[Bug] oom,torch.OutOfMemoryError: seems to only use one gpu on A800-80G,available 40g on each card

chuangzhidan commented 2 days ago

Checklist

[X] 1. I have searched related issues but cannot get the expected help.
[ ] 2. The bug has not been fixed in the latest version.
[X] 3. Please note that if the bug-related issue you submitted lacks corresponding environment info and a minimal reproducible demo, it will be challenging for us to reproduce and resolve the issue, reducing the likelihood of receiving feedback.
[X] 4. If the issue you raised is not a bug but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
[X] 5. Please use English, otherwise it will be closed.

Describe the bug

File "/usr/local/lib/python3.10/dist-packages/torch/utils/_device.py", line 79, in __torch_function__ return func(*args, **kwargs) torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.32 GiB. GPU 0 has a total capacity of 79.25 GiB of which 2.27 GiB is free. Including non-PyTorch memory, this process has 0 bytes memory in use. Of the allocated memory 36.46 GiB is allocated by PyTorch, and 20.98 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

Reproduction

docker run --gpus all -it -p 30000:30000 -v ~/.cache/huggingface:/root/.cache/huggingface -v /data/xgp:/workspace -v /data/llm:/workspace/model --env "HF_TOKEN=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" --env CUDA_VISIBLE_DEVICES=0,1 --ipc=host lmsysorg/sglang:latest python3 -m sglang.launch_server --model-path /workspace/model/Qwen2.5-72-int4 --host 0.0.0.0 --port 30000 --quantization gptq_marlin --mem-fraction-static 0.9 --disable-cuda-grap

Environment

Python: 3.12.4 | packaged by Anaconda, Inc. | (main, Jun 18 2024, 15:12:24) [GCC 11.2.0] CUDA available: True GPU 0,1: NVIDIA A800 80GB PCIe GPU 0,1 Compute Capability: 8.0 CUDA_HOME: /usr NVCC: Cuda compilation tools, release 11.5, V11.5.119 CUDA Driver Version: 560.35.03 PyTorch: 2.4.0+cu121 sglang: 0.3.0 flashinfer: Module Not Found triton: 3.0.0 transformers: 4.44.2 requests: 2.32.3 tqdm: 4.66.5 numpy: 1.26.4 aiohttp: 3.10.5 fastapi: 0.113.0 hf_transfer: Module Not Found huggingface_hub: 0.24.6 interegular: 0.3.3 packaging: 24.1 PIL: 10.4.0 psutil: 6.0.0 pydantic: 2.9.0 uvicorn: 0.30.6 uvloop: 0.20.0 zmq: 26.2.0 vllm: 0.6.0 multipart: Module Not Found openai: 1.43.1 anthropic: Module Not Found litellm: Module Not Found NVIDIA Topology: GPU0 GPU1 CPU Affinity NUMA Affinity GPU NUMA ID GPU0 X NODE 0-23,48-71 0 N/A GPU1 NODE X 0-23,48-71 0 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

ulimit soft: 1024

m0g1cian commented 2 days ago

You need to use --tp-size [#_OF_DEVICES] to enable tensor parallelization. Otherwise SGL will only use 1 GPU by default

chuangzhidan commented 2 days ago

You need to use --tp-size [#_OF_DEVICES] to enable tensor parallelization. Otherwise SGL will only use 1 GPU by default

thank you so much. it works like charm: not sure if i can do anything that can make it faster

[00:19:29] server_args=ServerArgs(model_path='/workspace/model/Qwen2.5-72-int4', tokenizer_path='/workspace/model/Qwen2.5-72-int4', tokenizer_mode='auto', skip_tokenizer_init=False, load_format='auto', dtype='auto', kv_cache_dtype='auto', trust_remote_code=False, context_length=None, quantization='gptq_marlin', served_model_name='/workspace/model/Qwen2.5-72-int4', chat_template=None, is_embedding=False, host='0.0.0.0', port=30000, additional_ports=[30001, 30002, 30003, 30004], mem_fraction_static=0.9, max_running_requests=None, max_num_reqs=None, max_total_tokens=None, chunked_prefill_size=8192, max_prefill_tokens=16384, schedule_policy='lpm', schedule_conservativeness=1.0, tp_size=2, stream_interval=1, random_seed=595017997, log_level='info', log_level_http=None, log_requests=False, show_time_cost=False, api_key=None, file_storage_pth='SGLang_storage', dp_size=1, load_balance_method='round_robin', disable_flashinfer=False, disable_flashinfer_sampling=False, disable_radix_cache=False, disable_regex_jump_forward=False, disable_cuda_graph=True, disable_cuda_graph_padding=False, disable_disk_cache=False, disable_custom_all_reduce=False, enable_mixed_chunk=False, enable_torch_compile=False, enable_p2p_check=False, enable_mla=False, triton_attention_reduce_in_fp32=False, nccl_init_addr=None, nnodes=1, node_rank=None)

