[Bug][PERF]: Qwen2.5 performance degradation 0.8.4 -> 0.8.5

[vadiklyutiy]

Your current environment

The output of python collect_env.py for v0.8.4

INFO 05-23 14:36:10 [__init__.py:239] Automatically detected platform cuda.
Collecting environment information...
==============================
        System Info
==============================
OS                           : Ubuntu 22.04.5 LTS (x86_64)
GCC version                  : (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version                : Could not collect
CMake version                : version 4.0.2
Libc version                 : glibc-2.35

==============================
       PyTorch Info
==============================
PyTorch version              : 2.6.0+cu124
Is debug build               : False
CUDA used to build PyTorch   : 12.4
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.10.12 (main, Feb  4 2025, 14:57:36) [GCC 11.4.0] (64-bit runtime)
Python platform              : Linux-5.15.0-130-generic-x86_64-with-glibc2.35

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : 12.8.61
CUDA_MODULE_LOADING set to   : LAZY
GPU models and configuration : GPU 0: NVIDIA H100 80GB HBM3
Nvidia driver version        : 550.127.08
cuDNN version                : Could not collect
HIP runtime version          : N/A
MIOpen runtime version       : N/A
Is XNNPACK available         : True

==============================
          CPU Info
==============================
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        43 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               16
On-line CPU(s) list:                  0-15
Vendor ID:                            GenuineIntel
Model name:                           Intel(R) Xeon(R) Platinum 8468
CPU family:                           6
Model:                                143
Thread(s) per core:                   2
Core(s) per socket:                   8
Socket(s):                            1
Stepping:                             8
BogoMIPS:                             4200.00
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology cpuid tsc_known_freq pni pclmulqdq dtes64 ssse3 fma cx16 pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch cpuid_fault invpcid_single ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx_vnni avx512_bf16 wbnoinvd arat avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b fsrm md_clear serialize tsxldtrk amx_bf16 avx512_fp16 amx_tile amx_int8 arch_capabilities
Hypervisor vendor:                    KVM
Virtualization type:                  full
L1d cache:                            512 KiB (16 instances)
L1i cache:                            512 KiB (16 instances)
L2 cache:                             32 MiB (8 instances)
L3 cache:                             16 MiB (1 instance)
NUMA node(s):                         1
NUMA node0 CPU(s):                    0-15
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:        Unknown: No mitigations
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 and seccomp
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 SW loop, KVM SW loop
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Mitigation; TSX disabled

==============================
Versions of relevant libraries
==============================
[pip3] numpy==2.2.6
[pip3] nvidia-cublas-cu12==12.4.5.8
[pip3] nvidia-cuda-cupti-cu12==12.4.127
[pip3] nvidia-cuda-nvrtc-cu12==12.4.127
[pip3] nvidia-cuda-runtime-cu12==12.4.127
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.2.1.3
[pip3] nvidia-curand-cu12==10.3.5.147
[pip3] nvidia-cusolver-cu12==11.6.1.9
[pip3] nvidia-cusparse-cu12==12.3.1.170
[pip3] nvidia-cusparselt-cu12==0.6.2
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] pyzmq==26.4.0
[pip3] torch==2.6.0
[pip3] torchaudio==2.6.0
[pip3] torchvision==0.21.0
[pip3] transformers==4.52.3
[pip3] triton==3.2.0
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
Neuron SDK Version           : N/A
vLLM Version                 : 0.8.4
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
        GPU0    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      0-15    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

==============================
     Environment Variables
==============================
LD_LIBRARY_PATH=/usr/local/cuda-12.8/lib64:/usr/mpi/gcc/openmpi-4.1.7a1/lib
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

The output of python collect_env.py for v0.8.5

INFO 05-23 14:38:32 [__init__.py:239] Automatically detected platform cuda.
Collecting environment information...
/home/vadim/.local/lib/python3.10/site-packages/_distutils_hack/__init__.py:30: UserWarning: Setuptools is replacing distutils. Support for replacing an already imported distutils is deprecated. In the future, this condition will fail. Register concerns at https://github.com/pypa/setuptools/issues/new?template=distutils-deprecation.yml
  warnings.warn(
==============================
        System Info
==============================
OS                           : Ubuntu 22.04.5 LTS (x86_64)
GCC version                  : (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version                : Could not collect
CMake version                : version 4.0.2
Libc version                 : glibc-2.35

==============================
       PyTorch Info
==============================
PyTorch version              : 2.6.0+cu124
Is debug build               : False
CUDA used to build PyTorch   : 12.4
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.10.12 (main, Feb  4 2025, 14:57:36) [GCC 11.4.0] (64-bit runtime)
Python platform              : Linux-5.15.0-130-generic-x86_64-with-glibc2.35

