Repository navigation
[ROCm][MiniMax-M3] MXFP4 mode for the fused norm / RoPE / KV insert - #64359
Closed
amd-dlimpus wants to merge 8 commits into
Closed
amd-dlimpus wants to merge 8 commits into
amd-dlimpus wants to merge 8 commits into
Conversation
… ROCm Squash of the MiniMax-M3.1 serving branch, rebased on 5d35ce3. NVIDIA: - Route NVFP4 routed experts in MXFP8 checkpoints to the CT NVFP4 MoE - NVFP4 KV cache on the MSA sparse attention path, served only by the CUTLASS Q8KV4 kernels (E4M3 query) on SM100/SM103/SM107 - NVFP4 index-K cache via the MSA Q8KV4/Q8KV8 indexers; opt-in decode context parallelism for the NVFP4 indexer at TP4 - Vendor fmha_sm100 from vllm-project/MSA dev 25c1447 - NVFP4 MSA decode past 512 sequences: Q8KV4 limits, bounded indexer and plan memory - Decode metadata before the forward: 13 ops -> 2; in-place indexer I/O, no counter reset, PDL chain for the NVFP4 sparse decode - DeepGEMM NVFP4 MegaMoE on the unified DeepGemmMegaMoEExperts layer; per-token FC1 input scales outside MegaMoE; split staging rows over programs - Fuse small-batch MoE finalize into the all-reduce + Gemma norm + MXFP8 quant - QKV+indexer MXFP8 GEMM with fused epilogue; SM100 fused dense GEMM1 - Fused QK-norm/RoPE/KV insert: persistent row-group kernel - Read aux hidden states after the deferred TP all-reduce - FA4: pass no FP8 descale for a unit per-tensor scale - Keep KV cache zeroing for mixed-precision UniformTypeKVCacheSpecs groups - Clear per-token A4 metadata for NVFP4 W4A16 MoE kernels (Marlin only) - Hand the Q8KV4 prefill a per-device split workspace - Align image preprocessing: long-side resize with detail tiers, fp32 patch embedding - NIXL KV-head re-sharding for per-head-contiguous NVFP4 pages; BLHNC layout for NIXL transfer - SimpleCPUOffload: 1 GiB HugeTLB host buffer + PreferCE batch copies Spec decode: - DSpark for MiniMax-M3.1: draft model, fixes and perf; default the draft to the checkpoint's dspark/ subfolder - Fuse the draft's q/k norm + RoPE and its context-K/V write - FLASHINFER_TS: prims_ts paged decode for non-causal draft blocks - Store the draft tokens' K/V under torch.compile - Markov sampling: one step helper, bias in the Gumbel kernel - DFlash: pad the prepared inputs in wide blocks on every platform ROCm: - Re-quantize the NVFP4 routed experts to MXFP8 (or MXFP4) at load; NVFP4 -> MXFP4 re-quantization and fused shared experts through the online quantization API - Mono decode (one FlyDSL launch per layer): native MXFP8 shared expert, top-16 sort/merge, aux hidden states, indexer CP - NVFP4 KV + NVFP4 indexer on gfx950, decode indexer CP - Tune small-M MXFP8 linears - Sequence-parallel prefill under pure TP (opt-in), with MXFP4 MoE support; TBO/DBO under pure TP; warm the microbatch threads' hipBLASLt handles at init - gfx942: avoid the load-time FP8 MoE quant crash; fall back from the fused QK-norm - Keep the NVFP4 KV-insert helpers CUDA-only so the fused kernel builds on HIP Co-authored-by: Taoyu Zhu <zhutaoyu97@gmail.com> Co-authored-by: Hongxia Yang <hongxia.yang@amd.com> Co-authored-by: qizixi <22851944+zixi-qi@users.noreply.github.com> Co-authored-by: indivats <Indira.Vats@amd.com> Co-authored-by: andyluo7 <43718156+andyluo7@users.noreply.github.com> Co-authored-by: Thien Tran <gau.nernst@yahoo.com.sg> Co-authored-by: Jee Jee Li <pandaleefree@gmail.com> Co-authored-by: fxmarty-amd <felmarty@amd.com> Co-authored-by: Jiejing Zhang <280536569+jiejingzhangamd@users.noreply.github.com> Co-authored-by: Prasun Gera <pgera@users.noreply.github.com> Co-authored-by: Dao Le <daole@inferact.ai> Co-authored-by: Isotr0py <mozf@inferact.ai> Co-authored-by: Neha Prakriya <neha.prakriya@amd.com> Co-authored-by: Summer Yang <girasoley@inferact.ai> Co-authored-by: Zhewen Li <jerven.vllm@gmail.com> Co-authored-by: Zijing Liu <liuzijing2014@users.noreply.github.com> Co-authored-by: xiaozhoupy <peiyuanzhou1994@gmail.com> Co-authored-by: v <v@inferact.ai> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
…wn PRs (vllm-project#63490) Signed-off-by: Taoyu Zhu <zhutaoyu97@gmail.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…ssage (vllm-project#63492) Signed-off-by: Seungmin Kim <8457324+ehfd@users.noreply.github.com>
Per-head MXFP4 KV: E2M1 data plus one E8M0 scale per 32 elements, stored as uint8 (head_size/2 + head_size/32 bytes per head row, 68 for 128), with its own KVQuantMode.MXFP4. Backends opt in through supported_kv_cache_dtypes. Signed-off-by: Limpus, David <dlimpus@amd.com>
Writes K, V and index-K as MXFP4 pages (E2M1 data, E8M0 scale per 32 dims, no global scale; 68 B per token and head) from the same FlyDSL launch as the NVFP4 insert. The data layout is NVFP4's; K and index-K scales are [token][4], V scales [group][token]. The scale is 2^ceil(log2(amax / 6)), 2^0 for an all-zero group. The bit-exact test runs both formats against a torch reference. Signed-off-by: Limpus, David <dlimpus@amd.com>
