search

vllm-project / vllm

A high-throughput and memory-efficient inference and serving engine for LLMs
https://docs.vllm.ai
Apache License 2.0
30.71k stars 4.66k forks source link

[Usage]: Benchmarking Issues: Low Success Rate and Tensor Parallel Size Constraints on 8x AMD MI300x GPUs #9070

Open Bihan opened 1 month ago

Bihan commented 1 month ago

Your current environment

The output of `python collect_env.py`
Collecting environment information...
WARNING 10-04 10:39:09 rocm.py:13] `fork` method is not supported by ROCm. VLLM_WORKER_MULTIPROC_METHOD is overridden to `spawn` instead.
WARNING 10-04 10:39:09 _custom_ops.py:18] Failed to import from vllm._C with ModuleNotFoundError("No module named 'vllm._C'")
WARNING 10-04 10:39:09 _custom_ops.py:18] Failed to import from vllm._C with ModuleNotFoundError("No module named 'vllm._C'")
PyTorch version: 2.4.1+rocm6.1
Is debug build: False
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: 6.1.40091-a8dbc0c19

OS: Ubuntu 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: 17.0.0 (https://github.com/RadeonOpenCompute/llvm-project roc-6.1.0 24103 7db7f5e49612030319346f900c08f474b1f9023a)
CMake version: version 3.26.4
Libc version: glibc-2.35

Python version: 3.10.14 (main, Mar 21 2024, 16:24:04) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-6.8.0-45-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: AMD Instinct MI300X (gfx942:sramecc+:xnack-)
Nvidia driver version: Could not collect
cuDNN version: Could not collect
HIP runtime version: 6.1.40093
MIOpen runtime version: 3.1.0
Is XNNPACK available: True

CPU:
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        46 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               208
On-line CPU(s) list:                  0-207
Vendor ID:                            GenuineIntel
Model name:                           Intel(R) Xeon(R) Platinum 8470
CPU family:                           6
Model:                                143
Thread(s) per core:                   2
Core(s) per socket:                   52
Socket(s):                            2
Stepping:                             8
CPU max MHz:                          3800.0000
CPU min MHz:                          800.0000
BogoMIPS:                             4000.00
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect user_shstk avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts vnmi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr ibt amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization:                       VT-x
L1d cache:                            4.9 MiB (104 instances)
L1i cache:                            3.3 MiB (104 instances)
L2 cache:                             208 MiB (104 instances)
L3 cache:                             210 MiB (2 instances)
NUMA node(s):                         2
NUMA node0 CPU(s):                    0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62,64,66,68,70,72,74,76,78,80,82,84,86,88,90,92,94,96,98,100,102,104,106,108,110,112,114,116,118,120,122,124,126,128,130,132,134,136,138,140,142,144,146,148,150,152,154,156,158,160,162,164,166,168,170,172,174,176,178,180,182,184,186,188,190,192,194,196,198,200,202,204,206
NUMA node1 CPU(s):                    1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63,65,67,69,71,73,75,77,79,81,83,85,87,89,91,93,95,97,99,101,103,105,107,109,111,113,115,117,119,121,123,125,127,129,131,133,135,137,139,141,143,145,147,149,151,153,155,157,159,161,163,165,167,169,171,173,175,177,179,181,183,185,187,189,191,193,195,197,199,201,203,205,207
Vulnerability Gather data sampling:   Not affected
Vulnerability Itlb multihit:          Not affected
Vulnerability L1tf:                   Not affected
Vulnerability Mds:                    Not affected
Vulnerability Meltdown:               Not affected
Vulnerability Mmio stale data:        Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed:               Not affected
Vulnerability Spec rstack overflow:   Not affected
Vulnerability Spec store bypass:      Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:             Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:             Mitigation; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI BHI_DIS_S
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] mypy==1.4.1
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.4
[pip3] pytorch-triton-rocm==3.0.0
[pip3] pyzmq==24.0.1
[pip3] torch==2.4.1+rocm6.1
[pip3] torchaudio==2.4.1+rocm6.1
[pip3] torchvision==0.16.1+fdea156
[pip3] transformers==4.45.1
[pip3] triton==3.0.0
[conda] No relevant packages
ROCM Version: 6.1.40091-a8dbc0c19
Neuron SDK Version: N/A
vLLM Version: N/A
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
============================ ROCm System Management Interface ============================
================================ Weight between two GPUs =================================
       GPU0         GPU1         GPU2         GPU3         GPU4         GPU5         GPU6         GPU7         
GPU0   0            15           15           15           15           15           15           15           
GPU1   15           0            15           15           15           15           15           15           
GPU2   15           15           0            15           15           15           15           15           
GPU3   15           15           15           0            15           15           15           15           
GPU4   15           15           15           15           0            15           15           15           
GPU5   15           15           15           15           15           0            15           15           
GPU6   15           15           15           15           15           15           0            15           
GPU7   15           15           15           15           15           15           15           0            

