search

vllm-project / vllm

A high-throughput and memory-efficient inference and serving engine for LLMs
https://docs.vllm.ai
Apache License 2.0
27.83k stars 4.11k forks source link

[Bug]: The speed of loading the qwen2 72b model, glm-4-9b-chat-1m model in v0.5.0 is much lower than that in v0.4.2. #5523

Closed majestichou closed 3 months ago

majestichou commented 3 months ago

Your current environment

I run python3 collect_env.py in the docker container started by vllm:0.5.0, and the output is as follows:

Collecting environment information...
PyTorch version: 2.3.0+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.29.5
Libc version: glibc-2.35

Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-3.10.0-1160.el7.x86_64-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:
GPU 0: NVIDIA A100-SXM4-40GB
GPU 1: NVIDIA A100-SXM4-40GB
GPU 2: NVIDIA A100-SXM4-40GB
GPU 3: NVIDIA A100-SXM4-40GB
GPU 4: NVIDIA A100-SXM4-40GB
GPU 5: NVIDIA A100-SXM4-40GB
GPU 6: NVIDIA A100-SXM4-40GB
GPU 7: NVIDIA A100-SXM4-40GB

Nvidia driver version: 535.86.10
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Address sizes:                   43 bits physical, 48 bits virtual
Byte Order:                      Little Endian
CPU(s):                          192
On-line CPU(s) list:             0-191
Vendor ID:                       AuthenticAMD
Model name:                      AMD EPYC 7642 48-Core Processor
CPU family:                      23
Model:                           49
Thread(s) per core:              2
Core(s) per socket:              48
Socket(s):                       2
Stepping:                        0
Frequency boost:                 enabled
CPU max MHz:                     2300.0000
CPU min MHz:                     1500.0000
BogoMIPS:                        4599.72
Flags:                           fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc art rep_good nopl nonstop_tsc extd_apicid aperfmperf eagerfpu pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_l2 cpb cat_l3 cdp_l3 hw_pstate sme retpoline_amd ssbd ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif umip overflow_recov succor smca
Virtualization:                  AMD-V
L1d cache:                       3 MiB (96 instances)
L1i cache:                       3 MiB (96 instances)
L2 cache:                        48 MiB (96 instances)
L3 cache:                        512 MiB (32 instances)
NUMA node(s):                    2
NUMA node0 CPU(s):               0-47,96-143
NUMA node1 CPU(s):               48-95,144-191
Vulnerability Itlb multihit:     Not affected
Vulnerability L1tf:              Not affected
Vulnerability Mds:               Not affected
Vulnerability Meltdown:          Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:        Mitigation; Load fences, usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:        Mitigation; Full retpoline, IBPB
Vulnerability Srbds:             Not affected
Vulnerability Tsx async abort:   Not affected

Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] nvidia-nccl-cu12==2.20.5
[pip3] torch==2.3.0
[pip3] triton==2.3.0
[conda] Could not collectROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.5.0
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    GPU1    GPU2    GPU3    GPU4    GPU5    GPU6    GPU7    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      NV12    NV12    NV12    NV12    NV12    NV12    NV12    0-47,96-143     0               N/A
GPU1    NV12     X      NV12    NV12    NV12    NV12    NV12    NV12    0-47,96-143     0               N/A
GPU2    NV12    NV12     X      NV12    NV12    NV12    NV12    NV12    0-47,96-143     0               N/A
GPU3    NV12    NV12    NV12     X      NV12    NV12    NV12    NV12    0-47,96-143     0               N/A
GPU4    NV12    NV12    NV12    NV12     X      NV12    NV12    NV12    48-95,144-191   1               N/A
GPU5    NV12    NV12    NV12    NV12    NV12     X      NV12    NV12    48-95,144-191   1               N/A
GPU6    NV12    NV12    NV12    NV12    NV12    NV12     X      NV12    48-95,144-191   1               N/A
GPU7    NV12    NV12    NV12    NV12    NV12    NV12    NV12     X      48-95,144-191   1               N/A

