Bump CuTeDSL to 4.6.2 (#34372)

python/pyproject.toml

@@ -46,7 +46,7 @@ dependencies = [
   "ninja",
   "numba==0.65.1",
   "numpy",
-  "nvidia-cutlass-dsl[cu13]==4.6.0",
+  "nvidia-cutlass-dsl[cu13]==4.6.2",
   "nvidia-mathdx==25.6.0",
   "nvidia-ml-py",
   "openai==2.6.1",
@@ -63,7 +63,7 @@ dependencies = [
   "pydantic",
   "python-multipart",
   "pyzmq>=25.1.2",
-  "quack-kernels>=0.6.1",
+  "quack-kernels==0.6.4",
   "requests",
   "scipy",
   "sentencepiece",

python/sglang/kernels/ops/attention/flash_attn/cute/interface.py

@@ -1,6 +1,4 @@
 # Copyright (c) 2025, Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, Tri Dao.
-# [2025-07-04] Version in Cute-DSL, for Hopper and Blackwell. You'll need install nvidia-cutlass-dsl==4.2.0.
-
 import math
 import os
 from dataclasses import dataclass

python/sglang/kernels/ops/attention/flash_attn/cute/pyproject.toml

@@ -22,17 +22,17 @@ classifiers = [
 ]
 
 dependencies = [
-    "nvidia-cutlass-dsl>=4.5.2",
+    "nvidia-cutlass-dsl==4.6.2",
     "torch",
     "einops",
     "typing_extensions",
     "apache-tvm-ffi>=0.1.5,<0.2",
     "torch-c-dlpack-ext",
-    "quack-kernels>=0.5.0",
+    "quack-kernels==0.6.4",
 ]
 
 [project.optional-dependencies]
-cu13 = ["nvidia-cutlass-dsl[cu13]>=4.5.2"]
+cu13 = ["nvidia-cutlass-dsl[cu13]==4.6.2"]
 dev = [
     "pytest",
     "pytest-xdist",