Repository navigation
Conversation
Advance MLX to 243bdf09c0cccbb85ca6a686dab3cc0102cf1f7e, rebase changed CPU/CUDA/Metal overlays against pristine upstream sources, retire the CUDA teardown overlay superseded by upstream #4514, and retain the gathered-row correctness fix while using upstream's Xcode 27 workaround. Retarget the ROCm core overlay and adapt move-only allocation ownership, custom releases and primitive APIs without changing the fork revision. Add host allocation lifetime contracts and preserve reviewed macOS filesystem test portability fixes for candidate CI. Document every retained overlay and hardware validation boundaries. Validation: 27 MLX pin/parser tests, version/dtype/dispatch/llama contracts, 123 Python tooling tests and formatting pass. Native CUDA compilation and candidate Metal execution are coordinated separately; ROCm GPU validation requires an AMD host. Refs #2303
inureyes
added
status:review
The rebased upstream gathered GEMM route called make_cta_tiler with two arguments, while the retained mlxcel selector needs the active CUDA device for its Blackwell tile cap. Pass encoder.device() as the existing quantized GEMM route does, preserving the row count, quantization group size and runtime tile overrides. The missing argument caused the first candidate CUDA build to fail and the later CuTe layout errors were cascading diagnostics. Validation: inspected both selector call sites and checked the diff. Native CUDA compilation is being rerun on the corrected source. Refs #2303
Add 384-wide sorted-prefill reference cases for MXFP4 and MXFP8 with float16 and bfloat16 activations. The existing 320-wide cases fail the new CUDA gathered RHS kernel's output-alignment admission, so they cannot validate its JIT path or the Blackwell 128-row tile policy. Keep those tail cases and reuse the existing expanded routing fixture, host reference and tolerances. Correct the CUDA ops overlay's stale provenance comment. Formatting and diff checks pass; candidate hardware execution follows in the coordinated validation run. Refs #2303
Upstream #4554 adds cp_async_wait<0>() and a CTA barrier immediately before the quantized GEMM epilogue overwrites the shared-memory A/B buffers. The former local #910 pair precedes only a register-fragment conversion, so upstream now supplies the same async-copy completion and cross-thread ordering for both ordinary and gathered RHS kernels. Retire the whole device header overlay after verifying that removing its provenance and local barrier hunk reproduces pristine 243bdf09 byte for byte. Record the proof and updated overlay inventory; retain the existing JIT module revision boundary. Validation: pristine-object equality assertion and diff check pass. The coordinated CUDA run will restore the retired cached header from the pinned Git object before rebuilding. Refs #2303
inureyes
marked this pull request as ready for review
October 14, 2026 01:04
This branch has not been deployed
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
1 participant
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.
Summary
Advance MLX from
59d600b5to captured upstream main243bdf09(111 commits, 312 files). Reconcile whole-file overlays to retain mlxcel's numerical behavior and runtime controls without overwriting upstream fixes.B * Mcorrection; use upstream #4594 for Xcode 27.75915908. Adapt move-only allocation ownership, borrowed-storage donation safety, compiled callbacks, SDPA sinks, gated-delta fallback, floor division, and odd inverse FFT. Add actual-header host ownership regressions; these do not prove HIP completion ordering.Validation
3aeccc2b: fmt/Clippy pass; 12,285 tests pass, 390 ignored; 51 FFmpeg video tests pass.53473340= production3aeccc2b+ independent fix: restore Kimi Linear CUDA inference #2296 mask fix1fec677b; that model fix is not included here. CLI SHA-256:b52d70388e0420582c4f60f3341d570c2afd736685b37e2c0c6e50ed9eaf5fae.ROCm compilation/runtime and M5-specific execution remain unavailable. Ordinary PR CI is tracked separately; queued checks are not passes. The requested whole-batch CUDA/Clippy and fresh main Metal integration gate remains deferred until the expanded follow-ups are ready. This PR does not independently close #1599, #2158, or #1986. English/Korean technical reports record provenance and remaining limits.
Closes #2303