qwen2-vl-72b-instruct and qwen-vl-72b-instruct on ali cloud L20， Vllm frame can`t start server incorrect

[xinzaifeixiang1992]

I met following questions:
INFO 09-22 21:48:03 api_server.py:495] vLLM API server version 0.6.1
INFO 09-22 21:48:03 api_server.py:496] args: Namespace(host='0.0.0.0', port=40116, 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=False, enable_auto_tool_choice=False, tool_call_parser=None, model='/data/llms/qwen/Qwen2-VL-72B-Instruct', 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='float16', kv_cache_dtype='auto', quantization_param_path=None, max_model_len=20000, guided_decoding_backend='outlines', distributed_executor_backend=None, worker_use_ray=False, pipeline_parallel_size=1, tensor_parallel_size=8, max_parallel_loading_workers=None, ray_workers_use_nsight=False, block_size=16, enable_prefix_caching=False, disable_sliding_window=False, use_v2_block_manager=False, num_lookahead_slots=0, seed=0, swap_space=4, cpu_offload_gb=0, gpu_memory_utilization=0.9, num_gpu_blocks_override=None, max_num_batched_tokens=None, max_num_seqs=256, max_logprobs=20, disable_log_stats=True, quantization=None, rope_scaling=None, rope_theta=None, enforce_eager=True, max_context_len_to_capture=None, max_seq_len_to_capture=8192, disable_custom_all_reduce=False, tokenizer_pool_size=0, tokenizer_pool_type='ray', tokenizer_pool_extra_config=None, limit_mm_per_prompt=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=1, scheduler_delay_factor=0.0, enable_chunked_prefill=None, speculative_model=None, speculative_model_quantization=None, num_speculative_tokens=None, 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=['Qwen2-VL-72B-Instruct'], qlora_adapter_name_or_path=None, otlp_traces_endpoint=None, collect_detailed_traces=None, disable_async_output_proc=False, override_neuron_config=None, engine_use_ray=False, disable_log_requests=False, max_log_len=None)
INFO 09-22 21:48:03 api_server.py:162] Multiprocessing frontend to use ipc:///tmp/fe6dac8c-587f-49d9-9547-ae65fa9976a0 for RPC Path.
INFO 09-22 21:48:03 api_server.py:178] Started engine process with PID 115210
WARNING 09-22 21:48:06 config.py:1650] Casting torch.bfloat16 to torch.float16.
INFO 09-22 21:48:06 config.py:897] Defaulting to use mp for distributed inference
WARNING 09-22 21:48:06 config.py:383] To see benefits of async output processing, enable CUDA graph. Since, enforce-eager is enabled, async output processor cannot be used
INFO 09-22 21:48:06 llm_engine.py:232] Initializing an LLM engine (v0.6.1) with config: model='/data/llms/qwen/Qwen2-VL-72B-Instruct', speculative_config=None, tokenizer='/data/llms/qwen/Qwen2-VL-72B-Instruct', 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=20000, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=8, pipeline_parallel_size=1, disable_custom_all_reduce=False, 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=Qwen2-VL-72B-Instruct, use_v2_block_manager=False, num_scheduler_steps=1, enable_prefix_caching=False, use_async_output_proc=False)
WARNING 09-22 21:48:06 multiproc_gpu_executor.py:56] Reducing Torch parallelism from 64 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
INFO 09-22 21:48:06 custom_cache_manager.py:17] Setting Triton cache manager to: vllm.triton_utils.custom_cache_manager:CustomCacheManager
(VllmWorkerProcess pid=115364) INFO 09-22 21:48:06 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
(VllmWorkerProcess pid=115365) INFO 09-22 21:48:06 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
(VllmWorkerProcess pid=115364) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] Exception in worker VllmWorkerProcess while processing method init_device: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method, Traceback (most recent call last):
(VllmWorkerProcess pid=115364) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/executor/multiproc_worker_utils.py", line 223, in _run_worker_process
(VllmWorkerProcess pid=115364) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] output = executor(*args, **kwargs)
(VllmWorkerProcess pid=115364) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/worker/worker.py", line 166, in init_device
(VllmWorkerProcess pid=115364) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch.cuda.set_device(self.device)
(VllmWorkerProcess pid=115364) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 420, in set_device
(VllmWorkerProcess pid=115364) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch._C._cuda_setDevice(device)
(VllmWorkerProcess pid=115364) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 300, in _lazy_init
(VllmWorkerProcess pid=115364) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] raise RuntimeError(
(VllmWorkerProcess pid=115364) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method
(VllmWorkerProcess pid=115364) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226]
(VllmWorkerProcess pid=115365) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] Exception in worker VllmWorkerProcess while processing method init_device: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method, Traceback (most recent call last):
(VllmWorkerProcess pid=115365) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/executor/multiproc_worker_utils.py", line 223, in _run_worker_process
(VllmWorkerProcess pid=115365) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] output = executor(*args, **kwargs)
(VllmWorkerProcess pid=115365) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/worker/worker.py", line 166, in init_device
(VllmWorkerProcess pid=115365) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch.cuda.set_device(self.device)
(VllmWorkerProcess pid=115365) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 420, in set_device
(VllmWorkerProcess pid=115365) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch._C._cuda_setDevice(device)
(VllmWorkerProcess pid=115365) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 300, in _lazy_init
(VllmWorkerProcess pid=115365) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] raise RuntimeError(
(VllmWorkerProcess pid=115365) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method
(VllmWorkerProcess pid=115365) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226]
(VllmWorkerProcess pid=115366) INFO 09-22 21:48:06 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
(VllmWorkerProcess pid=115366) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] Exception in worker VllmWorkerProcess while processing method init_device: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method, Traceback (most recent call last):
(VllmWorkerProcess pid=115366) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/executor/multiproc_worker_utils.py", line 223, in _run_worker_process
(VllmWorkerProcess pid=115366) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] output = executor(*args, **kwargs)