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

🐛 Describe the bug

I use vLLM (v0.5.0 docker image) to load Qwen/Qwen2-72B-Instruct model. I download the model weights to a local directory named /serving/data/models/. Run the following command to start the Docker container:

docker run --gpus 8 -p 8080:8000 --ipc=host -v /serving/data/models/:/data/models/ --env "TRANSFORMERS_OFFLINE=1" --env "HF_DATASET_OFFLINE=1" vllm/vllm-openai:v0.5.0 --model "/data/models/Qwen2-72B-Instruct/" --tensor-parallel-size 8

It took 40 minutes to start the model with vllm version 0.5.0((v0.5.0 docker image). During the boot process, I use nvitop to check the gpu memory usage of the machine. The usage of the display memory stays at 5.4% for a long time. The GPU memory usage stays at 5.4% for a long time, sometimes increases to 5.6%, and then decreases to 5.4%. Sometimes, the message "zombile process" is displayed on the screen of the nvitop. I used the https://docs.vllm.ai/en/stable/getting_started/debugging.html link to troubleshoot, everthing is ok, I didn't find any strange problems.

The docker log is as follows:

INFO 06-14 10:21:15 api_server.py:177] vLLM API server version 0.5.0
INFO 06-14 10:21:15 api_server.py:178] 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, chat_template=None, response_role='assistant', ssl_keyfile=None, ssl_certfile=None, ssl_ca_certs=None, ssl_cert_reqs=0, root_path=None, middleware=[], model='/data/models/Qwen2-72B-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', dtype='auto', 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=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, 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, 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, device='auto', image_input_type=None, image_token_id=None, image_input_shape=None, image_feature_size=None, image_processor=None, image_processor_revision=None, disable_image_processor=False, scheduler_delay_factor=0.0, enable_chunked_prefill=False, speculative_model=None, num_speculative_tokens=None, speculative_max_model_len=None, speculative_disable_by_batch_size=None, ngram_prompt_lookup_max=None, ngram_prompt_lookup_min=None, model_loader_extra_config=None, preemption_mode=None, served_model_name=None, qlora_adapter_name_or_path=None, engine_use_ray=False, disable_log_requests=False, max_log_len=None)
2024-06-14 10:21:19,072 INFO worker.py:1753 -- Started a local Ray instance.
INFO 06-14 10:21:29 config.py:623] Defaulting to use mp for distributed inference
INFO 06-14 10:21:29 llm_engine.py:161] Initializing an LLM engine (v0.5.0) with config: model='/data/models/Qwen2-72B-Instruct/', speculative_config=None, tokenizer='/data/models/Qwen2-72B-Instruct/', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, rope_scaling=None, rope_theta=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=8, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto, quantization_param_path=None, device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='outlines'), seed=0, served_model_name=/data/models/Qwen2-72B-Instruct/)
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
INFO 06-14 10:21:30 selector.py:50] Using XFormers backend.
(VllmWorkerProcess pid=10874) INFO 06-14 10:21:45 selector.py:50] Using XFormers backend.
(VllmWorkerProcess pid=10872) INFO 06-14 10:21:45 selector.py:50] Using XFormers backend.
(VllmWorkerProcess pid=10873) INFO 06-14 10:21:45 selector.py:50] Using XFormers backend.
(VllmWorkerProcess pid=10870) INFO 06-14 10:21:45 selector.py:50] Using XFormers backend.
(VllmWorkerProcess pid=10875) INFO 06-14 10:21:46 selector.py:50] Using XFormers backend.
(VllmWorkerProcess pid=10869) INFO 06-14 10:21:46 selector.py:50] Using XFormers backend.
