[WIP] [NV] Kimi fp4 configs

[ankursingh-nv]

Summary

Add Kimi K2.5 FP4 benchmark configuration for B200 using vLLM.

Changes

• New config kimik2.5-fp4-b200-vllm in nvidia-master.yaml
  • Model: nvidia/Kimi-K2.5-NVFP4
  • Image: vllm/vllm-openai:v0.16.0
  • Parallelism: TP=8/EP=1 (conc 4-128) and TP=4/EP=4 (conc 4-64)
  • Sequence lengths: 1k1k, 1k8k, 8k1k
• New benchmark script benchmarks/single_node/kimik2.5_fp4_b200.sh
  • Uses --reasoning-parser kimi_k2 and --tool-call-parser kimi_k2
  • Enables --compilation_config.pass_config.fuse_allreduce_rms true
  • Sets TORCH_CUDA_ARCH_LIST="10.0" for B200
  • GPU memory utilization: 0.90

[ankursingh-nv]

@claude update PR description and also add an entry to perf-changelog file.

[Klaud-Cold]

Claude finished @ankursingh-nv's task in 1m 27s —— View job

---

Completed

• Gather context on PR changes
• Update PR description
• Add entry to perf-changelog.yaml
• Commit and push changes

Done:

• Updated PR description with a summary of config and script changes (model, image, parallelism, sequence lengths, key flags)
• Added kimik2.5-fp4-b200-vllm entry to perf-changelog.yaml referencing this PR
  |

[functionstackx]

@claude update the PR description to match the acutal vllm image being used

[Klaud-Cold]

Claude finished @functionstackx's task in 44s —— View job

---

Completed

• Gather context on actual vLLM image used
• Update PR description to match

Done: Updated the PR description to reflect the actual vLLM image vllm/vllm-openai:v0.16.0 (was incorrectly listed as v0.15.1).

[functionstackx]

thanks for the PR! overview LGTM!

@ankursingh-nv @kedarpotdar-nv 1 small thing is if you can add documentations about nvfp4 version of kimi k2.5 nvidia/Kimi-K2.5-NVFP4 to the vllm recipes https://github.com/vllm-project/recipes/blob/main/moonshotai/Kimi-K2.5.md . Lets ensure that the documentation is first class such that the entire ml community can benefit from your hard work!

+viz @faradawn

[kedarpotdar-nv]

started PR here vllm-project/recipes#267

[cquil11]

@ankursingh-nv what is the hold up here? probably should update to vllm 0.17.0 before merging anyway since its up now