[Bug]: the throughput of qwen3moe is low for prompts above 2000 tokens

[aabbccddwasd]

Your current environment

The output of python collect_env.py

INFO 05-05 20:13:42 [__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: 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: Could not collect
Libc version: glibc-2.35

Python version: 3.12.8 | packaged by Anaconda, Inc. | (main, Dec 11 2024, 16:31:09) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-136-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.4.99
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA GeForce RTX 2080 Ti
GPU 1: NVIDIA GeForce RTX 2080 Ti
GPU 2: NVIDIA GeForce RTX 2080 Ti
GPU 3: NVIDIA GeForce RTX 2080 Ti
GPU 4: NVIDIA GeForce RTX 2080 Ti
GPU 5: NVIDIA GeForce RTX 2080 Ti

Nvidia driver version: 550.144.03
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.8.0
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, 48 bits virtual
Byte Order:                           Little Endian
CPU(s):                               56
On-line CPU(s) list:                  0-55
Vendor ID:                            GenuineIntel
Model name:                           Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz
CPU family:                           6
Model:                                79
Thread(s) per core:                   2
Core(s) per socket:                   14
Socket(s):                            2
Stepping:                             1
CPU max MHz:                          3300.0000
CPU min MHz:                          1200.0000
BogoMIPS:                             4800.01
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 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 cdp_l3 invpcid_single pti intel_ppin ssbd ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm rdt_a rdseed adx smap intel_pt xsaveopt cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts md_clear flush_l1d
Virtualization:                       VT-x
L1d cache:                            896 KiB (28 instances)
L1i cache:                            896 KiB (28 instances)
L2 cache:                             7 MiB (28 instances)
L3 cache:                             70 MiB (2 instances)
NUMA node(s):                         2
NUMA node0 CPU(s):                    0-13,28-41
NUMA node1 CPU(s):                    14-27,42-55
Vulnerability Gather data sampling:   Not affected
Vulnerability Itlb multihit:          KVM: Mitigation: VMX disabled
Vulnerability L1tf:                   Mitigation; PTE Inversion; VMX conditional cache flushes, SMT vulnerable
Vulnerability Mds:                    Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Meltdown:               Mitigation; PTI
Vulnerability Mmio stale data:        Mitigation; Clear CPU buffers; SMT vulnerable
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; Retpolines; IBPB conditional; IBRS_FW; STIBP conditional; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Mitigation; Clear CPU buffers; SMT vulnerable

