From 92f441eaa516939499742051487b53b275e137ba Mon Sep 17 00:00:00 2001 From: Goni Zahavy Date: Wed, 15 Jul 2026 20:52:49 +0300 Subject: [PATCH 1/9] Bump mlx for dense affine-4 BF16 QMM kernels Point the submodule at the Qwen3.6-27B dense 4-bit fused matvec/QMM path so Unsloth UD-MLX-4bit prefill/decode can stay on BF16 without staging through float32. --- mlx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/mlx b/mlx index df10f7b..0b51f65 160000 --- a/mlx +++ b/mlx @@ -1 +1 @@ -Subproject commit df10f7b8fa535bf4b07e6ff6e3082f38565bd20f +Subproject commit 0b51f651269cf2d9e7fc3d18d7e56b549726beec From 20104c540870dda8e79b6e982fc0c39e3bbb1d47 Mon Sep 17 00:00:00 2001 From: Goni Zahavy Date: Wed, 15 Jul 2026 21:48:12 +0300 Subject: [PATCH 2/9] Bump mlx for dense affine-4 BF16 QMM (no affine8 multi-col) Keep the 27B 4-bit fused path while avoiding the dense affine8 multi-column decode regression on small 8-bit models. --- mlx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/mlx b/mlx index 0b51f65..80814ae 160000 --- a/mlx +++ b/mlx @@ -1 +1 @@ -Subproject commit 0b51f651269cf2d9e7fc3d18d7e56b549726beec +Subproject commit 80814ae71efe1c74a68b7156180e2fb249578ae3 From e73a4e8c051f4f134893e99fc14f3f19c75004e7 Mon Sep 17 00:00:00 2001 From: Goni Zahavy Date: Wed, 15 Jul 2026 21:54:00 +0300 Subject: [PATCH 3/9] Bump mlx for dense coopmat affine-4 prefill QMM Point at fused coopmat 4-bit QMM to close more of the Qwen3.6-27B prefill gap toward llama.cpp Strix Halo targets (~270-350 / ~12 tok/s). --- mlx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/mlx b/mlx index 80814ae..a3e17ef 160000 --- a/mlx +++ b/mlx @@ -1 +1 @@ -Subproject commit 80814ae71efe1c74a68b7156180e2fb249578ae3 +Subproject commit a3e17ef42c44cfa29824005b4d34857faa17b836 From 49cf036c29172671926393e652655f1940adc396 Mon Sep 17 00:00:00 2001 From: Goni Zahavy Date: Wed, 15 Jul 2026 22:14:35 +0300 Subject: [PATCH 4/9] Bump mlx for F16 coopmat1 large BF16 GEMM path Accelerates Qwen3.6-27B GDN bf16 projections via F16 coopmat1 promote. --- mlx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/mlx b/mlx index a3e17ef..0a2164a 160000 --- a/mlx +++ b/mlx @@ -1 +1 @@ -Subproject commit a3e17ef42c44cfa29824005b4d34857faa17b836 +Subproject commit 0a2164a741323078490078b2cdd4b5a5d0d7bce0 From cd912e319c454a479764bc374fc4eccb61f58532 Mon Sep 17 00:00:00 2001 From: Goni Zahavy Date: Thu, 16 Jul 2026 05:24:43 +0300 Subject: [PATCH 5/9] Bump mlx for coop4 exponent prefetch and large matvec tune Prefill: one-shot BF16 exponent prep for dense coop4 QMM. Decode: NUM_ROWS=2 for large AMD RDNA BF16 matvecs (GDN). --- mlx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/mlx b/mlx index 0a2164a..cbb55b2 160000 --- a/mlx +++ b/mlx @@ -1 +1 @@ -Subproject commit 0a2164a741323078490078b2cdd4b5a5d0d7bce0 +Subproject commit cbb55b22776054ab7c5aed42ac2c7c67bf6400e6 From 3c4e5ffa8652742f55b1991efe258f8e1c340f17 Mon Sep 17 00:00:00 2001 From: Goni Zahavy Date: Thu, 16 Jul 2026 05:40:50 +0300 Subject: [PATCH 6/9] Bump mlx for coopmat/coop4 review-comment hardenings Addresses MLX PR #66 review feedback on aligned K gating, 16x16 f16 coopmat property checks, and coop4 / exponent subgroup sizing. --- mlx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/mlx b/mlx index cbb55b2..e6b880b 160000 --- a/mlx +++ b/mlx @@ -1 +1 @@ -Subproject commit cbb55b22776054ab7c5aed42ac2c7c67bf6400e6 +Subproject commit e6b880b1e7f2e5cc1657f45d7367ee4bc3bd8136 From a8b22890f0ec877ac9355848c6bf71fd43235b49 Mon Sep 17 00:00:00 2001 From: Goni Zahavy Date: Thu, 16 Jul 2026 06:41:47 +0300 Subject: [PATCH 7/9] Bump mlx for cached F16 B^T coopmat1 promote Avoids re-casting static BF16 weights to F16 on every large dense GEMM. --- mlx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/mlx b/mlx index e6b880b..6ae98f9 160000 --- a/mlx +++ b/mlx @@ -1 +1 @@ -Subproject commit e6b880b1e7f2e5cc1657f45d7367ee4bc3bd8136 +Subproject commit 6ae98f9c4cd8c93afc9c794e4035609a6fe9bd47 From e51ce91e2cc852ab8111d895471c2e651124d3f9 Mon Sep 17 00:00:00 2001 From: Goni Zahavy Date: Thu, 16 Jul 2026 07:04:51 +0300 Subject: [PATCH 8/9] Bump mlx for SDPA VJP fallback fix Fixes test_sdpa_grad SIGFPE on Vulkan/AMD (mlx-vulkan#66). --- mlx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/mlx b/mlx index 6ae98f9..05644d6 160000 --- a/mlx +++ b/mlx @@ -1 +1 @@ -Subproject commit 6ae98f9c4cd8c93afc9c794e4035609a6fe9bd47 +Subproject commit 05644d634e94acc2029821c462fae1535ed821d4 From d885f77b8a745b8977c115402a1239314725e852 Mon Sep 17 00:00:00 2001 From: Goni Zahavy Date: Thu, 16 Jul 2026 08:24:45 +0300 Subject: [PATCH 9/9] Bump mlx to merged PR #66 on feat/vulkan Point submodule at 430a327061 (dense affine-4 / coopmat / SDPA VJP fallback). --- mlx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/mlx b/mlx index 05644d6..430a327 160000 --- a/mlx +++ b/mlx @@ -1 +1 @@ -Subproject commit 05644d634e94acc2029821c462fae1535ed821d4 +Subproject commit 430a3270612c8b0de2eec29a0f52c18c4fdc81b9