(VllmWorkerProcess pid=10871) INFO 06-14 10:21:46 selector.py:50] Using XFormers backend.
(VllmWorkerProcess pid=10874) INFO 06-14 10:21:49 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
(VllmWorkerProcess pid=10872) INFO 06-14 10:21:49 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
(VllmWorkerProcess pid=10873) INFO 06-14 10:21:50 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
(VllmWorkerProcess pid=10870) INFO 06-14 10:21:50 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
(VllmWorkerProcess pid=10875) INFO 06-14 10:21:50 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
(VllmWorkerProcess pid=10869) INFO 06-14 10:21:50 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
(VllmWorkerProcess pid=10871) INFO 06-14 10:21:50 multiproc_worker_utils.py:215] Worker ready; awaiting tasks
INFO 06-14 10:21:54 utils.py:623] Found nccl from library libnccl.so.2
INFO 06-14 10:21:54 pynccl.py:65] vLLM is using nccl==2.20.5
(VllmWorkerProcess pid=10869) INFO 06-14 10:21:54 utils.py:623] Found nccl from library libnccl.so.2
(VllmWorkerProcess pid=10869) INFO 06-14 10:21:54 pynccl.py:65] vLLM is using nccl==2.20.5
(VllmWorkerProcess pid=10871) INFO 06-14 10:21:54 utils.py:623] Found nccl from library libnccl.so.2
(VllmWorkerProcess pid=10871) INFO 06-14 10:21:54 pynccl.py:65] vLLM is using nccl==2.20.5
(VllmWorkerProcess pid=10870) INFO 06-14 10:21:54 utils.py:623] Found nccl from library libnccl.so.2
(VllmWorkerProcess pid=10870) INFO 06-14 10:21:54 pynccl.py:65] vLLM is using nccl==2.20.5
(VllmWorkerProcess pid=10875) INFO 06-14 10:21:54 utils.py:623] Found nccl from library libnccl.so.2
(VllmWorkerProcess pid=10873) INFO 06-14 10:21:54 utils.py:623] Found nccl from library libnccl.so.2
(VllmWorkerProcess pid=10875) INFO 06-14 10:21:54 pynccl.py:65] vLLM is using nccl==2.20.5
(VllmWorkerProcess pid=10873) INFO 06-14 10:21:54 pynccl.py:65] vLLM is using nccl==2.20.5
(VllmWorkerProcess pid=10872) INFO 06-14 10:21:54 utils.py:623] Found nccl from library libnccl.so.2
(VllmWorkerProcess pid=10872) INFO 06-14 10:21:54 pynccl.py:65] vLLM is using nccl==2.20.5
(VllmWorkerProcess pid=10874) INFO 06-14 10:21:54 utils.py:623] Found nccl from library libnccl.so.2
(VllmWorkerProcess pid=10874) INFO 06-14 10:21:54 pynccl.py:65] vLLM is using nccl==2.20.5
Traceback (most recent call last):
  File "/usr/lib/python3.10/multiprocessing/resource_tracker.py", line 209, in main
    cache[rtype].remove(name)
KeyError: '/psm_d0f5318f'
Traceback (most recent call last):
  File "/usr/lib/python3.10/multiprocessing/resource_tracker.py", line 209, in main
    cache[rtype].remove(name)
KeyError: '/psm_d0f5318f'
Traceback (most recent call last):
  File "/usr/lib/python3.10/multiprocessing/resource_tracker.py", line 209, in main
    cache[rtype].remove(name)
KeyError: '/psm_d0f5318f'
Traceback (most recent call last):
  File "/usr/lib/python3.10/multiprocessing/resource_tracker.py", line 209, in main
    cache[rtype].remove(name)
KeyError: '/psm_d0f5318f'
Traceback (most recent call last):
  File "/usr/lib/python3.10/multiprocessing/resource_tracker.py", line 209, in main
    cache[rtype].remove(name)
KeyError: '/psm_d0f5318f'
Traceback (most recent call last):
  File "/usr/lib/python3.10/multiprocessing/resource_tracker.py", line 209, in main
    cache[rtype].remove(name)
KeyError: '/psm_d0f5318f'
Traceback (most recent call last):
  File "/usr/lib/python3.10/multiprocessing/resource_tracker.py", line 209, in main
    cache[rtype].remove(name)
KeyError: '/psm_d0f5318f'
INFO 06-14 10:21:58 custom_all_reduce_utils.py:169] generating GPU P2P access cache in /root/.config/vllm/gpu_p2p_access_cache_for_0,1,2,3,4,5,6,7.json

The docker logs are not updated for a long time.

