Segmented mm nax kernel#3419
Merged
Merged
Conversation
nastya236
approved these changes
Apr 16, 2026
nastya236
left a comment
nastya236
left a comment
Collaborator
There was a problem hiding this comment.
Looks great! I asked couple of questions as usual for my own understanding :)
| // Use NAX kernel if available | ||
| if (use_nax) { | ||
| int average_k = K / batch_size_out; | ||
| bm = 64; |
Collaborator
There was a problem hiding this comment.
Just out of curiosity, why 2 x 2 simdgroups and 64 x 64 tiles?
Member
Author
There was a problem hiding this comment.
Well this can be tuned for sure. I haven't actually ran extensive testing and it seemed like a good default 🤷♂️
| MTL::Size group_dims = MTL::Size(32, wn, wm); | ||
| MTL::Size grid_dims = | ||
| MTL::Size(params.tiles_n, params.tiles_m, batch_size_out); | ||
| MTL::Size grid_dims = (use_nax && bk == 64) |
Collaborator
There was a problem hiding this comment.
I am wondering should non nax kernel also swap dimensions when there are many small segments?
Member
Author
There was a problem hiding this comment.
Yeah it is possible. The idea was that we launch threadgroups in order x, y, z and since K is very small in these cases we want the different Ks to run one after the other since it is very likely that the 2nd matmul's K will be in the cache already.
1 task
ekryski
added a commit
to ekryski/mlx
that referenced
this pull request
May 6, 2026
…dLocalStream, linalg.det) Highlights: - ThreadLocalStream in C++ (ml-explore#3405) — joins StreamOrDevice variant - Segmented mm NAX kernel (ml-explore#3419) — new steel/gemm/kernels/steel_gemm_segmented{,_nax} - mlx.core.linalg additions: determinant, sign-log-determinant (ml-explore#3416) - NAX split-K tuning + addmm fix (ml-explore#3422) - gather_mm CUDA + qmm CUDA fixes (ml-explore#3414, ml-explore#3417, ml-explore#3445) - Scheduler::enqueue thread safety (ml-explore#3423) - Random kernel bytes_per_key truncation fix (ml-explore#3432) - Many smaller bug + CI fixes; nanobind 2.12.0 pin Conflict resolution (5 files; all auto-merged cleanly, no manual edits): - mlx/utils.h — `ThreadLocalStream` joins StreamOrDevice; our metal_retain_bound_buffers() helper preserved. - mlx/backend/metal/device.cpp — upstream's "no Metal device available" error message coexists with our signpost/MLX_METAL_PROFILE perf diagnostic + hazard-mode retain logic. - mlx/backend/metal/CMakeLists.txt — upstream's steel_gemm_segmented{,_nax} joins our kernel registrations. - mlx/backend/metal/kernels/CMakeLists.txt — same: upstream segmented + our fused_gate_activation kernel. - tests/scheduler_tests.cpp — upstream's "test thread local stream" joins our scheduler-test alignment for lazy command-encoder init. No disputed conflicts; the auto-merge produced a coherent result with both sides' work intact. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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
2 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.
As the title suggests it adds a NAX kernel for the
segmented_mm. It also fixes #3362.Micro benchmark numbers