search

vllm-project / vllm

A high-throughput and memory-efficient inference and serving engine for LLMs
https://docs.vllm.ai
Apache License 2.0
22.37k stars 3.16k forks source link

[Bug]: enable_chunked_prefill feature hangs on AMD Radeon PRO W7900 (gfx1100) #5013

Open hongxiayang opened 1 month ago

hongxiayang commented 1 month ago

Your current environment

Collecting environment information...
INFO 05-23 16:19:36 pynccl.py:58] Loading nccl from library librccl.so.1
/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/cuda/__init__.py:611: UserWarning: Can't initialize NVML
  warnings.warn("Can't initialize NVML")
PyTorch version: 2.1.1+git011de5c
Is debug build: False
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: 6.0.32830-d62f6a171

OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0
Clang version: 17.0.0 (https://github.com/RadeonOpenCompute/llvm-project roc-6.0.0 23483 7208e8d15fbf218deb74483ea8c549c67ca4985e)
CMake version: version 3.29.2
Libc version: glibc-2.31

Python version: 3.9.18 (main, Sep 11 2023, 13:41:44)  [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.19.0-42-generic-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: 10.1.243
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: AMD Radeon PRO W7900NoGCNArchNameOnOldPyTorch
Nvidia driver version: Could not collect
cuDNN version: Could not collect
HIP runtime version: 6.0.32830
MIOpen runtime version: 3.0.0
Is XNNPACK available: True

CPU:
Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Byte Order:                      Little Endian
Address sizes:                   48 bits physical, 48 bits virtual
CPU(s):                          64
On-line CPU(s) list:             0-63
Thread(s) per core:              2
Core(s) per socket:              32
Socket(s):                       1
NUMA node(s):                    1
Vendor ID:                       AuthenticAMD
CPU family:                      25
Model:                           8
Model name:                      AMD Ryzen Threadripper PRO 5975WX 32-Cores
Stepping:                        2
Frequency boost:                 enabled
CPU MHz:                         3600.000
CPU max MHz:                     7006.6401
CPU min MHz:                     1800.0000
BogoMIPS:                        7186.72
Virtualization:                  AMD-V
L1d cache:                       1 MiB
L1i cache:                       1 MiB
L2 cache:                        16 MiB
L3 cache:                        128 MiB
NUMA node0 CPU(s):               0-63
Vulnerability Itlb multihit:     Not affected
Vulnerability L1tf:              Not affected
Vulnerability Mds:               Not affected
Vulnerability Meltdown:          Not affected
Vulnerability Mmio stale data:   Not affected
Vulnerability Retbleed:          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; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected
Vulnerability Srbds:             Not affected
Vulnerability Tsx async abort:   Not affected
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 rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic 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_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin brs arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm

Versions of relevant libraries:
[pip3] mypy==1.4.1
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.22.4
[pip3] torch==2.1.1+git011de5c
[pip3] torchvision==0.16.1+fdea156
[pip3] triton==2.1.0
[conda] No relevant packagesROCM Version: 6.0.32830-d62f6a171
Neuron SDK Version: N/A
vLLM Version: 0.4.0.post1
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
Could not collect

🐛 Describe the bug

A user reported an issue with the chunked-prefill on Radeon PRO W7900 (gfx1100) , which uses python naive attention. Reproduced the issue on gfx1100.

Test script as below:

prompts = [
    "Hello, my name is",
    "The president of the United States is",
    "The capital of France is",
    "The future of AI is",
]
from vllm import LLM, SamplingParams

import time
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)

start_time = time.time()
llm = LLM(model="/path/to/model", enable_chunked_prefill=True )
outputs = llm.generate(prompts, sampling_params)
elapsed_time = time.time() - start_time
print(f"elapsed time {elapsed_time}")

# Print the outputs.

for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
hongxiayang commented 1 month ago

The feature seems no problem when I enabled it on MI250x.

simon-mo commented 1 month ago

Can you also paste the log? I think the main parameters need to be considered is the number of available blocks

hongxiayang commented 1 month ago

It seems ok for me now. Might because some other workload was running something at the same time using the same GPU when I experienced the problem.