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : 12.8.61
CUDA_MODULE_LOADING set to   : LAZY
GPU models and configuration : GPU 0: NVIDIA H100 80GB HBM3
Nvidia driver version        : 550.127.08
cuDNN version                : Could not collect
HIP runtime version          : N/A
MIOpen runtime version       : N/A
Is XNNPACK available         : True

==============================
          CPU Info
==============================
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        43 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               16
On-line CPU(s) list:                  0-15
Vendor ID:                            GenuineIntel
Model name:                           Intel(R) Xeon(R) Platinum 8468
CPU family:                           6
Model:                                143
Thread(s) per core:                   2
Core(s) per socket:                   8
Socket(s):                            1
Stepping:                             8
BogoMIPS:                             4200.00
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology cpuid tsc_known_freq pni pclmulqdq dtes64 ssse3 fma cx16 pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch cpuid_fault invpcid_single ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx_vnni avx512_bf16 wbnoinvd arat avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b fsrm md_clear serialize tsxldtrk amx_bf16 avx512_fp16 amx_tile amx_int8 arch_capabilities
Hypervisor vendor:                    KVM
Virtualization type:                  full
L1d cache:                            512 KiB (16 instances)
L1i cache:                            512 KiB (16 instances)
L2 cache:                             32 MiB (8 instances)
L3 cache:                             16 MiB (1 instance)
NUMA node(s):                         1
NUMA node0 CPU(s):                    0-15
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:        Unknown: No mitigations
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 and seccomp
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 SW loop, KVM SW loop
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Mitigation; TSX disabled

==============================
Versions of relevant libraries
==============================
[pip3] numpy==2.2.4
[pip3] nvidia-cublas-cu12==12.4.5.8
[pip3] nvidia-cuda-cupti-cu12==12.4.127
[pip3] nvidia-cuda-nvrtc-cu12==12.4.127
[pip3] nvidia-cuda-runtime-cu12==12.4.127
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.2.1.3
[pip3] nvidia-cufile-cu12==1.11.1.6
[pip3] nvidia-curand-cu12==10.3.5.147
[pip3] nvidia-cusolver-cu12==11.6.1.9
[pip3] nvidia-cusparse-cu12==12.3.1.170
[pip3] nvidia-cusparselt-cu12==0.6.2
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] pyzmq==26.4.0
[pip3] torch==2.6.0
[pip3] torchaudio==2.6.0
[pip3] torchvision==0.21.0
[pip3] transformers==4.52.3
[pip3] triton==3.2.0
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
Neuron SDK Version           : N/A
vLLM Version                 : 0.8.5.post1
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
        GPU0    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      0-15    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

==============================
     Environment Variables
==============================
LD_LIBRARY_PATH=/usr/local/cuda-12.8/lib64:/usr/mpi/gcc/openmpi-4.1.7a1/lib
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

I found a performance degradation for Qwen2.5 model 0.8.4 vs 0.8.5.

vllm serve run command:

vllm serve Qwen/Qwen2.5-VL-3B-Instruct --disable-log-requests --max-num-seqs 1024  --block-size 16 --max-num-batched-tokens 2048 --no-enable-prefix-caching

bench run command:

vllm bench serve  --model Qwen/Qwen2.5-VL-3B-Instruct --request-rate 200  --num-prompts 1000  --random-input-len 600 --random-output-len 125 --ignore-eos

Results of 7 runs
0.8.4

Throughput values: 62.75, 63.09, 62.81, 62.78, 62.90, 62.79, 61.42
Median throughput: 62.79 req/s
Standard error: 0.21 (0.33%)

0.8.5

Throughput values: 57.87, 57.91, 57.71, 57.74, 57.98, 57.80, 57.74
Median throughput: 57.80 req/s
Standard error: 0.04 (0.07%)

Degradation is 8.6%

Profile analysis.
I see sufficient increase in time we spend in attn.
0.8.4

0.8.5

Percentage we spend in attn increased from 14.2% to 38.8%

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[xinli-sw]

cc @DarkLight1337 for visibility

[DarkLight1337]

cc @LucasWilkinson

[Zerohertz]

I've tested it with scripts as below:

VLLM_TORCH_PROFILER_DIR=./profiler vllm serve Qwen/Qwen2.5-VL-3B-Instruct \
	--disable-log-requests --max-num-seqs 1024 \
	--block-size 16 \
	--max-num-batched-tokens 2048 \
	--no-enable-prefix-caching \
	--seed 42

vllm bench serve \
	--model Qwen/Qwen2.5-VL-3B-Instruct \
	--request-rate 200 \
	--num-prompts 1000 \
	--random-input-len 600 \
	--random-output-len 125 \
	--ignore-eos \
	--profile \
	--seed 42 \
	--percentile-metrics "ttft,tpot,itl,e2el" \
	--save-result

And, the benchmark results is:

[xinli-sw]

The absolute values of output throughput seem quite different (issue reports 57 but graph has 405), I believe @vadiklyutiy used a single H100 GPU, @Zerohertz could you tell me what GPU you were using, and if output throughput means request/s ?

