[DNM][AMD] agentx-v0.4

[seungrokj]

Summary

• Add minimaxm2.5-fp4-mi355x-vllm-agentic-lmcache and kimik2.5-fp4-mi355x-vllm-agentic-lmcache entries in amd-master.yaml
• Add minimaxm2.5_fp4_mi355x.sh agentic benchmark script with LMCache support
• Refactor kimik2.5_fp4_mi355x.sh: simplify env vars, build LMCache from source (ROCm HIP), tune LMCACHE_L1_SIZE_GB/TTL/chunk size
• Refactor qwen3.5_fp8_mi355x.sh: add HiCache offloading support, add 256k trace corpus cap via WEKA_LOADER_OVERRIDE
• Set LMCACHE_CHUNK_SIZE default to 32 for MiniMax agentic script

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes how long-running Slurm agentic jobs install KV offload (git-built LMCache, larger DRAM knobs with TODOs) and alter CI sweep matrices; misconfiguration could waste cluster time or skew comparability, but there is no production serving or auth impact.

Overview
Extends AMD master CI with new agentic-coding targets: minimaxm2.5-fp4-mi355x-vllm-agentic-lmcache (TP=1, GPU-only vs lmcache) and kimik2.5-fp4-mi355x-vllm-agentic-lmcache (TP=4, same pair, pinned to vLLM 0.21.0). The existing Qwen3.5 agentic HiCache entry bumps the SGLang image, moves the sweep to TP=4, and aligns hicache concurrency with the none grid instead of a high-conc-only hicache slice.

kimik2.5_fp4_mi355x.sh is largely rewritten: it drops the large inline ROCm LMCache / vLLM monkey-patch bundle (demand pinning, chunked KV load, scheduler KV-xfer fix, NIXL/CuPy guards) in favor of cloning and HIP-building LMCache at run time, scales CPU/L1 pools by TP (3 TB node budget), lengthens L1 read TTL, caps traces via WEKA_LOADER_OVERRIDE, and tunes vLLM (fp8 KV, block size 1, no explicit max-model-len). A new minimaxm2.5_fp4_mi355x.sh follows the same LMCache/CPU-offload pattern with MiniMax-specific server flags (e.g. chunk size 32, ROCM_AITER_FA, block size 32).

qwen3.5_fp8_mi355x.sh gains an OFFLOADING switch (none vs hicache with sized CPU hierarchical pools, warmup skip, graph cap) and the same 256k trace override, plus stricter required env vars for the agentic matrix.

^{Reviewed by Cursor Bugbot for commit fc0d0d4. 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]

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]

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.

[cursor]

Cursor Bugbot has reviewed your changes and found 3 potential issues.

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

^{Reviewed by Cursor Bugbot for commit dc999ef. Configure here.}

[cursor]

Trace override on wrong script

Medium Severity

The 256k WEKA_LOADER_OVERRIDE was added only in qwen3.5_fp8_mi355x.sh, but MI355X SGLang jobs prefer qwen3.5_fp8_mi355x_sglang.sh when it exists. Standard launches therefore skip the new corpus cap the PR describes for Qwen agentic HiCache.

 

^{Reviewed by Cursor Bugbot for commit dc999ef. Configure here.}

[cursor]

Missing AITER RMSNorm disable

Medium Severity

The new MiniMax agentic script enables AITER but never disables AITER RMSNorm for TP below 8. The matrix runs tp: 1, and the Kimi MI355X agentic script explicitly turns off VLLM_ROCM_USE_AITER_RMSNORM for accuracy at low TP.

 

^{Reviewed by Cursor Bugbot for commit dc999ef. Configure here.}