GLM5 H200 FP8 SGL single node

[functionstackx]

@claude GLM5 H200 SGL single node FP8 TP8 & DP attention true TP8 (-tp 8 , -dp 8, u also need to enable --enable-dp-attention )

model: zai-org/GLM-5-FP8
image: lmsysorg/sglang:glm5-hopper

following the official sglang cookbook https://cookbook.sglang.io/autoregressive/GLM/GLM-5

remember above python -m sglang.launch_server in the bash script to always put the you are following the official sglang cookbook and the link to it

[Klaud-Cold]

Claude finished @functionstackx's task in 6m 22s —— View job • claude/issue-844-20260302-0039 • Create PR ➔

---

Adding GLM-5 H200 FP8 SGLang Single Node Benchmark

• Gather context and understand repository structure
• Review existing benchmark configs and scripts for patterns
• Create benchmark script for GLM-5 FP8 H200 SGLang
• Add config entry to nvidia-master.yaml
• Update perf-changelog.yaml
• Commit and push changes

Summary

Created a new GLM-5 FP8 H200 SGLang single-node benchmark with the following:

Benchmark script (benchmarks/single_node/glm5_fp8_h200.sh):

• Model: zai-org/GLM-5-FP8
• Image: lmsysorg/sglang:glm5-hopper
• --tensor-parallel-size 8, --data-parallel-size 8, --enable-dp-attention
• EAGLE speculative decoding (--speculative-num-steps 3, --speculative-num-draft-tokens 4, --speculative-eagle-topk 1)
• GLM-specific parsers: --tool-call-parser glm47, --reasoning-parser glm45
• --mem-fraction-static 0.85
• Cookbook reference comment above python -m sglang.launch_server

Config entry (glm5-fp8-h200-sglang in nvidia-master.yaml):

• TP=8, concurrency 4–64
• Sequence lengths: 1k1k, 1k8k, 8k1k

|