diff --git a/benchmarks/edge_harness/profile_minicpmo_text_speech.py b/benchmarks/edge_harness/profile_minicpmo_text_speech.py index 3a2a54e35..18bed9ced 100644 --- a/benchmarks/edge_harness/profile_minicpmo_text_speech.py +++ b/benchmarks/edge_harness/profile_minicpmo_text_speech.py @@ -25,6 +25,7 @@ from vllm_omni.entrypoints.omni import Omni class MemorySampler: def __init__(self, interval_s: float = 0.5): self.interval_s = interval_s + self.scope = "wsl_guest" if "microsoft" in platform.release().lower() else "native_os" self.stop = threading.Event() self.samples: list[dict[str, int | float]] = [] self.thread = threading.Thread(target=self._run, name="memory-sampler", daemon=True) @@ -43,8 +44,8 @@ class MemorySampler: self.samples.append( { "monotonic_s": time.perf_counter(), - "wsl_used_bytes": memory.used, - "wsl_available_bytes": memory.available, + "system_used_bytes": memory.used, + "system_available_bytes": memory.available, "swap_used_bytes": swap.used, "process_tree_rss_sum_bytes": rss_sum, } @@ -101,6 +102,18 @@ def _run_request(omni: Omni, prompt: dict, memory_sampler: MemorySampler) -> dic if not text or not token_ids or not audio_samples or audio_rms == 0: raise RuntimeError("MiniCPM-o did not produce both nonempty text and nonzero audio") samples = [sample for sample in memory_sampler.samples if started <= sample["monotonic_s"] <= completed] + sampled_memory = { + "scope": memory_sampler.scope, + "n": len(samples), + "max_system_used_bytes": max((s["system_used_bytes"] for s in samples), default=None), + "min_system_available_bytes": min((s["system_available_bytes"] for s in samples), default=None), + "max_swap_used_bytes": max((s["swap_used_bytes"] for s in samples), default=None), + # Sum of RSS can double-count shared pages; it is not unique RAM. + "max_process_tree_rss_sum_bytes": max((s["process_tree_rss_sum_bytes"] for s in samples), default=None), + } + if memory_sampler.scope == "wsl_guest": + sampled_memory["max_wsl_used_bytes"] = sampled_memory["max_system_used_bytes"] + sampled_memory["min_wsl_available_bytes"] = sampled_memory["min_system_available_bytes"] return { "request_ids": sorted(request_ids), "wall_s": completed - started, @@ -111,14 +124,7 @@ def _run_request(omni: Omni, prompt: dict, memory_sampler: MemorySampler) -> dic "audio_samples": audio_samples, "audio_duration_s_at_24khz": audio_samples / 24000, "audio_rms_float": audio_rms, - "sampled_memory": { - "n": len(samples), - "max_wsl_used_bytes": max((s["wsl_used_bytes"] for s in samples), default=None), - "min_wsl_available_bytes": min((s["wsl_available_bytes"] for s in samples), default=None), - "max_swap_used_bytes": max((s["swap_used_bytes"] for s in samples), default=None), - # Sum of RSS can double-count shared pages; it is not unique RAM. - "max_process_tree_rss_sum_bytes": max((s["process_tree_rss_sum_bytes"] for s in samples), default=None), - }, + "sampled_memory": sampled_memory, } @@ -129,6 +135,8 @@ def main() -> None: parser.add_argument("--output", type=Path, required=True) parser.add_argument("--warmups", type=int, default=1) parser.add_argument("--repeats", type=int, default=20) + parser.add_argument("--init-timeout", type=int, default=600) + parser.add_argument("--stage-init-timeout", type=int, default=300) args = parser.parse_args() if args.warmups < 0 or args.repeats < 1: parser.error("warmups must be nonnegative and repeats must be positive") @@ -141,6 +149,8 @@ def main() -> None: model=str(args.model.resolve()), deploy_config=str(args.deploy_config.resolve()), trust_remote_code=True, + init_timeout=args.init_timeout, + stage_init_timeout=args.stage_init_timeout, ) startup_s = time.perf_counter() - started try: @@ -168,10 +178,21 @@ def main() -> None: "torch": torch.__version__, "vllm": vllm.__version__, "torch_cuda_available": torch.cuda.is_available(), + "memory_scope": sampler.scope, "environment": { name: os.environ.get(name) - for name in ("VLLM_TARGET_DEVICE", "VLLM_CPU_KVCACHE_SPACE", "OMP_NUM_THREADS", "CUDA_VISIBLE_DEVICES") + for name in ( + "VLLM_TARGET_DEVICE", + "VLLM_CPU_KVCACHE_SPACE", + "OMP_NUM_THREADS", + "MKL_NUM_THREADS", + "CUDA_VISIBLE_DEVICES", + "VLLM_USE_FLASHINFER_SAMPLER", + "VLLM_ENABLE_V1_MULTIPROCESSING", + ) }, + "init_timeout_s": args.init_timeout, + "stage_init_timeout_s": args.stage_init_timeout, "startup_s": startup_s, "warmups": args.warmups, "repeats": args.repeats, diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/README.md b/benchmarks/edge_harness/results/e2e_expansion_20260923/README.md index bb56c8594..72fc5379d 100644 --- a/benchmarks/edge_harness/results/e2e_expansion_20260923/README.md +++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/README.md @@ -1,10 +1,11 @@ # 2026-09-23 WSL CPU and native Windows expansion -This extends the [60-cell device × model matrix](../e2e_profiling_20260922/evidence/summary/README.md) with previously unrun WSL CPU and native Windows paths. The September 22 timing tables remain historical. MiniCPM-o now has a separate 20-request serial text-to-speech timing run; each InternVLA timing remains one observation. None is a sustained or full-modality profile. The [checkpoint manifests](../e2e_recovery_20260923/README.md) pin the real weights. No model, precision or context was silently substituted. +This extends the [60-cell device × model matrix](../e2e_profiling_20260922/evidence/summary/README.md) with previously unrun WSL CPU and native Windows paths. The September 22 timing tables remain historical. MiniCPM-o now has separate 20-request serial text-to-speech timing runs on WSL CPU and native Windows GPU; each InternVLA timing remains one observation. None is a sustained or full-modality profile. The [checkpoint manifests](../e2e_recovery_20260923/README.md) pin the real weights. No model, precision or context was silently substituted. | Device / model | Verified depth | Observed output and time | Remaining gate | |---|---|---|---| | WSL CPU / MiniCPM-o 4.5, revision `503e754207c94da6bb26850b4469f367c9ea3582` | Scoped complete three-stage text-to-text+speech path (P for this input only) | Coherent text and a 5.2 s mono 24 kHz WAV. First request 27.09 s after 73.65 s startup; separate 1-warmup/20-measured serial run: startup 73.89 s, request wall p50 21.86 s/p95 22.97 s. | Speech intelligibility/alignment, image/audio/video inputs, concurrency, interruption, power and thermal behavior. | +| Native Windows RTX 5090 Laptop / MiniCPM-o 4.5, same revision | Scoped complete three-stage text-to-text+speech path (P for this input only) | Coherent text with the same 32 thinker token IDs as WSL CUDA, and a nonzero 5.2 s mono 24 kHz WAV. First engine startup 297.20 s and request 78.75 s, including first-use graph/JIT work; separate 1-warmup/20-measured serial run: startup 239.05 s, request wall p50 18.38 s/p95 18.73 s. | Speech intelligibility/alignment, image/audio/video inputs, concurrency, interruption, loading peak and sustained behavior. | | WSL CPU / InternVLA-A1-3B Place_Markpen, revision `0a557c6a503e6545b369bff2f34c76a410bbd190` | Strict real-weight load and full policy forward on **synthetic zero** observations/noise | Finite float32 actions `[1, 50, 32]`; CPU-only PyTorch startup 31.78 s and