Repository navigation
[AMD][MI355X] DSv4.1 Flash vLLM: re-sweep with SWA bounded replay and --monokernel on the latest ROCm 10.0 nightly / [AMD][MI355X] DSv4.1 Flash vLLM：在最新 ROCm 10.0 nightly 上启用 SWA bounded replay 与 --monokernel 并重测 - #3959
Fangzhou-Ai wants to merge 2 commits into
Conversation
…replay and --monokernel / DSv4.1-Flash MI355X vLLM AgentX 启用 SWA bounded replay 与 --monokernel 并重测 Bump the image to nightly-rocm100-5555a2d8f7116b2b4877f66516f4128176cf722a, which contains vllm-project/vllm#59039 and #61082. Drop --no-swa-bounded-replay and replace VLLM_ROCM_MONO_DECODE=1 with --monokernel. 镜像更新至 nightly-rocm100-5555a2d8f7116b2b4877f66516f4128176cf722a，包含 vllm-project/vllm#59039 与 #61082。移除 --no-swa-bounded-replay，并用 --monokernel 取代 VLLM_ROCM_MONO_DECODE=1。 Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…angelog 条目指向 PR 3959 Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
|
Thanks for the contribution!
中文感谢你的贡献！
|
Fangzhou-Ai
added
full-sweep-fail-fast
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=38377181342 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=38377181342 |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Looks good, straightforward config change.
What was reviewed: the recipe image bump and its model.container/amd-master.yaml image sync, the flag swap from no-swa-bounded-replay to monokernel: true and removal of the now-defunct VLLM_ROCM_MONO_DECODE env var, and the appended perf-changelog.yaml entry (byte-append at tail, matches the established format/style of prior entries, correct PR link).
Extended reasoning...
Three-file, config-only diff: a srt-slurm recipe YAML, its matching amd-master.yaml entry, and an appended perf-changelog.yaml block for a DeepSeek-V4.1-Flash MI355X vLLM benchmark recipe. No source code, auth, or security-sensitive surface is touched. model.container and image stay in sync as required, the changelog entry is appended only at the tail in the same style as neighboring entries, and the flag/env-var swap is internally consistent with the PR's stated rationale (moving mono-decode control from an env var to a KernelConfig flag). The bug-hunting run reported zero bugs and the one candidate issue it considered was ruled out, and the change is mechanical and self-contained enough that no human look is strictly required.
This review covers commit 5244c2a, which is no longer the latest commit on this pull request; later commits are not covered by it.
|
recipe PR vllm-project/recipes#1104 |
This branch has not been deployed
1 participant
Follow-up to #3690. Re-sweeps the MI355X DeepSeek-V4.1-Flash vLLM DSpark AgentX recipe (
dsv41flash-fp4-mi355x-vllm-agentic-dspark) now that both vllm-project/vllm#59039 (decoder SWA bounded replay on ROCm) and vllm-project/vllm#61082 (mono decode on every backbone layer) are merged.Changes
nightly-rocm100-7d0b4e57aac4c4323225b0eee9ae960ce75ffffe→nightly-rocm100-5555a2d8f7116b2b4877f66516f4128176cf722a, the latest ROCm 10.0 nightly (pushed 2026-10-13 07:02 UTC). It is 6 commits past #59039's merge commit (c077dc663d92) and 8 past #61082's (cc2a6eea1a9f), so it contains both.model.containermatches.no-swa-bounded-replay: true. The flag was added because the ROCm sparse SWA path lacked the replay window clamp. #59039 enables bounded replay on ROCm, so it is on by default again at every point.VLLM_ROCM_MONO_DECODE: '1'withmonokernel: true(--monokernel). #61082 moves the switch toKernelConfig.monokerneland removes the environment variable (no match invllm/envs.pyat the pinned commit), so keeping the old setting would silently turn the kernels off. #61082 also runs all 40 backbone layers whole as two launches (30 before). The 48-row step limit is unchanged, so concurrency 1-8 with five drafts takes the mono path.For reviewers
infx.matrix.generate test-configexpands to the 14 points on the new image;validate_perf_changelogpasses.中文
作为 #3690 的后续，在 vllm-project/vllm#59039（ROCm 上的 decoder SWA bounded replay）与 vllm-project/vllm#61082（所有主干层启用 mono decode）均已合入后，重新测试 MI355X DeepSeek-V4.1-Flash vLLM DSpark AgentX 配置（
dsv41flash-fp4-mi355x-vllm-agentic-dspark）。改动
nightly-rocm100-7d0b4e57aac4c4323225b0eee9ae960ce75ffffe→nightly-rocm100-5555a2d8f7116b2b4877f66516f4128176cf722a，即最新的 ROCm 10.0 nightly（2026-10-13 07:02 UTC 推送）。它在 #59039 合入提交（c077dc663d92）之后 6 个提交、#61082 合入提交（cc2a6eea1a9f）之后 8 个提交，因此同时包含两者。model.container同步更新。no-swa-bounded-replay: true。该参数当初是因为 ROCm sparse SWA 路径缺少 replay 窗口钳制而加入的。#59039 在 ROCm 上启用了 bounded replay，因此所有点默认重新开启。VLLM_ROCM_MONO_DECODE: '1'替换为monokernel: true（--monokernel）。#61082 把开关改为KernelConfig.monokernel并移除了该环境变量（固定提交的vllm/envs.py中已无此项），保留旧设置会悄悄关闭内核。#61082 还让全部 40 个主干层整层走两次内核启动（此前为 30 层）。48 行的 decode 步上限不变，因此 5 个草稿 token 下并发 1-8 走 mono 路径。评审注意
infx.matrix.generate test-config在新镜像上展开为 14 个点；validate_perf_changelog通过。AI model disclosure
Prepared with Claude Sonnet 5.5 (
claude-sonnet-5-5) in Claude Code: the recipe, master-config and changelog edits, the nightly tag lookup and ancestry check against vllm#59039 and #61082, and this description.🤖 Generated with Claude Code