Closed cjackal closed 2 months ago
youkaichao
commented
2 months ago How do you install pytorch? Normal installation of pytorch should work. pytorch will load libcudart.so when you import it.
In addition, if you want to run it, you should add --disable-custom-all-reduce , we don't have --enable-custom-all-reduce flag.
cjackal
commented
2 months ago How do you install pytorch? Normal installation of pytorch should work. pytorch will load
libcudart.sowhen you import it.
By pip install torch==2.4.0+cu121 --index-url https://download.pytorch.org/whl/cu121, nothing special.
In addition, if you want to run it, you should add
--disable-custom-all-reduce, we don't have--enable-custom-all-reduceflag.
Right, my bad. Indeed I have checked that after adding --disable-custom-all-reduce option the vLLM server runs okay.
youkaichao
commented
2 months ago can you have a try:
import torch
from vllm.distributed.device_communicators.cuda_wrapper import find_loaded_library
print(find_loaded_library("libcudart.so"))see what happens?
cjackal
commented
2 months ago can you have a try:
import torch from vllm.distributed.device_communicators.cuda_wrapper import find_loaded_library print(find_loaded_library("libcudart.so"))see what happens?
it prints None. FYI, if I print the full /proc/self/maps, torch's libcudart is loaded from /home/jovyan/git/vllm-serving/.venv/lib/python3.10/site-packages/torch/lib/libcudart-9335f6a2.so.12.
cjackal
commented
2 months ago FYI: in my torch lib, there are 5 shared objects suffixed by hash:
Seems library search should be updated to use some sort of glob patterns?
youkaichao
commented
2 months ago libcudart-9335f6a2.so.12
interesting, I will update the glob patterns.
youkaichao
commented
2 months ago @cjackal can you try #7620 ?
cjackal
commented
2 months ago @youkaichao Thanks for prompt response, I have left a suggested change there in the PR. It works nicely indeed.
Your current environment
The output of `python collect_env.py`
```text Collecting environment information... PyTorch version: 2.4.0+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 3.30.2 Libc version: glibc-2.35 Python version: 3.10.12 (main, Mar 22 2024, 16:50:05) [GCC 11.4.0] (64-bit runtime) Python platform: Linux-4.18.0-372.19.1.el8_6.x86_64-x86_64-with-glibc2.35 Is CUDA available: True CUDA runtime version: 12.1.105 CUDA_MODULE_LOADING set to: LAZY GPU models and configuration: GPU 0: NVIDIA A100 80GB PCIe GPU 1: NVIDIA A100 80GB PCIe Nvidia driver version: 515.86.01 cuDNN version: Probably one of the following: /usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.3 /usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.3 /usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.3 /usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.3 /usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.3 /usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.3 /usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.3 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: 46 bits physical, 57 bits virtual Byte Order: Little Endian CPU(s): 112 On-line CPU(s) list: 0-111 Vendor ID: GenuineIntel Model name: Intel(R) Xeon(R) Gold 6348 CPU @ 2.60GHz CPU family: 6 Model: 106 Thread(s) per core: 2 Core(s) per socket: 28 Socket(s): 2 Stepping: 6 CPU max MHz: 3500.0000 CPU min MHz: 800.0000 BogoMIPS: 5200.00 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 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 invpcid_single intel_ppin ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi 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 wbnoinvd dtherm ida arat pln pts avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid fsrm md_clear pconfig flush_l1d arch_capabilities Virtualization: VT-x L1d cache: 2.6 MiB (56 instances) L1i cache: 1.8 MiB (56 instances) L2 cache: 70 MiB (56 instances) L3 cache: 84 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 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 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; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Enhanced IBRS, IBPB conditional, RSB filling Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected Versions of relevant libraries: [pip3] mypy-extensions==1.0.0 [pip3] numpy==1.26.4 [pip3] torch==2.4.0+cu121 [pip3] torchvision==0.19.0+cu121 [pip3] triton==3.0.0 [conda] Could not collect ```🐛 Describe the bug
It seems the GPU P2P check script in
custom_all_reduce_utils.pyhas trouble finding the cuda runtime; after some boot-up logic, vLLM CLI fails shortly with the error messageAssertionError: libcudart.so is not loaded in the current process.It raises regardless of the GPU topology (whether NVX or PCIe, I mean) and vLLM server runs well with adding
--enable-custom-all-reduce=Falseoption or manually creating the~/.cache/vllm/gpu_p2p_access_cache_for_0,1.jsoncache file, so I think the current library search method ingpu_p2p_access_checkshould be updated somehow.The CLI arguments I used is as follow:
cf. there's another open issue about P2P check #3688 but it doesn't look relevant.
full vLLM CLI log and traceback
``` INFO 08-16 11:21:48 api_server.py:339] vLLM API server version 0.5.4 INFO 08-16 11:21:48 api_server.py:340] args: Namespace(model_tag='/mnt/model-vol-1/model/META-LLAMA-3.1-70B-INSTRUCT/', host=None, port=8080, 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='/notebook/cjackal/cjackal-vscode/proxy/8080', middleware=[], return_tokens_as_token_ids=False, disable_frontend_multiprocessing=False, model='/mnt/model-vol-1/model/META-LLAMA-3.1-70B-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=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=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=False, quantization='fp8', 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, enable_prompt_adapter=False, max_prompt_adapters=1, max_prompt_adapter_token=0, device='auto', scheduler_delay_factor=0.0, enable_chunked_prefill=None, speculative_model=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=['meta-llama/meta-llama-3.1-70b-instruct'], qlora_adapter_name_or_path=None, otlp_traces_endpoint=None, engine_use_ray=False, disable_log_requests=False, max_log_len=None, dispatch_function=