diff --git a/worklogs/2026-07-13-initial-run.md b/worklogs/2026-07-13-initial-run.md index 5225c76..435c0fd 100644 --- a/worklogs/2026-07-13-initial-run.md +++ b/worklogs/2026-07-13-initial-run.md @@ -268,3 +268,22 @@ - Increasing concurrency does not improve aggregate decode throughput and worsens per-request TPS. - Low average GPU utilization suggests scheduling, TP synchronization, MoE kernel, paged attention, or xFormers backend limitations before raw compute saturation. - Next focused experiment: disable chunked prefill and retry `--num-scheduler-steps 4` as a decode-only diagnostic variant. + +## 2026-07-14 Multi-Step Scheduling Diagnostic + +- Goal: verify whether `--num-scheduler-steps 4` can improve decode throughput. +- Appended details to: + - `worklogs/decode_analysis_report_2026-07-14.md` +- Synced remote failure logs to: + - `worklogs/remote_results/2026-07-14-scheduler/` +- Experiments: + - No chunked prefill + `max_model_len=100000` failed because `max_num_batched_tokens=8192` is smaller than `max_model_len=100000`. + - No chunked prefill + `max_model_len=4096` still failed because `Multi-Step not supported for attention backend: xformers`. + - Forced `VLLM_ATTENTION_BACKEND=FLASHINFER` failed because `BatchDecodeWithPagedKVCacheWrapper` was `None`, indicating FlashInfer dependencies/wrappers are unavailable or incompatible. +- Conclusion: + - Multi-step scheduling cannot currently be tested or used on this server because the available attention backend is xFormers. + - The next practical direction is not more `num_scheduler_steps` tuning, but either making FlashAttention/FlashInfer backend available or profiling the current xFormers + MoE + TP decode path. +- Recovery: + - Restored full parser service with chunked prefill and prefix cache. + - Recovery log: `/root/work/logs/server_exp_seq2_b8192_fullparser_restored_after_schedtest.log`. + - `/health` returned `200`. diff --git a/worklogs/decode_analysis_report_2026-07-14.md b/worklogs/decode_analysis_report_2026-07-14.md index d11b239..3a31ab4 100644 --- a/worklogs/decode_analysis_report_2026-07-14.md +++ b/worklogs/decode_analysis_report_2026-07-14.md @@ -297,3 +297,137 @@ GPU 利用率和功耗都偏低。虽然瞬时 GPU-Util 能到 100%，但平均 4. 若没有提升，进入 qwen3_moe / fused_moe / attention 的代码级 profiling 本轮最重要的事实是：GPU 平均利用率只有约 `29%`，所以先不要把问题简单归因为“卡算不动”。更像是当前 decode 执行路径没有把四张卡持续喂满。 + +## 七、调度方向验证实验：multi-step decode + +追加日期：2026-07-14 + +### 目标 + +验证 `--num-scheduler-steps 4` 是否能作为 decode 吞吐提升方向。 + +由于上一轮实验显示并发升高没有带来 aggregate TPS 增益，且平均 GPU 利用率只有约 `29%`，multi-step scheduling 是最直接的调度侧候选优化：它理论上可以减少每 token 调度往返和 Python/worker 协调开销，让 decode 连续执行多个 step。 + +### 实验 1：关闭 chunked prefill，直接启用 multi-step + +启动变体： + +- `--num-scheduler-steps 4` +- 不加 `--enable-chunked-prefill` +- `--max-model-len 100000` +- `--max-num-batched-tokens 8192` +- 其它参数沿用完整 parser 服务配置 + +结果： + +- 服务未启动。 +- 远端日志：`/root/work/logs/server_exp_sched4_nochunk_fullparser.log` +- 本地归档：`worklogs/remote_results/2026-07-14-scheduler/server_exp_sched4_nochunk_fullparser.log` + +失败原因： + +```text +ValueError: max_num_batched_tokens (8192) is smaller than max_model_len (100000). +``` + +解释： + +关闭 chunked prefill 后，vLLM 要求 `max_num_batched_tokens >= max_model_len`，否则实际最大可处理序列会被 `max_num_batched_tokens` 限制。这个配置不能用于官方长上下文，也无法进入 decode 压测。 + +### 实验 2：decode-only 诊断配置 + +为绕过实验 1 的限制，临时降低上下文长度，只用于短 prompt decode 诊断。 + +启动变体： + +- `--max-model-len 4096` +- `--max-seq-len-to-capture 4096` +- `--max-num-batched-tokens 8192` +- `--num-scheduler-steps 4` +- 不加 `--enable-chunked-prefill` +- attention backend 仍为自动选择 + +结果： + +- 服务未启动。 +- 远端日志：`/root/work/logs/server_exp_sched4_nochunk_len4096_fullparser.log` +- 本地归档：`worklogs/remote_results/2026-07-14-scheduler/server_exp_sched4_nochunk_len4096_fullparser.log` + +关键日志： + +```text +Using XFormers backend. +ValueError: Multi-Step not supported for attention backend: xformers. +Set VLLM_ATTENTION_BACKEND to a value from ['flash-attn', 'rocm-flash-attn', 'flashinfer']. +``` + +解释： + +这说明当前环境不是“chunked prefill 与 multi-step 的组合不支持”这么简单，而是 `xformers` attention backend 本身不支持 multi-step worker。只要 attention backend 仍然落到 xFormers，multi-step 调度就无法启用。 + +### 实验 3：强制 FlashInfer backend + +启动变体： + +- 环境变量：`VLLM_ATTENTION_BACKEND=FLASHINFER` +- `--max-model-len 4096` +- `--num-scheduler-steps 4` +- 不加 `--enable-chunked-prefill` +- 其它参数同实验 2 + +结果： + +- 服务未启动。 +- 远端日志：`/root/work/logs/server_exp_sched4_nochunk_len4096_flashinfer.log` +- 本地归档：`worklogs/remote_results/2026-07-14-scheduler/server_exp_sched4_nochunk_len4096_flashinfer.log` + +关键日志： + +```text +TypeError: 'NoneType' object is not callable +... +self._decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(...) +``` + +解释： + +`vllm/attention/backends/flashinfer.py` 会导入： + +- `flashinfer.BatchDecodeWithPagedKVCacheWrapper` +- `flashinfer.decode.CUDAGraphBatchDecodeWithPagedKVCacheWrapper` +- `flashinfer.prefill.BatchPrefillWithPagedKVCacheWrapper` +- `ixformer.contrib.vllm_flash_attn.flash_attn_varlen_func` + +当前环境中至少有关键 FlashInfer wrapper 没导入成功，导致 wrapper 为 `None`，在 profiling 阶段调用时报错。 + +### 实验结论 + +本轮没有进入 c1/c2/c4 decode 曲线压测，因为 multi-step 服务在启动阶段就失败。 + +结论不是“调度一定无效”，而是： + +1. 当前 xFormers backend 下，multi-step 调度不可用。 +2. 当前 FlashInfer backend 依赖不完整或与 CoreX 环境不兼容，不能直接替代 xFormers。 +3. 自动 FlashAttention 也不可用；此前日志已显示 `vllm_flash_attn` 包缺失，因此自动回落到 xFormers。 + +因此，调度优化如果要继续推进，前置任务是 attention backend 适配： + +- 路线 A：补齐/修复 CoreX 环境里的 FlashAttention 或 FlashInfer backend。 +- 路线 B：改造 xFormers backend 或 MultiStepModelRunner，使其支持当前 xFormers 路径。 +- 路线 C：绕过 multi-step，直接 profile xFormers decode、paged attention、MoE 与 TP 通信开销。 + +### 对下一步方向的影响 + +短期内，继续调 `--num-scheduler-steps` 没意义；它被 backend 卡住了。 + +下一步建议改为两条线并行： + +1. **backend 可用性线** + - 检查 `flashinfer` 和 `ixformer.contrib.vllm_flash_attn` 在服务器上的实际导入错误。 + - 确认官方镜像/包中是否本应包含 `vllm_flash_attn`。 + - 如果能补齐依赖，再重跑 multi-step decode 曲线。 + +2. **代码 profiling 线** + - 直接在当前可用 xFormers 路径插桩。 + - 重点记录每 token decode 中 attention、MoE、sampler、TP 同步的耗时。 + - 当前平均 GPU 利用率低，profiling 比继续盲调参数更有价值。 diff --git a/worklogs/remote_results/2026-07-14-scheduler/server_exp_sched4_nochunk_fullparser.log b/worklogs/remote_results/2026-07-14-scheduler/server_exp_sched4_nochunk_fullparser.log new file mode 100644 index 0000000..a776058 --- /dev/null +++ b/worklogs/remote_results/2026-07-14-scheduler/server_exp_sched4_nochunk_fullparser.log @@ -0,0 +1,41 @@ +/usr/local/corex/lib/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you. + import pynvml # type: ignore[import] +INFO 07-14 09:19:28 importing.py:10] Triton not installed; certain GPU-related functions will not be available. +2026-07-14 09:19:30.064377: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`. +2026-07-14 09:19:30.115922: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations. +To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags. +WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them. +INFO 07-14 09:19:35 api_server.py:530] vLLM API server version 0.6.3 +INFO 07-14 09:19:35 api_server.py:531] args: Namespace(host='0.0.0.0', port=1111, uvicorn_log_level='info', allow_credentials=False, allowed_origins=['*'], allowed_methods=['*'], allowed_headers=['*'], api_key=None, lora_modules=None, prompt_adapters=None, chat_template=None, response_role='assistant', ssl_keyfile=None, ssl_certfile=None, ssl_ca_certs=None, ssl_cert_reqs=0, root_path=None, middleware=[], return_tokens_as_token_ids=False, disable_frontend_multiprocessing=True, enable_auto_tool_choice=True, tool_call_parser='qwen3_coder', tool_parser_plugin='', reasoning_parser='qwen3', model='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', tokenizer=None, skip_tokenizer_init=False, revision=None, code_revision=None, tokenizer_revision=None, tokenizer_mode='auto', trust_remote_code=True, download_dir=None, load_format='auto', config_format='auto', dtype='auto', kv_cache_dtype='auto', quantization_param_path=None, max_model_len=100000, guided_decoding_backend='outlines', distributed_executor_backend=None, worker_use_ray=False, pipeline_parallel_size=1, tensor_parallel_size=4, max_parallel_loading_workers=None, ray_workers_use_nsight=False, block_size=16, enable_prefix_caching=True, disable_sliding_window=False, use_v2_block_manager=True, num_lookahead_slots=0, seed=0, swap_space=4, cpu_offload_gb=0, gpu_memory_utilization=0.95, num_gpu_blocks_override=None, max_num_batched_tokens=8192, max_num_seqs=2, max_logprobs=20, disable_log_stats=False, quantization=None, rope_scaling=None, rope_theta=None, enforce_eager=False, max_context_len_to_capture=None, max_seq_len_to_capture=32768, disable_custom_all_reduce=False, tokenizer_pool_size=0, tokenizer_pool_type='ray', tokenizer_pool_extra_config=None, limit_mm_per_prompt=None, mm_processor_kwargs=None, enable_lora=False, max_loras=1, max_lora_rank=16, lora_extra_vocab_size=256, lora_dtype='auto', long_lora_scaling_factors=None, max_cpu_loras=None, fully_sharded_loras=False, enable_prompt_adapter=False, max_prompt_adapters=1, max_prompt_adapter_token=0, device='auto', num_scheduler_steps=4, multi_step_stream_outputs=True, scheduler_delay_factor=0.0, enable_chunked_prefill=None, speculative_model=None, speculative_model_quantization=None, num_speculative_tokens=None, speculative_disable_mqa_scorer=False, speculative_draft_tensor_parallel_size=None, speculative_max_model_len=None, speculative_disable_by_batch_size=None, ngram_prompt_lookup_max=None, ngram_prompt_lookup_min=None, spec_decoding_acceptance_method='rejection_sampler', typical_acceptance_sampler_posterior_threshold=None, typical_acceptance_sampler_posterior_alpha=None, disable_logprobs_during_spec_decoding=None, model_loader_extra_config=None, ignore_patterns=[], preemption_mode=None, served_model_name=['llm'], qlora_adapter_name_or_path=None, otlp_traces_endpoint=None, collect_detailed_traces=None, disable_async_output_proc=False, override_neuron_config=None, scheduling_policy='fcfs', disable_log_requests=True, max_log_len=None, disable_fastapi_docs=False) +INFO 07-14 09:19:35 config.py:1670] Downcasting torch.float32 to torch.float16. +INFO 07-14 09:19:46 config.py:887] Defaulting to use mp for distributed inference +WARNING 07-14 09:19:46 arg_utils.py:963] The model has a long context length (100000). This may cause OOM errors during the initial memory profiling phase, or result in low performance due to small KV cache space. Consider setting --max-model-len to a smaller value. +Traceback (most recent call last): + File "/usr/local/lib/python3.10/runpy.py", line 196, in _run_module_as_main + return _run_code(code, main_globals, None, + File "/usr/local/lib/python3.10/runpy.py", line 86, in _run_code + exec(code, run_globals) + File "/usr/local/corex/lib/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 595, in + uvloop.run(run_server(args)) + File "/usr/local/lib/python3.10/site-packages/uvloop/__init__.py", line 82, in run + return loop.run_until_complete(wrapper()) + File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete + File "/usr/local/lib/python3.10/site-packages/uvloop/__init__.py", line 61, in wrapper + return await main + File "/usr/local/corex/lib/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 562, in run_server + async with build_async_engine_client(args) as engine_client: + File "/usr/local/lib/python3.10/contextlib.py", line 199, in __aenter__ + return await anext(self.gen) + File "/usr/local/corex/lib/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 108, in build_async_engine_client + async with build_async_engine_client_from_engine_args( + File "/usr/local/lib/python3.10/contextlib.py", line 199, in __aenter__ + return await anext(self.gen) + File "/usr/local/corex/lib/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 130, in build_async_engine_client_from_engine_args + engine_config = engine_args.create_engine_config() + File "/usr/local/corex/lib/python3/dist-packages/vllm/engine/arg_utils.py", line 1021, in create_engine_config + scheduler_config = SchedulerConfig( + File "/usr/local/corex/lib/python3/dist-packages/vllm/config.py", line 1021, in __init__ + self._verify_args() + File "/usr/local/corex/lib/python3/dist-packages/vllm/config.py", line 1026, in _verify_args + raise ValueError( +ValueError: max_num_batched_tokens (8192) is smaller than max_model_len (100000). This effectively limits the maximum sequence length to max_num_batched_tokens and makes vLLM reject longer sequences. Please increase max_num_batched_tokens or decrease max_model_len. diff --git a/worklogs/remote_results/2026-07-14-scheduler/server_exp_sched4_nochunk_len4096_flashinfer.log b/worklogs/remote_results/2026-07-14-scheduler/server_exp_sched4_nochunk_len4096_flashinfer.log new file mode 100644 index 0000000..12a3935 --- /dev/null +++ b/worklogs/remote_results/2026-07-14-scheduler/server_exp_sched4_nochunk_len4096_flashinfer.log @@ -0,0 +1,212 @@ +/usr/local/corex/lib/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you. + import pynvml # type: ignore[import] +INFO 07-14 09:43:16 importing.py:10] Triton not installed; certain GPU-related functions will not be available. +2026-07-14 09:43:17.917732: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`. +2026-07-14 09:43:17.968433: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations. +To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags. +WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them. +INFO 07-14 09:43:23 api_server.py:530] vLLM