Versions of relevant libraries:
[pip3] numpy==2.0.1
[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-ml-py==12.570.86
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] onnxruntime==1.20.1
[pip3] pyzmq==26.2.0
[pip3] rapidocr-onnxruntime==1.3.24
[pip3] torch==2.6.0
[pip3] torchaudio==2.6.0
[pip3] torchvision==0.21.0
[pip3] transformers==4.51.3
[pip3] triton==3.2.0
[conda] blas                      1.0                         mkl  
[conda] cuda-cudart               12.4.127                      0    nvidia
[conda] cuda-cupti                12.4.127                      0    nvidia
[conda] cuda-libraries            12.4.1                        0    nvidia
[conda] cuda-nvrtc                12.4.127                      0    nvidia
[conda] cuda-nvtx                 12.4.127                      0    nvidia
[conda] cuda-opencl               12.8.55                       0    nvidia
[conda] cuda-runtime              12.4.1                        0    nvidia
[conda] cuda-version              12.8                          3    nvidia
[conda] ffmpeg                    4.3                  hf484d3e_0    pytorch
[conda] libcublas                 12.4.5.8                      0    nvidia
[conda] libcufft                  11.2.1.3                      0    nvidia
[conda] libcufile                 1.13.0.11                     0    nvidia
[conda] libcurand                 10.3.9.55                     0    nvidia
[conda] libcusolver               11.6.1.9                      0    nvidia
[conda] libcusparse               12.3.1.170                    0    nvidia
[conda] libjpeg-turbo             2.0.0                h9bf148f_0    pytorch
[conda] libnpp                    12.2.5.30                     0    nvidia
[conda] libnvfatbin               12.8.55                       0    nvidia
[conda] libnvjitlink              12.4.127                      0    nvidia
[conda] libnvjpeg                 12.3.1.117                    0    nvidia
[conda] mkl                       2023.1.0         h213fc3f_46344  
[conda] mkl-service               2.4.0           py312h5eee18b_2  
[conda] mkl_fft                   1.3.11          py312h5eee18b_0  
[conda] mkl_random                1.2.8           py312h526ad5a_0  
[conda] numpy                     2.0.1           py312hc5e2394_1  
[conda] numpy-base                2.0.1           py312h0da6c21_1  
[conda] nvidia-ml-py              12.570.86                pypi_0    pypi
[conda] pytorch-cuda              12.4                 hc786d27_7    pytorch
[conda] pytorch-mutex             1.0                        cuda    pytorch
[conda] pyzmq                     26.2.0          py312h6a678d5_0  
[conda] torch                     2.6.0                    pypi_0    pypi
[conda] torchaudio                2.6.0                    pypi_0    pypi
[conda] torchvision               0.21.0                   pypi_0    pypi
[conda] transformers              4.51.3                   pypi_0    pypi
[conda] triton                    3.2.0                    pypi_0    pypi
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	GPU1	GPU2	GPU3	GPU4	GPU5	CPU Affinity	NUMA Affinity	GPU NUMA ID
GPU0	 X 	NV2	SYS	SYS	SYS	SYS	0-13,28-41	0		N/A
GPU1	NV2	 X 	SYS	SYS	SYS	SYS	0-13,28-41	0		N/A
GPU2	SYS	SYS	 X 	NV2	PHB	PHB	14-27,42-55	1		N/A
GPU3	SYS	SYS	NV2	 X 	PHB	PHB	14-27,42-55	1		N/A
GPU4	SYS	SYS	PHB	PHB	 X 	NV2	14-27,42-55	1		N/A
GPU5	SYS	SYS	PHB	PHB	NV2	 X 	14-27,42-55	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

LD_LIBRARY_PATH=:/usr/local/cuda-12.4/lib64
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

this is my start script:

#!/bin/bash
export CUDA_VISIBLE_DEVICES=2,3,4,5
export VLLM_ATTENTION_BACKEND=XFORMERS
# export VLLM_USE_V1=1

source /usr/local/anaconda3/bin/activate /home/aabbccddwasd/.conda/envs/vllm-latest
python -m vllm.entrypoints.openai.api_server \
    --model /data-HV-1t/AI-stuff/qwen-models/Qwen3-30B-A3B \
    --tensor-parallel-size=4 \
    --enable-expert-parallel \
    --gpu-memory-utilization 0.98 \
    --max-model-len 32768 \
    --enable-prefix-caching \
    --served-model-name qwen3-30ba3b \
    --swap-space 8 \
    --disable-log-requests \
    --dtype float16 \
    --port 8000 \

I found it can run up to about 75 tokens/s before 2000 token than suddenly drop to 10 tokens/s after 2000 token, at the same time, the utilization of one gpu drops to about 13%.

could it be a problem with GPU communication? but why the speed drops so fast（suddenly from 75t/s to 10t/s）

someone help me please

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.