synchronized forward 3.25 s. Compared with a same-checkpoint CUDA run, relative L2 difference 1.3025% and cosine similarity 0.999917. | Real Place_Markpen observations and reference actions, task tolerance, units/timestamps, repeated latency, memory and deadlines. | | Native Windows CPU / InternVLA-A1-3B Place_Markpen, same revision | Strict real-weight load and full policy forward on **synthetic zero** observations/noise; policy and actions checked on CPU | Finite float32 actions `[1, 50, 32]`; placement-checked startup 61.66 s and forward 3.36 s. Relative L2 difference from WSL CPU 1.0824%; cosine 0.999942. | Real observations/reference actions and quality tolerance; repeated latency, memory, units/timestamps and deadlines. This is the direct policy CPU path, not native vLLM CPU attention. | | Native Windows RTX 5090 Laptop / InternVLA-A1-3B Place_Markpen, same revision | Strict real-weight load and full policy forward on **synthetic zero** observations/noise; policy and actions checked on CUDA | Finite float32 actions `[1, 50, 32]`; placement-checked startup 75.54 s and synchronized forward 1.12 s. Relative L2 difference from WSL CUDA 0.8315%; cosine 0.999966. | Real observations/reference actions and quality tolerance; repeated latency, memory, units/timestamps and deadlines. | @@ -13,6 +14,10 @@ Both runs used Ryzen AI 9 HX 370 on Ubuntu 26.04/WSL2 kernel `6.18.33.2-microsof The separate [20-request profile](evidence/minicpmo/minicpmo_cpu_profile_20.json) records every measured output and sampled memory; the [JSONL](evidence/minicpmo/minicpmo_cpu_profile_20.jsonl) survives interruption, and the [run log](evidence/minicpmo/minicpmo_cpu_profile_20.log) retains worker details. One warmup was excluded. All 20 measured requests returned the same thinker token sequence, nonzero 124,800-sample audio, and request wall times 20.95–23.33 s; nearest-rank p50/p95 were 21.86/22.97 s (real-time factor 4.20/4.42 for 5.2 s of audio). Sampled WSL available RAM reached a minimum of 3.80 GB and swap use a maximum of 5.75 GB; summed process-tree RSS reached 29.04 GB but double-counts shared pages. These are 0.5 s samples during requests, not load-peak or unique-resident-memory measurements. Audio and text were returned together at request completion, so the profile cannot establish playable-stream latency, and no concurrency or 30-minute thermal run was performed. +Native Windows MiniCPM-o used the same [constrained CUDA deployment config](../e2e_recovery_20260923/evidence/minicpmo/minicpmo_24gb_offline.yaml) as the WSL CUDA run: BF16, eager vLLM 0.29.0, 2,048-token thinker / 512-token talker limits, text input only, one request and 5 GiB thinker host offload. The local Windows 11 build 26200 host has 63.12 GiB visible RAM, an RTX 5090 Laptop GPU (driver 610.71), Python 3.12.10, PyTorch 2.13.0+cu130, Transformers 5.14.1, `stepaudio2-minicpmo` 0.1.1 and `pyarrow` 25.0.1. The checkpoint was read through a local WSL UNC path. The example's then-default 300 s overall startup timeout [expired](evidence/minicpmo_windows/minicpmo_native_gpu.log) after stage initialization had nearly completed; rerunning with `--init-timeout 900 --stage-init-timeout 600` [completed](evidence/minicpmo_windows/minicpmo_native_gpu_timeout900.log) in 297.20 s startup and one 78.75 s request. The example default has since been raised to 600 s, matching Omni's default. The text [and token IDs](evidence/minicpmo_windows/0_b4835d8d-f4f8-48b9-85d2-bb7ef06d9249.txt) exactly match the WSL CUDA result. The [WAV](evidence/minicpmo_windows/output_0_b4835d8d-f4f8-48b9-85d2-bb7ef06d9249.wav) has 124,800 mono PCM16 frames at 24 kHz; 124,560 samples are nonzero. The [artifact report](evidence/minicpmo_windows/minicpmo_windows_artifact_report.json) records its checksum and signal statistics. One first-use request is not a latency distribution or speech-quality assessment; the distinct Windows timing must not be transferred to WSL or CPU. + +The separate native Windows [20-request profile](evidence/minicpmo_windows/minicpmo_windows_gpu_profile20.json), [per-request JSONL](evidence/minicpmo_windows/minicpmo_windows_gpu_profile20.jsonl), and [raw log](evidence/minicpmo_windows/minicpmo_windows_gpu_profile20.log) used one warmup followed by 20 measured serial requests in one engine. All measured requests returned one consistent thinker token sequence and 124,800-sample audio with nonzero RMS. Nearest-rank request wall p50/p95 were 18.38/18.73 s (real-time factor 3.53/3.60 for 5.2 s of audio), with observed range 18.22–19.79 s. Engine startup was 239.05 s on this later, warm-filesystem run. During measured requests, 0.5 s samples found native host available RAM as low as 0.78 GB and pagefile use as high as 11.17 GB; summed process-tree RSS reached 21.08 GB but double-counts shared pages. These are host-wide samples, not model-only RAM or loading peak. Neither a playable-stream timing nor speech intelligibility, concurrency, interruption, power or a 30-minute thermal run was measured. + The CPU-only InternVLA run used the same BF16 policy checkpoint, synthetic six-camera history/state input and zero diffusion noise as the CUDA reference. The raw [CPU JSON](evidence/internvla/internvla_pure_cpu_synthetic.json), [CPU log](evidence/internvla/internvla_pure_cpu_synthetic.log), [CPU actions](evidence/internvla/internvla_pure_cpu_actions.npy), [CUDA JSON](evidence/internvla/internvla_cuda_synthetic_with_actions.json), [CUDA actions](evidence/internvla/internvla_cuda_actions.npy) and [numerical comparison](evidence/internvla/internvla_pure_cpu_cuda_parity.json) are retained. The CUDA reference used the WSL RTX 5090 Laptop and PyTorch 2.13.0+cu130; its one forward took 1.33 s. The action difference has maximum absolute error 0.00993 and mean absolute error 0.00117. No reference-task tolerance is available, so this comparison is evidence of numerical proximity, not an accuracy pass. The native Windows runs used Windows 11 build 26200, NVIDIA driver 610.71 for the GPU path, Python 3.12.10, PyTorch 2.13.0+cu130, vLLM 0.29.0, Transformers 5.14.1, and the current Omni source. The CPU run explicitly selected the CPU policy despite the CUDA-capable wheel. The placement-checked reruns assert that all policy parameters and the output action tensor remain on the requested device. Both loaded the same checkpoint from the local WSL filesystem through a UNC path; Windows CPU and GPU runs used the same synthetic input and zero diffusion noise as the WSL reference. The native [CPU JSON](evidence/internvla/internvla_windows_cpu_synthetic.json), [GPU JSON](evidence/internvla/internvla_windows_gpu_synthetic.json), [CPU log](evidence/internvla/internvla_windows_cpu_synthetic.log), [GPU log](evidence/internvla/internvla_windows_gpu_synthetic.log), [CPU actions](evidence/internvla/internvla_windows_cpu_actions.npy), [GPU actions](evidence/internvla/internvla_windows_gpu_actions.npy), [CPU/WSL comparison](evidence/internvla/internvla_windows_cpu_wsl_cpu_parity.json), and [GPU/WSL comparison](evidence/internvla/internvla_windows_gpu_wsl_cuda_parity.json) are retained. Native vLLM import initially failed on a Linux-specific CuMem allocator patch; the added Windows guard resolved the