[MUSA] Use MUSA-optimized operators in piecewise CUDA graph (#23633)

Signed-off-by: popsiclexu <zhenxuexu@gmail.com>

sgl-kernel/pyproject_musa.toml

@@ -3,7 +3,7 @@ requires = [
   "setuptools>=75.0",
   "scikit-build-core>=0.10",
   "torch",
-  "torchada>=0.1.54",
+  "torchada>=0.1.55",
   "wheel",
 ]
 build-backend = "setuptools.build_meta"

sgl-kernel/python/sgl_kernel/utils.py

@@ -56,7 +56,7 @@ def cache_once(fn):
 
 @cache_once
 def is_arch_support_pdl() -> bool:
-    if bool(torch.version.hip):
+    if getattr(torch.version, "hip", None) or getattr(torch.version, "musa", None):
         return False
     try:
         device = torch.cuda.current_device()