API server version 0.6.3 +INFO 07-14 09:43:23 api_server.py:531] args: Namespace(host='0.0.0.0', port=1111, uvicorn_log_level='info', allow_credentials=False, allowed_origins=['*'], allowed_methods=['*'], allowed_headers=['*'], api_key=None, lora_modules=None, prompt_adapters=None, chat_template=None, response_role='assistant', ssl_keyfile=None, ssl_certfile=None, ssl_ca_certs=None, ssl_cert_reqs=0, root_path=None, middleware=[], return_tokens_as_token_ids=False, disable_frontend_multiprocessing=True, enable_auto_tool_choice=True, tool_call_parser='qwen3_coder', tool_parser_plugin='', reasoning_parser='qwen3', model='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', tokenizer=None, skip_tokenizer_init=False, revision=None, code_revision=None, tokenizer_revision=None, tokenizer_mode='auto', trust_remote_code=True, download_dir=None, load_format='auto', config_format='auto', dtype='auto', kv_cache_dtype='auto', quantization_param_path=None, max_model_len=4096, guided_decoding_backend='outlines', distributed_executor_backend=None, worker_use_ray=False, pipeline_parallel_size=1, tensor_parallel_size=4, max_parallel_loading_workers=None, ray_workers_use_nsight=False, block_size=16, enable_prefix_caching=True, disable_sliding_window=False, use_v2_block_manager=True, num_lookahead_slots=0, seed=0, swap_space=4, cpu_offload_gb=0, gpu_memory_utilization=0.95, num_gpu_blocks_override=None, max_num_batched_tokens=8192, max_num_seqs=2, max_logprobs=20, disable_log_stats=False, quantization=None, rope_scaling=None, rope_theta=None, enforce_eager=False, max_context_len_to_capture=None, max_seq_len_to_capture=4096, disable_custom_all_reduce=False, tokenizer_pool_size=0, tokenizer_pool_type='ray', tokenizer_pool_extra_config=None, limit_mm_per_prompt=None, mm_processor_kwargs=None, enable_lora=False, max_loras=1, max_lora_rank=16, lora_extra_vocab_size=256, lora_dtype='auto', long_lora_scaling_factors=None, max_cpu_loras=None, fully_sharded_loras=False, enable_prompt_adapter=False, max_prompt_adapters=1, max_prompt_adapter_token=0, device='auto', num_scheduler_steps=4, multi_step_stream_outputs=True, scheduler_delay_factor=0.0, enable_chunked_prefill=None, speculative_model=None, speculative_model_quantization=None, num_speculative_tokens=None, speculative_disable_mqa_scorer=False, speculative_draft_tensor_parallel_size=None, speculative_max_model_len=None, speculative_disable_by_batch_size=None, ngram_prompt_lookup_max=None, ngram_prompt_lookup_min=None, spec_decoding_acceptance_method='rejection_sampler', typical_acceptance_sampler_posterior_threshold=None, typical_acceptance_sampler_posterior_alpha=None, disable_logprobs_during_spec_decoding=None, model_loader_extra_config=None, ignore_patterns=[], preemption_mode=None, served_model_name=['llm'], qlora_adapter_name_or_path=None, otlp_traces_endpoint=None, collect_detailed_traces=None, disable_async_output_proc=False, override_neuron_config=None, scheduling_policy='fcfs', disable_log_requests=True, max_log_len=None, disable_fastapi_docs=False) +INFO 07-14 09:43:23 config.py:1670] Downcasting torch.float32 to torch.float16. +INFO 07-14 09:43:34 config.py:887] Defaulting to use mp for distributed inference +WARNING 07-14 09:43:34 config.py:380] To see benefits of async output processing, enable CUDA graph. Since, enforce-eager is enabled, async output processor cannot be used +INFO 07-14 09:43:34 llm_engine.py:237] Initializing an LLM engine (v0.6.3) with config: model='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', speculative_config=None, tokenizer='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, rope_scaling=None, rope_theta=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.float16, max_seq_len=4096, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=4, pipeline_parallel_size=1, disable_custom_all_reduce=True, quantization=None, enforce_eager=True, kv_cache_dtype=auto, quantization_param_path=None, device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='outlines'), observability_config=ObservabilityConfig(otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=0, served_model_name=llm, use_v2_block_manager=True, num_scheduler_steps=4, chunked_prefill_enabled=False multi_step_stream_outputs=True, enable_prefix_caching=True, use_async_output_proc=False, use_cached_outputs=False, mm_processor_kwargs=None) +INFO 07-14 09:43:34 custom_cache_manager.py:17] Setting Triton cache manager to: vllm.triton_utils.custom_cache_manager:CustomCacheManager +INFO 07-14 09:43:34 selector.py:141] Using Flashinfer backend. +WARNING 07-14 09:43:34 registry.py:205] `mm_limits` has already been set for model=/root/public-storage/models/Qwen/Qwen3.6-35B-A3B, and will be overwritten by the new values. +/usr/local/corex/lib/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you. + import pynvml # type: ignore[import] +/usr/local/corex/lib/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you. + import pynvml # type: ignore[import] +/usr/local/corex/lib/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you. + import pynvml # type: ignore[import] +INFO 07-14 09:43:36 importing.py:10] Triton not installed; certain GPU-related functions will not be available. +INFO 07-14 09:43:36 importing.py:10] Triton not installed; certain GPU-related functions will not be available. +INFO 07-14 09:43:36 importing.py:10] Triton not installed; certain GPU-related functions will not be available. +WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them. +WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them. +WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them. +(VllmWorkerProcess pid=6761) INFO 07-14 09:43:44 selector.py:141] Using Flashinfer backend. +(VllmWorkerProcess pid=6761) WARNING 07-14 09:43:44 registry.py:205] `mm_limits` has already been set for model=/root/public-storage/models/Qwen/Qwen3.6-35B-A3B, and will be overwritten by the new values. +(VllmWorkerProcess pid=6761) INFO 07-14 09:43:44 multiproc_worker_utils.py:216] Worker ready; awaiting tasks +(VllmWorkerProcess pid=6760) INFO 07-14 09:43:44 selector.py:141] Using Flashinfer backend. +(VllmWorkerProcess pid=6760) WARNING 07-14 09:43:44 registry.py:205] `mm_limits` has already been set for model=/root/public-storage/models/Qwen/Qwen3.6-35B-A3B, and will be overwritten by the new values. +(VllmWorkerProcess pid=6760) INFO 07-14 09:43:44 multiproc_worker_utils.py:216] Worker ready; awaiting tasks +(VllmWorkerProcess pid=6759) INFO 07-14 09:43:44 selector.py:141] Using Flashinfer backend. +(VllmWorkerProcess pid=6759) WARNING 07-14 09:43:44 registry.py:205] `mm_limits` has already been set for model=/root/public-storage/models/Qwen/Qwen3.6-35B-A3B, and will be overwritten by the new values. +(VllmWorkerProcess pid=6759) INFO 07-14 09:43:44 multiproc_worker_utils.py:216] Worker ready; awaiting tasks +INFO 07-14 09:43:45 shm_broadcast.py:242] vLLM message queue communication handle: Handle(connect_ip='127.0.0.1', local_reader_ranks=[1, 2, 3], buffer=, local_subscribe_port=60197, remote_subscribe_port=None) +INFO 07-14 09:43:45 model_runner.py:1065] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B... +(VllmWorkerProcess pid=6759) INFO 07-14 09:43:45 model_runner.py:1065] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B... +(VllmWorkerProcess pid=6760) INFO 07-14 09:43:45 model_runner.py:1065] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B... +(VllmWorkerProcess pid=6761) INFO 07-14 09:43:45 model_runner.py:1065] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B... +(VllmWorkerProcess pid=6759) INFO 07-14 09:43:45 selector.py:141] Using Flashinfer backend. +(VllmWorkerProcess pid=6761) INFO 07-14 09:43:45 selector.py:141] Using Flashinfer backend. +(VllmWorkerProcess pid=6760) INFO 07-14 09:43:45 selector.py:141] Using Flashinfer backend. +INFO 07-14 09:43:45 selector.py:141] Using Flashinfer backend. + Loading safetensors checkpoint shards: 0% Completed | 0/26 [00:00 + uvloop.run(run_server(args)) + File "/usr/local/lib/python3.10/site-packages/uvloop/__init__.py", line 82, in run + return loop.run_until_complete(wrapper()) + File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete + File "/usr/local/lib/python3.10/site-packages/uvloop/__init__.py", line 61, in wrapper + return await main + File "/usr/local/corex/lib/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 562, in run_server + async with build_async_engine_client(args) as engine_client: + File "/usr/local/lib/python3.10/contextlib.py", line 199, in __aenter__ + return await anext(self.gen) + File "/usr/local/corex/lib/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 108, in build_async_engine_client + async with build_async_engine_client_from_engine_args( + File "/usr/local/lib/python3.10/contextlib.py", line 199, in __aenter__ + return await anext(self.gen) + File "/usr/local/corex/lib/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 142, in build_async_engine_client_from_engine_args + engine_client = await asyncio.get_running_loop().run_in_executor( + File "/usr/local/lib/python3.10/concurrent/futures/thread.py", line 58, in run + result = self.fn(*self.args, **self.kwargs) + File "/usr/local/corex/lib/python3/dist-packages/vllm/engine/async_llm_engine.py", line 674, in from_engine_args + engine = cls( + File "/usr/local/corex/lib/python3/dist-packages/vllm/engine/async_llm_engine.py", line 569, in __init__ + self.engine = self._engine_class(*args, **kwargs) + File "/usr/local/corex/lib/python3/dist-packages/vllm/engine/async_llm_engine.py", line 265, in __init__ + super().