[aabbccddwasd]

May 05 20:19:16 server-2080ti sudo[2121399]: INFO 05-05 20:19:16 [metrics.py:486] Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 63.5 tokens/s, Running: 1 reqs, Swapped: 0 reqs, Pending: 0 reqs, GPU KV cache usage: 3.5%, CPU KV cache usage: 0.0%.
May 05 20:19:16 server-2080ti sudo[2121399]: INFO 05-05 20:19:16 [metrics.py:502] Prefix cache hit rate: GPU: 0.00%, CPU: 0.00%
May 05 20:19:21 server-2080ti sudo[2121399]: INFO 05-05 20:19:21 [metrics.py:486] Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 64.1 tokens/s, Running: 1 reqs, Swapped: 0 reqs, Pending: 0 reqs, GPU KV cache usage: 3.7%, CPU KV cache usage: 0.0%.
May 05 20:19:21 server-2080ti sudo[2121399]: INFO 05-05 20:19:21 [metrics.py:502] Prefix cache hit rate: GPU: 0.00%, CPU: 0.00%
May 05 20:19:26 server-2080ti sudo[2121399]: INFO 05-05 20:19:26 [metrics.py:486] Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 31.0 tokens/s, Running: 1 reqs, Swapped: 0 reqs, Pending: 0 reqs, GPU KV cache usage: 3.8%, CPU KV cache usage: 0.0%.
May 05 20:19:26 server-2080ti sudo[2121399]: INFO 05-05 20:19:26 [metrics.py:502] Prefix cache hit rate: GPU: 0.00%, CPU: 0.00%
May 05 20:19:31 server-2080ti sudo[2121399]: INFO 05-05 20:19:31 [metrics.py:486] Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10.5 tokens/s, Running: 1 reqs, Swapped: 0 reqs, Pending: 0 reqs, GPU KV cache usage: 3.8%, CPU KV cache usage: 0.0%.
May 05 20:19:31 server-2080ti sudo[2121399]: INFO 05-05 20:19:31 [metrics.py:502] Prefix cache hit rate: GPU: 0.00%, CPU: 0.00%

[aabbccddwasd]

by the way, it's even slower than a dense 72b model. I think it is not a problem with GPU communication

[syskn]

I encountered similar phenomenon - with very few tokens it runs okay, but with as the input gets larger it gets unreasonably slow, like the old days of Transformers before Flash Attention. The issue is much worse with multi GPU setup, while single GPU is also slower than expected.

Image: vllm/vllm-openai:v0.8.5.post1
Hardware: 8xA100 80GB NVLink
CUDA 12.6
Prefill size is 20000~24000 tokens

Tested with V0 Engine, enforce-eager on and off, backend FLASHINFER and FA2, expert-parallelism on and off, but yieled the same result with all combinations.

Qwen3-30B-A3B/tp1
vllm serve /models/qwen/Qwen3-30B-A3B --port 7777 --tensor-parallel-size 1 --max-num-batched-tokens 131072 --max-model-len 32768 --enable-prefix-caching
21.0 tokens generated/sec on vllm

Qwen3-30B-A3B/tp8
vllm serve /models/qwen/Qwen3-30B-A3B --port 7777 --tensor-parallel-size 8 --max-num-batched-tokens 131072 --max-model-len 32768 --enable-prefix-caching
4.5 tokens generated/sec on vllm
80.0 tokens generated/sec on SGLang

Qwen3-235B-A22B/tp8
vllm serve /models/qwen/Qwen3-235B-A22B --port 7777 --tensor-parallel-size 8 --max-num-batched-tokens 131072 --max-model-len 32768 --enable-prefix-caching
2.2 tokens generated/sec on vllm
36.0 tokens generated/se on SGLang

[Some-random]

I'm observing the same issue

[chensiye-csy]

Same issue

[Beckham007]

