perf(agentx): refresh Kimi-K3 MI355X LMCache curve / 刷新 Kimi-K3 MI355X LMCache 曲线

[hyukjlee]

Description / 说明

• Refreshes the canonical kimik3-fp4-mi355x-vllm-agentic-mtp curve while preserving the tuning established by merged PR perf(agentx): retune Kimi-K3 FP4 MI355X vLLM agentic MTP recipe #2787.

• Bumps vLLM ROCm to nightly-7c5dc571cbd1064ecc8a9b1045637ff647aa22cb.

• Pins LMCache to 0.5.5rc3+rocm7.2 and enables LMCache DRAM KV offload for C1, C8, C14, C40, C44, and C48.

• Keeps C1/C8/C14 on TP8 with DSpark MTP and sets --gpu-memory-utilization 0.88 for all six points.

• Uses the C40 TP8/DCP8 no-spec profile for C40/C44/C48: LMCache chunk size 12288, DCP KV-cache interleave size 1536, max-num-seqs 80, 16K batched tokens, and full CUDA graphs through size 4096.

• 刷新标准的 kimik3-fp4-mi355x-vllm-agentic-mtp 曲线，并保留已合并 PR perf(agentx): retune Kimi-K3 FP4 MI355X vLLM agentic MTP recipe #2787 中验证过的调优配置。

• 将 vLLM ROCm 镜像更新到 nightly-7c5dc571cbd1064ecc8a9b1045637ff647aa22cb。

• 固定 LMCache 版本为 0.5.5rc3+rocm7.2，并为 C1、C8、C14、C40、C44 和 C48 启用 LMCache DRAM KV offload。

• C1/C8/C14 保持 TP8 和 DSpark MTP 配置，全部六个测试点统一使用 --gpu-memory-utilization 0.88。

• C40/C44/C48 使用相同的 TP8/DCP8 无投机解码配置：LMCache chunk size 12288、DCP KV-cache interleave size 1536、max-num-seqs 80、16K batched tokens，以及最大到 4096 的完整 CUDA graph。

Duplicate-work check / 重复工作检查

This does not duplicate open PR #2795, which removes LMCache and targets a different no-offload C52 profile. It also supersedes the old LMCache 0.5.4rc work in PR #2598 by updating the already-merged canonical PR #2787 recipe to LMCache rc3 and the new vLLM image.

本 PR 不与开放的 PR #2795 重复；该 PR 移除了 LMCache，并针对不同的无 offload C52 配置。本 PR 也取代了 PR #2598 中旧的 LMCache 0.5.4rc 工作，将已合并 PR #2787 的标准配置更新到 LMCache rc3 和新的 vLLM 镜像。

Type of Change / 变更类型

• Bug fix / 缺陷修复
• New feature / 新功能
• Configuration change / 配置变更
• Documentation update / 文档更新
• Other / 其他

Checklist / 检查清单

• I have tested my changes locally / 已在本地测试变更
• I have updated documentation if necessary / 已按需更新文档
• For every performance-affecting benchmark change, I appended an entry to the physical end of perf-changelog.yaml without editing historical entries / 每项影响性能的基准变更均已在 perf-changelog.yaml 文件末尾追加记录，且未修改历史条目
• Before merging via reuse, an authorized maintainer has commented /reuse-sweep-run after a final all-green sweep with evals passing / 通过复用方式合并前，授权维护者需在最终完整 sweep 和 eval 全部通过后评论 /reuse-sweep-run

AI assistance was used to prepare and validate this change. The submitting human must review every changed line and confirm the benchmark results before merge.

本变更使用了 AI 辅助。提交者在合并前必须审阅每一处改动，并确认基准测试结果。

---

Note

Low Risk
Benchmark and sweep configuration only; no production runtime or auth paths change, though refreshed numbers will replace the published Kimi-K3 MI355X curve.

Overview
Refreshes the kimik3-fp4-mi355x-vllm-agentic-mtp AgentX curve: bumps the vLLM ROCm image to nightly-7c5dc571, pins LMCache to 0.5.5rc3+rocm7.2, and records the sweep in perf-changelog.yaml.

amd-master.yaml narrows the MTP + LMCache DRAM offload matrix to C1, C8, C14 and adds TP8/DCP8 points at C40, C44, C48 with spec-decoding: none. A thin kimik3_fp4_mi355x.sh wrapper sets SPEC_DECODING=none and delegates to the main script.

kimik3_fp4_mi355x_mtp.sh keys tuning on ${SPEC_DECODING}:${CONC}: MTP arms use gpu-memory-utilization 0.88; high-concurrency no-spec runs get a dedicated server profile (fixed max-num-seqs 80, expanded CUDA-graph capture through 4096, --stream-interval, --prefix-match-unit 128, prefill FP8 quant in attention config, extra compilation custom ops). LMCache uses chunk-size 12288, exports backend metadata, and a lighter import sanity check; with DCP and LMCache, --cp-kv-cache-interleave-size 1536, ROCM_AITER_MLA, and related env flags replace the prior TRITON-only DCP path.

^{Reviewed by Cursor Bugbot for commit a720a04. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33614613326
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=33614613326

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33618719560
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=33618719560

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33630491494
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=33630491494

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33631260867
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=33631260867

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33631260867
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=33631260867

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33631260867
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=33631260867