__init__(*args, **kwargs) + File "/usr/local/corex/lib/python3/dist-packages/vllm/engine/llm_engine.py", line 349, in __init__ + self._initialize_kv_caches() + File "/usr/local/corex/lib/python3/dist-packages/vllm/engine/llm_engine.py", line 484, in _initialize_kv_caches + self.model_executor.determine_num_available_blocks()) + File "/usr/local/corex/lib/python3/dist-packages/vllm/executor/distributed_gpu_executor.py", line 39, in determine_num_available_blocks + num_blocks = self._run_workers("determine_num_available_blocks", ) + File "/usr/local/corex/lib/python3/dist-packages/vllm/executor/multiproc_gpu_executor.py", line 192, in _run_workers + driver_worker_output = driver_worker_method(*args, **kwargs) + File "/usr/local/corex/lib/python3/dist-packages/torch/utils/_contextlib.py", line 115, in decorate_context + return func(*args, **kwargs) + File "/usr/local/corex/lib/python3/dist-packages/vllm/worker/worker.py", line 223, in determine_num_available_blocks + self.model_runner.profile_run() + File "/usr/local/corex/lib/python3/dist-packages/vllm/worker/multi_step_model_runner.py", line 661, in profile_run + return self._base_model_runner.profile_run() + File "/usr/local/corex/lib/python3/dist-packages/torch/utils/_contextlib.py", line 115, in decorate_context + return func(*args, **kwargs) + File "/usr/local/corex/lib/python3/dist-packages/vllm/worker/model_runner.py", line 1314, in profile_run + self.execute_model(model_input, kv_caches, intermediate_tensors) + File "/usr/local/corex/lib/python3/dist-packages/torch/utils/_contextlib.py", line 115, in decorate_context + return func(*args, **kwargs) + File "/usr/local/corex/lib/python3/dist-packages/vllm/worker/model_runner.py", line 1641, in execute_model + self.attn_state.begin_forward(model_input) + File "/usr/local/corex/lib/python3/dist-packages/vllm/attention/backends/flashinfer.py", line 261, in begin_forward + model_input.attn_metadata.decode_wrapper = state._get_decode_wrapper() + File "/usr/local/corex/lib/python3/dist-packages/vllm/attention/backends/flashinfer.py", line 130, in _get_decode_wrapper + self._decode_wrapper = BatchDecodeWithPagedKVCacheWrapper( +TypeError: 'NoneType' object is not callable +ERROR 07-14 09:44:24 multiproc_worker_utils.py:117] Worker VllmWorkerProcess pid 6760 died, exit code: -15 +ERROR 07-14 09:44:24 multiproc_worker_utils.py:117] Worker VllmWorkerProcess pid 6761 died, exit code: -15 +INFO 07-14 09:44:24 multiproc_worker_utils.py:121] Killing local vLLM worker processes +/usr/local/lib/python3.10/multiprocessing/resource_tracker.py:224: UserWarning: resource_tracker: There appear to be 3 leaked semaphore objects to clean up at shutdown + warnings.warn('resource_tracker: There appear to be %d ' +/usr/local/lib/python3.10/multiprocessing/resource_tracker.py:224: UserWarning: resource_tracker: There appear to be 1 leaked shared_memory objects to clean up at shutdown + warnings.warn('resource_tracker: There appear to be %d ' diff --git a/worklogs/remote_results/2026-07-14-scheduler/server_exp_sched4_nochunk_len4096_fullparser.log b/worklogs/remote_results/2026-07-14-scheduler/server_exp_sched4_nochunk_len4096_fullparser.log new file mode 100644 index 0000000..dc3f339 --- /dev/null +++ b/worklogs/remote_results/2026-07-14-scheduler/server_exp_sched4_nochunk_len4096_fullparser.log @@ -0,0 +1,72 @@ +/usr/local/corex/lib/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you. + import pynvml # type: ignore[import] +INFO 07-14 09:23:27 importing.py:10] Triton not installed; certain GPU-related functions will not be available. +2026-07-14 09:23:29.177777: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`. +2026-07-14 09:23:29.228965: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations. +To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags. +WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them. +INFO 