430f80bebc322a38ab74",
|
||||
"buffers": [],
|
||||
"terminal": true,
|
||||
"error": null
|
||||
}
|
||||
}
|
||||
],
|
||||
"standalone_pcm_parity": [
|
||||
false,
|
||||
false
|
||||
],
|
||||
"status": "failed",
|
||||
"error": "RuntimeError: CPU Omni PCM differs from the same checkpoint's standalone output",
|
||||
"ledger_after_shutdown": {
|
||||
"capacities": {
|
||||
"host_ram": 17179869184
|
||||
},
|
||||
"reserved": {
|
||||
"host_ram": 0
|
||||
},
|
||||
"owners": [],
|
||||
"quarantined": []
|
||||
}
|
||||
}
|
||||
Traceback (most recent call last):
|
||||
File "/home/zhout/project/edge_infer/vllm-omni-edge/benchmarks/edge_harness/experiments/probe_omni_qwen_tts_cpu.py", line 216, in <module>
|
||||
asyncio.run(main())
|
||||
File "/home/zhout/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/asyncio/runners.py", line 195, in run
|
||||
return runner.run(main)
|
||||
^^^^^^^^^^^^^^^^
|
||||
File "/home/zhout/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/asyncio/runners.py", line 118, in run
|
||||
return self._loop.run_until_complete(task)
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
File "/home/zhout/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/asyncio/base_events.py", line 691, in run_until_complete
|
||||
return future.result()
|
||||
^^^^^^^^^^^^^^^
|
||||
File "/home/zhout/project/edge_infer/vllm-omni-edge/benchmarks/edge_harness/experiments/probe_omni_qwen_tts_cpu.py", line 172, in main
|
||||
raise RuntimeError("CPU Omni PCM differs from the same checkpoint's standalone output")
|
||||
RuntimeError: CPU Omni PCM differs from the same checkpoint's standalone output
|
||||
+117
@@ -0,0 +1,117 @@
|
||||
{
|
||||
"scope": "complete resident native Windows CPU Qwen3-TTS text-to-WAV through Omni",
|
||||
"warmup_count": 0,
|
||||
"measured_count": 1,
|
||||
"memory_budget": {
|
||||
"capacities": {
|
||||
"host_ram": 17179869184
|
||||
},
|
||||
"demands": {
|
||||
"host_ram": 10737418240
|
||||
}
|
||||
},
|
||||
"startup_s": 20.292010181990918,
|
||||
"execution_plan": {
|
||||
"backend": "external.qwen_tts.cpu.v1",
|
||||
"stage_id": 0,
|
||||
"worker_generation": "11b36820eb34430f80bebc322a38ab74",
|
||||
"worker_pid": 3642346,
|
||||
"worker_script_sha256": "862177bec7221114ec93f496602434dbefd36ca903d009ea35ffdca37b529205",
|
||||
"artifact_sha256": {
|
||||
"python_sha256": "021044895e95be79dc2f110367607e684119afbc8ce75f6f0eec94844e0acec7",
|
||||
"talker_sha256": "bc3c7e785eb961179c25450d1acff03f839e0002f2f3a5aeb67b5735c0fa2adb",
|
||||
"tokenizer_sha256": "836b7b357f5ea43e889936a3709af68dfe3751881acefe4ecf0dbd30ba571258"
|
||||
},
|
||||
"model_dir": "/mnt/c/Users/zhout/w2/models/Qwen3-TTS-0.6B-85e237c",
|
||||
"placement": "CPU BF16 SDPA, 8 threads; isolated Transformers 4.57.3",
|
||||
"worker_props": {
|
||||
"model_dir": "/mnt/c/Users/zhout/w2/models/Qwen3-TTS-0.6B-85e237c",
|
||||
"placement": "cpu",
|
||||
"torch": "2.13.0+cu130",
|
||||
"transformers": "4.57.3",
|
||||
"qwen_tts": "0.1.1",
|
||||
"dtype": "bfloat16",
|
||||
"attention": "sdpa",
|
||||
"threads": 8,
|
||||
"sample_rate": 24000,
|
||||
"speaker": "Ryan",
|
||||
"max_new_tokens": 64,
|
||||
"load_s": 3.402708667970728
|
||||
},
|
||||
"loaded_rss_bytes": 2125488128,
|
||||
"loaded_private_bytes": 2125488128,
|
||||
"loaded_process_tree": [
|
||||
{
|
||||
"pid": 3642346,
|
||||
"rss_bytes": 2125488128,
|
||||
"private_bytes": 2125488128
|
||||
}
|
||||
],
|
||||
"reserved_bytes": {
|
||||
"host_ram": 10737418240
|
||||
},
|
||||
