fix: dD_acc sum via axis=0 not scalar.reshape(1)

mamba3_patches.py

@@ -51,7 +51,9 @@ old = """            dD_acc += tl.dot(
                 tl.full([1, CHUNK_SIZE], 1, dtype=tl.float32),
                 dQK_dot_block
             ).reshape(1)"""
-new = """            dD_acc += tl.sum(dQK_dot_block.to(tl.float32)).reshape(1)"""
+# tl.sum without axis returns scalar which can't .reshape(1) in triton 3.4.
+# Sum over axis=0 on [CHUNK_SIZE, 1] yields [1] directly.
+new = """            dD_acc += tl.sum(dQK_dot_block.to(tl.float32), axis=0)"""
 assert old in s, "bwd.py pattern B not found"
 s = s.replace(old, new)