Repository navigation
[Klaud Cold] Update Qwen3.5 FP8 MI355X ATOM MTP image to v0.1.7 / [Klaud Cold] 将 Qwen3.5 FP8 MI355X ATOM MTP 镜像更新至 v0.1.7 - #3918
Conversation
Move qwen3.5-fp8-mi355x-atom-mtp from rocm/atom atom0.1.2.post to the digest-pinned rocm/atom-dev:nightly_202610080328 build of ATOM v0.1.7. 将 qwen3.5-fp8-mi355x-atom-mtp 从 rocm/atom atom0.1.2.post 更新为按摘要固定的 ATOM v0.1.7 构建 rocm/atom-dev:nightly_202610080328。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Thanks for the contribution!
中文感谢你的贡献！
|
|
Initial attempt · Passed · Run 38190921586 / attempt 1 · 2026-10-12 01:38 UTC
Note: All rows: request errors unavailable.
Next: Append the changelog entry and run the final full sweep. 中文初次尝试 · 已通过 · Run 38190921586 / attempt 1 · 2026-10-12 01:38 UTC |
为 Qwen3.5 FP8 MI355X ATOM MTP 的 v0.1.7 镜像更新追加 perf-changelog 条目。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Klaud-Cold
added
the
full-sweep-fail-fast
|
Final full sweep · Passed · Run 38192728482 / attempt 1 · 2026-10-12 02:41 UTC
Note: All rows: request errors unavailable.
Note: 1k/1k c4 TP4 EP1 7303f8, 1k/1k c4 TP8 EP1 dd674d, 1k/1k c8 TP4 EP1 a1a381, 1k/1k c8 TP8 EP1 a553cc, 1k/1k c16 TP4 EP1 d7c0e2, 1k/1k c16 TP8 EP1 a4d7e1, 1k/1k c32 TP4 EP1 6fb399, 1k/1k c32 TP8 EP1 f35014, 1k/1k c64 TP4 EP1 191187, 1k/1k c64 TP8 EP1 457676, 1k/1k c128 TP4 EP1 163ecc, 1k/1k c128 TP8 EP1 55e615, 1k/1k c256 TP4 EP1 50d77c, 1k/1k c256 TP8 EP1 902ea9: Single-turn 1k1k retired 2026-07-17 in #2263 after this baseline; excluded from coverage. Next: Repair the in-run JIT stalls by keeping prefix caching off, then rerun the final sweep. 中文最终完整 sweep · 已通过 · Run 38192728482 / attempt 1 · 2026-10-12 02:41 UTC |
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=38198420473 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=38198420473 |
Klaud-Cold
removed
the
full-sweep-fail-fast
ATOM v0.1.7 enables prefix caching by default; the old image did not. Cache hits on reused benchmark prompts triggered in-run AITER JIT builds. ATOM v0.1.7 默认开启前缀缓存，而旧镜像未开启。基准测试复用提示词导致的缓存命中 在测量期间触发了 AITER JIT 编译，因此保持前缀缓存关闭。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Repair 1/5 · Passed · Run 38196729419 / attempt 1 · 2026-10-12 03:05 UTC
Note: All rows: request errors unavailable.
Next: Run the final full sweep. 中文修复 1/5 · 已通过 · Run 38196729419 / attempt 1 · 2026-10-12 03:05 UTC |
Keep this PR's perf-changelog block at the physical tail after #3819. 合并 origin/main，并将本 PR 的 perf-changelog 条目保留在文件末尾。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Klaud-Cold
added
the
full-sweep-fail-fast
|
Final full sweep · Passed · Run 38198420473 / attempt 1 · 2026-10-12 03:56 UTC
Note: All rows: request errors unavailable.