================================= Hops between two GPUs ==================================
       GPU0         GPU1         GPU2         GPU3         GPU4         GPU5         GPU6         GPU7         
GPU0   0            1            1            1            1            1            1            1            
GPU1   1            0            1            1            1            1            1            1            
GPU2   1            1            0            1            1            1            1            1            
GPU3   1            1            1            0            1            1            1            1            
GPU4   1            1            1            1            0            1            1            1            
GPU5   1            1            1            1            1            0            1            1            
GPU6   1            1            1            1            1            1            0            1            
GPU7   1            1            1            1            1            1            1            0            

=============================== Link Type between two GPUs ===============================
       GPU0         GPU1         GPU2         GPU3         GPU4         GPU5         GPU6         GPU7         
GPU0   0            XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         
GPU1   XGMI         0            XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         
GPU2   XGMI         XGMI         0            XGMI         XGMI         XGMI         XGMI         XGMI         
GPU3   XGMI         XGMI         XGMI         0            XGMI         XGMI         XGMI         XGMI         
GPU4   XGMI         XGMI         XGMI         XGMI         0            XGMI         XGMI         XGMI         
GPU5   XGMI         XGMI         XGMI         XGMI         XGMI         0            XGMI         XGMI         
GPU6   XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         0            XGMI         
GPU7   XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         0            

======================================= Numa Nodes =======================================
GPU[0]          : (Topology) Numa Node: 0
GPU[0]          : (Topology) Numa Affinity: 0
GPU[1]          : (Topology) Numa Node: 0
GPU[1]          : (Topology) Numa Affinity: 0
GPU[2]          : (Topology) Numa Node: 0
GPU[2]          : (Topology) Numa Affinity: 0
GPU[3]          : (Topology) Numa Node: 0
GPU[3]          : (Topology) Numa Affinity: 0
GPU[4]          : (Topology) Numa Node: 1
GPU[4]          : (Topology) Numa Affinity: 1
GPU[5]          : (Topology) Numa Node: 1
GPU[5]          : (Topology) Numa Affinity: 1
GPU[6]          : (Topology) Numa Node: 1
GPU[6]          : (Topology) Numa Affinity: 1
GPU[7]          : (Topology) Numa Node: 1
GPU[7]          : (Topology) Numa Affinity: 1
================================== End of ROCm SMI Log ===================================

How would you like to use vllm

I am currently working on creating benchmarking metrics using 8 x MI300x GPUs and have encountered a couple of issues while using your framework. I wanted to bring these to your attention and seek your guidance.

Issue 1: Low Number of Successful Requests During Benchmarking

Configuration used to run the server:

ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m vllm.entrypoints.openai.api_server \
  --model=meta-llama/Llama-3.1-8B-Instruct \
  --tensor-parallel-size=8 \
  --dtype=float16 \
  --disable-log-requests \
  --disable-frontend-multiprocessing

Benchmarking configuration used:

python benchmark_serving.py \
  --backend vllm \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --dataset-name sharegpt \
  --dataset-path "/root/data/ShareGPT_V3_unfiltered_cleaned_split.json" \
  --num-prompt=4096

Bench Marking Output

Traffic request rate: inf
100%|█████████████████████████████████████████| 4096/4096 [01:27<00:00, 46.61it/s]
============ Serving Benchmark Result ============
Successful requests:                     979       
Benchmark duration (s):                  87.88     
Total input tokens:                      208599    
Total generated tokens:                  193672    
Request throughput (req/s):              11.14     
Output token throughput (tok/s):         2203.92   
Total Token throughput (tok/s):          4577.70   
---------------Time to First Token----------------
Mean TTFT (ms):                          32684.85  
Median TTFT (ms):                        28503.74  
P99 TTFT (ms):                           70015.18  
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          83.84     
Median TPOT (ms):                        87.04     
P99 TPOT (ms):                           147.36    
---------------Inter-token Latency----------------
Mean ITL (ms):                           80.93     
Median ITL (ms):                         76.72     
P99 ITL (ms):                            441.34    
==================================================

Out of 4,096 requests, only 979 were successful. I wonder whether this low success rate is due to the model's inability to respond to all requests or if it's an issue with the vLLM inference server. Could you please advise on potential causes or configurations that might improve the success rate?

