Repository navigation
Refresh MLX through v0.32.4 and preserve owned runtime patches - #7
Merged
Merged
Conversation
Co-authored-by: Cheng <git@zcbenz.com>
…explore#4212) Co-authored-by: x14ngch3n <x14ngch3n@users.noreply.github.com> Co-authored-by: Cheng <git@zcbenz.com>
…e#4139) Co-authored-by: Cheng <git@zcbenz.com>
Co-authored-by: Cheng <git@zcbenz.com>
ml-explore#4357) Co-authored-by: Anastasiia Filippova <a_filippova@apple.com>
…xplore#3842) Co-authored-by: Cheng <git@zcbenz.com>
Co-authored-by: Jake Bowhay <60778417+j-bowhay@users.noreply.github.com>
Co-authored-by: Cheng <git@zcbenz.com>
Co-authored-by: Anastasiia Filippova <a_filippova@apple.com>
Co-authored-by: Cheng <git@zcbenz.com>
…4266) Co-authored-by: Cheng <git@zcbenz.com>
…lore#4284) Co-authored-by: Cheng <git@zcbenz.com>
Co-authored-by: Daniel Hiltgen <daniel.hiltgen@ollama.com> Co-authored-by: Cheng <git@zcbenz.com>
Co-authored-by: Cheng <git@zcbenz.com>
Co-authored-by: Cheng <git@zcbenz.com>
…ass_list workflow (ml-explore#4667)
…l-explore#4661) Co-authored-by: Cheng <git@zcbenz.com>
Co-authored-by: Cheng <git@zcbenz.com>
Co-authored-by: Cheng <git@zcbenz.com>
Co-authored-by: Cheng <git@zcbenz.com>
Co-authored-by: Cheng <git@zcbenz.com>
Co-authored-by: Cheng <git@zcbenz.com>
…e#4679) Co-authored-by: Cheng <git@zcbenz.com>
Co-authored-by: “Ronald <“Ronaldmannak@me.com”>
Co-authored-by: Cheng <git@zcbenz.com>
This was referenced Oct 27, 2026
a-github-name
marked this pull request as ready for review
October 28, 2026 17:50
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Refresh the owned core to upstream MLX v0.32.4 (
36969c094354086e15d2a327a9c50feb20b9cf11), preserving the patch inventory inFORK_PATCHES.md. Final core revision:735d48ca1436dc2648073c548715af3aea5d32c7.Restore upstream NAX device availability for floating-point matmul, attention, and gated delta. Keep generation-17 NAX disabled locally in all four quantized dispatch/alignment decisions. The former global guard caused the matched BF16 long-context Python failure; the scoped implementation returns the correct code with all 325 generated tokens identical to stock upstream. Add wide-output FP16/BF16 activation regressions across 1/2/4/8-bit ordinary and gathered quantized matmul.
Preserve owned CUDA 1-bit execution, custom quantized headers, NVFP4 staging, H3/M4 tuning, and Apple/Linux compatibility. Upstream replaces overlapping shared 1-bit support, CUDA encoder lifetime, and obsolete QMV template handling. CPU transposed 1-bit accumulation uses float after input-dtype rounding; wider affine arithmetic and existing tolerances remain unchanged. The CPU reduction alias explicitly spells
typenamefor Linux compiler compatibility.Validation on an M5 Max:
MLX_ENABLE_TF32=0), following upstream's runner. Production inference defaults are unchanged.git diff --checkpasses.Downstream mere.run #599 pins the Swift refresh. All three installed native BF16 tests and 50 LFM2 fixtures pass. Its native release 128K probe now passes after aligning the terminal prompt-token forward with upstream: exactly TANGERINE-4827, 295 output tokens, finish_reason=stop, under the unchanged 2,048-token budget. Downstream also corrects the custom affine-Q4 batching bias sum to promote BF16 operands before addition, matching serial QMV. All 22 application quantized/binary-linear GPU checks now pass, including the two formerly failing bit-exact tests and a cancellation regression.
The downstream Ubuntu 22.04 / Swift 6.1 Linux CLI job, macOS gate and app bundle, and iOS CI all passed at mere.run ml-explore#599 head
472893026a7126b4004cf7a5a3d4f4f5b8fea2bd. This core and the wide-head follow-up in #8 are shipped in mere.run v0.66.1. CUDA execution, actual Swift 6.0, and older Apple hardware/OS remain unvalidated.AI usage disclosure: implementation and validation assisted by Codex.