I will check with the client to get some more information.

hongxiayang commented 1 month ago

Here is one log for latency benchmarking, which I observed that the avg latency is much worse than not using the chunked-prefill. Maybe some parameters need tuning ?:

INFO 05-23 21:41:39 pynccl.py:58] Loading nccl from library librccl.so.1
Namespace(model='/dockerx/data/llama-2-13b-chat-hf', tokenizer=None, quantization=None, tensor_parallel_size=1, input_len=1024, output_len=256, batch_size=1, n=1, use_beam_search=False, num_iters_warmup=10, num_iters=10, trust_remote_code=False, dtype='auto', enforce_eager=False, kv_cache_dtype='auto', quantization_param_path=None, profile=False, profile_result_dir=None, device='cuda', block_size=16, enable_chunked_prefill=True, ray_workers_use_nsight=False, download_dir=None)
INFO 05-23 21:41:40 config.py:581] Chunked prefill is enabled (EXPERIMENTAL).
INFO 05-23 21:41:40 llm_engine.py:84] Initializing an LLM engine (v0.4.0.post1) with config: model='/dockerx/data/llama-2-13b-chat-hf', speculative_config=None, tokenizer='/dockerx/data/llama-2-13b-chat-hf', tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.float16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, 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)
/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/cuda/__init__.py:611: UserWarning: Can't initialize NVML
  warnings.warn("Can't initialize NVML")
INFO 05-23 21:41:41 selector.py:59] flash_atten is not supported on NAVI GPUs.
INFO 05-23 21:41:41 selector.py:38] Using ROCmFlashAttention backend.
INFO 05-23 21:41:50 model_runner.py:169] Loading model weights took 24.2835 GB
INFO 05-23 21:41:52 gpu_executor.py:61] # GPU blocks: 1270, # CPU blocks: 327
INFO 05-23 21:41:52 model_runner.py:967] Capturing the model for CUDA graphs. This may lead to unexpected consequences if the model is not static. To run the model in eager mode, set 'enforce_eager=True' or use '--enforce-eager' in the CLI.
INFO 05-23 21:41:52 model_runner.py:971] CUDA graphs can take additional 1~3 GiB memory per GPU. If you are running out of memory, consider decreasing `gpu_memory_utilization` or enforcing eager mode. You can also reduce the `max_num_seqs` as needed to decrease memory usage.
INFO 05-23 21:41:57 model_runner.py:1048] Graph capturing finished in 5 secs.
SamplingParams(n=1, best_of=1, presence_penalty=0.0, frequency_penalty=0.0, repetition_penalty=1.0, temperature=1.0, top_p=1.0, top_k=-1, min_p=0.0, seed=None, use_beam_search=False, length_penalty=1.0, early_stopping=False, stop=[], stop_token_ids=[], include_stop_str_in_output=False, ignore_eos=True, max_tokens=256, min_tokens=0, logprobs=None, prompt_logprobs=None, skip_special_tokens=True, spaces_between_special_tokens=True, truncate_prompt_tokens=None)
Warming up...
Warmup iterations: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 10/10 [05:01<00:00, 30.15s/it]
Profiling iterations: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 10/10 [05:01<00:00, 30.12s/it]
Avg latency: xxx seconds
hongxiayang commented 1 month ago

Below is the log from the client: 

