Rebase DBraun/sequence-layers mlx.001 onto main; bump to 0.3.dev1#15
Closed
JulianSlzr wants to merge 17 commits into
Closed
Rebase DBraun/sequence-layers mlx.001 onto main; bump to 0.3.dev1#15JulianSlzr wants to merge 17 commits into
JulianSlzr wants to merge 17 commits into
Conversation
commit f2f2198 Author: David Braun <2096055+DBraun@users.noreply.github.com> Date: Thu Feb 26 09:06:30 2026 -0500 remove depthformer and moshi_demo.py commit 74669fc Author: Kehang Han <kehanghan@gmail.com> Date: Wed Feb 25 15:15:26 2026 -0800 This commit finally makes MLX generate good music! commit 9e4a452 Author: Kehang Han <kehanghan@gmail.com> Date: Wed Feb 25 12:47:09 2026 -0800 Changes to make temporal output match between jax and mlx commit 8d267f2 Author: David Braun <2096055+DBraun@users.noreply.github.com> Date: Wed Feb 25 12:45:53 2026 -0500 Update weight_converter.py commit 68add71 Author: David Braun <2096055+DBraun@users.noreply.github.com> Date: Wed Feb 25 12:41:52 2026 -0500 MLX attention: add per_dim_scale commit b4ae905 Author: David Braun <2096055+DBraun@users.noreply.github.com> Date: Wed Feb 25 11:31:02 2026 -0500 Update normalization.py commit a3e84cf Author: Kehang Han <kehanghan@gmail.com> Date: Tue Feb 24 16:04:27 2026 -0800 To make encoded approx bit level compatible commit c6596c7 Author: Kehang Han <kehanghan@gmail.com> Date: Tue Feb 24 14:35:16 2026 -0800 Updates to MLX port of SL to support magenta-rt commit 08077bb Author: David Braun <2096055+DBraun@users.noreply.github.com> Date: Tue Feb 24 08:54:43 2026 -0500 Create moshi_demo.py
use mx.fast.scaled_dot_product_attention
use mx.fast.rope
These tests exercise the DotProductSelfAttention KV cache ring buffer with time > max_past_horizon, causing buffer wrap-around. They expose a write-before-read bug: the current code overwrites the oldest key in the attention window before queries can attend to it, even with block_size=1. Failures: - test_use_kv_cache_ringbuffer: 33.3% element mismatch at time=6 - LocalDotProductSelfAttention.test_layer: 19.4% element mismatch at time=10 The following commit fixes these by implementing write-after-read.
Includes: - Expanded sweep for KV cache ringbuffer test with max_future_horizon=0. - Restored informative comments about ring buffer and mask logic.
TAG=agy CONV=21ada17b-3411-4090-8450-e69d8ebfeae6
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
4 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.
Rebases @DBraun's MLX fork onto sequence-layers 0.02 as an initial next version.
Temporary measure before the refactor (#14) for downstream SL MLX users.