Then I booted the model on the same machine with vllm version 0.4.2. docker run --gpus 8 -p 8080:8000 --ipc=host -v /serving/data/models/:/data/models/ --env "TRANSFORMERS_OFFLINE=1" --env "HF_DATASET_OFFLINE=1" vllm/vllm-openai:v0.4.2 --model "/data/models/Qwen2-72B-Instruct/" --tensor-parallel-size 8 It took 20 minutes to start the model with vllm version 0.4.2((v0.4.2 docker image).

The speed of loading the qwen2 72b model in v0.5.0 is much lower than that in v0.4.2.

I repeated the above process many times, with the same conclusion.

NiuBlibing commented 3 months ago

same issue. The "generating GPU P2P access cache" progress is very slow.

majestichou commented 3 months ago

I also tested the loading of the glm-4-9b-chat-1m model. The speed of loading the glm-4-9b-chat-1m model in v0.5.0 is also much lower than that in v0.4.2. The "generating GPU P2P access cache" progress is very slow.

NiuBlibing commented 3 months ago

The first "generating GPU P2P access cache" is very slow which is not related with specify models I think, if the file exists, it will reading from the file directly.

youkaichao commented 3 months ago

I understand "generating GPU P2P access cache" can be slow, because it needs to test the p2p of each gpu pair. I don't think it can take 20 minutes though. In my experience that's kind of 1~2 minutes.

majestichou commented 3 months ago

I understand "generating GPU P2P access cache" can be slow, because it needs to test the p2p of each gpu pair. I don't think it can take 20 minutes though. In my experience that's kind of 1~2 minutes.

can you reproduce my step? And you will see what i said.

youkaichao commented 3 months ago

Can you try https://github.com/vllm-project/vllm/pull/5528 and give some feedback? I think that should solve this issue.

majestichou commented 3 months ago

Can you try #5528 and give some feedback? I think that should solve this issue.

What should I try? I use a docker image to start the service instead of compiling the source code. Should I modify the code in the docker container of vllm version 0.5.0 as the PR #5528 ? Then restart the container. Or should I pull the latest v0.5.0.post1 docker image and load the model with the latest image?

NiuBlibing commented 3 months ago

Can you try #5528 and give some feedback? I think that should solve this issue.

What should I try? I use a docker image to start the service instead of compiling the source code. Should I modify the code in the docker container of vllm version 0.5.0 as the PR #5528 ? Then restart the container. Or should I pull the latest v0.5.0.post1 docker image and load the model with the latest image?

@youkaichao Maybe we need a ci to build nightly package for pr? Locally build and install is too boring and slow.

This is my test results:

v0.5.0: 8min50s

INFO 06-14 14:22:10 custom_all_reduce_utils.py:169] generating GPU P2P access cache in /root/.config/vllm/gpu_p2p_access_cache_for_0,1,2,3,4,5,6,7.json
INFO 06-14 14:31:00 custom_all_reduce_utils.py:179] reading GPU P2P access cache from /root/.config/vllm/gpu_p2p_access_cache_for_0,1,2,3,4,5,6,7.json

pr #5528 (I just remove the vllm config /root/.config/vllm/gpu_p2p_access_cache_for_0,1,2,3,4,5,6,7.json): 44s

INFO 06-14 18:44:29 custom_all_reduce_utils.py:184] generating GPU P2P access cache in /root/.config/vllm/gpu_p2p_access_cache_for_0,1,2,3,4,5,6,7.json
INFO 06-14 18:45:13 custom_all_reduce_utils.py:199] reading GPU P2P access cache from /root/.config/vllm/gpu_p2p_access_cache_for_0,1,2,3,4,5,6,7.json
Elissa0723 commented 3 months ago

I try to load qwen2_57b_a14b_instruct and met the same issue. I had already updated to the latest code and used os.environ['NCCL_IGNORE_DISABLED_P2P'] = '1'`

youkaichao commented 3 months ago

@NiuBlibing glad it helps, the PR should be merged recently.

Maybe we need a ci to build nightly package for pr?

It's ongoing. Since most PR only touches Python files, you can also manually replace these files to test.