Issue 2: Tensor Parallel Sizes Other Than 1 and 8 Not Working

When using tensor-parallel-size=1 or tensor-parallel-size=8, the server operates as expected.

With tensor-parallel-size=8:

ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m vllm.entrypoints.openai.api_server \
  --model=meta-llama/Llama-3.1-8B-Instruct \
  --tensor-parallel-size=8 \
  --dtype=float16 \
  --disable-log-requests \
  --disable-frontend-multiprocessing

GPU Utilization (rocm-smi --showuse):

============================ ROCm System Management Interface ============================
=================================== % time GPU is busy ===================================
GPU[0]: GPU use (%): 44
GPU[1]: GPU use (%): 39
GPU[2]: GPU use (%): 38
GPU[3]: GPU use (%): 38
GPU[4]: GPU use (%): 37
GPU[5]: GPU use (%): 37
GPU[6]: GPU use (%): 37
GPU[7]: GPU use (%): 37

With tensor-parallel-size=1:

ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m vllm.entrypoints.openai.api_server \
  --model=meta-llama/Llama-3.1-8B-Instruct \
  --tensor-parallel-size=1 \
  --dtype=float16 \
  --disable-log-requests \
  --disable-frontend-multiprocessing

GPU Utilization (rocm-smi --showuse):

============================ ROCm System Management Interface ============================
=================================== % time GPU is busy ===================================
GPU[0]: GPU use (%): 87
GPU[1]: GPU use (%): 0
GPU[2]: GPU use (%): 0
GPU[3]: GPU use (%): 0
GPU[4]: GPU use (%): 0
GPU[5]: GPU use (%): 0
GPU[6]: GPU use (%): 0
GPU[7]: GPU use (%): 0

However, when attempting to use tensor-parallel-size=2, 4, or 6, the server fails to operate properly.

Is there a specific reason why only tensor parallel sizes of 1 and 8 are functioning? Are additional configurations required for other tensor parallel sizes?

Before submitting a new issue...

kliuae commented 1 month ago

It is possible that issue 1 was caused by your system running out of file handles for 4096 http requests. If it has not been configured otherwise, the default number of open files on a linux system is 1024, which is very close to the 979 successful requests you had. Could you help confirm if the errors reported by the failed requests are on the lines of OSError 24 and check the number of open files in your system (ulimit -n)? If possible you can also try to increase the open files limit to some higher values (e.g. ulimit -n 8192) and see if this would help the rest of the requests to come through.

For issue 2, running with TP=1,2,4,8 all work fine for me. I'm on 8x MI250 GCDs. Could you share the error logs of the server failure you saw?

