Call .wait_tensor() in compiled region for dist.Work created in eager…

… region (pytorch#2485)

Summary:

In compiled region, instead of calling `dist.Work.wait()`, we will call `torch.ops._c10d_functional.wait_tensor()` on the dist.Work's output tensor. This way, we can capture the `wait_tensor()` op within the torch.compile graph (instead of graph-breaking on `dist.Work.wait()`), and the tensor will be waited on properly within the graph.

This diff also depends on pytorch/pytorch#137763 to function properly.

Reviewed By: Microve

Differential Revision: D64275115

torchrec/distributed/comm_ops.py

@@ -124,6 +124,15 @@ def _wait_impl(self) -> W:
         return ret
 
 
+def wait_req(req: Request[W]) -> None:
+    if is_torchdynamo_compiling():
+        assert req.tensor is not None
+        torch.ops._c10d_functional.wait_tensor(req.tensor)
+    else:
+        assert isinstance(req.req, dist.Work)
+        req.req.wait()
+
+
 @dataclass
 class All2AllPooledInfo(object):
     """
@@ -1334,7 +1343,7 @@ def backward(ctx, *unused) -> Tuple[None, None, None, Tensor]:
         myreq = ctx.myreq
         a2ai = myreq.a2ai
         assert myreq.req is not None
-        myreq.req.wait()
+        wait_req(myreq)
         myreq.req = None
         grad_output = myreq.tensor
         dim_sum_per_rank = a2ai.dim_sum_per_rank
@@ -1368,7 +1377,7 @@ def forward(
         a2ai = myreq.a2ai
         ctx.a2ai = a2ai
         assert myreq.req is not None
-        myreq.req.wait()
+        wait_req(myreq)
         sharded_output_embeddings = myreq.tensor
         myreq.req = None
         myreq.tensor = None
@@ -1573,9 +1582,9 @@ def backward(ctx, *unused) -> Tuple[None, None, None, Tensor]:
         myreq = ctx.myreq
         a2ai = myreq.a2ai
         assert myreq.req is not None
-        myreq.req.wait()
+        wait_req(myreq)
         if isinstance(myreq.req, dist.Work):
-            myreq.req.wait()
+            wait_req(myreq)
 
         myreq.req = None
         grad_output = myreq.tensor
@@ -1606,7 +1615,7 @@ def forward(
         ctx.a2ai = a2ai
         assert myreq.req is not None
         if isinstance(myreq.req, dist.Work):
-            myreq.req.wait()
+            wait_req(myreq)
         sharded_output_embeddings = myreq.tensor
         myreq.req = None
         myreq.tensor = None
@@ -1797,7 +1806,7 @@ def backward(ctx, *unused) -> Tuple[None, None, None, Tensor]:
             a2ai.permuted_lengths_after_sparse_data_all2all
         )
         assert myreq.req is not None
-        myreq.req.wait()
+        wait_req(myreq)
         sharded_grad_input = myreq.tensor
         if a2ai.codecs is not None:
             codecs = none_throws(a2ai.codecs)
@@ -1845,7 +1854,7 @@ def forward(
         D = a2ai.embedding_dim
         ctx.a2ai = a2ai
         assert myreq.req is not None
-        myreq.req.wait()
+        wait_req(myreq)
         myreq.req = None
         sharded_output_embeddings = myreq.tensor
         myreq.tensor = None
@@ -1952,7 +1961,7 @@ def forward(
     def backward(ctx, *grad_output):
         a2ai = ctx.a2ai
         myreq = ctx.myreq
-        myreq.req.wait()
+        wait_req(myreq)
         myreq.req = None
         grad_input = myreq.tensor
         if a2ai.codecs is not None:
@@ -1980,7 +1989,7 @@ def forward(
         a2ai = myreq.a2ai
         ctx.a2ai = a2ai
         assert myreq.req is not None
-        myreq.req.wait()
+        wait_req(myreq)
         myreq.req = None
         output = myreq.tensor
         myreq.tensor = None
@@ -2067,7 +2076,7 @@ def forward(
     def backward(ctx, *unused: Tensor) -> Tuple[Optional[Tensor], ...]:
         myreq = ctx.myreq
         assert myreq.req is not None
-        myreq.req.wait()
+        wait_req(myreq)
         myreq.req = None
         grad_inputs = list(myreq.tensor)
         rsi = myreq.rsi
@@ -2095,7 +2104,7 @@ def forward(
         *dummy_tensor: Tensor,
     ) -> Tensor:
         assert myreq.req is not None
-        myreq.req.wait()
+        wait_req(myreq)
         myreq.req = None
         output = myreq.tensor
         myreq.tensor = None
@@ -2174,7 +2183,7 @@ def forward(
     # pyre-fixme[2]: Parameter must be annotated.
     def backward(ctx, *unused: Tensor) -> Tuple[Optional[Tensor], ...]:
         myreq = ctx.myreq
-        myreq.req.wait()
+        wait_req(myreq)
         myreq.req = None
         grad_inputs = myreq.tensor
         rsi = myreq.rsi
@@ -2199,7 +2208,7 @@ def forward(
         *dummy_Tensor: Tensor,
     ) -> Tensor:
         assert myreq.req is not None
-        myreq.req.wait()
+        wait_req(myreq)
         myreq.req = None
         output = myreq.tensor
         myreq.tensor = None
@@ -2270,7 +2279,7 @@ def forward(
     def backward(ctx, *unused: Tensor) -> Tuple[Optional[Tensor], ...]:
         myreq = ctx.myreq
         assert myreq.req is not None
-        myreq.req.wait()
+        wait_req(myreq)
         myreq.req = None
         agi = myreq.agi
         grad_input = myreq.tensor
@@ -2296,7 +2305,7 @@ def forward(
         *dummy_tensor: Tensor,
     ) -> Tensor:
         assert myreq.req is not None
-        myreq.req.wait()
+        wait_req(myreq)
         myreq.req = None
         outputs = myreq.tensor
         myreq.tensor = None
@@ -2382,7 +2391,7 @@ def forward(
     def backward(ctx, *unused: Tensor) -> Tuple[Optional[Tensor], ...]:
         myreq = ctx.myreq
         assert myreq.req is not None
-        myreq.req.wait()
+        wait_req(myreq)
         myreq.req = None
         grad_input = myreq.tensor
         rsi = myreq.rsi
@@ -2407,7 +2416,7 @@ def forward(
         *dummy_tensor: Tensor,
     ) -> Tensor:
         assert myreq.req is not None
-        myreq.req.wait()
+        wait_req(myreq)
         myreq.req = None
         # pyre-ignore
         output: torch.Tensor = myreq.tensor