"memory_overhead_bytes": 6442450944,
|
||||
"max_text_bytes": 4096,
|
||||
"max_wav_bytes": 4194304,
|
||||
"max_audio_s": 10.0,
|
||||
"voice": "Ryan",
|
||||
"seed": 42,
|
||||
"request_capacity": 1,
|
||||
"stateful_session": "Qwen PyTorch worker-owned; complete WAV per request",
|
||||
"evidence": "B"
|
||||
},
|
||||
"checks": [
|
||||
{
|
||||
"text": "Hello from the local computer.",
|
||||
"wall_s": 12.656105556990951,
|
||||
"frames": 82560,
|
||||
"duration_s": 3.44,
|
||||
"pcm_sha256": "66ca76a468807381c208b21ba458e5005510c752efa89323ee909749047ab1db",
|
||||
"stage_event": {
|
||||
"request_id": "cpu-tts-434866642434220",
|
||||
"stage_id": 0,
|
||||
"epoch": 1,
|
||||
"seq": 1,
|
||||
"kind": "audio",
|
||||
"worker_generation": "11b36820eb34430f80bebc322a38ab74",
|
||||
"buffers": [],
|
||||
"terminal": true,
|
||||
"error": null
|
||||
}
|
||||
},
|
||||
{
|
||||
"text": "The blue car is parked beside the library.",
|
||||
"wall_s": 8.73547542700544,
|
||||
"frames": 80640,
|
||||
"duration_s": 3.36,
|
||||
"pcm_sha256": "c661953310ee1b6b39806156f4180a0793d4372425d13f58bc1440b1d027eeb0",
|
||||
"stage_event": {
|
||||
"request_id": "cpu-tts-434879299154073",
|
||||
"stage_id": 0,
|
||||
"epoch": 2,
|
||||
"seq": 1,
|
||||
"kind": "audio",
|
||||
"worker_generation": "11b36820eb34430f80bebc322a38ab74",
|
||||
"buffers": [],
|
||||
"terminal": true,
|
||||
"error": null
|
||||
}
|
||||
}
|
||||
],
|
||||
"standalone_pcm_parity": [
|
||||
false,
|
||||
false
|
||||
],
|
||||
"status": "failed",
|
||||
"error": "RuntimeError: CPU Omni PCM differs from the same checkpoint's standalone output",
|
||||
"ledger_after_shutdown": {
|
||||
"capacities": {
|
||||
"host_ram": 17179869184
|
||||
},
|
||||
"reserved": {
|
||||
"host_ram": 0
|
||||
},
|
||||
"owners": [],
|
||||
"quarantined": []
|
||||
}
|
||||
}
|
||||
+23
@@ -0,0 +1,23 @@
|
||||
|
||||
********
|
||||
Warning: flash-attn is not installed. Will only run the manual PyTorch version. Please install flash-attn for faster inference.
|
||||
********
|
||||
|
||||
/bin/sh: 1: sox: not found
|
||||
SoX could not be found!
|
||||
|
||||
If you do not have SoX, proceed here:
|
||||
- - - http://sox.sourceforge.net/ - - -
|
||||
|
||||
If you do (or think that you should) have SoX, double-check your
|
||||
path variables.
|
||||
|
||||
qwen-tts-cpu-worker ready port=36531 props={"model_dir": "/mnt/c/Users/zhout/w2/models/Qwen3-TTS-0.6B-85e237c", "placement": "cpu", "torch": "2.13.0+cu130", "transformers": "4.57.3", "qwen_tts": "0.1.1", "dtype": "bfloat16", "attention": "sdpa", "threads": 8, "sample_rate": 24000, "speaker": "Ryan", "max_new_tokens": 64, "load_s": 3.402708667970728}
|
||||
127.0.0.1 - - [24/Sep/2026 00:27:25] "GET /health HTTP/1.1" 200 -
|
||||
127.0.0.1 - - [24/Sep/2026 00:27:25] "GET /props HTTP/1.1" 200 -
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:27:38] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=12.619
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:27:47] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.360 wall_s=8.729
|
||||
+1
@@ -0,0 +1 @@
|
||||
raise ImportError("optional kernels disabled for CPU-only Qwen3-TTS worker")
|
||||
+51
@@ -0,0 +1,51 @@
|
||||
INFO 09-24 00:33:52 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
|
||||
INFO 09-24 00:33:52 [patch.py:510] inductor factorable-divisibility patch: installed.
|
||||
INFO 09-24 00:33:52 [patch.py:589] [cumem-cuda] CuMemAllocator._python_free_callback patched: asleep guard extended to all platforms.
