[FRONTEND] fix default max_num_imprecise_acc (triton-lang#2804)

Author: ptillet

python/triton/compiler/backends/cuda.py

@@ -78,7 +78,7 @@ def __init__(self, device_type: tuple) -> None:
     def parse_options(self, opts) -> Any:
         args = {k: opts[k] for k in CUDAOptions.__dataclass_fields__.keys() if k in opts}
         args["allow_fp8e4nv"] = self.capability >= 89
-        args["max_num_imprecise_acc_default"] = 0 if self.capability >= 89 else None
+        args["max_num_imprecise_acc_default"] = 2**30 if self.capability == 90 else 0
         return CUDAOptions(**args)
 
     @staticmethod

python/triton/language/semantic.py

@@ -1276,11 +1276,10 @@ def assert_dtypes_valid(lhs_dtype, rhs_dtype, options):
         assert acc.type == ret_ty
 
     # max_num_imprecise_acc only applies to fp8 -> fp32 dot on sm_90
-    max_num_imprecise_acc = 0
-    if lhs.dtype.is_fp8() and rhs.dtype.is_fp8():
+    if lhs.dtype.is_fp8() and rhs.dtype.is_fp8() and max_num_imprecise_acc is None:
         max_num_imprecise_acc = builder.options.max_num_imprecise_acc_default
-        if max_num_imprecise_acc is None:
-            max_num_imprecise_acc = 2**30
+    else:
+        max_num_imprecise_acc = 0
 
     return tl.tensor(builder.create_dot(lhs.handle, rhs.handle, acc_handle, allow_tf32, max_num_imprecise_acc), ret_ty)