[BC breaking] Add MulticastTensor support to hl.signal & hl.wait (as_ptrs)

examples/all_gather_matmul.py

@@ -96,10 +96,6 @@ def helion_matmul_w_progress(
                 tile_m.begin // (M_per_rank // SPLITS_PER_RANK),
             ],
             signal=1,
-            update=None,
-            op="ld",
-            scope="gpu",
-            sem="acquire",
         )
         for tile_k in hl.tile(K):
             # TODO(joydddd): use a_shared and skip barrier when data is available on local rank.

helion/language/multicast_tensor.py

@@ -9,6 +9,8 @@
 from . import _decorators
 
 if TYPE_CHECKING:
+    from typing import Sequence
+
     from .._compiler.type_propagation import TypeInfo
     from .._compiler.variable_origin import Origin
 
@@ -79,6 +81,11 @@ def __setitem__(  # pyright ignore[reportIncompatibleMethodOverride]
     ) -> None:
         raise exc.NotInsideKernel
 
+    def new_empty(
+        self, *args: Sequence[int | torch.SymInt], **kwargs: dict
+    ) -> torch.Tensor:
+        return self.tensor_like.new_empty(*args, **kwargs)  # pyright: ignore[reportCallIssue]
+
 
 def multicast_like(
     tensor_like: torch.Tensor,