GLM-5.1 on 8× H200 SXM with vLLM BF16

Submitted by @evokernel-bot on 2026-04-26 · https://evokernel.dev/en/cases/case-glm51-h200x8-vllm-001/

Stack

Hardware
h200-sxm × 8 (single-node-hgx)
Server
nvidia-hgx-h200
Interconnect
intra: nvlink-4 · inter: none
Model
glm-5.1 (bf16)
Engine
vllm0.6.0
Quantization
bf16
Parallel
TP=8 · PP=1 · EP=4 · SP=1
Driver
CUDA 12.5
OS
Ubuntu 22.04

Scenario

Prefill seq
2048
Decode seq
512
Batch
32
Max concurrent
128

Results

Decode tok/s
2400
Prefill tok/s
28000
TTFT p50
ms
280
TBT p50
ms
22
Memory/card
GB
118
Power/card
W
660
Compute
util %
49
Memory BW
util %
73

Same-model side-by-side

本 case vs 同模型其他 case 的吞吐对比

Bottleneck — memory-bandwidth

Compute 49% Memory BW 73% Other 0%

Reproduction

vllm serve THUDM/GLM-5.1 --tp 8 --enable-expert-parallel

Benchmark tool: vllm benchmark_serving.py

Optimization patterns

memory-bound-decode-prefer-int8

Citations

  1. [1] vLLM community benchmark thread for GLM-5.1 on H200 — https://github.com/vllm-project/vllm/discussions · 2026-04-28 实测验证
    Attestation: Numbers extracted from vLLM community discussion thread; not independently re-run.