(VllmWorkerProcess pid=115366) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/worker/worker.py", line 166, in init_device
(VllmWorkerProcess pid=115366) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch.cuda.set_device(self.device)
(VllmWorkerProcess pid=115366) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 420, in set_device
(VllmWorkerProcess pid=115366) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch._C._cuda_setDevice(device)
(VllmWorkerProcess pid=115366) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 300, in _lazy_init
(VllmWorkerProcess pid=115366) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] raise RuntimeError(
(VllmWorkerProcess pid=115366) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method
(VllmWorkerProcess pid=115366) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226]
(VllmWorkerProcess pid=115367) INFO 09-22 21:48:06 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
(VllmWorkerProcess pid=115367) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] Exception in worker VllmWorkerProcess while processing method init_device: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method, Traceback (most recent call last):
(VllmWorkerProcess pid=115367) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/executor/multiproc_worker_utils.py", line 223, in _run_worker_process
(VllmWorkerProcess pid=115367) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] output = executor(*args, **kwargs)
(VllmWorkerProcess pid=115367) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/worker/worker.py", line 166, in init_device
(VllmWorkerProcess pid=115367) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch.cuda.set_device(self.device)
(VllmWorkerProcess pid=115367) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 420, in set_device
(VllmWorkerProcess pid=115367) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch._C._cuda_setDevice(device)
(VllmWorkerProcess pid=115367) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 300, in _lazy_init
(VllmWorkerProcess pid=115367) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] raise RuntimeError(
(VllmWorkerProcess pid=115367) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method
(VllmWorkerProcess pid=115367) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226]
(VllmWorkerProcess pid=115368) INFO 09-22 21:48:06 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
(VllmWorkerProcess pid=115368) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] Exception in worker VllmWorkerProcess while processing method init_device: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method, Traceback (most recent call last):
(VllmWorkerProcess pid=115368) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/executor/multiproc_worker_utils.py", line 223, in _run_worker_process
(VllmWorkerProcess pid=115368) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] output = executor(*args, **kwargs)
(VllmWorkerProcess pid=115368) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/worker/worker.py", line 166, in init_device
(VllmWorkerProcess pid=115368) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch.cuda.set_device(self.device)
(VllmWorkerProcess pid=115368) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 420, in set_device
(VllmWorkerProcess pid=115368) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch._C._cuda_setDevice(device)
(VllmWorkerProcess pid=115368) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 300, in _lazy_init
(VllmWorkerProcess pid=115368) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] raise RuntimeError(
(VllmWorkerProcess pid=115368) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method
(VllmWorkerProcess pid=115368) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226]
(VllmWorkerProcess pid=115369) INFO 09-22 21:48:06 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
(VllmWorkerProcess pid=115370) INFO 09-22 21:48:06 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
(VllmWorkerProcess pid=115370) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] Exception in worker VllmWorkerProcess while processing method init_device: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method, Traceback (most recent call last):
(VllmWorkerProcess pid=115370) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/executor/multiproc_worker_utils.py", line 223, in _run_worker_process
(VllmWorkerProcess pid=115370) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] output = executor(*args, **kwargs)
(VllmWorkerProcess pid=115370) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/worker/worker.py", line 166, in init_device
(VllmWorkerProcess pid=115370) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch.cuda.set_device(self.device)
(VllmWorkerProcess pid=115370) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 420, in set_device
(VllmWorkerProcess pid=115370) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch._C._cuda_setDevice(device)
(VllmWorkerProcess pid=115370) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 300, in _lazy_init
(VllmWorkerProcess pid=115370) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] raise RuntimeError(
(VllmWorkerProcess pid=115370) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method
(VllmWorkerProcess pid=115370) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226]
(VllmWorkerProcess pid=115369) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] Exception in worker VllmWorkerProcess while processing method init_device: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method, Traceback (most recent call last):
(VllmWorkerProcess pid=115369) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/executor/multiproc_worker_utils.py", line 223, in _run_worker_process
(VllmWorkerProcess pid=115369) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] output = executor(*args, **kwargs)
(VllmWorkerProcess pid=115369) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/vllm/worker/worker.py", line 166, in init_device
(VllmWorkerProcess pid=115369) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch.cuda.set_device(self.device)
(VllmWorkerProcess pid=115369) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 420, in set_device
(VllmWorkerProcess pid=115369) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] torch._C._cuda_setDevice(device)
(VllmWorkerProcess pid=115369) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] File "/data/anaconda3/envs/qwen-vl/lib/python3.10/site-packages/torch/cuda/init.py", line 300, in _lazy_init
(VllmWorkerProcess pid=115369) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] raise RuntimeError(
(VllmWorkerProcess pid=115369) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226] RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method
(VllmWorkerProcess pid=115369) ERROR 09-22 21:48:06 multiproc_worker_utils.py:226]

