From 7ea567a2458a4800a6a0e3e0a6ba41fcbd00d146 Mon Sep 17 00:00:00 2001
From: Martin Vit <martin@voipmonitor.org>
Date: Sat, 18 Jul 2026 15:53:47 +0000
Subject: [PATCH] build(cuda): pin CUTLASS DSL 4.5.3 for SM120

---
 requirements/cuda.txt | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/requirements/cuda.txt b/requirements/cuda.txt
index 8bd497eb2e56..f80c46c07fd5 100644
--- a/requirements/cuda.txt
+++ b/requirements/cuda.txt
@@ -24,7 +24,9 @@ nvtx==0.2.15
 fastsafetensors >= 0.3.2
 
 # QuACK and Cutlass DSL for FA4 (cute-DSL implementation)
-nvidia-cutlass-dsl[cu13]==4.6.0
+# 4.6.0 spills the SM120 B12X W4A16 prefill kernel (255 registers and a
+# 256-byte stack per thread); 4.5.3 restores spill-free code generation.
+nvidia-cutlass-dsl[cu13]==4.5.3
 quack-kernels>=0.4.0  # Required for tml-fa4
 
 # Tokenspeed_MLA for faster mla with spec decode