[ROCm] Document Kimi-K2.5 MXFP4 MI355X benchmark reproduction

[haic0]

Summary

• Add a Kimi-K2.5 MXFP4 MI355X benchmark reproduction subsection based on SemiAnalysisAI/InferenceX PR #2213.
• Replace the older fixed-sequence throughput note with the PR #2213 TP4, fp8 KV, scheduler, and AITER env settings.
• Keep the default Kimi-K2.5 MXFP4 serving command unchanged; document the InferenceX lane as a separate high-concurrency benchmark configuration.

Test plan

• python YAML parse and top-level schema-order check for models/moonshotai/Kimi-K2.5.yaml
• git diff --check -- models/moonshotai/Kimi-K2.5.yaml
• node scripts/build-recipes-api.mjs

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview, Comment	Jul 23, 2026 8:00am

[gemini-code-assist]

Code Review

This pull request updates the Kimi-K2.5.yaml model configuration file to include detailed instructions, environment variables, and a run command for reproducing the MXFP4 benchmark on the InferenceX MI355X platform. The review feedback recommends adding a default fallback value for the $MAX_MODEL_LEN environment variable in the reproduction command to prevent execution failures if the variable is not defined.

[gemini-code-assist]

The environment variable $MAX_MODEL_LEN is not defined or exported in the preceding setup instructions. If a user copies and runs this command directly, the variable will evaluate to an empty string, which can cause the vLLM server to fail or behave unexpectedly. Providing a default fallback value like ${MAX_MODEL_LEN:-9472} ensures the command is robust and runnable out-of-the-box.

    --max-model-len "${MAX_MODEL_LEN:-9472}" \

[chunfangamd]

SemiAnalysisAI/InferenceX#2213

[functionstackx]

@haic0 @chunfangamd u gotta add ur extra envs here too or else the recipes production website doesnt show it in quick select

https://recipes.vllm.ai/moonshotai/Kimi-K2.5?hardware=mi355x&variant=mxfp4