Skip to content

[Klaud Cold] Update kimik3-fp4-mi355x-atom-agentic-mtp ATOM image to kimi_k3_agentic_0924 - #3456

Open
functionstackx wants to merge 1 commit into
mainfrom
klaud/kimik3-fp4-mi355x-atom-agentic-mtp-0924
Open

functionstackx wants to merge 1 commit into
mainfrom
klaud/kimik3-fp4-mi355x-atom-agentic-mtp-0924

Conversation

@functionstackx

Copy link
Copy Markdown
Collaborator

Summary

Update the ATOM image from rocm/atom-dev:ubuntu24.04_py3.12_pytorch_release_2.10.0_kimi_k3_agentic_0911 to …_kimi_k3_agentic_0924 (latest Kimi-K3 AgentX build on Docker Hub, pushed 2026-09-24) for the Kimi-K3 FP4 MI355X ATOM AgentX config.

Recipes touched: kimik3-fp4-mi355x-atom-agentic-mtp

Changes:

  • configs/amd-master.yaml: image bump. This also moves the DCP8 LMCache arms, which still run the legacy kimik3_fp4_mi355x_atom_mtp.sh script (srtctl reserves ATOM's kv-transfer-config for disaggregated workers).
  • srt-slurm-recipes/kimik3/atom/mi355x-fp4-mtp/agentic.yaml: container bump for the TP8 GPU-resident arm (conc 1, 4). The shared atomesh frontend container_image digest is unchanged.
  • perf-changelog.yaml: new entry.

Image-only. #3407 targets the same config and carries ATOM env tuning (FlyDSL FP8 prefill attention, prefill-decode interval from conc 16); that tuning is not included here.

Test plan

  • full-sweep-enabled sweep passes.

🤖 Generated with Claude Code

…kimi_k3_agentic_0924

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@github-actions

Copy link
Copy Markdown
Contributor

Thanks for the contribution!

  • Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
  • PR verification: Sweeps only run on labeled PRs. Add full-sweep-fail-fast (strongly recommended); use full-sweep-enabled only when matrix jobs should continue after a failure.
  • After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.
中文

感谢你的贡献!

  • **审阅:**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER,请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md,再通过 Slack 联系核心维护者。必须严格遵循模板,并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have,才能触发签核验证。
  • **PR 验证:**扫描仅在带有标签的 PR 上运行。强烈建议添加 full-sweep-fail-fast;仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。
  • **合并后:**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复;参见重新运行失败任务的说明。

@claude claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.

Beyond the inline pr-link finding, I checked whether this speculative-decoding (MTP) image bump needed a draft-model-precision justification paragraph in the perf-changelog entry itself, since the prior entry for this same recipe included one — it doesn't: CONTRIBUTING.md's Draft-model precision requirement is enforced via the CODEOWNER sign-off's additional-detail section, not via changelog-entry text, so the shorter new entry isn't itself a defect on that count.

Extended reasoning...

The diff is a coupled container-image tag bump across the srt-slurm recipe YAML and configs/amd-master.yaml (kept identical as required) plus one appended perf-changelog.yaml entry; no security-sensitive code paths are touched. The confirmed, already-queued inline finding is that the new entry's pr-link is the literal placeholder PRLINK_PLACEHOLDER rather than a real PR URL, breaking traceability that every other entry in the file provides. I additionally verified, from CONTRIBUTING.md's Draft-model precision section, that the missing per-entry draft-precision justification (present in the prior entry but absent here) is not a required field of perf-changelog.yaml — that confirmation lives in the CODEOWNER checklist's additional-detail section instead — so it is not a separate bug in this diff.

Additional findings (outside the current diff — GitHub can't attach inline comments there):

  • 🟡 perf-changelog.yaml — The new perf-changelog entry ships with pr-link: PRLINK_PLACEHOLDER instead of a real PR URL, so anyone auditing perf history via pr-link can't trace this image bump back to its PR. Every other entry in the file uses a real https://github.com/.../pull/N link (see line 8965 above), and CONTRIBUTING.md's template at CONTRIBUTING.md:190 also expects that format. Fix: replace PRLINK_PLACEHOLDER with the actual merged PR URL before/at merge time, e.g. by a pre-merge check or bot substitution that fails the build if the literal placeholder reaches main.

    Why this was flagged

    The entry appended at perf-changelog.yaml:8964-8971 has pr-link set to the literal string PRLINK_PLACEHOLDER (line 8971) rather than a URL. This file is described in AGENTS.md as append-only and used to trace which PR caused a perf change. On the base branch, and in every prior entry (e.g. line 8965's pull/3419), pr-link resolves to a real GitHub PR. If this literal reaches main unmodified, tooling or humans grepping perf-changelog.yaml for provenance get a dead placeholder instead of a working link, breaking traceability for this image bump.

    Verification: nit. The new perf-changelog entry ends with a literal placeholder rather than a real PR URL: perf-changelog.yaml:8970 pr-link: PRLINK_PLACEHOLDER. This is a real, present defect — the diff appends exactly this line (git diff shows + pr-link: PRLINK_PLACEHOLDER), and it is the only occurrence of that token in any YAML in the repo (grep confirms one hit). Every prior entry resolves pr-link to…

@gbyu-amd

Copy link
Copy Markdown
Collaborator

The image has been updated to rocm/atom-dev:nightly_202609251613 in #3407.

@github-actions

Copy link
Copy Markdown
Contributor

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

Status: No status

Development

Successfully merging this pull request may close these issues.

2 participants