07-14 09:23:34 api_server.py:530] vLLM API server version 0.6.3 +INFO 07-14 09:23:34 api_server.py:531] args: Namespace(host='0.0.0.0', port=1111, uvicorn_log_level='info', allow_credentials=False, allowed_origins=['*'], allowed_methods=['*'], allowed_headers=['*'], api_key=None, lora_modules=None, prompt_adapters=None, chat_template=None, response_role='assistant', ssl_keyfile=None, ssl_certfile=None, ssl_ca_certs=None, ssl_cert_reqs=0, root_path=None, middleware=[], return_tokens_as_token_ids=False, disable_frontend_multiprocessing=True, enable_auto_tool_choice=True, tool_call_parser='qwen3_coder', tool_parser_plugin='', reasoning_parser='qwen3', model='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', tokenizer=None, skip_tokenizer_init=False, revision=None, code_revision=None, tokenizer_revision=None, tokenizer_mode='auto', trust_remote_code=True, download_dir=None, load_format='auto', config_format='auto', dtype='auto', kv_cache_dtype='auto', quantization_param_path=None, max_model_len=4096, guided_decoding_backend='outlines', distributed_executor_backend=None, worker_use_ray=False, pipeline_parallel_size=1, tensor_parallel_size=4, max_parallel_loading_workers=None, ray_workers_use_nsight=False, block_size=16, enable_prefix_caching=True, disable_sliding_window=False, use_v2_block_manager=True, num_lookahead_slots=0, seed=0, swap_space=4, cpu_offload_gb=0, gpu_memory_utilization=0.95, num_gpu_blocks_override=None, max_num_batched_tokens=8192, max_num_seqs=2, max_logprobs=20, disable_log_stats=False, quantization=None, rope_scaling=None, rope_theta=None, enforce_eager=False, max_context_len_to_capture=None, max_seq_len_to_capture=4096, disable_custom_all_reduce=False, tokenizer_pool_size=0, tokenizer_pool_type='ray', tokenizer_pool_extra_config=None, limit_mm_per_prompt=None, mm_processor_kwargs=None, enable_lora=False, max_loras=1, max_lora_rank=16, lora_extra_vocab_size=256, lora_dtype='auto', long_lora_scaling_factors=None, max_cpu_loras=None, fully_sharded_loras=False, enable_prompt_adapter=False, max_prompt_adapters=1, max_prompt_adapter_token=0, device='auto', num_scheduler_steps=4, multi_step_stream_outputs=True, scheduler_delay_factor=0.0, enable_chunked_prefill=None, speculative_model=None, speculative_model_quantization=None, num_speculative_tokens=None, speculative_disable_mqa_scorer=False, speculative_draft_tensor_parallel_size=None, speculative_max_model_len=None, speculative_disable_by_batch_size=None, ngram_prompt_lookup_max=None, ngram_prompt_lookup_min=None, spec_decoding_acceptance_method='rejection_sampler', typical_acceptance_sampler_posterior_threshold=None, typical_acceptance_sampler_posterior_alpha=None, disable_logprobs_during_spec_decoding=None, model_loader_extra_config=None, ignore_patterns=[], preemption_mode=None, served_model_name=['llm'], qlora_adapter_name_or_path=None, otlp_traces_endpoint=None, collect_detailed_traces=None, disable_async_output_proc=False, override_neuron_config=None, scheduling_policy='fcfs', disable_log_requests=True, max_log_len=None, disable_fastapi_docs=False) +INFO 07-14 09:23:34 config.py:1670] Downcasting torch.float32 to torch.float16. +INFO 07-14 09:23:45 config.py:887] Defaulting to use mp for distributed inference +WARNING 07-14 09:23:45 config.py:380] To see benefits of async output processing, enable CUDA graph. Since, enforce-eager is enabled, async output processor cannot be used +INFO 07-14 09:23:45 llm_engine.py:237] Initializing an LLM engine (v0.6.3) with config: model='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', speculative_config=None, tokenizer='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, rope_scaling=None, rope_theta=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.float16, max_seq_len=4096, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=4, pipeline_parallel_size=1, disable_custom_all_reduce=True, quantization=None, enforce_eager=True, kv_cache_dtype=auto, quantization_param_path=None, device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='outlines'), observability_config=ObservabilityConfig(otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=0, served_model_name=llm, use_v2_block_manager=True, num_scheduler_steps=4, chunked_prefill_enabled=False multi_step_stream_outputs=True, enable_prefix_caching=True, use_async_output_proc=False, use_cached_outputs=False, mm_processor_kwargs=None) +INFO 07-14 09:23:46 