import. Native default GBK source decoding also required `PYTHONUTF8=1`. Both platform differences are setup requirements, not model-quality evidence. Power mode and thermal state were not captured. @@ -42,3 +47,21 @@ export CUDA_VISIBLE_DEVICES= VLLM_TARGET_DEVICE=cpu VLLM_CPU_KVCACHE_SPACE=1 VLL ``` For native Windows InternVLA, set `PYTHONUTF8=1`, put this checkout on `PYTHONPATH`, and run `benchmarks/edge_harness/probe_internvla_actions.py` with the same model/processor paths as above but Windows-accessible UNC paths; use `--device cpu` or `--device cuda` and separate output paths. + +The native Windows MiniCPM-o profile used this PowerShell command (run the example `end2end.py` instead for the single WAV-producing probe): + +```powershell +$root = '\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer' +$env:PYTHONUTF8 = '1' +$env:PYTHONPATH = "$root\vllm-omni-edge" +$env:OMP_NUM_THREADS = '8' +$env:MKL_NUM_THREADS = '8' +$env:TOKENIZERS_PARALLELISM = 'false' +$env:VLLM_USE_FLASHINFER_SAMPLER = '0' +$env:VLLM_ENABLE_V1_MULTIPROCESSING = '0' +& 'C:\Users\zhout\w2\omni029venv\Scripts\python.exe' "$root\vllm-omni-edge\benchmarks\edge_harness\profile_minicpmo_text_speech.py" ` + --model "$root\models\MiniCPM-o-4_5" ` + --deploy-config "$root\vllm-omni-edge\benchmarks\edge_harness\results\e2e_recovery_20260923\evidence\minicpmo\minicpmo_24gb_offline.yaml" ` + --warmups 1 --repeats 20 --init-timeout 900 --stage-init-timeout 600 ` + --output 'C:\Users\zhout\w2\minicpmo_windows_gpu_profile20.json' +``` diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/minicpmo_windows/0_b4835d8d-f4f8-48b9-85d2-bb7ef06d9249.txt b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/minicpmo_windows/0_b4835d8d-f4f8-48b9-85d2-bb7ef06d9249.txt new file mode 100644 index 000000000..1e65f7dd2 --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/minicpmo_windows/0_b4835d8d-f4f8-48b9-85d2-bb7ef06d9249.txt @@ -0,0 +1,15 @@ +Prompt: +<|im_start|>system +You are MiniCPM-o, a helpful multimodal assistant that can understand images, audio and video, and respond in text and speech.<|im_end|> +<|im_start|>user +Say hello, then introduce vLLM-Omni in one sentence.<|im_end|> +<|im_start|>assistant + + + + +<|tts_bos|> +vllm_text_output: +Hello, I am vLLM-Omni, a large multimodal model capable of understanding and generating text, images, audio, and video. +vllm_token_ids: +[9707, 11, 358, 1079, 348, 4086, 44, 23620, 76, 7751, 11, 264, 3460, 79049, 57597, 1614, 12875, 315, 8660, 323, 23163, 1467, 11, 5335, 11, 7699, 11, 323, 2766, 13, 151704, 151645] diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/minicpmo_windows/minicpmo_native_gpu.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/minicpmo_windows/minicpmo_native_gpu.log new file mode 100644 index 000000000..23b8bb25e --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/minicpmo_windows/minicpmo_native_gpu.log @@ -0,0 +1,261 @@ +INFO 09-23 16:24:18 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed. +INFO 09-23 16:24:18 [patch.py:510] inductor factorable-divisibility patch: installed. +==================== + vllm version: 0.29.0 + ==================== +INFO 09-23 16:24:21 [omni_base.py:216] [Omni] Initializing with model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5 +INFO 09-23 16:24:21 [async_omni_engine.py:172] [AsyncOmniEngine] Initializing with model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5 +INFO 09-23 16:24:22 [async_omni_engine.py:290] [AsyncOmniEngine] Launching Orchestrator thread with 3 stages +INFO 09-23 16:24:22 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn +INFO 09-23 16:24:23 [initialization.py:356] Loaded OmniTransferConfig with 1 connector configurations +WARNING 09-23 16:24:23 [utils.py:501] Dropping unknown OmniEngineArgs field 'inline_diffusion' (not declared on the dataclass) +WARNING 09-23 16:24:23 [utils.py:501] Dropping unknown OmniEngineArgs field 'shm_threshold_bytes' (not declared on the dataclass) +WARNING 09-23 16:24:23 [arg_utils.py:1801] The global random seed is set to 0. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. +INFO 09-23 16:24:38 [model.py:684] Resolved architecture: MiniCPMO45OmniForConditionalGeneration +INFO 09-23 16:24:38 [model.py:2021] Using max model len 2048 +WARNING 09-23 16:24:38 [model.py:981] Model does not support mm_device_do_normalize, forcing mm_device_do_normalize = False. +INFO 09-23 16:24:39 [registry.py:141] All limits of multimodal modalities supported by the model are set to 0, running in text-only mode. +INFO 09-23 16:24:39 [model.py:887] Disabled mm_prefix attention mode because multimodal inputs are configuration-disabled. Attention backends without mm_prefix support may now be selected. +INFO 09-23 16:24:39 [scheduler.py:277] Chunked prefill is enabled with max_num_batched_tokens=2048. +WARNING 09-23 16:24:39 [vllm.py:1371] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none +WARNING 09-23 16:24:39 [vllm.py:1406] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. +INFO 09-23 16:24:39 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +INFO 09-23 16:24:39 [vllm.py:1585] Cudagraph is disabled under eager mode +INFO 09-23 16:24:39 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant +WARNING 09-23 16:24:39 [utils.py:501] Dropping unknown OmniEngineArgs field 'inline_diffusion' (not declared on the dataclass) +WARNING 09-23 16:24:39 [utils.py:501] Dropping unknown OmniEngineArgs field 'shm_threshold_bytes' (not declared on the dataclass) +WARNING 09-23 16:24:40 [arg_utils.py:1801] The global random seed is set to 0. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. +INFO 09-23 16:24:40 [model.py:684] Resolved architecture: MiniCPMO45OmniForConditionalGeneration +INFO 09-23 16:24:40 [model.py:2021] Using max model len 512 +WARNING 09-23 16:24:40 [model.py:981] Model does not support mm_device_do_normalize, forcing mm_device_do_normalize = False. +INFO 09-23 16:24:40 [model.py:2021] Using max model len 512 +INFO 09-23 16:24:41 [scheduler.py:277] Chunked prefill is enabled with max_num_batched_tokens=512. +INFO 09-23 16:24:41 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +WARNING 09-23 16:24:41 [utils.py:501] Dropping unknown OmniEngineArgs field 'inline_diffusion' (not declared on the dataclass) +WARNING 09-23 16:24:41 [utils.py:501] Dropping unknown OmniEngineArgs field 'shm_threshold_bytes' (not declared on the dataclass) +WARNING 09-23 16:24:41 [arg_utils.py:1801] The global random seed is set to 0. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. +INFO 09-23 16:24:56 [model.py:684] Resolved architecture: MiniCPMO45Code2Wav +INFO 09-23 16:24:56 [model.py:2021] Using max model len 4096 +WARNING 09-23 16:24:56 [arg_utils.py:2780] This model does not officially support disabling chunked prefill. Disabling this manually may cause the engine to crash or produce incorrect outputs. +INFO 09-23 16:24:56 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +INFO 09-23 16:24:56 [stage_runtime.py:853] [stage_init] Stage-0 set runtime devices: 0 +INFO 09-23 16:25:08 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed. +INFO 09-23 16:25:08 [patch.py:510] inductor factorable-divisibility patch: installed. +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:12 [core.py:123] Initializing a V1 LLM engine (v0.29.0) with config: model='\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\MiniCPM-o-4_5', speculative_config=None, tokenizer='\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\MiniCPM-o-4_5', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, per_request_spec_decode_metrics='none', kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5, enable_prefix_caching=False, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'enable_qk_norm_rope_fusion': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False, 'fuse_qk_norm_rope_kvcache': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, enable_jit_warmup=True, enable_bf16x3_router_gemm=False, moe_backend='auto', linear_backend='auto') +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:14 [registry.py:141] All limits of multimodal modalities supported by the model are set to 0, running in text-only mode. +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:14 [parallel_state.py:1775] world_size=1 rank=0 local_rank=0 distributed_init_method=file://C:\Users\zhout\AppData\Local\Temp/vllm_dist_2d36997f7ac0457d963dd5c17559f334 backend=nccl +(StageEngineCoreProc_stage0_replica0 pid=8840) WARNING 09-23 16:25:14 [parallel_state.py:1788] Distributed backend nccl is not available; falling back to gloo. +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:14 [parallel_state.py:2119] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A +(StageEngineCoreProc_stage0_replica0 pid=8840) WARNING 09-23 16:25:15 [gpu_ar_worker.py:120] OMNI GPUARWorker forces v1 model runner for omni hooks. +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:15 [topk_topp_sampler.py:46] FlashInfer top-p/top-k sampling disabled via VLLM_USE_FLASHINFER_SAMPLER=0. +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:15 [base.py:138] Offloader set to UVAOffloader +(StageEngineCoreProc_stage0_replica0 pid=8840) WARNING 09-23 16:25:15 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage0_replica0 pid=8840) WARNING 09-23 16:25:15 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:15 [gpu_model_runner.py:5379] Starting to load model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5... +(StageEngineCoreProc_stage0_replica0 pid=8840) WARNING 09-23 16:25:15 [vllm.py:1371] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none +(StageEngineCoreProc_stage0_replica0 pid=8840) WARNING 09-23 16:25:15 [vllm.py:1406] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:15 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:15 [vllm.py:1585] Cudagraph is disabled under eager mode +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:15 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:15 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:15 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:15 [cuda.py:492] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'TRITON_ATTN', 'FLEX_ATTENTION']. +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:15 [flash_attn.py:897] Using FlashAttention version 2 +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:17 [uva.py:65] Total CPU offloaded parameters: 5.03 +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:17 [weight_utils.py:863] Filesystem type for checkpoints: unknown. Checkpoint size: 17.46 GiB. Available RAM: 21.93 GiB. +(StageEngineCoreProc_stage0_replica0 pid=8840) INFO 09-23 16:25:17 [weight_utils.py:886] Auto-prefetch is disabled because the filesystem (unknown) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch. +(StageEngineCoreProc_stage0_replica0 pid=8840) +Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [512], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'enable_qk_norm_rope_fusion': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False, 'fuse_qk_norm_rope_kvcache': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, enable_jit_warmup=True, enable_bf16x3_router_gemm=False, moe_backend='auto', linear_backend='auto') +(StageEngineCoreProc_stage1_replica0 pid=57896) INFO 09-23 16:28:18 [parallel_state.py:1775] world_size=1 rank=0 local_rank=0 distributed_init_method=file://C:\Users\zhout\AppData\Local\Temp/vllm_dist_c46fc169d2b24ce5bf1909728c14d88d backend=nccl +(StageEngineCoreProc_stage1_replica0 pid=57896) WARNING 09-23 16:28:18 [parallel_state.py:1788] Distributed backend nccl is not available; falling back to gloo. +(StageEngineCoreProc_stage1_replica0 pid=57896) INFO 09-23 16:28:18 [parallel_state.py:2119] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A +(StageEngineCoreProc_stage1_replica0 pid=57896) WARNING 09-23 16:28:19 [gpu_ar_worker.py:120] OMNI GPUARWorker forces v1 model runner for omni hooks. +(StageEngineCoreProc_stage1_replica0 pid=57896) INFO 09-23 16:28:19 [topk_topp_sampler.py:46] FlashInfer top-p/top-k sampling disabled via VLLM_USE_FLASHINFER_SAMPLER=0. +(StageEngineCoreProc_stage1_replica0 pid=57896) [transformers] `MiniCPMOProcessor` defines `image_processor_class = 'AutoImageProcessor'`, which is deprecated. Register the correct mapping in `AutoImageProcessor` instead. +(StageEngineCoreProc_stage1_replica0 pid=57896) [transformers] Requested torchvision backend is not available. Falling back to pil backend. +(StageEngineCoreProc_stage1_replica0 pid=57896) [transformers] `MiniCPMOProcessor` defines `feature_extractor_class = 'WhisperFeatureExtractor'`, which is deprecated. Register the correct mapping in `AutoFeatureExtractor` instead. +(StageEngineCoreProc_stage1_replica0 pid=57896) [transformers] Image processor MiniCPMVImageProcessor: kwargs ['max_slice_nums'] were applied for backward compatibility. To avoid this warning, add them to valid_kwargs: create a custom TypedDict extending ImagesKwargs with these keys and set it as the `valid_kwargs` class attribute. +(StageEngineCoreProc_stage1_replica0 pid=57896) INFO 09-23 16:28:36 [factory.py:46] Created connector: SharedMemoryConnector +(StageEngineCoreProc_stage1_replica0 pid=57896) WARNING 09-23 16:28:36 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage1_replica0 pid=57896) WARNING 09-23 16:28:36 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage1_replica0 pid=57896) INFO 09-23 16:28:36 [gpu_model_runner.py:5379] Starting to load model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5... +(StageEngineCoreProc_stage1_replica0 pid=57896) WARNING 09-23 16:28:36 [vllm.py:1371] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none +(StageEngineCoreProc_stage1_replica0 pid=57896) WARNING 09-23 16:28:36 [vllm.py:1406] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. +(StageEngineCoreProc_stage1_replica0 pid=57896) INFO 09-23 16:28:36 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage1_replica0 pid=57896) INFO 09-23 16:28:36 [vllm.py:1585] Cudagraph is disabled under eager mode +(StageEngineCoreProc_stage1_replica0 pid=57896) INFO 09-23 16:28:36 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant +(StageEngineCoreProc_stage1_replica0 pid=57896) INFO 09-23 16:28:36 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage1_replica0 pid=57896) INFO 09-23 16:28:37 [cuda.py:492] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'TRITON_ATTN', 'FLEX_ATTENTION']. +(StageEngineCoreProc_stage1_replica0 