Note: 1k/1k c4 TP4 EP1 7303f8, 1k/1k c4 TP8 EP1 dd674d, 1k/1k c8 TP4 EP1 a1a381, 1k/1k c8 TP8 EP1 a553cc, 1k/1k c16 TP4 EP1 d7c0e2, 1k/1k c16 TP8 EP1 a4d7e1, 1k/1k c32 TP4 EP1 6fb399, 1k/1k c32 TP8 EP1 f35014, 1k/1k c64 TP4 EP1 191187, 1k/1k c64 TP8 EP1 457676, 1k/1k c128 TP4 EP1 163ecc, 1k/1k c128 TP8 EP1 55e615, 1k/1k c256 TP4 EP1 50d77c, 1k/1k c256 TP8 EP1 902ea9: Single-turn 1k1k retired 2026-07-17 in #2263 after this baseline; excluded from coverage. Next: Mark ready for maintainer review. 中文最终完整 sweep · 已通过 · Run 38198420473 / attempt 1 · 2026-10-12 03:56 UTC |
|
validated · Repairs: 1 · Runs: 38190921586, 38192708421, 38192728482, 38196707365, 38196728215, 38196729419, 38198402715, 38198420473 中文validated · 修复次数：1 · 运行：38190921586, 38192708421, 38192728482, 38196707365, 38196728215, 38196729419, 38198402715, 38198420473 |
|
/use 38198420473 |
|
Claude finished @Klaud-Cold's task in 17s —— View job LGTM - no blocking issues found |
|
@Klaud-Cold staged run 38198420473: https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-10-12~r38198420473 This run remains available across future |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
I reviewed this PR and didn't find any bugs. Because this bumps the ATOM serving container for an MTP (speculative-decoding) recipe to an unpinned-release nightly build, and inferencex-e2e/configs/amd-master.yaml is owned by a non-core CODEOWNER group (requiring sign-off per CONTRIBUTING.md), a human/CODEOWNER look would still be worthwhile.
What was reviewed: the image-bump pairing between the recipe YAML's model.container and amd-master.yaml's image (identical new digest in both, satisfying the "must match" invariant); the new no-enable_prefix_caching: true flag as a plausible behavior-preserving addition alongside the image bump; and the perf-changelog.yaml entry (appended at the tail, correct config-key, English-only, with a PR link). Not visible in the diff or PR text is the CODEOWNER "additional detail section" CONTRIBUTING.md's draft-model-precision rule requires for MTP changes (draft checkpoint/revision, shipped precision, and confirmation the new ATOM nightly's default handling still matches it) — worth a CODEOWNER confirming before merge, since the rule explicitly treats image-only bumps as not exempt.
Extended reasoning...
Config-only diff (recipe YAML + master config + changelog) for an AMD MI355X MTP benchmark; no application code, auth, or injection surface touched. The decisive facts: amd-master.yaml carries a non-core CODEOWNER list requiring sign-off, and CONTRIBUTING.md's draft-model-precision rule specifically calls out image-only bumps on speculative-decoding recipes as requiring a reviewer-visible precision confirmation, which isn't present in what's available here, so I deferred rather than approved.
Resolve conflicts: take main's refactored recipes and re-apply this PR's non-image edits; take main's perf-changelog.yaml and re-append this PR's entry. 中文：合并 origin/main；配方采用 main 重构后的版本并重新应用本 PR 的非镜像改动，perf-changelog.yaml 采用 main 版本并在末尾重新追加本 PR 的条目。
Resolve conflicts: take main's refactored recipes and re-apply this PR's non-image edits; take main's perf-changelog.yaml and re-append this PR's entry. 中文：合并 origin/main；配方采用 main 重构后的版本并重新应用本 PR 的非镜像改动，perf-changelog.yaml 采用 main 版本并在末尾重新追加本 PR 的条目。
|
/reuse-sweep-run 38198420473 |
Goal: Update ATOM image from
rocm/atom:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.2.posttorocm/atom-dev:nightly_202610080328@sha256:9c3070bb7d2600d2e1bcfd03cf5f71e4dc0941f3e1d20d3a5845056aae2334eb.Baseline: 2026-05-22 ·
rocm/atom:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.2.postMean latency · Sources: API 1, API 2, API 3
Note: 1k/1k c4 TP4 EP1 7303f8, 1k/1k c4 TP8 EP1 dd674d, 1k/1k c8 TP4 EP1 a1a381, 1k/1k c8 TP8 EP1 a553cc, 1k/1k c16 TP4 EP1 d7c0e2, 1k/1k c16 TP8 EP1 a4d7e1, 1k/1k c32 TP4 EP1 6fb399, 1k/1k c32 TP8 EP1 f35014, 1k/1k c64 TP4 EP1 191187, 1k/1k c64 TP8 EP1 457676, 1k/1k c128 TP4 EP1 163ecc, 1k/1k c128 TP8 EP1 55e615, 1k/1k c256 TP4 EP1 50d77c, 1k/1k c256 TP8 EP1 902ea9: Single-turn 1k1k retired 2026-07-17 in #2263 after this baseline; excluded from coverage.
中文
**目标：**将 ATOM 镜像从
rocm/atom:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.2.post更新为rocm/atom-dev:nightly_202610080328@sha256:9c3070bb7d2600d2e1bcfd03cf5f71e4dc0941f3e1d20d3a5845056aae2334eb。**基线：**2026-05-22 ·
rocm/atom:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.2.post平均延迟 · 来源： API 1, API 2, API 3；数值及异常说明见上表。