|
||||
INFO 09-24 00:33:53 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn
|
||||
WARNING 09-24 00:33:53 [stage_init_utils.py:1887] [stage_init] Failed to load transfer config: [Errno 21] Is a directory: '/home/zhout/project/edge_infer/vllm-omni-edge'
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] [StageRuntime] Stage initialization failed; shutting down 0 initialized client(s)
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] Traceback (most recent call last):
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/stage_runtime.py", line 252, in initialize
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] initialized_clients = self._initialize_stage_replicas(stage_plans, self._stage_init_timeout)
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/stage_runtime.py", line 589, in _initialize_stage_replicas
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] raise primary_exc
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/stage_runtime.py", line 550, in _init_group
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] client = self._initialize_replica(
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] ^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/stage_runtime.py", line 817, in _initialize_replica
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] return create_graph_client(
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] ^^^^^^^^^^^^^^^^^^^^
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/backends/__init__.py", line 23, in create_graph_client
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] return QwenTTSCPUStageClient(metadata, backend, ledger, reservation)
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] File "/home/zhout/project/edge_infer/vllm-omni-edge/vllm_omni/engine/backends/qwen_tts_cpu.py", line 89, in __init__
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] raise ResourceUnavailable("CPU TTS weights plus state/workspace/headroom exceed reservation")
|
||||
ERROR 09-24 00:34:06 [stage_runtime.py:265] vllm_omni.engine.resource_ledger.ResourceUnavailable: CPU TTS weights plus state/workspace/headroom exceed reservation
|
||||
{
|
||||
"scope": "complete resident local CPU Qwen3-TTS text-to-WAV through Omni",
|
||||
"host_os": "Linux-6.18.33.2-microsoft-standard-WSL2-x86_64-with-glibc2.43",
|
||||
"allow_platform_variation": true,
|
||||
"warmup_count": 1,
|
||||
"measured_count": 20,
|
||||
"memory_budget": {
|
||||
"capacities": {
|
||||
"host_ram": 17179869184
|
||||
},
|
||||
"demands": {
|
||||
"host_ram": 4294967296
|
||||
}
|
||||
},
|
||||
"admission_refusal": "ResourceUnavailable: CPU TTS weights plus state/workspace/headroom exceed reservation",
|
||||
"status": "passed",
|
||||
"ledger_after_shutdown": {
|
||||
"capacities": {
|
||||
"host_ram": 17179869184
|
||||
},
|
||||
"reserved": {
|
||||
"host_ram": 0
|
||||
},
|
||||
"owners": [],
|
||||
"quarantined": []
|
||||
}
|
||||
}
|
||||
+27
@@ -0,0 +1,27 @@
|
||||
{
|
||||
"scope": "complete resident local CPU Qwen3-TTS text-to-WAV through Omni",
|
||||
"host_os": "Linux-6.18.33.2-microsoft-standard-WSL2-x86_64-with-glibc2.43",
|
||||
"allow_platform_variation": true,
|
||||
"warmup_count": 1,
|
||||
"measured_count": 20,
|
||||
"memory_budget": {
|
||||
"capacities": {
|
||||
"host_ram": 17179869184
|
||||
},
|
||||
"demands": {
|
||||
"host_ram": 4294967296
|
||||
}
|
||||
},
|
||||
"admission_refusal": "ResourceUnavailable: CPU TTS weights plus state/workspace/headroom exceed reservation",
|
||||
"status": "passed",
|
||||
"ledger_after_shutdown": {
|
||||
"capacities": {
|
||||
"host_ram": 17179869184
|
||||
},
|
||||
"reserved": {
|
||||
"host_ram": 0
|
||||
},
|
||||
"owners": [],
|
||||
"quarantined": []
|
||||
}
|
||||
}
|
||||
+40
@@ -0,0 +1,40 @@
|
||||
{
|
||||
"scope": "WSL CPU Qwen worker-only Python overlay; base Omni interpreter is separate",
|
||||
"packages": {
|
||||
"qwen-tts": {
|
||||
"version": "0.1.1",
|
||||
"path": "qwen_tts/__init__.py",
|
||||
"sha256": "ea52de59d070fde366467a6902d0edcfc1b0575b8c570a0c71020c41d6a593ed"
|
||||
},
|
||||
"transformers": {
|
||||
"version": "4.57.3",
|
||||
"path": "transformers/__init__.py",
|
||||
"sha256": "7eb0743ed843f24c1d4e8b8daf4b18249cb81403b4571a70c00be0a4c0a67bd4"
|
||||
},
|
||||
"huggingface-hub": {
|
||||
"version": "0.36.2",
|
||||
"path": "huggingface_hub/__init__.py",
|
||||
"sha256": "e586b64612482128fd24119364bc4dc6dca938407134e454a5e876341b364e35"
|
||||
},
|
||||
"sox": {
|
||||
"version": "1.5.0",
|
||||
"path": "sox/__init__.py",
|
||||
"sha256": "0c3e15258ce1f95ca050e045af634d1b16956af2b47ff7670f78b08567d19d47"
|
||||
},
|
||||
"optional-kernels-stub": {
|
||||
"version": "local",
|
||||
"path": "kernels/__init__.py",
|
||||
"sha256": "e155388b1045b573082534132778828a71203522f95a22680f8245362f1e06ee"
|
||||
}
|
||||
},
|
||||
"base_interpreter_packages": {
|
||||
"torch": "2.13.0",
|
||||
"torchaudio": "2.11.0",
|
||||
"numpy": "2.3.5",
|
||||
"tokenizers": "0.22.2",
|
||||
"accelerate": "1.12.0",
|
||||
"safetensors": "0.8.0",
|
||||
"librosa": "1.0.0",
|
||||
"soundfile": "0.14.0"
|
||||
}
|
||||
}
|
||||
+6390
File diff suppressed because it is too large.
