From cbb673d2b80346d08b78ce2cec5c4a0534f9500f Mon Sep 17 00:00:00 2001 From: Goni Zahavy Date: Wed, 15 Jul 2026 11:59:53 +0300 Subject: [PATCH 1/4] Point mlx submodule at Qwen3.6 decode matvec optimizations Updates mlx to cursor/qwen36-decode-opt-088d (~30.4 tok/s decode on Qwen3.6-35B-A3B-8bit with prefill held near 683 tok/s). --- mlx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/mlx b/mlx index 501c4b5..36d0a1b 160000 --- a/mlx +++ b/mlx @@ -1 +1 @@ -Subproject commit 501c4b5571a1e1be85f11e244473a67123b50465 +Subproject commit 36d0a1b7522ad703eaa1ccbccc86fa1b2e0bf92e From c6301c342aa69c900e3c8dd7770ffc83a48213cf Mon Sep 17 00:00:00 2001 From: Goni Zahavy Date: Wed, 15 Jul 2026 12:23:08 +0300 Subject: [PATCH 2/4] Bump mlx submodule for subgroup partial sizing fix Addresses Codex review on goniz/mlx#65; decode still ~30.4 tok/s. --- mlx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/mlx b/mlx index 36d0a1b..994b472 160000 --- a/mlx +++ b/mlx @@ -1 +1 @@ -Subproject commit 36d0a1b7522ad703eaa1ccbccc86fa1b2e0bf92e +Subproject commit 994b472b8239bd43344aa3184b7ad50a660eaa2f From 0aa7b9754ca1ee1c128fd8e471948b28160ecfda Mon Sep 17 00:00:00 2001 From: Goni Zahavy Date: Wed, 15 Jul 2026 12:36:08 +0300 Subject: [PATCH 3/4] Bump mlx for submit-tail barrier fix Keeps Qwen3.6-35B-A3B-8bit around 684 prefill / 31 decode after addressing Codex P1 on deferred submit memory visibility. --- mlx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/mlx b/mlx index 994b472..0386ff1 160000 --- a/mlx +++ b/mlx @@ -1 +1 @@ -Subproject commit 994b472b8239bd43344aa3184b7ad50a660eaa2f +Subproject commit 0386ff1cdc0e924753dacc6eb3b64625a88a0ecb From 273cd7881643b00984ff929725777edc8cd1d2b1 Mon Sep 17 00:00:00 2001 From: Goni Zahavy Date: Wed, 15 Jul 2026 13:37:00 +0300 Subject: [PATCH 4/4] Point mlx submodule at merged PR #65 on feat/vulkan Updates to merge commit df10f7b8f (goniz/mlx#65 Qwen3.6 decode opts). --- mlx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/mlx b/mlx index 0386ff1..df10f7b 160000 --- a/mlx +++ b/mlx @@ -1 +1 @@ -Subproject commit 0386ff1cdc0e924753dacc6eb3b64625a88a0ecb +Subproject commit df10f7b8fa535bf4b07e6ff6e3082f38565bd20f