[Klaud Cold] Add minimaxm3-fp4-mi355x-atom (upstream branch for full-sweep validation)#1813
Merged
Merged
Conversation
Smoke-tested on MI355X (mia1-p01-g07): TP4 conc-1 1k1k served and benched clean (mean TPOT 6.8ms). KV cache left at default dtype — amd/MiniMax-M3-MXFP4 has no calibrated FP8 KV scales, so --kv_cache_dtype fp8 asserts in the MSA fused_qknorm kernel.
andyluo7
requested review from
1am9trash,
billishyahao,
chunfangamd,
seungrokj and
yctseng0211
as code owners
Collaborator
Author
|
/sweep test-config --config-keys minimaxm3-fp4-mi355x-atom --config-files .github/configs/amd-master.yaml |
Contributor
|
@andyluo7 Kicking off a sweep. Run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27718177974 |
Contributor
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27732861480 |
Collaborator
|
/reuse-sweep-run |
Collaborator
|
/merge-prs |
Collaborator
|
@functionstackx @cquil11 can you plz approve this ? |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Upstream-branch mirror of #1812 (originally from
indianspeedster/InferenceMAX:feat/minimaxm3-fp4-mi355x-atom) so the GPU full-sweep PR validation can run — fork PRs can't access the self-hosted runners/secrets thatrun-sweep.ymlneeds.Original author: @indianspeedster. Mirrors PR #1812 at commit
a68c303(includes the Bugbot fix "use matrix MAX_MODEL_LEN").Summary
Adds the
minimaxm3-fp4-mi355x-atomconfig — MiniMax-M3 MXFP4 (amd/MiniMax-M3-MXFP4) on MI355X, single-node atom engine — for the 1k/1k and 8k/1k fixed-seq-len cells, TP4. Follows the ROCm/ATOM MiniMax-M3 recipe (FP4 on 4×MI355 section)..github/configs/amd-master.yaml: new config entry + search space (TP4, conc 1→128, imagerocm/atom-dev:M3).benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh: atom serve script —--block-size 128(mandatory for MiniMax MSA),--gpu-memory-utilization 0.8,--trust-remote-code,--max-model-len $MAX_MODEL_LEN. KV cache left at default dtype (MXFP4 checkpoint ships no calibrated FP8 KV scales).runners/launch_mi355x-amds.sh: routeamd/MiniMax-M3*weights to the NFS cache.Validation
generate_sweep_configs.py test-config→ 16 configs (minimaxm3_1k1k,minimaxm3_8k1k, TP4 conc 1–128).Adding
full-sweep-enabledto run the full PR validation sweep.Closes/supersedes #1812 once validated.
Note
Low Risk
Benchmark and CI config only; no changes to auth, data handling, or production serving paths.
Overview
Adds a day-zero
minimaxm3-fp4-mi355x-atomsweep for MiniMax-M3 MXFP4 (amd/MiniMax-M3-MXFP4) on MI355X using the ATOM engine, aligned with the ROCm/ATOM MiniMax-M3 recipe (TP4,--block-size128 for MSA)..github/configs/amd-master.yamldefines fixed-seq-len cells at 1k/1k and 8k/1k with TP4 and concurrency 1→128, imagerocm/atom-dev:M3.benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.shstartsatom.entrypoints.openai_serverwith matrixMAX_MODEL_LEN, mandatory block size 128, and default KV cache dtype (no FP8 KV — the MXFP4 checkpoint has no calibrated scales). Optional eval and standard serving benchmark follow.runners/launch_mi355x-amds.shroutesamd/MiniMax-M3*weights to the NFS Hugging Face cache, same as existing MiniMaxAI M3 paths.perf-changelog.yamldocuments the new config key.Reviewed by Cursor Bugbot for commit c251a03. Bugbot is set up for automated code reviews on this repo. Configure here.