Load diff
BIN
Binary file not shown.
+16
@@ -0,0 +1,16 @@
|
||||
{
|
||||
"scope": "local Whisper tiny.en ASR proxy for generated speech; no perceptual quality claim",
|
||||
"asr_model": "openai/whisper-tiny.en",
|
||||
"asr_revision": "87c7102498dcde7456f24cfd30239ca606ed9063",
|
||||
"asr_weights_sha256": "db59695928ded6043adaef491a53ef4e12da9611184d77c53baa691a60b958ad",
|
||||
"transformers": "4.57.3",
|
||||
"torch": "2.4.1+cpu",
|
||||
"wav_sha256": "f1b5348ed3fe95e3ce1518656383e631cb88a4ec59fe80eacf8b8fc6b4023566",
|
||||
"input_sample_rate": 24000,
|
||||
"input_duration_s": 3.44,
|
||||
"reference_text": "Hello from the local computer.",
|
||||
"transcript": "Hello from the local computer!",
|
||||
"word_error_rate": 0.0,
|
||||
"asr_load_s": 0.8476594999665394,
|
||||
"asr_inference_s": 0.8502050000242889
|
||||
}
|
||||
+6395
File diff suppressed because it is too large.
Load diff
+87
@@ -0,0 +1,87 @@
|
||||
|
||||
********
|
||||
Warning: flash-attn is not installed. Will only run the manual PyTorch version. Please install flash-attn for faster inference.
|
||||
********
|
||||
|
||||
/bin/sh: 1: sox: not found
|
||||
SoX could not be found!
|
||||
|
||||
If you do not have SoX, proceed here:
|
||||
- - - http://sox.sourceforge.net/ - - -
|
||||
|
||||
If you do (or think that you should) have SoX, double-check your
|
||||
path variables.
|
||||
|
||||
qwen-tts-cpu-worker ready port=42833 props={"model_dir": "/mnt/c/Users/zhout/w2/models/Qwen3-TTS-0.6B-85e237c", "placement": "cpu", "torch": "2.13.0+cu130", "transformers": "4.57.3", "qwen_tts": "0.1.1", "dtype": "bfloat16", "attention": "sdpa", "threads": 8, "sample_rate": 24000, "speaker": "Ryan", "max_new_tokens": 64, "load_s": 3.237298524007201}
|
||||
127.0.0.1 - - [24/Sep/2026 00:29:33] "GET /health HTTP/1.1" 200 -
|
||||
127.0.0.1 - - [24/Sep/2026 00:29:33] "GET /props HTTP/1.1" 200 -
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:29:45] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=12.176
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:29:54] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.360 wall_s=8.703
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:30:02] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.348
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:30:11] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.261
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:30:19] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.223
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:30:27] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.163
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:30:35] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.229
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:30:44] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.398
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:30:52] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.129
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:31:00] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.251
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:31:08] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.431
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:31:17] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.461
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:31:25] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.501
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:31:34] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.250
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:31:42] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.414
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:31:51] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.367
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:31:59] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.346
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:32:07] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.326
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:32:16] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.388
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:32:24] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.201
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:32:32] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.207
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:32:40] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.165
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:32:48] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=8.282
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
+49
@@ -0,0 +1,49 @@
|
||||
INFO 09-24 00:34:52 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
|
||||
INFO 09-24 00:34:52 [patch.py:510] inductor factorable-divisibility patch: installed.
|
||||
INFO 09-24 00:34:52 [patch.py:589] [cumem-cuda] CuMemAllocator._python_free_callback patched: asleep guard extended to all platforms.
|
||||
INFO 09-24 00:34:53 [omni_base.py:216] [AsyncOmni] Initializing with model /mnt/c/Users/zhout/w2/models/Qwen3-TTS-0.6B-85e237c
|
||||
INFO 09-24 00:34:53 [async_omni_engine.py:172] [AsyncOmniEngine] Initializing with model /mnt/c/Users/zhout/w2/models/Qwen3-TTS-0.6B-85e237c
|
||||
ERROR 09-24 00:34:53 [repo_utils.py:119] Error parsing config for /mnt/c/Users/zhout/w2/models/Qwen3-TTS-0.6B-85e237c: The checkpoint you are trying to load has model type `qwen3_tts` but Transformers does not recognize this architecture. This could be because of an issue with the checkpoint, or because your version of Transformers is out of date.
