build: pin mlx and mlx-c to the JACCL fixes and the owned bootstrap bridge

[jonathan308]

Summary

Two pin moves on current main (6923a80f) that bring the two-Mac (JACCL, RDMA over Thunderbolt) fixes into the Swift stack.

Commit	Path	From	To	Source
0656737	Source/Cmlx/mlx	cb77239b (fork main)	d0f8b0a8	Layr-Labs/mlx#33, every source and test change. The one later commit there only reformats a test CMake file
27ca572	Source/Cmlx/mlx-c and the two mirrored copies of distributed_group.h	02cf6f4d	489e9654	head of Layr-Labs/mlx-c#14

Status: draft. Both pins are heads of draft branches. They must move to the merged commits before this is merged. Order: Layr-Labs/mlx#33 and Layr-Labs/mlx-c#14 first, then this PR, then the pin of libs/mlx-swift and libs/mlx in Layr-Labs/d-inference#1407.

This PR contains and extends #33. #33 has the same mlx-c pin and the same two header copies on an older base (0f4fe403), and pins mlx to the send-frame commit alone on an older mlx base, which is behind the mlx that main pins now. #33 can be closed when this one is accepted.

What the mlx pin brings (0656737)

• Upstream JACCL and distributed fixes: the ring all_gather slice (ml-explore/mlx#4443), the ReduceScatter::eval_cpu assertion (ml-explore/mlx#4557), and a failed backend initialization that was cached (ml-explore/mlx#4558).
• The unused tail of each JACCL send frame is set to zero (Gajesh Naik's commit of Preserve JACCL scratch-frame clearing from the cluster workspace mlx#26).
• A progress guard. A JACCL call that gets no completion for longer than JACCL_PROGRESS_TIMEOUT_MS (default 120000) fails with an error and closes its group. Before, a rank whose peer had died polled at 100% CPU with no limit.
• The queue pair sends from the GID that was advertised to the peer.

Nothing else changes between the two mlx commits: no Metal kernel, no mirrored header and no generated source. Source/Cmlx/include, include-framework and mlx-generated are not touched by this commit. The JACCL sources are compiled only for a macOS 26.2 deployment target (mlx-conditional/jaccl_conditional.cpp). Other targets keep the stub.

What the mlx-c pin brings (27ca572)

mlx_distributed_init_jaccl_with_bootstrap: a strict JACCL initialization whose bootstrap exchange goes through a callback of the caller instead of the TCP side channel. One commit by Gajesh Naik on top of the mlx-c commit that main pins.

• The two mirrored headers are the header of that commit, copied as tools/update-mlx.sh copies them. They are byte-identical to the copies in Preserve Swift Cmlx bootstrap exports and captured dependency pins #33.
• The new source mlx/c/distributed_bootstrap.cpp is compiled by the existing Cmlx target. Package.swift does not change.
• xcode/MLX.xcodeproj is not updated.

What was run

Check	Result
Two Macs over Thunderbolt RDMA, two-rank mesh, with the three JACCL fix commits before they were moved onto the upstream backports, built through this package for macOS 26.2	Normal traffic exact. A rank whose peer ended reported the timeout at the limit and exited, with and without a model loaded. Neither Mac kept a process or wired memory. Details in Layr-Labs/mlx#33.
Simulated-verbs JACCL tests at mlx d0f8b0a8	8 of 8 CTest tests pass with ASan and UBSan.
mlx/c/distributed_bootstrap.cpp at mlx-c 489e9654 against mlx d0f8b0a8	Syntax check passes.
Header copies	Identical to #33.
CI build-and-test on this pull request (swift build --build-tests, default deployment target)	distributed_bootstrap.cpp compiles against the pinned mlx. 927 tests run, 0 failures, 1 skipped.

Not shown

• A build of this package at exactly these two pins for macOS 26.2, where the real JACCL sources are compiled. CI builds the default target, which keeps the stub. The hardware runs used mlx at the three fix commits without the backports, and mlx-c at 02cf6f4d. (The MLX CMake build for 26.2 at this mlx commit passes in JACCL: zeroed send-frame tails, a progress guard, the advertised GID and the upstream JACCL fixes mlx#33.)
• Any call of mlx_distributed_init_jaccl_with_bootstrap. No caller exists in this repository. The provider side that will call it is not written yet.
• The Xcode project build.

How AI was used

Claude Code (Anthropic) prepared the two commits, the commit messages and this description, and ran the checks above, at the direction of the account owner, who asked for this draft. The mlx-c commit and the header text are by Gajesh Naik. Nothing is proposed to ml-explore/mlx-swift.

Update 2026-10-10

64070c4 moves Source/Cmlx/mlx to fdae3174 (draft Layr-Labs/mlx#33 head): the 64-bit ring all_gather slice fix, the revert of the NAX quantised-MoE gather_qmm simdgroup skip (upstream c7ff35d97, which lost JACCL completions on a two-Mac pair with one NAX GPU; output unchanged), and guard v2 (30 s progress limit, teardown exit with memory release, status 75). Gitlink only; no source files added or removed. Two-Mac evidence is in the mlx#33 description.

🤖 Generated with Claude Code