pid=57896) INFO 09-23 16:28:37 [flash_attn.py:897] Using FlashAttention version 2 +(StageEngineCoreProc_stage1_replica0 pid=57896) INFO 09-23 16:28:37 [weight_utils.py:863] Filesystem type for checkpoints: unknown. Checkpoint size: 17.46 GiB. Available RAM: 15.41 GiB. +(StageEngineCoreProc_stage1_replica0 pid=57896) INFO 09-23 16:28:37 [weight_utils.py:893] Auto-prefetch is disabled because the filesystem (unknown) is not a recognized network FS (NFS/Lustre) and the checkpoint size (17.46 GiB) exceeds 90% of available RAM (15.41 GiB). +(StageEngineCoreProc_stage1_replica0 pid=57896) +Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [4096], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'enable_qk_norm_rope_fusion': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False, 'fuse_qk_norm_rope_kvcache': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, enable_jit_warmup=True, enable_bf16x3_router_gemm=False, moe_backend='auto', linear_backend='auto') +(StageEngineCoreProc_stage2_replica0 pid=6400) INFO 09-23 16:29:09 [parallel_state.py:1775] world_size=1 rank=0 local_rank=0 distributed_init_method=file://C:\Users\zhout\AppData\Local\Temp/vllm_dist_e5dbe2b9f92549ba9f9566b6a436c49d backend=nccl +(StageEngineCoreProc_stage2_replica0 pid=6400) WARNING 09-23 16:29:09 [parallel_state.py:1788] Distributed backend nccl is not available; falling back to gloo. +(StageEngineCoreProc_stage2_replica0 pid=6400) INFO 09-23 16:29:09 [parallel_state.py:2119] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A +(StageEngineCoreProc_stage2_replica0 pid=6400) WARNING 09-23 16:29:09 [gpu_generation_worker.py:103] OMNI GPUGenerationWorker forces v1 model runner for omni hooks. +(StageEngineCoreProc_stage2_replica0 pid=6400) INFO 09-23 16:29:09 [topk_topp_sampler.py:46] FlashInfer top-p/top-k sampling disabled via VLLM_USE_FLASHINFER_SAMPLER=0. +(StageEngineCoreProc_stage2_replica0 pid=6400) INFO 09-23 16:29:09 [factory.py:46] Created connector: SharedMemoryConnector +(StageEngineCoreProc_stage2_replica0 pid=6400) WARNING 09-23 16:29:09 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage2_replica0 pid=6400) WARNING 09-23 16:29:09 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage2_replica0 pid=6400) INFO 09-23 16:29:09 [gpu_model_runner.py:5379] Starting to load model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5... +(StageEngineCoreProc_stage2_replica0 pid=6400) INFO 09-23 16:29:10 [weight_utils.py:863] Filesystem type for checkpoints: unknown. Checkpoint size: 17.46 GiB. Available RAM: 13.21 GiB. +(StageEngineCoreProc_stage2_replica0 pid=6400) INFO 09-23 16:29:10 [weight_utils.py:893] Auto-prefetch is disabled because the filesystem (unknown) is not a recognized network FS (NFS/Lustre) and the checkpoint size (17.46 GiB) exceeds 90% of available RAM (13.21 GiB). +(StageEngineCoreProc_stage2_replica0 pid=6400) +Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00 + main(args) + File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\examples\offline_inference\minicpmo\end2end.py", line 283, in main + omni = Omni(**omni_kwargs) + ^^^^^^^^^^^^^^^^^^^ + File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\entrypoints\omni_base.py", line 223, in __init__ + self.engine = AsyncOmniEngine( + ^^^^^^^^^^^^^^^^ + File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\async_omni_engine.py", line 305, in __init__ + self._wait_for_orchestrator_init(startup_future, startup_timeout) + File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\async_omni_engine.py", line 537, in _wait_for_orchestrator_init + raise TimeoutError(f"Orchestrator did not become ready within {startup_timeout}s") +TimeoutError: Orchestrator did not become ready within 300s diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/minicpmo_windows/minicpmo_native_gpu_timeout900.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/minicpmo_windows/minicpmo_native_gpu_timeout900.log new file mode 100644 index 000000000..8d46caefb --- /dev/null +++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/minicpmo_windows/minicpmo_native_gpu_timeout900.log @@ -0,0 +1,267 @@ +INFO 09-23 16:30:42 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed. +INFO 09-23 16:30:42 [patch.py:510] inductor factorable-divisibility patch: installed. +==================== + vllm version: 0.29.0 + ==================== +INFO 09-23 16:30:47 [omni_base.py:216] [Omni] Initializing with model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5 +INFO 09-23 16:30:47 [async_omni_engine.py:172] [AsyncOmniEngine] Initializing with model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5 +INFO 09-23 16:30:48 [async_omni_engine.py:290] [AsyncOmniEngine] Launching Orchestrator thread with 3 stages +INFO 09-23 16:30:48 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn +INFO 09-23 16:30:48 [initialization.py:356] Loaded OmniTransferConfig with 1 connector configurations +WARNING 09-23 16:30:48 [utils.py:501] Dropping unknown OmniEngineArgs field 'inline_diffusion' (not declared on the dataclass) +WARNING 09-23 16:30:48 [utils.py:501] Dropping unknown OmniEngineArgs field 'shm_threshold_bytes' (not declared on the dataclass) +WARNING 09-23 16:30:48 [arg_utils.py:1801] The global random seed is set to 0. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. +INFO 09-23 16:30:48 [model.py:684] Resolved architecture: MiniCPMO45OmniForConditionalGeneration +INFO 09-23 16:30:48 [model.py:2021] Using max model len 2048 +WARNING 09-23 16:30:48 [model.py:981] Model does not support mm_device_do_normalize, forcing mm_device_do_normalize = False. +INFO 09-23 16:30:51 [registry.py:141] All limits of multimodal modalities supported by the model are set to 0, running in text-only mode. +INFO 09-23 16:30:51 [model.py:887] Disabled mm_prefix attention mode because multimodal inputs are configuration-disabled. Attention backends without mm_prefix support may now be selected. +INFO 09-23 16:30:51 [scheduler.py:277] Chunked prefill is enabled with max_num_batched_tokens=2048. +WARNING 09-23 16:30:51 [vllm.py:1371] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none +WARNING 09-23 16:30:51 [vllm.py:1406] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. +INFO 09-23 16:30:51 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +INFO 09-23 16:30:51 [vllm.py:1585] Cudagraph is disabled under eager mode +INFO 09-23 16:30:51 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant +WARNING 09-23 16:30:51 [utils.py:501] Dropping unknown OmniEngineArgs field 'inline_diffusion' (not declared on the dataclass) +WARNING 09-23 16:30:51 [utils.py:501] Dropping unknown OmniEngineArgs field 'shm_threshold_bytes' (not declared on the dataclass) +WARNING 09-23 16:30:52 [arg_utils.py:1801] The global random seed is set to 0. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. +INFO 09-23 16:30:52 [model.py:684] Resolved architecture: MiniCPMO45OmniForConditionalGeneration +INFO 09-23 16:30:52 [model.py:2021] Using max model len 512 +WARNING 09-23 16:30:52 [model.py:981] Model does not support mm_device_do_normalize, forcing mm_device_do_normalize = False. +INFO 09-23 16:30:52 [model.py:2021] Using max model len 512 +INFO 09-23 16:30:52 [scheduler.py:277] Chunked prefill is enabled with max_num_batched_tokens=512. +INFO 09-23 16:30:52 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +WARNING 09-23 16:30:52 [utils.py:501] Dropping unknown OmniEngineArgs field 'inline_diffusion' (not declared on the dataclass) +WARNING 09-23 16:30:52 [utils.py:501] Dropping unknown OmniEngineArgs field 'shm_threshold_bytes' (not declared on the dataclass) +WARNING 09-23 16:30:52 [arg_utils.py:1801] The global random seed is set to 0. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. +INFO 09-23 16:30:52 [model.py:684] Resolved architecture: MiniCPMO45Code2Wav +INFO 09-23 16:30:52 [model.py:2021] Using max model len 4096 +WARNING 09-23 16:30:52 [arg_utils.py:2780] This model does not officially support disabling chunked prefill. Disabling this manually may cause the engine to crash or produce incorrect outputs. +INFO 09-23 16:30:53 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +INFO 09-23 16:30:53 [stage_runtime.py:853] [stage_init] Stage-0 set runtime devices: 0 +INFO 09-23 16:31:08 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed. +INFO 09-23 16:31:08 [patch.py:510] inductor factorable-divisibility patch: installed. +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:12 [core.py:123] Initializing a V1 LLM engine (v0.29.0) with config: model='\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\MiniCPM-o-4_5', speculative_config=None, tokenizer='\\\\wsl.localhost\\Ubuntu\\home\\zhout\\project\\edge_infer\\models\\MiniCPM-o-4_5', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, per_request_spec_decode_metrics='none', kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5, enable_prefix_caching=False, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'enable_qk_norm_rope_fusion': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False, 'fuse_qk_norm_rope_kvcache': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, enable_jit_warmup=True, enable_bf16x3_router_gemm=False, moe_backend='auto', linear_backend='auto') +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:15 [registry.py:141] All limits of multimodal modalities supported by the model are set to 0, running in text-only mode. +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:15 [parallel_state.py:1775] world_size=1 rank=0 local_rank=0 distributed_init_method=file://C:\Users\zhout\AppData\Local\Temp/vllm_dist_53967da026804fc7a1199a7d00af3066 backend=nccl +(StageEngineCoreProc_stage0_replica0 pid=62176) WARNING 09-23 16:31:15 [parallel_state.py:1788] Distributed backend nccl is not available; falling back to gloo. +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:15 [parallel_state.py:2119] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A +(StageEngineCoreProc_stage0_replica0 pid=62176) WARNING 09-23 16:31:16 [gpu_ar_worker.py:120] OMNI GPUARWorker forces v1 model runner for omni hooks. +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:16 [topk_topp_sampler.py:46] FlashInfer top-p/top-k sampling disabled via VLLM_USE_FLASHINFER_SAMPLER=0. +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:16 [base.py:138] Offloader set to UVAOffloader +(StageEngineCoreProc_stage0_replica0 pid=62176) WARNING 09-23 16:31:16 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage0_replica0 pid=62176) WARNING 09-23 16:31:16 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:16 [gpu_model_runner.py:5379] Starting to load model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5... +(StageEngineCoreProc_stage0_replica0 pid=62176) WARNING 09-23 16:31:16 [vllm.py:1371] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none +(StageEngineCoreProc_stage0_replica0 pid=62176) WARNING 09-23 16:31:16 [vllm.py:1406] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:16 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:16 [vllm.py:1585] Cudagraph is disabled under eager mode +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:16 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:16 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:16 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:17 [cuda.py:492] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'TRITON_ATTN', 'FLEX_ATTENTION']. +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:17 [flash_attn.py:897] Using FlashAttention version 2 +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:20 [uva.py:65] Total CPU offloaded parameters: 5.03 +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:20 [weight_utils.py:863] Filesystem type for checkpoints: unknown. Checkpoint size: 17.46 GiB. Available RAM: 27.79 GiB. +(StageEngineCoreProc_stage0_replica0 pid=62176) INFO 09-23 16:31:20 [weight_utils.py:886] Auto-prefetch is disabled because the filesystem (unknown) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch. +(StageEngineCoreProc_stage0_replica0 pid=62176) +Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [512], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'enable_qk_norm_rope_fusion': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False, 'fuse_qk_norm_rope_kvcache': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, enable_jit_warmup=True, enable_bf16x3_router_gemm=False, moe_backend='auto', linear_backend='auto') +(StageEngineCoreProc_stage1_replica0 pid=58484) INFO 09-23 16:34:26 [parallel_state.py:1775] world_size=1 rank=0 local_rank=0 distributed_init_method=file://C:\Users\zhout\AppData\Local\Temp/vllm_dist_a346e62d6aed4005983a9bf545a9460f backend=nccl +(StageEngineCoreProc_stage1_replica0 pid=58484) WARNING 09-23 16:34:26 [parallel_state.py:1788] Distributed backend nccl is not available; falling back to gloo. +(StageEngineCoreProc_stage1_replica0 pid=58484) INFO 09-23 16:34:26 [parallel_state.py:2119] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A +(StageEngineCoreProc_stage1_replica0 pid=58484) WARNING 09-23 16:34:27 [gpu_ar_worker.py:120] OMNI GPUARWorker forces v1 model runner for omni hooks. +(StageEngineCoreProc_stage1_replica0 pid=58484) INFO 09-23 16:34:27 [topk_topp_sampler.py:46] FlashInfer top-p/top-k sampling disabled via VLLM_USE_FLASHINFER_SAMPLER=0. +(StageEngineCoreProc_stage1_replica0 pid=58484) [transformers] `MiniCPMOProcessor` defines `image_processor_class = 'AutoImageProcessor'`, which is deprecated. Register the correct mapping in `AutoImageProcessor` instead. +(StageEngineCoreProc_stage1_replica0 pid=58484) [transformers] Requested torchvision backend is not available. Falling back to pil backend. +(StageEngineCoreProc_stage1_replica0 pid=58484) [transformers] `MiniCPMOProcessor` defines `feature_extractor_class = 'WhisperFeatureExtractor'`, which is deprecated. Register the correct mapping in `AutoFeatureExtractor` instead. +(StageEngineCoreProc_stage1_replica0 pid=58484) [transformers] Image processor MiniCPMVImageProcessor: kwargs ['max_slice_nums'] were applied for backward compatibility. To avoid this warning, add them to valid_kwargs: create a custom TypedDict extending ImagesKwargs with these keys and set it as the `valid_kwargs` class attribute. +(StageEngineCoreProc_stage1_replica0 pid=58484) INFO 09-23 16:34:43 [factory.py:46] Created connector: SharedMemoryConnector +(StageEngineCoreProc_stage1_replica0 