|
||||
ERROR 09-24 00:34:53 [repo_utils.py:119]
|
||||
ERROR 09-24 00:34:53 [repo_utils.py:119] You can update Transformers with the command `pip install --upgrade transformers`. If this does not work, and the checkpoint is very new, then there may not be a release version that supports this model yet. In this case, you can get the most up-to-date code by installing Transformers from source with the command `pip install git+https://github.com/huggingface/transformers.git`, retrying 1 of 2
|
||||
ERROR 09-24 00:34:55 [repo_utils.py:117] Error parsing config for /mnt/c/Users/zhout/w2/models/Qwen3-TTS-0.6B-85e237c: The checkpoint you are trying to load has model type `qwen3_tts` but Transformers does not recognize this architecture. This could be because of an issue with the checkpoint, or because your version of Transformers is out of date.
|
||||
ERROR 09-24 00:34:55 [repo_utils.py:117]
|
||||
ERROR 09-24 00:34:55 [repo_utils.py:117] You can update Transformers with the command `pip install --upgrade transformers`. If this does not work, and the checkpoint is very new, then there may not be a release version that supports this model yet. In this case, you can get the most up-to-date code by installing Transformers from source with the command `pip install git+https://github.com/huggingface/transformers.git`
|
||||
INFO 09-24 00:34:55 [async_omni_engine.py:290] [AsyncOmniEngine] Launching Orchestrator thread with 1 stages
|
||||
INFO 09-24 00:34:55 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn
|
||||
INFO 09-24 00:34:55 [initialization.py:356] Loaded OmniTransferConfig with 0 connector configurations
|
||||
INFO 09-24 00:35:13 [orchestrator.py:604] [Orchestrator] Starting event loop
|
||||
INFO 09-24 00:35:13 [async_omni_engine.py:323] [AsyncOmniEngine] Orchestrator ready with 1 stages
|
||||
INFO 09-24 00:35:13 [omni_base.py:235] [AsyncOmni] AsyncOmniEngine initialized in 20.28 seconds
|
||||
INFO 09-24 00:35:13 [omni_base.py:255] [AsyncOmni] Initialized with 1 stages for model /mnt/c/Users/zhout/w2/models/Qwen3-TTS-0.6B-85e237c
|
||||
INFO 09-24 00:35:25 [omni_base.py:781] [AsyncOmni] Shutting down
|
||||
INFO 09-24 00:35:25 [async_omni_engine.py:1807] [AsyncOmniEngine] Shutting down Orchestrator
|
||||
INFO 09-24 00:35:25 [orchestrator.py:705] [Orchestrator] Received shutdown signal
|
||||
INFO 09-24 00:35:25 [orchestrator.py:3092] [Orchestrator] Shutting down all 1 client(s)
|
||||
INFO 09-24 00:35:25 [stage_pool.py:1354] [StagePool] Stage 0 replica 0 shut down
|
||||
{
|
||||
"entrypoint": "AsyncOmni.generate",
|
||||
"pipeline": "qwen3_tts_cpu_whole",
|
||||
"startup_s": 20.290129358007107,
|
||||
"request_wall_s": 12.097587361990009,
|
||||
"outputs": [
|
||||
{
|
||||
"request_id": "public-cpu-tts-1",
|
||||
"sample_rate": 24000,
|
||||
"frames": 82560,
|
||||
"pcm_sha256": "66ca76a468807381c208b21ba458e5005510c752efa89323ee909749047ab1db",
|
||||
"stage_event": {
|
||||
"request_id": "public-cpu-tts-1-843808ee",
|
||||
"stage_id": 0,
|
||||
"epoch": 1,
|
||||
"seq": 1,
|
||||
"kind": "audio",
|
||||
"worker_generation": "bc022f43fee646f18572e4d6cef602a6",
|
||||
"buffers": [],
|
||||
"terminal": true,
|
||||
"error": null
|
||||
}
|
||||
}
|
||||
],
|
||||
"status": "passed"
|
||||
}
|
||||
+26
@@ -0,0 +1,26 @@
|
||||
{
|
||||
"entrypoint": "AsyncOmni.generate",
|
||||
"pipeline": "qwen3_tts_cpu_whole",
|
||||
"startup_s": 20.290129358007107,
|
||||
"request_wall_s": 12.097587361990009,
|
||||
"outputs": [
|
||||
{
|
||||
"request_id": "public-cpu-tts-1",
|
||||
"sample_rate": 24000,
|
||||
"frames": 82560,
|
||||
"pcm_sha256": "66ca76a468807381c208b21ba458e5005510c752efa89323ee909749047ab1db",
|
||||
"stage_event": {
|
||||
"request_id": "public-cpu-tts-1-843808ee",
|
||||
"stage_id": 0,
|
||||
"epoch": 1,
|
||||
"seq": 1,
|
||||
"kind": "audio",
|
||||
"worker_generation": "bc022f43fee646f18572e4d6cef602a6",
|
||||
"buffers": [],
|
||||
"terminal": true,
|
||||
"error": null
|
||||
}
|
||||
}
|
||||
],
|
||||
"status": "passed"
|
||||
}
|
||||
+20
@@ -0,0 +1,20 @@
|
||||
|
||||
********
|
||||
Warning: flash-attn is not installed. Will only run the manual PyTorch version. Please install flash-attn for faster inference.