INFO 05-06 02:57:08 [api_server.py:1043] vLLM API server version 0.8.5.post1
INFO 05-06 02:57:08 [api_server.py:1044] args: Namespace(subparser='serve', model_tag='/models/checkpoints/Qwen3-235B-A22B-FP8', config='', host='0.0.0.0', port=8000, uvicorn_log_level='info', disable_uvicorn_access_log=False, allow_credentials=False, allowed_origins=[''], allowed_methods=[''], allowed_headers=['*'], api_key=None, lora_modules=None, prompt_adapters=None, chat_template=None, chat_template_content_format='auto', response_role='assistant', ssl_keyfile=None, ssl_certfile=None, ssl_ca_certs=None, enable_ssl_refresh=False, ssl_cert_reqs=0, root_path=None, middleware=[], return_tokens_as_token_ids=False, disable_frontend_multiprocessing=False, enable_request_id_headers=False, enable_auto_tool_choice=True, tool_call_parser='hermes', tool_parser_plugin='', model='/models/checkpoints/Qwen3-235B-A22B-FP8', task='auto', tokenizer=None, hf_config_path=None, skip_tokenizer_init=False, revision=None, code_revision=None, tokenizer_revision=None, tokenizer_mode='auto', trust_remote_code=False, allowed_local_media_path=None, load_format='auto', download_dir=None, model_loader_extra_config={}, use_tqdm_on_load=True, config_format=<ConfigFormat.AUTO: 'auto'>, dtype='auto', max_model_len=32768, guided_decoding_backend='auto', reasoning_parser='deepseek_r1', logits_processor_pattern=None, model_impl='auto', distributed_executor_backend=None, pipeline_parallel_size=1, tensor_parallel_size=8, data_parallel_size=1, enable_expert_parallel=True, max_parallel_loading_workers=None, ray_workers_use_nsight=False, disable_custom_all_reduce=False, block_size=None, gpu_memory_utilization=0.9, swap_space=4, kv_cache_dtype='auto', num_gpu_blocks_override=None, enable_prefix_caching=None, prefix_caching_hash_algo='builtin', cpu_offload_gb=0, calculate_kv_scales=False, disable_sliding_window=False, use_v2_block_manager=True, seed=None, max_logprobs=20, disable_log_stats=False, quantization=None, rope_scaling=None, rope_theta=None, hf_token=None, hf_overrides=None, enforce_eager=True, max_seq_len_to_capture=8192, tokenizer_pool_size=0, tokenizer_pool_type='ray', tokenizer_pool_extra_config={}, limit_mm_per_prompt={}, mm_processor_kwargs=None, disable_mm_preprocessor_cache=False, enable_lora=None, enable_lora_bias=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=None, max_prompt_adapters=1, max_prompt_adapter_token=0, device='auto', speculative_config=None, ignore_patterns=[], served_model_name=['Qwen3-235B-A22B-FP8'], qlora_adapter_name_or_path=None, show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, disable_async_output_proc=False, max_num_batched_tokens=None, max_num_seqs=None, max_num_partial_prefills=1, max_long_partial_prefills=1, long_prefill_token_threshold=0, num_lookahead_slots=0, scheduler_delay_factor=0.0, preemption_mode=None, num_scheduler_steps=1, multi_step_stream_outputs=True, scheduling_policy='fcfs', enable_chunked_prefill=None, disable_chunked_mm_input=False, scheduler_cls='vllm.core.scheduler.Scheduler', override_neuron_config=None, override_pooler_config=None, compilation_config=None, kv_transfer_config=None, worker_cls='auto', worker_extension_cls='', generation_config='auto', override_generation_config=None, enable_sleep_mode=False, additional_config=None, enable_reasoning=True, disable_cascade_attn=False, disable_log_requests=False, max_log_len=None, disable_fastapi_docs=False, enable_prompt_tokens_details=False, enable_server_load_tracking=False, dispatch_function=<function ServeSubcommand.cmd at 0x7fc28c84b6a0>)

the performce on 8*4090-48G:
INFO 05-06 03:03:16 [loggers.py:111] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7.4 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 0.0%

[zzlgreat]

8*5090d, 9.7t/s,Same issue

[thesillystudent]

same 4*A10G getting 7-8 tokens/sec on 32k max model len, actual input is around 20k tokens

[VastOcean-Yang]

so？ we need to use sglang？

[aabbccddwasd]

so？ we need to use sglang？

maybe。。。。。

[VastOcean-Yang]

So, does this mean we need to use SGLang?

