AMD][MI35X] Qwen3.5-fp8 SGLang mtp single-node benchmark#1671
AMD][MI35X] Qwen3.5-fp8 SGLang mtp single-node benchmark#16711am9trash wants to merge 4 commits into
Conversation
1am9trash
requested review from
chunfangamd,
seungrokj and
yctseng0211
as code owners
|
Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook 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 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[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 50efc8f. Configure here.
| if [ "${EVAL_ONLY}" = "true" ]; then | ||
| setup_eval_context | ||
| EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN" | ||
| else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH" |
There was a problem hiding this comment.
Missing throughput context length
High Severity
For normal benchmark runs (EVAL_ONLY not true), the server launch no longer receives --context-length. The script drops CONTEXT_LENGTH and the else branch that passed ISL + OSL + 20, so EVAL_CONTEXT_ARGS stays empty while throughput jobs still use fixed ISL/OSL from CI.
Reviewed by Cursor Bugbot for commit 50efc8f. Configure here.
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
LGTM — benchmark image bump and AITER launch-flag retune for a single AMD MI355X recipe.
Extended reasoning...
Overview
This PR touches only AMD MI355X benchmark configuration for the qwen3.5-fp8-mi355x-sglang-mtp recipe: a docker image bump in .github/configs/amd-master.yaml (...20260517 → v0.5.12.post1...20260528), a retuned launcher script (qwen3.5_fp8_mi355x_mtp.sh) switching to the AITER attention backend with adjusted serving flags, and a corresponding entry in perf-changelog.yaml.
Security risks
None. No production code paths, auth, crypto, or permissions are touched. Changes are confined to benchmark recipes that run in CI sweeps.
Level of scrutiny
Low. This is a pure benchmark-config change scoped to a single recipe key — well-established pattern for image/flag retuning PRs in this repo. The dropped --context-length (non-EVAL path), --max-prefill-tokens, and --cuda-graph-max-bs are intentional per the PR description and replaced by --chunked-prefill-size and --max-running-requests.
Other factors
- No bugs reported by the bug hunting system.
- Changes are self-contained and follow the same pattern as adjacent recipe entries in the same file.
- The recipe reminder bot comment is the standard reminder, not a blocker.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26993294925 |
2 participants
Note
Low Risk
Benchmark-only image and launch-flag changes; no application or security-sensitive code paths.
Overview
Updates the Qwen3.5 FP8 MI355X SGLang MTP single-node benchmark to a newer ROCm image and retunes the launch recipe for AITER.
The
qwen3.5-fp8-mi355x-sglang-mtpconfig in.github/configs/amd-master.yamlnow points atlmsysorg/sglang-rocm:v0.5.12.post1-rocm720-mi35x-20260528(was...20260517).benchmarks/single_node/fixed_seq_len/qwen3.5_fp8_mi355x_mtp.shenablesSGLANG_USE_AITER/ unified AITER attention, switches--attention-backendfrom triton to aiter, and adjusts serving limits: fixed--max-running-requests 512,--chunked-prefill-size 32768,--mem-fraction-static 0.9,--page-size 16, multithreaded model load—while dropping the prior ISL/OSL-derived--context-length,--max-prefill-tokens, and concurrency-tied--cuda-graph-max-bs.perf-changelog.yamldocuments the config key change.Reviewed by Cursor Bugbot for commit 50efc8f. Bugbot is set up for automated code reviews on this repo. Configure here.