Skip to content

fix(metal): build the GPT-OSS MXFP4 prefill tile into the metallib #24

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Gajesh2007 merged 2 commits into from
Sep 28, 2026
Merged

fix(metal): build the GPT-OSS MXFP4 prefill tile into the metallib #24

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4d66c31
fix(metal): build the GPT-OSS MXFP4 prefill tile into the metallib
davidtai Sep 25, 2026
5621947
docs(forkdiff): describe the ahead-of-time prefill tile kernels
davidtai Sep 25, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 3 additions & 0 deletions fork.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,6 +75,9 @@ def:
E=32, K=2880, N=2880|5760, FP32/BF16 shape; `MLX_GPTOSS_MXFP4_DECODE_FAST_TAIL=0`
restores the original route and `MLX_GPTOSS_MXFP4_PREFILL_TILE=m32n32k32` opts into a
separately measured prefill tile (`gptoss_mxfp4_policy.h`). (PR #14)
`fp_quantized.metal` also builds the two m32n32k32 prefill-tile kernels (float and
bfloat16) ahead of time, so CMake builds with the default `MLX_METAL_JIT=OFF` can load
them. (PR #24)
globs:
- "mlx/backend/metal/gptoss_mxfp4_policy.h"
- "mlx/backend/metal/kernels/fp_quantized.h"
Expand Down
4 changes: 4 additions & 0 deletions mlx/backend/metal/kernels/fp_quantized.metal
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -226,3 +226,7 @@ instantiate_quantized_types(float16_t)
// Exact GPT-OSS MXFP4 gathered-vector tail experiment.
instantiate_quantized(mxfp4, gather_qmv_fast_tail, float, 32, 4)
instantiate_quantized(mxfp4, gather_qmv_fast_tail, bfloat16_t, 32, 4)

// GPT-OSS MXFP4 prefill tile m32n32k32, selected by MLX_GPTOSS_MXFP4_PREFILL_TILE.
instantiate_gather_qmm_rhs(fp_gather_qmm_rhs, gather_qmm_rhs_nt, float, 32, 32, 32, 2, 2, true, mxfp4, 32, 4)
instantiate_gather_qmm_rhs(fp_gather_qmm_rhs, gather_qmm_rhs_nt, bfloat16_t, 32, 32, 32, 2, 2, true, mxfp4, 32, 4)
Loading