When I was using vLLM 0.85, with Qwen3 235B in FP8 format, the speed was around 20 tokens. However, it suddenly dropped by about 0.7. After that, only a small number of HTTP requests could be responded to.

So, I switched the project to use SGLang, but the same problem occurred.

[VastOcean-Yang]

when i use Evalscope i get this resualt

Processing:  36%|████████████████████████████████████████████████▉                                                                                       | 9/25 [01:18<01:42,  6.41s/it]
2025-05-19 06:13:22,735 - evalscope - INFO - {
  "Time taken for tests (s)": 83.4406,
  "Number of concurrency": 5,
  "Total requests": 10,
  "Succeed requests": 10,
  "Failed requests": 0,
  "Output token throughput (tok/s)": 135.0302,
  "Total token throughput (tok/s)": 138.4218,
  "Request throughput (req/s)": 0.1198,
  "Average latency (s)": 35.9143,
  "Average time to first token (s)": 1.2511,
  "Average time per output token (s)": 0.0305,
  "Average input tokens per request": 28.3,
  "Average output tokens per request": 1126.7,
  "Average package latency (s)": 0.0309,
  "Average package per request": 1119.2
}
Processing:  60%|█████████████████████████████████████████████████████████████████████████████████                                                      | 15/25 [02:11<01:15,  7.60s/it]

After the 15 request, the background showed that the GPU had stopped computing.

Every 1.0s: nvidia-smilingma02: Mon May 19 14:17:40 2025

Mon May 19 14:17:41 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.127.08Driver Version: 550.127.08     CUDA Version: 12.4     |
|-----------------------------------------+------------------------+----------------------+
| GPU  NamePersistence-M | Bus-IdDisp.A | Volatile Uncorr. ECC |
| Fan  Temp   PerfPwr:Usage/Cap |Memory-Usage | GPU-Util  Compute M. |
|||MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA H20Off |   00000000:0F:00.0 Off |0 |
| N/A   35C    P0114W /  500W |   92400MiB /  97871MiB |  0%  Default |
|||Disabled |
+-----------------------------------------+------------------------+----------------------+
|   1  NVIDIA H20Off |   00000000:34:00.0 Off |0 |
| N/A   33C    P0114W /  500W |   90136MiB /  97871MiB |  0%  Default |
|||Disabled |
+-----------------------------------------+------------------------+----------------------+
|   2  NVIDIA H20Off |   00000000:48:00.0 Off |0 |
| N/A   36C    P0123W /  500W |   90010MiB /  97871MiB |  0%  Default |
|||Disabled |
+-----------------------------------------+------------------------+----------------------+
|   3  NVIDIA H20Off |   00000000:5A:00.0 Off |0 |
| N/A   33C    P0115W /  500W |   89178MiB /  97871MiB |  0%  Default |
|||Disabled |
+-----------------------------------------+------------------------+----------------------+
|   4  NVIDIA H20Off |   00000000:87:00.0 Off |0 |
| N/A   36C    P0120W /  500W |   91087MiB /  97871MiB |  0%  Default |
|||Disabled |
+-----------------------------------------+------------------------+----------------------+
|   5  NVIDIA H20Off |   00000000:AE:00.0 Off |0 |
| N/A   33C    P0116W /  500W |   94523MiB /  97871MiB |  0%  Default |
|||Disabled |
+-----------------------------------------+------------------------+----------------------+
|   6  NVIDIA H20Off |   00000000:C2:00.0 Off |0 |
| N/A   35C    P0117W /  500W |   96196MiB /  97871MiB |  0%  Default |
|||Disabled |
+-----------------------------------------+------------------------+----------------------+
|   7  NVIDIA H20Off |   00000000:D7:00.0 Off |0 |
| N/A   32C    P0122W /  500W |   89341MiB /  97871MiB |  0%  Default |
|||Disabled |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:|
|  GPU   GI   CIPID   Type   Process nameGPU Memory |

[thesillystudent]

For me the GPU Utilization(Not GPU memory) is considerably lower than SGLang.