Bihan commented 1 month ago

@kliuae Regarding the issue 2, below is the error log

root@ENC1-CLS01-SVR13:/workflow/vllm# ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m vllm.entrypoints.openai.api_server   --model=meta-llama/Llama-3.1-8B-Instruct   --tensor-parallel-size=2   --dtype=float16   --disable-log-requests   --disable-frontend-multiprocessing
WARNING 10-07 12:58:19 rocm.py:13] `fork` method is not supported by ROCm. VLLM_WORKER_MULTIPROC_METHOD is overridden to `spawn` instead.
/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torchvision-0.16.1+fdea156-py3.10-linux-x86_64.egg/torchvision/io/image.py:13: UserWarning: Failed to load image Python extension: '/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torchvision-0.16.1+fdea156-py3.10-linux-x86_64.egg/torchvision/image.so: undefined symbol: _ZN3c1017RegisterOperatorsD1Ev'If you don't plan on using image functionality from `torchvision.io`, you can ignore this warning. Otherwise, there might be something wrong with your environment. Did you have `libjpeg` or `libpng` installed before building `torchvision` from source?
  warn(
INFO 10-07 12:58:22 api_server.py:527] vLLM API server version 0.6.3.dev116+g151ef4ef
INFO 10-07 12:58:22 api_server.py:528] args: Namespace(host=None, port=8000, 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=False, tool_call_parser=None, tool_parser_plugin='', model='meta-llama/Llama-3.1-8B-Instruct', tokenizer=None, skip_tokenizer_init=False, revision=None, code_revision=None, tokenizer_revision=None, tokenizer_mode='auto', trust_remote_code=False, download_dir=None, load_format='auto', config_format='auto', dtype='float16', kv_cache_dtype='auto', quantization_param_path=None, max_model_len=None, guided_decoding_backend='outlines', distributed_executor_backend=None, worker_use_ray=False, pipeline_parallel_size=1, tensor_parallel_size=2, 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=True, 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=False, quantization=None, rope_scaling=None, rope_theta=None, enforce_eager=False, 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, 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=1, 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=None, 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)
WARNING 10-07 12:58:22 config.py:1646] Casting torch.bfloat16 to torch.float16.
INFO 10-07 12:58:33 config.py:875] Defaulting to use mp for distributed inference
INFO 10-07 12:58:33 config.py:904] Disabled the custom all-reduce kernel because it is not supported on AMD GPUs.
WARNING 10-07 12:58:33 arg_utils.py:954] Chunked prefill is enabled by default for models with max_model_len > 32K. Currently, chunked prefill might not work with some features or models. If you encounter any issues, please disable chunked prefill by setting --enable-chunked-prefill=False.
INFO 10-07 12:58:33 config.py:993] Chunked prefill is enabled with max_num_batched_tokens=512.
INFO 10-07 12:58:33 llm_engine.py:237] Initializing an LLM engine (v0.6.3.dev116+g151ef4ef) with config: model='meta-llama/Llama-3.1-8B-Instruct', speculative_config=None, tokenizer='meta-llama/Llama-3.1-8B-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=False, dtype=torch.float16, max_seq_len=131072, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=2, pipeline_parallel_size=1, disable_custom_all_reduce=True, quantization=None, enforce_eager=False, 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=meta-llama/Llama-3.1-8B-Instruct, use_v2_block_manager=True, num_scheduler_steps=1, chunked_prefill_enabled=True multi_step_stream_outputs=True, enable_prefix_caching=False, use_async_output_proc=True, use_cached_outputs=False, mm_processor_kwargs=None)
WARNING 10-07 12:58:34 multiproc_gpu_executor.py:53] Reducing Torch parallelism from 104 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
INFO 10-07 12:58:34 custom_cache_manager.py:17] Setting Triton cache manager to: vllm.triton_utils.custom_cache_manager:CustomCacheManager
INFO 10-07 12:58:34 selector.py:121] Using ROCmFlashAttention backend.
/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torchvision-0.16.1+fdea156-py3.10-linux-x86_64.egg/torchvision/io/image.py:13: UserWarning: Failed to load image Python extension: '/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torchvision-0.16.1+fdea156-py3.10-linux-x86_64.egg/torchvision/image.so: undefined symbol: _ZN3c1017RegisterOperatorsD1Ev'If you don't plan on using image functionality from `torchvision.io`, you can ignore this warning. Otherwise, there might be something wrong with your environment. Did you have `libjpeg` or `libpng` installed before building `torchvision` from source?
  warn(
(VllmWorkerProcess pid=6924) INFO 10-07 12:58:38 selector.py:121] Using ROCmFlashAttention backend.
(VllmWorkerProcess pid=6924) INFO 10-07 12:58:38 multiproc_worker_utils.py:216] Worker ready; awaiting tasks
(VllmWorkerProcess pid=6924) INFO 10-07 12:58:38 utils.py:1005] Found nccl from library librccl.so.1
(VllmWorkerProcess pid=6924) INFO 10-07 12:58:38 pynccl.py:63] vLLM is using nccl==2.18.6
INFO 10-07 12:58:38 utils.py:1005] Found nccl from library librccl.so.1
INFO 10-07 12:58:38 pynccl.py:63] vLLM is using nccl==2.18.6
INFO 10-07 12:58:38 shm_broadcast.py:241] vLLM message queue communication handle: Handle(connect_ip='127.0.0.1', local_reader_ranks=[1], buffer=<vllm.distributed.device_communicators.shm_broadcast.ShmRingBuffer object at 0x72f84cd84d60>, local_subscribe_port=44365, remote_subscribe_port=None)
