Point mlx at Qwen3.6 decode matvec optimizations (~30 tok/s)

[goniz]

Summary

Updates the mlx submodule to the merged feat/vulkan tip from goniz/mlx#65:

df10f7b8f — Merge pull request #65 from goniz/cursor/qwen36-decode-opt-088d

Vulkan decode optimizations for MoE gather matvec + safer deferred barriers (subgroup partial sizing + submit-tail barrier included).

Benchmarks (mlx-community/Qwen3.6-35B-A3B-8bit)

	Prefill	Decode
Baseline	~682	~21.9
This change	683.9	31.0

Averages: prompt_tps=683.926, generation_tps=31.044, peak_memory=39.664

Merged MLX PR: goniz/mlx#65

  

[chatgpt-codex-connector]

💡 Codex Review

https://github.com/goniz/mlx-vulkan/blob/0aa7b9754ca1ee1c128fd8e471948b28160ecfda/mlx/mlx/backend/vulkan/kernels/mul_mv_affine8.comp#L82
Keep affine8 weight offsets byte-accurate

When an 8-bit affine weight row is not 4-byte aligned, such as a quantized layer with K/packed_row_bytes not divisible by 4, this floors the byte offset before indexing the new uint view. For col > 0 the row can start in the middle of a word, but both the packed fast path and the scalar fallback then read byte 0 of the floored word as k=0, mixing bytes from the previous row/column and producing incorrect matvec results; keep byte indexing or account for the intra-word byte offset.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[goniz]

Codex P2 (affine8 weight offsets) — no code change

packed_row_bytes is set as w.strides(-2) * sizeof(uint32_t) in quantized.cpp, so it is always a multiple of 4 by construction. Therefore col * packed_row_bytes never has an intra-word byte offset, and (col * packed_row_bytes) >> 2 is an exact word index for every column.

The same applies to gather (w_matrix_stride_bytes is also stride * sizeof(uint32_t)). Misaligned rows are not representable on this path for affine uint8 packs.

Happy to add an assert in the C++ dispatcher if we want a belt-and-suspenders check, but the shader flooring is safe under the current packing contract.