custom_cache_manager.py:17] Setting Triton cache manager to: vllm.triton_utils.custom_cache_manager:CustomCacheManager +INFO 07-14 09:23:46 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default). +INFO 07-14 09:23:46 selector.py:115] Using XFormers backend. +WARNING 07-14 09:23:46 registry.py:205] `mm_limits` has already been set for model=/root/public-storage/models/Qwen/Qwen3.6-35B-A3B, and will be overwritten by the new values. +Traceback (most recent call last): + File "/usr/local/lib/python3.10/runpy.py", line 196, in _run_module_as_main + return _run_code(code, main_globals, None, + File "/usr/local/lib/python3.10/runpy.py", line 86, in _run_code + exec(code, run_globals) + File "/usr/local/corex/lib/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 595, in + uvloop.run(run_server(args)) + File "/usr/local/lib/python3.10/site-packages/uvloop/__init__.py", line 82, in run + return loop.run_until_complete(wrapper()) + File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete + File "/usr/local/lib/python3.10/site-packages/uvloop/__init__.py", line 61, in wrapper + return await main + File "/usr/local/corex/lib/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 562, in run_server + async with build_async_engine_client(args) as engine_client: + File "/usr/local/lib/python3.10/contextlib.py", line 199, in __aenter__ + return await anext(self.gen) + File "/usr/local/corex/lib/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 108, in build_async_engine_client + async with build_async_engine_client_from_engine_args( + File "/usr/local/lib/python3.10/contextlib.py", line 199, in __aenter__ + return await anext(self.gen) + File "/usr/local/corex/lib/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 142, in build_async_engine_client_from_engine_args + engine_client = await asyncio.get_running_loop().run_in_executor( + File "/usr/local/lib/python3.10/concurrent/futures/thread.py", line 58, in run + result = self.fn(*self.args, **self.kwargs) + File "/usr/local/corex/lib/python3/dist-packages/vllm/engine/async_llm_engine.py", line 674, in from_engine_args + engine = cls( + File "/usr/local/corex/lib/python3/dist-packages/vllm/engine/async_llm_engine.py", line 569, in __init__ + self.engine = self._engine_class(*args, **kwargs) + File "/usr/local/corex/lib/python3/dist-packages/vllm/engine/async_llm_engine.py", line 265, in __init__ + super().__init__(*args, **kwargs) + File "/usr/local/corex/lib/python3/dist-packages/vllm/engine/llm_engine.py", line 335, in __init__ + self.model_executor = executor_class( + File "/usr/local/corex/lib/python3/dist-packages/vllm/executor/multiproc_gpu_executor.py", line 215, in __init__ + super().__init__(*args, **kwargs) + File "/usr/local/corex/lib/python3/dist-packages/vllm/executor/distributed_gpu_executor.py", line 26, in __init__ + super().__init__(*args, **kwargs) + File "/usr/local/corex/lib/python3/dist-packages/vllm/executor/executor_base.py", line 47, in __init__ + self._init_executor() + File "/usr/local/corex/lib/python3/dist-packages/vllm/executor/multiproc_gpu_executor.py", line 108, in _init_executor + self.driver_worker = self._create_worker( + File "/usr/local/corex/lib/python3/dist-packages/vllm/executor/gpu_executor.py", line 105, in _create_worker + return create_worker(**self._get_create_worker_kwargs( + File "/usr/local/corex/lib/python3/dist-packages/vllm/executor/gpu_executor.py", line 24, in create_worker + wrapper.init_worker(**kwargs) + File "/usr/local/corex/lib/python3/dist-packages/vllm/worker/worker_base.py", line 449, in init_worker + self.worker = worker_class(*args, **kwargs) + File "/usr/local/corex/lib/python3/dist-packages/vllm/worker/multi_step_worker.py", line 28, in __init__ + self.model_runner = MultiStepModelRunner( + File "/usr/local/corex/lib/python3/dist-packages/vllm/worker/multi_step_model_runner.py", line 317, in __init__ + raise ValueError( +ValueError: Multi-Step not supported for attention backend: xformers. Set VLLM_ATTENTION_BACKEND to a value from ['flash-attn', 'rocm-flash-attn', 'flashinfer']. +ERROR 07-14 09:23:46 multiproc_worker_utils.py:117] Worker VllmWorkerProcess pid 6686 died, exit code: -15 +ERROR 07-14 09:23:46 multiproc_worker_utils.py:117] Worker VllmWorkerProcess pid 6687 died, exit code: -15 +ERROR 07-14 09:23:46 multiproc_worker_utils.py:117] Worker VllmWorkerProcess pid 6688 died, exit code: -15 +INFO 07-14 09:23:46 multiproc_worker_utils.py:121] Killing local vLLM worker processes