|
👋 Hi! Thank you for contributing to the vLLM project. 💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in PRs do not trigger a full CI run by default. Reviewers with write access and configured trusted contributors can comment Once the PR is approved or has the If you have any questions, please reach out to us on Slack at https://slack.vllm.ai. Agent GuidelinesIMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban. 🚀 |
15 is MXFP4_DS_MLA in vllm-project#64120 and 16 is MXFP4 in vllm-project#64198. Signed-off-by: Limpus, David <dlimpus@amd.com> (cherry picked from commit b3a9d3e)
zyongye
force-pushed
the
minimax-m31-flash
branch
from
October 29, 2026 07:00
a0ea941 to
d0ebc06
Compare
mergify
Bot
added
ci/build
deepseek
Contributor
|
This pull request has merge conflicts that must be resolved before it can be |
zyongye
force-pushed
the
minimax-m31-flash
branch
3 times, most recently
from
October 29, 2026 17:08
1f5ea52 to
f4ca759
Compare
… scale FA4 reads a missing descale as 1.0, and it copies every expanded descale view into a contiguous tensor: three copy kernels per layer and step. The scales are fixed once the weights are loaded, so a per-tensor scale of 1.0 (an FP8 KV cache without calibrated scales) passes no descale. Calibrated and per-head scales keep theirs; a (1, 1) expanded view gets canonical strides, because FA4 treats it as contiguous and its layout check then rejects the zero stride. Co-authored-by: qizixi <22851944+zixi-qi@users.noreply.github.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
khluu
force-pushed
the
minimax-m31-flash
branch
from
October 29, 2026 20:59
f4ca759 to
f4bf304
Compare
2 of 6 tasks
The base branch was reset onto main (2abde62 plus one FA4 commit), so the old vllm-project#63232 squash (a0ea941) is gone and the merge base fell back to 5d35ce3, turning every MiniMax-M3 file into an add/add conflict. Take the base branch's tree and re-apply this PR's eight-file diff on top. Only test_minimax_m3_norm_rope_nvfp4_insert.py conflicted, on an argument reformat; keep main's explicit self.w[i] form and add mxfp4=. Signed-off-by: Limpus, David <dlimpus@amd.com>
zyongye
force-pushed
the
minimax-m31-flash
branch
from
October 31, 2026 01:27
f4bf304 to
3e48944
Compare
5 of 6 tasks
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
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.
Overview
Adds an MXFP4 mode to MiniMax-M3's fused FlyDSL norm / RoPE / KV insert on ROCm (gfx950). The same launch that writes NVFP4 pages can write K, V and index-K as MXFP4 pages: E2M1 data with one E8M0 scale per 32 dims and no global scale. That takes 68 B per token and head, against NVFP4's 72.
Stacked on #63232 (this PR targets its branch) and #64358 (the
mxfp4KV dtype; its commit is the first one here). Review the last commit.Changes
norm_rope_nvfp4_insert/kernel.py:compile_norm_rope_insert(..., mx=True)swaps the scale step and the scale addresses. Norm, RoPE, the E2M1 data layout, V's octet runs and the index-q output are unchanged.cvt_scalef32_pk_fp4_f32the NVFP4 path already uses.[token][4], V[group][token]; per head and page, 8192 B of data plus 512 B of scales.host.py:mxfp4=keyword. The slot width comes frommxfp4_kv_cache_full_dim/nvfp4_kv_cache_full_dim, andk_scale/v_scaleare ignored in MXFP4 mode.mx=False).The readers (sparse decode / prefill and the index kernels on scaled MFMA) and the backend's
customize_specfollow in later PRs. Nothing callsmxfp4=Trueyet.Validation
tests/kernels/attention/test_minimax_m3_norm_rope_nvfp4_insert.pyruns on MI355X: 45 passed. That's the existing 31 NVFP4 cases, plus the MXFP4 cases for the two-kernel comparison, the all-padding batch and graph replay. Every output byte is compared, stale page bytes included. The cases cover a single decode token, octet- and page-crossing runs, DSpark decode rows, unaligned prefill chunks and 4K-token prefill, at TP4 and TP1 head counts.The reference (
minimax_m3_nvfp4_ref.py) gains an MXFP4 quantizer written withtorch.frexp, independent of the kernel's bit arithmetic. On CPU it matches a brute-force OCP MXFP4 quantizer on random data at three magnitudes, exact ties, all-zero groups and the 2^-126 clamp.Accuracy of MXFP4 KV for M3.1
The same values were emulated by rounding K, V and index-K to MXFP4 before the FP8 KV insert. Against the NVFP4-KV baseline (Quark MXFP4 experts, seed 42, paired exact McNemar): GSM8K 94.31% vs 94.31% (p = 1.0), GPQA-Diamond 83.33% vs 80.30% (p = 0.39), RULER NIAH multikey-2 32k 100% vs 100% (p = 1.0).
Pull Request Checklist
I used vLLM's
/pr-checklistskill. (Mandatory for agents, optional for humans).AI assistance was used during the creation of this PR.
Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.
Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.
Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.
Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.