|
||||
********
|
||||
|
||||
/bin/sh: 1: sox: not found
|
||||
SoX could not be found!
|
||||
|
||||
If you do not have SoX, proceed here:
|
||||
- - - http://sox.sourceforge.net/ - - -
|
||||
|
||||
If you do (or think that you should) have SoX, double-check your
|
||||
path variables.
|
||||
|
||||
qwen-tts-cpu-worker ready port=48689 props={"model_dir": "/mnt/c/Users/zhout/w2/models/Qwen3-TTS-0.6B-85e237c", "placement": "cpu", "torch": "2.13.0+cu130", "transformers": "4.57.3", "qwen_tts": "0.1.1", "dtype": "bfloat16", "attention": "sdpa", "threads": 8, "sample_rate": 24000, "speaker": "Ryan", "max_new_tokens": 64, "load_s": 3.1872041929746047}
|
||||
127.0.0.1 - - [24/Sep/2026 00:35:13] "GET /health HTTP/1.1" 200 -
|
||||
127.0.0.1 - - [24/Sep/2026 00:35:13] "GET /props HTTP/1.1" 200 -
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:35:25] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=12.089
|
||||
BIN
Binary file not shown.
+1436
File diff suppressed because it is too large.
Load diff
+31
@@ -0,0 +1,31 @@
|
||||
|
||||
********
|
||||
Warning: flash-attn is not installed. Will only run the manual PyTorch version. Please install flash-attn for faster inference.
|
||||
********
|
||||
|
||||
'\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer'
|
||||
用作为当前目录的以上路径启动了 CMD.EXE。
|
||||
UNC 路径不受支持。默认值设为 Windows 目录。
|
||||
'sox' 不是内部或外部命令，也不是可运行的程序
|
||||
或批处理文件。
|
||||
SoX could not be found!
|
||||
|
||||
If you do not have SoX, proceed here:
|
||||
- - - http://sox.sourceforge.net/ - - -
|
||||
|
||||
If you do (or think that you should) have SoX, double-check your
|
||||
path variables.
|
||||
|
||||
qwen-tts-cpu-worker ready port=61654 props={"model_dir": "C:\\Users\\zhout\\w2\\models\\Qwen3-TTS-0.6B-85e237c", "placement": "cpu", "torch": "2.13.0+cu130", "transformers": "4.57.3", "qwen_tts": "0.1.1", "dtype": "bfloat16", "attention": "sdpa", "threads": 8, "sample_rate": 24000, "speaker": "Ryan", "max_new_tokens": 64, "load_s": 1.6466022999957204}
|
||||
127.0.0.1 - - [24/Sep/2026 00:38:46] "GET /health HTTP/1.1" 200 -
|
||||
127.0.0.1 - - [24/Sep/2026 00:38:46] "GET /props HTTP/1.1" 200 -
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:39:02] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=4.560 wall_s=15.215
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:39:13] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=3.440 wall_s=11.343
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
127.0.0.1 - - [24/Sep/2026 00:39:28] "POST /v1/audio/speech HTTP/1.1" 200 -
|
||||
tts request duration_s=4.560 wall_s=15.133
|
||||
Setting `pad_token_id` to `eos_token_id`:2150 for open-end generation.
|
||||
@@ -2,13 +2,19 @@
|
||||
|
||||
All 60 pairings have a disposition. The timing tables below remain the 2026-09-22 measurements. The configuration matrix and affected per-cell reasons incorporate the [2026-09-23 scoped recovery runs](../../../e2e_recovery_20260923/README.md) and [WSL CPU/native Windows expansion](../../../e2e_expansion_20260923/README.md). The expansion adds separate 20-request serial MiniCPM-o text-to-speech profiles on WSL CPU and native Windows GPU, a Spark output-head component profile on the HX370 AMD NPU, complete Spark text-request profiles through standalone llama.cpp and a bounded Omni whole-session stage on native Windows CPU and Radeon 890M, native Windows CPU Qwen3-TTS text-to-WAV profiles through both standalone PyTorch and a bounded Omni whole-session stage, and a hybrid CPU+Radeon 890M Qwen3-TTS profile likewise repeated through a bounded Omni whole-session stage. A pinned GGUF MiniCPM-o route adds combined image+audio to text+speech runs on WSL CPU, native Windows CPU and native Windows Radeon 890M+CPU. All three now pass a bounded complete request through an Omni graph stage; the WSL CPU path additionally passed one warmup plus 20 serial measured requests. These are scoped synthetic cases, not full release qualification.