[00:23:56 TP0] Load weight end. type=Qwen2ForCausalLM, dtype=torch.float16, avail mem=14.86 GB [00:23:56 TP1] Load weight end. type=Qwen2ForCausalLM, dtype=torch.float16, avail mem=18.11 GB [00:23:58 TP1] Memory pool end. avail mem=9.63 GB [00:23:58 TP0] Memory pool end. avail mem=6.32 GB [00:24:01 TP0] max_total_num_tokens=74912, max_prefill_tokens=16384, max_running_requests=2047, context_len=32768 [00:24:01 TP1] max_total_num_tokens=74912, max_prefill_tokens=16384, max_running_requests=2047, context_len=32768 INFO: Started server process [1] INFO: Waiting for application startup. INFO: Application startup complete. INFO: Uvicorn running on http://0.0.0.0:30000 (Press CTRL+C to quit) INFO: 127.0.0.1:40562 - "GET /get_model_info HTTP/1.1" 200 OK [00:24:03 TP0] Prefill batch. #new-seq: 1, #new-token: 6, #cached-token: 0, cache hit rate: 0.00%, #running-req: 0, #queue-req: 0 INFO: 127.0.0.1:40566 - "POST /generate HTTP/1.1" 200 OK [00:24:12] The server is fired up and ready to roll! INFO: 172.17.0.1:35430 - "GET /health HTTP/1.1" 200 OK [00:24:30 TP0] Prefill batch. #new-seq: 1, #new-token: 42, #cached-token: 0, cache hit rate: 0.00%, #running-req: 0, #queue-req: 0 [00:24:33 TP0] Decode batch. #running-req: 1, #token: 75, token usage: 0.00, gen throughput (token/s): 1.23, #queue-req: 0 [00:24:36 TP0] Decode batch. #running-req: 1, #token: 115, token usage: 0.00, gen throughput (token/s): 13.48, #queue-req: 0 INFO: 192.168.28.101:51561 - "POST /v1/chat/completions HTTP/1.1" 200 OK [00:24:59 TP0] Prefill batch. #new-seq: 1, #new-token: 1, #cached-token: 41, cache hit rate: 45.56%, #running-req: 0, #queue-req: 0 [00:25:00 TP0] Decode batch. #running-req: 1, #token: 56, token usage: 0.00, gen throughput (token/s): 1.65, #queue-req: 0 [00:25:06 TP0] Decode batch. #running-req: 1, #token: 96, token usage: 0.00, gen throughput (token/s): 7.54, #queue-req: 0 INFO: 172.17.0.1:55890 - "GET /health HTTP/1.1" 200 OK [00:25:11 TP0] Decode batch. #running-req: 1, #token: 136, token usage: 0.00, gen throughput (token/s): 7.06, #queue-req: 0 INFO: 192.168.28.101:51595 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 172.17.0.1:53822 - "GET /health HTTP/1.1" 200 OK INFO: 172.17.0.1:34902 - "GET /health HTTP/1.1" 200 OK INFO: 172.17.0.1:37092 - "GET /health HTTP/1.1" 200 OK [00:27:20 TP0] Prefill batch. #new-seq: 3, #new-token: 30, #cached-token: 42, cache hit rate: 51.23%, #running-req: 0, #queue-req: 0 [00:27:21 TP0] Prefill batch. #new-seq: 16, #new-token: 16, #cached-token: 368, cache hit rate: 82.60%, #running-req: 3, #queue-req: 81 [00:27:21 TP0] Prefill batch. #new-seq: 4, #new-token: 4, #cached-token: 92, cache hit rate: 84.58%, #running-req: 19, #queue-req: 77 [00:27:21 TP0] Prefill batch. #new-seq: 2, #new-token: 2, #cached-token: 46, cache hit rate: 85.36%, #running-req: 23, #queue-req: 75 [00:27:23 TP0] Prefill batch. #new-seq: 1, #new-token: 1, #cached-token: 23, cache hit rate: 85.71%, #running-req: 24, #queue-req: 74 INFO: 10.0.18.15:35886 - "POST /v1/chat/completions HTTP/1.1" 200 OK [00:27:23 TP0] Prefill batch. #new-seq: 1, #new-token: 1, #cached-token: 23, cache hit rate: 86.04%, #running-req: 25, #queue-req: 73 INFO: 10.0.18.15:35718 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35738 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35786 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35840 - "POST /v1/chat/completions HTTP/1.1" 200 OK [00:27:24 TP0] Prefill batch. #new-seq: 16, #new-token: 16, #cached-token: 368, cache hit rate: 89.39%, #running-req: 2, #queue-req: 57 INFO: 10.0.18.15:35660 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35758 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35824 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35828 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35862 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35872 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35900 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35912 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35696 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35706 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35670 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35686 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35730 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35754 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35774 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35802 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35808 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35830 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35856 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35896 - "POST /v1/chat/completions HTTP/1.1" 200 OK [00:27:24 TP0] Prefill batch. #new-seq: 6, #new-token: 6, #cached-token: 138, cache hit rate: 90.13%, #running-req: 18, #queue-req: 51 [00:27:24 TP0] Prefill batch. #new-seq: 2, #new-token: 2, #cached-token: 46, cache hit rate: 90.33%, #running-req: 24, #queue-req: 49 [00:27:25 TP0] Decode batch. #running-req: 26, #token: 174, token usage: 0.00, gen throughput (token/s): 5.31, #queue-req: 49 [00:27:25 TP0] Prefill batch. #new-seq: 1, #new-token: 1, #cached-token: 23, cache hit rate: 90.43%, #running-req: 26, #queue-req: 48 [00:27:26 TP0] Prefill batch. #new-seq: 1, #new-token: 1, #cached-token: 23, cache hit rate: 90.53%, #running-req: 25, #queue-req: 47 INFO: 10.0.18.15:35924 - "POST /v1/chat/completions HTTP/1.1" 200 OK INFO: 10.0.18.15:35934 - "POST /v1/chat/completions HTTP/1.1" 200 OK [00:27:26 TP0] Prefill batch. #new-seq: 1, #new-token: 1, #cached-token: 23, cache hit rate: 90.62%, #running-req: 26, #queue-req: 46

m0g1cian commented 2 days ago

See hyperparameter_tuning.md on tuning hyperparameters for better performance.

chuangzhidan commented 11 hours ago

See hyperparameter_tuning.md on tuning hyperparameters for better performance.

thank you!

sgl-project / sglang