---

machine config informations as followings:
一、basic env config

1. hardware config
   Ali Cloud L20 848G, actually 845G.
   NVIDIA-SMI 550.90.07 Driver Version: 550.90.07 CUDA Version: 12.4
2. python env
   python 3.10.13
3. packages config
   torch 2.4.0
   torchvision 0.19.0
   transformers 4.45.0.dev0
   vllm 0.6.1
   二、start shell
   `#!/bin/bash
   model_path="/data/llms/qwen/Qwen2-VL-72B-Instruct"
   model_server_name="Qwen2-VL-72B-Instruct"
   max_model_len=20000
   port=40116
   local_ip="0.0.0.0"

切换gcc版本到10.2

source /opt/rh/devtoolset-10/enable

激活环境

source activate /data/anaconda3/envs/qwen-vl

log_file="${model_server_name}_${port}.log"

启动服务

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m vllm.entrypoints.openai.api_server
--model ${model_path}
--trust-remote-code
--served-model-name ${model_server_name}
--enforce-eager
--dtype float16
--gpu-memory-utilization 0.9
--tensor-parallel-size 8
--host ${local_ip}
--max-model-len ${max_model_len}
--disable-log-stats
--port ${port} > "${log_file}" 2>&1 &`

Is there any great God who can help solve the problem? Thank you very much.....

[QwertyJack]

As it complains, you might need to specify VLLM_WORKER_MULTIPROC_METHOD=spawn python -m ...

After that, you probably will encounter a value error issue. See: #231

[xinzaifeixiang1992]

As it complains, you might need to specify VLLM_WORKER_MULTIPROC_METHOD=spawn python -m ...

After that, you probably will encounter a value error issue. See: #231

Okay, thanks very much!