pid=58484) WARNING 09-23 16:34:43 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage1_replica0 pid=58484) WARNING 09-23 16:34:43 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage1_replica0 pid=58484) INFO 09-23 16:34:43 [gpu_model_runner.py:5379] Starting to load model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5... +(StageEngineCoreProc_stage1_replica0 pid=58484) WARNING 09-23 16:34:43 [vllm.py:1371] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none +(StageEngineCoreProc_stage1_replica0 pid=58484) WARNING 09-23 16:34:43 [vllm.py:1406] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. +(StageEngineCoreProc_stage1_replica0 pid=58484) INFO 09-23 16:34:43 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage1_replica0 pid=58484) INFO 09-23 16:34:43 [vllm.py:1585] Cudagraph is disabled under eager mode +(StageEngineCoreProc_stage1_replica0 pid=58484) INFO 09-23 16:34:44 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant +(StageEngineCoreProc_stage1_replica0 pid=58484) INFO 09-23 16:34:44 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage1_replica0 pid=58484) INFO 09-23 16:34:44 [cuda.py:492] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'TRITON_ATTN', 'FLEX_ATTENTION']. +(StageEngineCoreProc_stage1_replica0 pid=58484) INFO 09-23 16:34:44 [flash_attn.py:897] Using FlashAttention version 2 +(StageEngineCoreProc_stage1_replica0 pid=58484) INFO 09-23 16:34:44 [weight_utils.py:863] Filesystem type for checkpoints: unknown. Checkpoint size: 17.46 GiB. Available RAM: 26.39 GiB. +(StageEngineCoreProc_stage1_replica0 pid=58484) INFO 09-23 16:34:44 [weight_utils.py:886] Auto-prefetch is disabled because the filesystem (unknown) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch. +(StageEngineCoreProc_stage1_replica0 pid=58484) +Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [4096], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'enable_qk_norm_rope_fusion': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False, 'fuse_qk_norm_rope_kvcache': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, enable_jit_warmup=True, enable_bf16x3_router_gemm=False, moe_backend='auto', linear_backend='auto') +(StageEngineCoreProc_stage2_replica0 pid=59372) INFO 09-23 16:35:17 [parallel_state.py:1775] world_size=1 rank=0 local_rank=0 distributed_init_method=file://C:\Users\zhout\AppData\Local\Temp/vllm_dist_136def56607542418cb1ca141c7c28fb backend=nccl +(StageEngineCoreProc_stage2_replica0 pid=59372) WARNING 09-23 16:35:17 [parallel_state.py:1788] Distributed backend nccl is not available; falling back to gloo. +(StageEngineCoreProc_stage2_replica0 pid=59372) INFO 09-23 16:35:17 [parallel_state.py:2119] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A +(StageEngineCoreProc_stage2_replica0 pid=59372) WARNING 09-23 16:35:17 [gpu_generation_worker.py:103] OMNI GPUGenerationWorker forces v1 model runner for omni hooks. +(StageEngineCoreProc_stage2_replica0 pid=59372) INFO 09-23 16:35:17 [topk_topp_sampler.py:46] FlashInfer top-p/top-k sampling disabled via VLLM_USE_FLASHINFER_SAMPLER=0. +(StageEngineCoreProc_stage2_replica0 pid=59372) INFO 09-23 16:35:17 [factory.py:46] Created connector: SharedMemoryConnector +(StageEngineCoreProc_stage2_replica0 pid=59372) WARNING 09-23 16:35:17 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage2_replica0 pid=59372) WARNING 09-23 16:35:17 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage2_replica0 pid=59372) INFO 09-23 16:35:17 [gpu_model_runner.py:5379] Starting to load model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5... +(StageEngineCoreProc_stage2_replica0 pid=59372) INFO 09-23 16:35:18 [weight_utils.py:863] Filesystem type for checkpoints: unknown. Checkpoint size: 17.46 GiB. Available RAM: 23.92 GiB. +(StageEngineCoreProc_stage2_replica0 pid=59372) INFO 09-23 16:35:18 [weight_utils.py:886] Auto-prefetch is disabled because the filesystem (unknown) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch. +(StageEngineCoreProc_stage2_replica0 pid=59372) +Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00._kernel. This causes a latency spike; consider extending warmup to cover this shape/config. +(StageEngineCoreProc_stage2_replica0 pid=59372) INFO 09-23 16:37:03 [cuda_graph_wrapper.py:98] Captured HiFT CUDA Graph for shape (1, 18, 3840) +Processed prompts: 0%| | 0/1 [00:00, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'enable_qk_norm_rope_fusion': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False, 'fuse_qk_norm_rope_kvcache': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, enable_jit_warmup=True, enable_bf16x3_router_gemm=False, moe_backend='auto', linear_backend='auto') +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:43 [registry.py:141] All limits of multimodal modalities supported by the model are set to 0, running in text-only mode. +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:43 [parallel_state.py:1775] world_size=1 rank=0 local_rank=0 distributed_init_method=file://C:\Users\zhout\AppData\Local\Temp/vllm_dist_f3f857b1fafb42afad38d136cada03d8 backend=nccl +(StageEngineCoreProc_stage0_replica0 pid=51364) WARNING 09-23 16:39:43 [parallel_state.py:1788] Distributed backend nccl is not available; falling back to gloo. +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:43 [parallel_state.py:2119] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A +(StageEngineCoreProc_stage0_replica0 pid=51364) WARNING 09-23 16:39:44 [gpu_ar_worker.py:120] OMNI GPUARWorker forces v1 model runner for omni hooks. +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:44 [topk_topp_sampler.py:46] FlashInfer top-p/top-k sampling disabled via VLLM_USE_FLASHINFER_SAMPLER=0. +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:44 [base.py:138] Offloader set to UVAOffloader +(StageEngineCoreProc_stage0_replica0 pid=51364) WARNING 09-23 16:39:44 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage0_replica0 pid=51364) WARNING 09-23 16:39:44 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:44 [gpu_model_runner.py:5379] Starting to load model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5... +(StageEngineCoreProc_stage0_replica0 pid=51364) WARNING 09-23 16:39:44 [vllm.py:1371] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none +(StageEngineCoreProc_stage0_replica0 pid=51364) WARNING 09-23 16:39:44 [vllm.py:1406] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:44 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:44 [vllm.py:1585] Cudagraph is disabled under eager mode +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:44 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:44 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:44 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:45 [cuda.py:492] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'TRITON_ATTN', 'FLEX_ATTENTION']. +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:45 [flash_attn.py:897] Using FlashAttention version 2 +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:46 [uva.py:65] Total CPU offloaded parameters: 5.03 +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:47 [weight_utils.py:863] Filesystem type for checkpoints: unknown. Checkpoint size: 