|
||||
|
||||
## Configuration × model
|
||||
## Configuration × model
|
||||
|
||||
The WSL CPU Qwen3-TTS cell also has a separate bounded Omni whole-session
|
||||
text-to-complete-WAV profile with one warmup, 20 measured requests, a public
|
||||
audio request, pinned ASR proxy, memory refusal and cancellation evidence.
|
||||
It does not replace the earlier short-stream finding or establish playable
|
||||
streaming.
|
||||
|
||||
NOT E2E means no complete device/model workload was verified. “Scoped E2E” identifies only the named input/output path; “standalone” means the whole model ran outside Omni. The Spark GGUF Omni stage uses a complete-request control path; it has no incremental token stream or restart-after-cancel contract yet. “Synthetic policy pass” does not establish a real robot-task result. See the per-cell reasons below.
|
||||
|
||||
| Device | Spark-X2.5 | Qwen3-TTS 0.6B CustomVoice | Qwen3.8-27B | MiniCPM-o 4.5 | InternVLA-A1 |
|
||||
|---|---|---|---|---|---|
|
||||
| PC HX370: CPU / WSL | profile completed | scoped E2E: short stream | NOT E2E | scoped E2E: Omni GGUF combined image+audio → text+speech | synthetic policy pass |
|
||||
| PC HX370: CPU / WSL | profile completed | scoped E2E: Omni CPU text → complete WAV | NOT E2E | scoped E2E: Omni GGUF combined image+audio → text+speech | synthetic policy pass |
|
||||
| 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 | synthetic policy pass |
|
||||
| PC HX370 + RTX 5090 Laptop / WSL | profile completed | profile completed | scoped E2E: text + image | scoped E2E: text/image/audio → speech | synthetic policy pass |
|
||||
| PC HX370 + RTX 5090 Laptop / Windows | profile completed | profile completed | NOT E2E | scoped E2E: text/image/audio → speech | synthetic policy pass |
|
||||
@@ -26,7 +32,7 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i
|
||||
| Device | Model | Current profiling disposition | Reason / remaining work | Next step |
|
||||
|---|---|---|---|---|
|
||||
| PC HX370: CPU / WSL | Spark-X2.5 | PROFILE_COMPLETED | not fully qualified: reference-quality and remaining reliability/performance gates must be reviewed | Qualify a matching vLLM 0.29 CPU build and token parity separately. |
|
||||
| PC HX370: CPU / WSL | Qwen3-TTS 0.6B CustomVoice | SCOPED_E2E | The 2026-09-23 compatibility fix enabled real-weight two-frame PCM and four short-stream requests, including cancellation recovery. One measured request generated 3.04 s of audio in 17.02 s (RTF 5.60) with four simulated playback underruns. The CPU wheel remains vLLM 0.28 versus Omni 0.29. | Qualify a matching CPU runtime, speech quality, uninterrupted playback and sustained load; [recovery evidence](../../../e2e_recovery_20260923/README.md). |
|
||||
| PC HX370: CPU / WSL | Qwen3-TTS 0.6B CustomVoice | SCOPED_E2E_OMNI_COMPLETE_REQUEST | The earlier vLLM 0.28 CPU path passed four short streams but had four simulated playback underruns in one measured request; its wheel remains mismatched with Omni 0.29. Separately, the pinned BF16 0.6B CustomVoice checkpoint ran two named text-to-complete-WAV requests through an Omni 0.29 whole-session stage with an isolated CPU worker on WSL. One warmup plus 20 serial measured requests returned identical 3.44 s audio; nearest-rank complete-request wall p50/p95 was 8.27/8.47 s (RTF 2.40/2.46), excluding 19.82 s startup. Pinned Whisper tiny.en transcribed the first measured WAV exactly (WER 0, one-output intelligibility proxy). Public `AsyncOmni` audio output passed; the stage verified weights, interpreter and CPU placement, reserved 10 GiB host RAM, refused 4 GiB before launch, and drained in-flight cancellation with zero remaining reservation. WSL PCM differs from the same-checkpoint native Windows outputs despite the declared settings; its cause is unqualified. Sampled process-tree RSS reached 3.39 GB, not a loading peak. This worker delivers only a complete WAV, not a playable stream. | Qualify cross-OS numerical/quality difference, playable streaming without underruns, post-cancel restart, loading peak, concurrency and sustained power/thermal behavior; [WSL Omni raw evidence](../../../e2e_expansion_20260923/evidence/qwen_tts_omni_wsl_cpu/README.md), [earlier stream evidence](../../../e2e_recovery_20260923/README.md). |
|
||||
| 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 | The Place_Markpen checkpoint strict-loaded in CPU-only PyTorch and produced finite `[1, 50, 32]` actions from zero observations/noise in one 3.25 s forward. Compared with the same CUDA checkpoint/input, relative L2 error was 1.30% and cosine similarity 0.999917; no task-quality tolerance is established. | Validate real observations against reference actions, units/timestamps, repeated latency and memory before robot-task qualification; [CPU evidence](../../../e2e_expansion_20260923/README.md). |
|
||||
|
||||
@@ -193,6 +193,17 @@ The 2026-09-23 [Spark GGUF Omni follow-up](../../benchmarks/edge_harness/results
|
||||
|
||||
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.