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231] Exception in worker VllmWorkerProcess while processing method init_device: tuple index out of range, Traceback (most recent call last):
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]   File "/workflow/vllm/vllm/executor/multiproc_worker_utils.py", line 224, in _run_worker_process
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]     output = executor(*args, **kwargs)
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]   File "/workflow/vllm/vllm/worker/worker.py", line 180, in init_device
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]     set_random_seed(self.model_config.seed)
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]   File "/workflow/vllm/vllm/model_executor/utils.py", line 10, in set_random_seed
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]     seed_everything(seed)
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]   File "/workflow/vllm/vllm/utils.py", line 393, in seed_everything
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]     torch.cuda.manual_seed_all(seed)
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/cuda/random.py", line 127, in manual_seed_all
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]     _lazy_call(cb, seed_all=True)
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/cuda/__init__.py", line 244, in _lazy_call
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]     callable()
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/cuda/random.py", line 124, in cb
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231]     default_generator = torch.cuda.default_generators[i]
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231] IndexError: tuple index out of range
(VllmWorkerProcess pid=6924) ERROR 10-07 12:58:38 multiproc_worker_utils.py:231] 
[rank0]: Traceback (most recent call last):
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 196, in _run_module_as_main
[rank0]:     return _run_code(code, main_globals, None,
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 86, in _run_code
[rank0]:     exec(code, run_globals)
[rank0]:   File "/workflow/vllm/vllm/entrypoints/openai/api_server.py", line 581, in <module>
[rank0]:     uvloop.run(run_server(args))
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/uvloop/__init__.py", line 82, in run
[rank0]:     return loop.run_until_complete(wrapper())
[rank0]:   File "uvloop/loop.pyx", line 1517, in uvloop.loop.Loop.run_until_complete
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/uvloop/__init__.py", line 61, in wrapper
[rank0]:     return await main
[rank0]:   File "/workflow/vllm/vllm/entrypoints/openai/api_server.py", line 548, in run_server
[rank0]:     async with build_async_engine_client(args) as engine_client:
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/contextlib.py", line 199, in __aenter__
[rank0]:     return await anext(self.gen)
[rank0]:   File "/workflow/vllm/vllm/entrypoints/openai/api_server.py", line 106, in build_async_engine_client
[rank0]:     async with build_async_engine_client_from_engine_args(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/contextlib.py", line 199, in __aenter__
[rank0]:     return await anext(self.gen)
[rank0]:   File "/workflow/vllm/vllm/entrypoints/openai/api_server.py", line 140, in build_async_engine_client_from_engine_args
[rank0]:     engine_client = await asyncio.get_running_loop().run_in_executor(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/concurrent/futures/thread.py", line 58, in run
[rank0]:     result = self.fn(*self.args, **self.kwargs)
[rank0]:   File "/workflow/vllm/vllm/engine/async_llm_engine.py", line 674, in from_engine_args
[rank0]:     engine = cls(
[rank0]:   File "/workflow/vllm/vllm/engine/async_llm_engine.py", line 569, in __init__
[rank0]:     self.engine = self._engine_class(*args, **kwargs)
[rank0]:   File "/workflow/vllm/vllm/engine/async_llm_engine.py", line 265, in __init__
[rank0]:     super().__init__(*args, **kwargs)
[rank0]:   File "/workflow/vllm/vllm/engine/llm_engine.py", line 335, in __init__
[rank0]:     self.model_executor = executor_class(
[rank0]:   File "/workflow/vllm/vllm/executor/multiproc_gpu_executor.py", line 215, in __init__
[rank0]:     super().__init__(*args, **kwargs)
[rank0]:   File "/workflow/vllm/vllm/executor/distributed_gpu_executor.py", line 26, in __init__
[rank0]:     super().__init__(*args, **kwargs)
[rank0]:   File "/workflow/vllm/vllm/executor/executor_base.py", line 47, in __init__
[rank0]:     self._init_executor()
[rank0]:   File "/workflow/vllm/vllm/executor/multiproc_gpu_executor.py", line 110, in _init_executor
[rank0]:     self._run_workers("init_device")
[rank0]:   File "/workflow/vllm/vllm/executor/multiproc_gpu_executor.py", line 196, in _run_workers
[rank0]:     ] + [output.get() for output in worker_outputs]
[rank0]:   File "/workflow/vllm/vllm/executor/multiproc_gpu_executor.py", line 196, in <listcomp>
[rank0]:     ] + [output.get() for output in worker_outputs]
[rank0]:   File "/workflow/vllm/vllm/executor/multiproc_worker_utils.py", line 55, in get
[rank0]:     raise self.result.exception
[rank0]: IndexError: tuple index out of range
root@ENC1-CLS01-SVR13:/workflow/vllm# /opt/conda/envs/py_3.10/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 '