17.46 GiB. Available RAM: 23.26 GiB. +(StageEngineCoreProc_stage0_replica0 pid=51364) INFO 09-23 16:39:47 [weight_utils.py:886] Auto-prefetch is disabled because the filesystem (unknown) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch. +(StageEngineCoreProc_stage0_replica0 pid=51364) +Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [512], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'enable_qk_norm_rope_fusion': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False, 'fuse_qk_norm_rope_kvcache': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, enable_jit_warmup=True, enable_bf16x3_router_gemm=False, moe_backend='auto', linear_backend='auto') +(StageEngineCoreProc_stage1_replica0 pid=42376) INFO 09-23 16:42:16 [parallel_state.py:1775] world_size=1 rank=0 local_rank=0 distributed_init_method=file://C:\Users\zhout\AppData\Local\Temp/vllm_dist_3dcd7151d4db4756a5d71f8cc6beb2e4 backend=nccl +(StageEngineCoreProc_stage1_replica0 pid=42376) WARNING 09-23 16:42:16 [parallel_state.py:1788] Distributed backend nccl is not available; falling back to gloo. +(StageEngineCoreProc_stage1_replica0 pid=42376) INFO 09-23 16:42:16 [parallel_state.py:2119] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A +(StageEngineCoreProc_stage1_replica0 pid=42376) WARNING 09-23 16:42:16 [gpu_ar_worker.py:120] OMNI GPUARWorker forces v1 model runner for omni hooks. +(StageEngineCoreProc_stage1_replica0 pid=42376) INFO 09-23 16:42:16 [topk_topp_sampler.py:46] FlashInfer top-p/top-k sampling disabled via VLLM_USE_FLASHINFER_SAMPLER=0. +(StageEngineCoreProc_stage1_replica0 pid=42376) [transformers] `MiniCPMOProcessor` defines `image_processor_class = 'AutoImageProcessor'`, which is deprecated. Register the correct mapping in `AutoImageProcessor` instead. +(StageEngineCoreProc_stage1_replica0 pid=42376) [transformers] Requested torchvision backend is not available. Falling back to pil backend. +(StageEngineCoreProc_stage1_replica0 pid=42376) [transformers] `MiniCPMOProcessor` defines `feature_extractor_class = 'WhisperFeatureExtractor'`, which is deprecated. Register the correct mapping in `AutoFeatureExtractor` instead. +(StageEngineCoreProc_stage1_replica0 pid=42376) [transformers] Image processor MiniCPMVImageProcessor: kwargs ['max_slice_nums'] were applied for backward compatibility. To avoid this warning, add them to valid_kwargs: create a custom TypedDict extending ImagesKwargs with these keys and set it as the `valid_kwargs` class attribute. +(StageEngineCoreProc_stage1_replica0 pid=42376) INFO 09-23 16:42:28 [factory.py:46] Created connector: SharedMemoryConnector +(StageEngineCoreProc_stage1_replica0 pid=42376) WARNING 09-23 16:42:29 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage1_replica0 pid=42376) WARNING 09-23 16:42:29 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage1_replica0 pid=42376) INFO 09-23 16:42:29 [gpu_model_runner.py:5379] Starting to load model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5... +(StageEngineCoreProc_stage1_replica0 pid=42376) WARNING 09-23 16:42:29 [vllm.py:1371] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none +(StageEngineCoreProc_stage1_replica0 pid=42376) WARNING 09-23 16:42:29 [vllm.py:1406] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. +(StageEngineCoreProc_stage1_replica0 pid=42376) INFO 09-23 16:42:29 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage1_replica0 pid=42376) INFO 09-23 16:42:29 [vllm.py:1585] Cudagraph is disabled under eager mode +(StageEngineCoreProc_stage1_replica0 pid=42376) INFO 09-23 16:42:29 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant +(StageEngineCoreProc_stage1_replica0 pid=42376) INFO 09-23 16:42:29 [kernel.py:369] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) +(StageEngineCoreProc_stage1_replica0 pid=42376) INFO 09-23 16:42:29 [cuda.py:492] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'TRITON_ATTN', 'FLEX_ATTENTION']. +(StageEngineCoreProc_stage1_replica0 pid=42376) INFO 09-23 16:42:29 [flash_attn.py:897] Using FlashAttention version 2 +(StageEngineCoreProc_stage1_replica0 pid=42376) INFO 09-23 16:42:29 [weight_utils.py:863] Filesystem type for checkpoints: unknown. Checkpoint size: 17.46 GiB. Available RAM: 14.27 GiB. +(StageEngineCoreProc_stage1_replica0 pid=42376) INFO 09-23 16:42:29 [weight_utils.py:893] Auto-prefetch is disabled because the filesystem (unknown) is not a recognized network FS (NFS/Lustre) and the checkpoint size (17.46 GiB) exceeds 90% of available RAM (14.27 GiB). +(StageEngineCoreProc_stage1_replica0 pid=42376) +Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [4096], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'enable_qk_norm_rope_fusion': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False, 'fuse_qk_norm_rope_kvcache': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, enable_jit_warmup=True, enable_bf16x3_router_gemm=False, moe_backend='auto', linear_backend='auto') +(StageEngineCoreProc_stage2_replica0 pid=55720) INFO 09-23 16:42:58 [parallel_state.py:1775] world_size=1 rank=0 local_rank=0 distributed_init_method=file://C:\Users\zhout\AppData\Local\Temp/vllm_dist_6fe666f458a54616af06bcd6b9a314b2 backend=nccl +(StageEngineCoreProc_stage2_replica0 pid=55720) WARNING 09-23 16:42:58 [parallel_state.py:1788] Distributed backend nccl is not available; falling back to gloo. +(StageEngineCoreProc_stage2_replica0 pid=55720) INFO 09-23 16:42:58 [parallel_state.py:2119] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A +(StageEngineCoreProc_stage2_replica0 pid=55720) WARNING 09-23 16:42:58 [gpu_generation_worker.py:103] OMNI GPUGenerationWorker forces v1 model runner for omni hooks. +(StageEngineCoreProc_stage2_replica0 pid=55720) INFO 09-23 16:42:58 [topk_topp_sampler.py:46] FlashInfer top-p/top-k sampling disabled via VLLM_USE_FLASHINFER_SAMPLER=0. +(StageEngineCoreProc_stage2_replica0 pid=55720) INFO 09-23 16:42:58 [factory.py:46] Created connector: SharedMemoryConnector +(StageEngineCoreProc_stage2_replica0 pid=55720) WARNING 09-23 16:42:58 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage2_replica0 pid=55720) WARNING 09-23 16:42:58 [base.py:173] [LLM Worker 0] Sleep Mode DISABLED. +(StageEngineCoreProc_stage2_replica0 pid=55720) INFO 09-23 16:42:58 [gpu_model_runner.py:5379] Starting to load model \\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\models\MiniCPM-o-4_5... +(StageEngineCoreProc_stage2_replica0 pid=55720) INFO 09-23 16:42:59 [weight_utils.py:863] Filesystem type for checkpoints: unknown. Checkpoint size: 17.46 GiB. Available RAM: 12.14 GiB. +(StageEngineCoreProc_stage2_replica0 pid=55720) INFO 09-23 16:42:59 [weight_utils.py:893] Auto-prefetch is disabled because the filesystem (unknown) is not a recognized network FS (NFS/Lustre) and the checkpoint size (17.46 GiB) exceeds 90% of available RAM (12.14 GiB). +(StageEngineCoreProc_stage2_replica0 pid=55720) +Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00._kernel. This causes a latency spike; consider extending warmup to cover this shape/config. +(StageEngineCoreProc_stage2_replica0 pid=55720) INFO 09-23 16:43:43 [cuda_graph_wrapper.py:98] Captured HiFT CUDA Graph for shape (1, 18, 3840) +Processed prompts: 0%| | 0/1 [00:00