Skip to content

[AMD] Refresh DeepSeek-R1 MI355X SGLang image / [AMD] 更新 DeepSeek-R1 MI355X SGLang 镜像 - #2691

Open
Oseltamivir wants to merge 2 commits into
mainfrom
refresh/dsr1-mi355x-sglang-image
Open

[AMD] Refresh DeepSeek-R1 MI355X SGLang image / [AMD] 更新 DeepSeek-R1 MI355X SGLang 镜像#2691
Oseltamivir wants to merge 2 commits into
mainfrom
refresh/dsr1-mi355x-sglang-image

Conversation

@Oseltamivir

Copy link
Copy Markdown
Collaborator

Summary

Validation

  • git diff --check
  • Parsed configs/amd-master.yaml, configs/runners.yaml, and perf-changelog.yaml with PyYAML.
  • Generated the exact config key with generate_sweep_configs.py test-config; all emitted points use the new image and preserve the existing topology/search space.
  • Verified the Docker Hub tag and image digest.
  • Verified the image-pinned AITER commit descends from both required upstream fixes.

中文说明

验证

  • git diff --check
  • 使用 PyYAML 成功解析 configs/amd-master.yamlconfigs/runners.yamlperf-changelog.yaml
  • 使用 generate_sweep_configs.py test-config 生成目标配置;所有生成点均使用新镜像,并保留原有拓扑和搜索空间。
  • 已验证 Docker Hub 标签与镜像摘要。
  • 已确认镜像固定的 AITER commit 包含上述两个上游修复。

将 DeepSeek-R1 MI355X SGLang 分离式 MTP 配置升级到包含 AITER 准确性修复的新镜像,并追加性能变更日志。
@github-actions

Copy link
Copy Markdown
Contributor

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

将性能变更日志链接更新为替代 PR #2691
@Oseltamivir
Oseltamivir marked this pull request as ready for review August 20, 2026 05:28
@Oseltamivir
Oseltamivir requested a review from a team August 20, 2026 05:28
@Oseltamivir Oseltamivir added full-sweep-fail-fast all-evals Expand eval selection to every fixed-sequence config labels Aug 20, 2026
@claude

claude Bot commented Aug 20, 2026

Copy link
Copy Markdown
Contributor

Claude finished @Oseltamivir's task in 1m 52s —— View job


PR Review

  • Gather context and read the PR diff
  • Validate configs/amd-master.yaml change (image format, config key, prefix)
  • Validate perf-changelog.yaml entry (present, appended at end of file)
  • Verify the new image tag reference
  • Post review summary

LGTM - no blocking issues found

Review notes:

  • The change is a single image bump for dsr1-fp4-mi355x-sglang-disagg-8k1k-mtp (configs/amd-master.yaml:772); model/topology/search-space are untouched.
  • perf-changelog.yaml was updated in the same PR, the new entry is correctly appended at the very end of the file, references the right config key, and links back to this PR.
  • The new image lmsysorg/sglang-rocm:v0.5.17-rocm720-mi35x-20260820 uses the public Docker Hub format consistent with existing lmsysorg/sglang-rocm tags in the repo. (I could not independently query Docker Hub from this sandbox — network access is not permitted — but the author reports having verified the tag and digest.)
  • Config key prefix dsr1 and model-prefix: dsr1 are valid.
  • The old tag remains only in configs/deprecated/ files, which this PR intentionally does not touch.
  • The full-sweep-fail-fast + all-evals labels are already applied, which is appropriate since this bump targets an accuracy fix (moe_kernels.py, stage-2 reduce path: small SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK silently corrupts output (gsm8k=0) at low concurrency sgl-project/sglang#27194) — the eval results from the sweep will be the real confirmation.

@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.

LGTM — straightforward image tag bump with no bugs found.

What was reviewed: confirmed the diff touches only the dsr1-fp4-mi355x-sglang-disagg-8k1k-mtp image tag in configs/amd-master.yaml and the matching perf-changelog.yaml entry; no other search-space, topology, or launch-flag fields changed.

Extended reasoning...

Overview

The PR bumps the Docker image for the dsr1-fp4-mi355x-sglang-disagg-8k1k-mtp recipe in configs/amd-master.yaml from lmsysorg/sglang-rocm:v0.5.12.post1-rocm720-mi35x-20260529 to v0.5.17-rocm720-mi35x-20260820, and adds a corresponding entry to perf-changelog.yaml documenting the AITER fixes picked up (masked-reduce and persistent-MLA synchronization) that address a known accuracy-corruption issue tracked upstream. This is a config-only change touching 2 files with a net diff of one modified line and one new changelog block; no launch scripts, topology, or search-space parameters are touched.

Security risks

None. This is a data-only YAML config change (image tag string + changelog metadata). No code execution paths, credentials, or user input handling are affected.

Level of scrutiny

Low. This is the standard, frequently-repeated 'refresh inference-engine image' pattern used throughout this repo (e.g. the adjacent Qwen3.5 and MiniMax refreshes visible in the same file). The change is mechanical, self-contained, and validated by the PR's own described process (YAML parsing, sweep-config generation, image digest/AITER-commit verification). The label is applied, so CI will exercise the new image against the existing search space before merge.

Other factors

The bug-hunting pass found no issues, and my own reading of the diff confirms only the image tag and changelog changed — the topology (prefill/decode worker counts, tp/ep, dp-attn, concurrency lists) is untouched, matching the PR description's claim of preserving the existing search space.

@github-actions

Copy link
Copy Markdown
Contributor

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

all-evals Expand eval selection to every fixed-sequence config full-sweep-fail-fast

Projects

Status: No status

Development

Successfully merging this pull request may close these issues.

1 participant