[Do Not Merge] kimik2.5-fp4-b300-vllm: align server launch with B200 recipe

[RohitNagraj]

Bring the B300 fixed-seq-len server launch to parity with the B200 vLLM recipe. The B300 script (benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh) reuses the B200 recipe but was missing these serve args, so add them:

• --kv-cache-dtype fp8
• --max-cudagraph-capture-size 2048
• --max-num-batched-tokens "$((ISL * 2))"
• --stream-interval 20

Appends a perf-changelog entry.

---

Note

Low Risk
Benchmark-only launch and changelog changes; no application or auth paths, only affects how the B300 vLLM server is started for sweeps.

Overview
Tunes Kimi-K2.5 FP4 B300 fixed-seq-len vLLM launch in kimik2.5_fp4_b300.sh so serving matches the tuned B200-style recipe and documents it under kimik2.5-fp4-b300-vllm in perf-changelog.yaml.

Sets VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0 so vLLM’s CUDA-graph memory profiler does not pre-reserve GPU memory and shrink the KV cache budget under --gpu-memory-utilization 0.90. Adds --kv-cache-dtype fp8, --max-cudagraph-capture-size 8192, --max-num-batched-tokens 8192 (both aligned to 8192, not the B200 script’s 2048 / ISL*2), --stream-interval 20, and keeps --no-enable-prefix-caching on the serve line.

^{Reviewed by Cursor Bugbot for commit 993085b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sglang/tree/main/docs_new

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers.

If additional help is needed, PR authors can reach out to core maintainers over Slack.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27239289414
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27239289414

[cursor]

Missing CUDA graph profiler disable

High Severity

The B300 launch adds --max-cudagraph-capture-size 2048 and related B200 serve flags but omits export VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0, which the paired kimik2.5_fp4_b200.sh sets before the same CUDA-graph tuning with --gpu-memory-utilization 0.90.

 

^{Reviewed by Cursor Bugbot for commit 62c377e. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

There are 2 total unresolved issues (including 1 from previous review).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 3274c53. Configure here.}

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27248802525
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27248802525

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27250490190
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27250490190