root@26a65a87bce7:/data/vllm-latest/vllm-240507/benchmarks# python benchmark_throughput.py --dataset /data/ShareGPT_V3_unfiltered_cleaned_split.json \
> --model /data/THUDM/.chatglm2-6b --enable-chunked-prefill --max-num-batched-tokens 1024 --trust-remote-code --num-prompts 100
Namespace(backend='vllm', dataset='/data/ShareGPT_V3_unfiltered_cleaned_split.json', input_len=None, output_len=None, model='/data/THUDM/chatglm2-6b', tokenizer='/data/THUDM/chatglm2-6b', quantization=None, tensor_parallel_size=1, n=1, use_beam_search=False, num_prompts=100, seed=0, hf_max_batch_size=None, trust_remote_code=True, max_model_len=None, dtype='auto', gpu_memory_utilization=0.9, enforce_eager=False, kv_cache_dtype='auto', quantization_param_path=None, device='cuda', enable_prefix_caching=False, enable_chunked_prefill=True, max_num_batched_tokens=1024, download_dir=None)
INFO 05-24 02:35:06 config.py:627] Chunked prefill is enabled (EXPERIMENTAL).
INFO 05-24 02:35:06 llm_engine.py:100] Initializing an LLM engine (v0.4.2) with config: model='/data/THUDM/chatglm2-6b', speculative_config=None, tokenizer='/data/THUDM/chatglm2-6b', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.float16, max_seq_len=32768, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=1, 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/THUDM/chatglm2-6b)
WARNING 05-24 02:35:06 tokenizer.py:126] Using a slow tokenizer. This might cause a significant slowdown. Consider using a fast tokenizer instead.
INFO 05-24 02:35:06 utils.py:660] Found nccl from library /opt/rocm-6.0.0/lib/librccl.so.1
/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/cuda/__init__.py:611: UserWarning: Can't initialize NVML
  warnings.warn("Can't initialize NVML")
INFO 05-24 02:35:09 selector.py:63] flash_atten is not supported on NAVI GPUs.
INFO 05-24 02:35:09 selector.py:37] Using ROCmFlashAttention backend.
INFO 05-24 02:35:23 model_runner.py:175] Loading model weights took 11.6608 GB
INFO 05-24 02:35:25 gpu_executor.py:114] # GPU blocks: 65192, # CPU blocks: 9362
INFO 05-24 02:35:27 model_runner.py:937] Capturing the model for CUDA graphs. This may lead to unexpected consequences if the model is not static. To run the model in eager mode, set 'enforce_eager=True' or use '--enforce-eager' in the CLI.
INFO 05-24 02:35:27 model_runner.py:941] CUDA graphs can take additional 1~3 GiB memory per GPU. If you are running out of memory, consider decreasing `gpu_memory_utilization` or enforcing eager mode. You can also reduce the `max_num_seqs` as needed to decrease memory usage.
INFO 05-24 02:35:30 model_runner.py:1017] Graph capturing finished in 4 secs.
Processed prompts:   0%|                                                                                                                                                          | 0/100 [00:00<?, ?it/s]

And python collect_env.py stdout

root@26a65a87bce7:/app/vllm# python collect_env.py
Collecting environment information...
/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/cuda/__init__.py:611: UserWarning: Can't initialize NVML
  warnings.warn("Can't initialize NVML")
PyTorch version: 2.1.1+git011de5c
Is debug build: False
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: 6.0.32830-d62f6a171

OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0
Clang version: 17.0.0 (https://github.com/RadeonOpenCompute/llvm-project roc-6.0.0 23483 7208e8d15fbf218deb74483ea8c549c67ca4985e)
CMake version: version 3.29.2
Libc version: glibc-2.31

Python version: 3.9.18 (main, Sep 11 2023, 13:41:44)  [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-92-generic-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: 10.1.243
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: AMD Radeon PRO W7900NoGCNArchNameOnOldPyTorch
Nvidia driver version: Could not collect
cuDNN version: Could not collect
HIP runtime version: 6.0.32830
MIOpen runtime version: 3.0.0
Is XNNPACK available: True

CPU:
Architecture:                       x86_64
CPU op-mode(s):                     32-bit, 64-bit
Byte Order:                         Little Endian
Address sizes:                      46 bits physical, 57 bits virtual
CPU(s):                             64
On-line CPU(s) list:                0-63
Thread(s) per core:                 1
Core(s) per socket:                 32
Socket(s):                          2
NUMA node(s):                       2
Vendor ID:                          GenuineIntel
CPU family:                         6
Model:                              106
Model name:                         Intel(R) Xeon(R) Platinum 8338C CPU @ 2.60GHz
Stepping:                           6
CPU MHz:                            800.000
CPU max MHz:                        3500.0000
CPU min MHz:                        800.0000
BogoMIPS:                           5200.00
Virtualization:                     VT-x
L1d cache:                          3 MiB
L1i cache:                          2 MiB
L2 cache:                           80 MiB
L3 cache:                           96 MiB
NUMA node0 CPU(s):                  0-31
NUMA node1 CPU(s):                  32-63
Vulnerability Gather data sampling: Mitigation; Microcode
Vulnerability Itlb multihit:        Not affected
Vulnerability L1tf:                 Not affected
Vulnerability Mds:                  Not affected
Vulnerability Meltdown:             Not affected
Vulnerability Mmio stale data:      Mitigation; Clear CPU buffers; SMT disabled
Vulnerability Retbleed:             Not affected
Vulnerability Spec rstack overflow: 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, PBRSB-eIBRS SW sequence
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Not affected
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 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

Versions of relevant libraries:
[pip3] mypy==1.4.1
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.22.4
[pip3] torch==2.1.1+git011de5c
[pip3] torchvision==0.16.1+fdea156
[pip3] triton==2.1.0
[conda] No relevant packagesROCM Version: 6.0.32830-d62f6a171
Neuron SDK Version: N/A
vLLM Version: 0.4.1
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
Could not collect