|
||||
|
||||
The [WSL CPU Qwen3-TTS Omni follow-up](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_wsl_cpu/README.md)
|
||||
adds the same bounded whole-session CPU stage on Ubuntu/WSL with a pinned
|
||||
worker-only dependency overlay. One warmup plus 20 serial complete requests
|
||||
produced identical 3.44 s WAVs at nearest-rank wall p50/p95 8.27/8.47 s,
|
||||
excluding worker startup. The public audio path, one-output Whisper proxy
|
||||
(WER 0), 10 GiB admission, 4 GiB refusal and cancellation drain passed. WSL
|
||||
PCM differs from the same-checkpoint native Windows result; its numerical
|
||||
cause remains open. The existing vLLM 0.28 short-stream path still has playback
|
||||
underruns. These complete-WAV findings advance M2 desktop coverage but do not
|
||||
establish playable streaming, RTF below one, broad quality or sustained use.
|
||||
|
||||
VLA qualification here ends at action data. Physical robot control, collision
|
||||
avoidance and actuator safety belong to the external controller.
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
"""Native Windows CPU Qwen3-TTS whole-session stage with isolated dependencies.
|
||||
"""Local CPU Qwen3-TTS whole-session stage with isolated dependencies.
|
||||
|
||||
The worker alone imports Qwen's PyTorch wrapper and Transformers 4.57.3.
|
||||
This adapter reuses the bounded complete-WAV request, fencing and ownership
|
||||
@@ -57,7 +57,12 @@ class QwenTTSCPUStageClient(CrispTTSStageClient):
|
||||
demand = reservation.demands[self._memory_pool]
|
||||
if self._max_wav_bytes > demand:
|
||||
raise ResourceUnavailable("CPU TTS output bound exceeds stage reservation")
|
||||
self._python = Path(config["python_bin"]).resolve(strict=True)
|
||||
# On POSIX, a virtualenv's python is often a symlink to the base
|
||||
# interpreter. Launch via the declared path so Python discovers
|
||||
# the virtualenv; _sha256 still follows the link for verification.
|
||||
self._python = Path(config["python_bin"]).absolute()
|
||||
if not self._python.is_file():
|
||||
raise FileNotFoundError(f"CPU TTS interpreter absent: {self._python}")
|
||||
self._model_dir = Path(config["model_dir"]).resolve(strict=True)
|
||||
self._overlay = Path(config["overlay_dir"]).resolve(strict=True)
|
||||
self._talker = self._model_dir / "model.safetensors"
|
||||
@@ -70,6 +75,11 @@ class QwenTTSCPUStageClient(CrispTTSStageClient):
|
||||
}
|
||||
if not (self._overlay / "transformers" / "__init__.py").is_file():
|
||||
raise FileNotFoundError("pinned Transformers overlay is absent")
|
||||
overlay_marker_hash = config.get("overlay_marker_sha256")
|
||||
if overlay_marker_hash is not None:
|
||||
marker = self._overlay / "kernels" / "__init__.py"
|
||||
if not marker.is_file() or _sha256(marker) != str(overlay_marker_hash).lower():
|
||||
raise ValueError("CPU TTS optional-kernels overlay differs from declared hash")
|
||||
for label, (path, digest) in expected.items():
|
||||
if _sha256(path) != digest:
|
||||
raise ValueError(f"CPU TTS {label} differs from the declared artifact hash")
|
||||
@@ -152,6 +162,7 @@ class QwenTTSCPUStageClient(CrispTTSStageClient):
|
||||
"worker_pid": self._proc.pid,
|
||||
"worker_script_sha256": _sha256(self._worker),
|
||||
"artifact_sha256": {label: digest for label, (_, digest) in expected.items()},
|
||||
"overlay_marker_sha256": overlay_marker_hash,
|
||||
"model_dir": str(self._model_dir),
|
||||
"placement": "CPU BF16 SDPA, 8 threads; isolated Transformers 4.57.3",
|
||||
"worker_props": props,
|
||||
|
||||
Reference in new issue
Block a user