Repository navigation
[Metal] Add fast route for dequantizing fp4 to half in fp_qmm_t - #4709
Merged
Merged
Conversation
zcbenz
reviewed
Oct 19, 2026
| int BK = 32, | ||
| int BN = 32> | ||
| int BN = 32, | ||
| bool half_tile = false> |
Member
There was a problem hiding this comment.
Rather than having the caller pass the half_tile parameter, can we just deduce inside fp_qmm_t?
constexpr half_tile =
(metal::is_same_v<T, float16> && bits == ... && group_size == ...) ||
(metal::is_same_v<T, bfloat16> && bits == ... && group_size == ...);
dhiltgen
force-pushed
the
qmm-nvfp4-half-tile
branch
from
October 19, 2026 15:56
12f0150 to
e075da1
Compare
dhiltgen
added a commit
to dhiltgen/ollama
that referenced
this pull request
Oct 20, 2026
Half-tile fp4 quantized matmul, which speeds up 4-bit prefill. Applied to the pinned MLX until the pin includes it; drop this commit then.
dhiltgen
added a commit
to dhiltgen/ollama
that referenced
this pull request
Oct 20, 2026
Quantized matmul tiled for 6-16 rows, the shape of speculative verify, rebased onto ml-explore/mlx#4709. Applied to the pinned MLX until the pin includes it; drop this commit then.
zcbenz
force-pushed
the
qmm-nvfp4-half-tile
branch
from
October 21, 2026 05:05
e075da1 to
ff91940
Compare
zcbenz
approved these changes
Oct 21, 2026
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
This change speeds up prompt processing for fp4 models on older generations. Instead of dequantizing to f32, applying group scale, then converting to bf16, decode to f16 tiles and scale with one op, and leverage bigger tiles for long prompts. For nvfp4 there is no precision loss. Same for mxfp4 fp16 based models. For mxfp4 bf16 based models to avoid precision loss, those are left on the prior code path.
mlx_lm.benchmark, 2048-token prompt, mlx-community Qwen3.8-27B-nvfp4 prompt tps(generation tps unchanged).