Open
Description
Proposal to improve performance
No response
Report of performance regression
I am running vllm on a single H100 NVL node with 95GB VRAM and vllm 0.8.5.post1.
Deployed Qwen3 32B fp16 using:
vllm serve "Qwen/Qwen3-32B" \
--host 0.0.0.0 \
--port 8000 \
--download-dir /workspace \
--dtype bfloat16 \
--max-model-len 32000 \
--enable-chunked-prefill \
--enable-prefix-caching \
--seed 42 \
--max-num-seqs 32 \
--disable-log-requests
and then ran the benchmark_serving script via:
python3 benchmark_serving.py --dataset-path /root/ShareGPT_V3_unfiltered_cleaned_split.json --host 0.0.0.0 --port 8000 --model "Qwen/Qwen3-32B" --temperature 0.7 --top-k 20 --top-p 0.8
and the results:
============ Serving Benchmark Result ============
Successful requests: 1000
Benchmark duration (s): 306.78
Total input tokens: 217393
Total generated tokens: 200516
Request throughput (req/s): 3.26
Output token throughput (tok/s): 653.61
Total Token throughput (tok/s): 1362.23
---------------Time to First Token----------------
Mean TTFT (ms): 146809.12
Median TTFT (ms): 148069.33
P99 TTFT (ms): 286485.34
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms): 47.06
Median TPOT (ms): 46.24
P99 TPOT (ms): 66.91
---------------Inter-token Latency----------------
Mean ITL (ms): 46.74
Median ITL (ms): 40.27
P99 ITL (ms): 193.03
==================================================
Deployed Qwen3 32B fp8 using:
vllm serve "Qwen/Qwen3-32B-FP8" \
--host 0.0.0.0 \
--port 8000 \
--download-dir /workspace \
--max-model-len 32000 \
--enable-chunked-prefill \
--enable-prefix-caching \
--seed 42 \
--max-num-seqs 32 \
--disable-log-requests
and the benchmark serving script:
python3 benchmark_serving.py --dataset-path /root/ShareGPT_V3_unfiltered_cleaned_split.json --host 0.0.0.0 --port 8000 --model "Qwen/Qwen3-32B-FP8" --temperature 0.7 --top-k 20 --top-p 0.8
and results:
============ Serving Benchmark Result ============
Successful requests: 1000
Benchmark duration (s): 228.73
Total input tokens: 217393
Total generated tokens: 201115
Request throughput (req/s): 4.37
Output token throughput (tok/s): 879.25
Total Token throughput (tok/s): 1829.67
---------------Time to First Token----------------
Mean TTFT (ms): 109054.64
Median TTFT (ms): 109874.45
P99 TTFT (ms): 212661.56
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms): 34.96
Median TPOT (ms): 34.36
P99 TPOT (ms): 48.82
---------------Inter-token Latency----------------
Mean ITL (ms): 34.65
Median ITL (ms): 30.50
P99 ITL (ms): 134.48
==================================================
Misc discussion on performance
Any reason why the ~1.6x speedup in throughput expected with fp8 vs fp16 was not seen here (as per https://docs.vllm.ai/en/v0.5.4/quantization/fp8.html)? The output token throughput with fp8 (879) here is only 1.3x of the fp16 number (653).
Same with latency: was expecting a much lower inter-token latency number with fp8 (34ms) vs fp16 (46ms).
Is this because vllm doesn't automatically use the fp8 kernels? Assuming half the memory bandwidth would be required, I would assume nearly double the number of tokens generated in the same unit of time.
Your current environment (if you think it is necessary)
python3 collect_env.py
INFO 05-07 11:59:48 [__init__.py:239] Automatically detected platform cuda.
Collecting environment information...
PyTorch version: 2.6.0+cu124
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A
OS: Debian GNU/Linux 12 (bookworm) (x86_64)
GCC version: (Debian 12.2.0-14) 12.2.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.36
Python version: 3.12.10 (main, Apr 28 2025, 22:12:29) [GCC 12.2.0] (64-bit runtime)
Python platform: Linux-6.8.0-40-generic-x86_64-with-glibc2.36
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA H100 NVL
Nvidia driver version: 550.107.02
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: 52 bits physical, 57 bits virtual
Byte Order: Little Endian
CPU(s): 128
On-line CPU(s) list: 0-127
Vendor ID: AuthenticAMD
Model name: AMD EPYC 9374F 32-Core Processor
CPU family: 25
Model: 17
Thread(s) per core: 2
Core(s) per socket: 32
Socket(s): 2
Stepping: 1
Frequency boost: enabled
CPU(s) scaling MHz: 48%
CPU max MHz: 4304.9312
CPU min MHz: 1500.0000
BogoMIPS: 7699.51
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 amd_lbr_v2 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 hw_pstate ssbd mba perfmon_v2 ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local user_shstk avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor smca fsrm flush_l1d debug_swap
Virtualization: AMD-V
L1d cache: 2 MiB (64 instances)
L1i cache: 2 MiB (64 instances)
L2 cache: 64 MiB (64 instances)
L3 cache: 512 MiB (16 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-31,64-95
NUMA node1 CPU(s): 32-63,96-127
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: Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Mitigation; Safe RET
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; Enhanced / Automatic IBRS; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Versions of relevant libraries:
[pip3] flashinfer-python==0.2.1.post2+cu124torch2.6
[pip3] numpy==2.2.5
[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==4.52.0.dev0
[pip3] triton==3.2.0
[conda] Could not collect
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 32-63,96-127 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
VLLM_CONFIGURE_LOGGING=1
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY
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.