[Zerohertz]

I used H100 in this benchmark test.
And output throughput is tok/s.

vllm/vllm/benchmarks/serve.py

Lines 440 to 441 in 57fd13a

	print("{:<40} {:<10.2f}".format("Output token throughput (tok/s):",
	metrics.output_throughput))

[xinli-sw]

Thanks! Do you still happen to have the request/s number for the benchmark above? I'm curious to know the inconsistency comes from the numbers in 0.8.5 or 0.8.4

[Zerohertz]

Of course!

Also I saved profile of each experiments too. But I have no idea how can I check as below:

[Zerohertz]

This is raw data of each experiments:

$ cat v0.8.4/*.json | jq | grep -i throughput
  "request_throughput": 3.062013260121091,
  "output_throughput": 382.7516575151364,
  "total_token_throughput": 2219.959613587791,
  "request_throughput": 3.2740703993211615,
  "output_throughput": 409.2587999151452,
  "total_token_throughput": 2373.701039507842,
  "request_throughput": 2.83955220944161,
  "output_throughput": 354.94402618020126,
  "total_token_throughput": 2058.6753518451674,

$ cat v0.8.5.post1/*.json | jq | grep -i throughput
  "request_throughput": 3.217340651749765,
  "output_throughput": 402.1675814687206,
  "total_token_throughput": 2332.5719725185795,
  "request_throughput": 3.219648200287808,
  "output_throughput": 402.456025035976,
  "total_token_throughput": 2334.2449452086607,
  "request_throughput": 3.2989874921542586,
  "output_throughput": 412.3734365192823,
  "total_token_throughput": 2391.765931811837,

[xinli-sw]

Thank you so much @Zerohertz for the update.

The grapbs on kernel break done are done with nsight system.

There seems to be quite a big difference between the two results though, req/s ~60 vs ~3.

Generally for a small model like Qwen3B, and relatively short output length around 128, we expect perf to be above 50 with an H100 system at high concurrency values.

[vadiklyutiy]

Number of request per second around 3 looks a quite weird. I had around 60 reqs/s and also use H100...

[vadiklyutiy]

Does VLLM_TORCH_PROFILER_DIR enable some kind of profiling?

Seems it enable torch.profiler. It likely brings a huge overhead to results.

[Zerohertz]

You're right, now that I think about it...
At first, I assumed kernel profiling was done using the profiler provided by vLLM and used the --profile option during benchmarking, but it caused significant delays.
The following experimental results were obtained without the profiler enabled. 😭

============ Serving Benchmark Result ============
Successful requests:                     1000      
Benchmark duration (s):                  18.74     
Total input tokens:                      600000    
Total generated tokens:                  125000    
Request throughput (req/s):              53.38     
Output token throughput (tok/s):         6671.99   
Total Token throughput (tok/s):          38697.52  
---------------Time to First Token----------------
Mean TTFT (ms):                          4431.91   
Median TTFT (ms):                        4282.04   
P99 TTFT (ms):                           9926.42   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          43.39     
Median TPOT (ms):                        48.04     
P99 TPOT (ms):                           50.27     
---------------Inter-token Latency----------------
Mean ITL (ms):                           43.39     
Median ITL (ms):                         47.61     
P99 ITL (ms):                            84.04     
----------------End-to-end Latency----------------
Mean E2EL (ms):                          9812.17   
Median E2EL (ms):                        10377.61  
P99 E2EL (ms):                           12454.34  
==================================================

$ cat openai-comp-200.0qps-Qwen2.5-VL-3B-Instruct-20250525-085912.json | jq | grep -i throughput
  "request_throughput": 53.37588503285498,
  "output_throughput": 6671.985629106873,
  "total_token_throughput": 38697.51664881986,

[vadiklyutiy]

The latest results look much better. 53reqs/s is closer to what I got.
One note. Usually the 1st run (mean without rerun vllm serve) a bit slow than next runs.