Add an auto_expand option to SinusoidalPositionalEmbedding (#5555)

fairseq/modules/positional_embedding.py

@@ -14,6 +14,7 @@ def PositionalEmbedding(
     embedding_dim: int,
     padding_idx: int,
     learned: bool = False,
+    auto_expand: bool = True,
 ):
     if learned:
         # if padding_idx is specified then offset the embedding ids by
@@ -31,5 +32,6 @@ def PositionalEmbedding(
             embedding_dim,
             padding_idx,
             init_size=num_embeddings + padding_idx + 1,
+            auto_expand=auto_expand,
         )
     return m

fairseq/modules/sinusoidal_positional_embedding.py

@@ -18,14 +18,19 @@ class SinusoidalPositionalEmbedding(nn.Module):
     Padding symbols are ignored.
     """
 
-    def __init__(self, embedding_dim, padding_idx, init_size=1024):
+    def __init__(self, embedding_dim, padding_idx, init_size=1024, auto_expand=True):
         super().__init__()
         self.embedding_dim = embedding_dim
         self.padding_idx = padding_idx if padding_idx is not None else 0
-        self.register_buffer("weights", SinusoidalPositionalEmbedding.get_embedding(
-            init_size, embedding_dim, padding_idx
-        ), persistent=False)
+        self.register_buffer(
+            "weights",
+            SinusoidalPositionalEmbedding.get_embedding(
+                init_size, embedding_dim, padding_idx
+            ),
+            persistent=False,
+        )
         self.max_positions = int(1e5)
+        self.auto_expand = auto_expand
         self.onnx_trace = False
 
     def prepare_for_onnx_export_(self):
@@ -75,28 +80,36 @@ def forward(
         bspair = torch.onnx.operators.shape_as_tensor(input)
         bsz, seq_len = bspair[0], bspair[1]
         max_pos = self.padding_idx + 1 + seq_len
+        weights = self.weights
+
         if max_pos > self.weights.size(0):
-            # expand embeddings if needed
-            self.weights = SinusoidalPositionalEmbedding.get_embedding(
+            # If the input is longer than the number of pre-computed embeddings,
+            # compute the extra embeddings on the fly.
+            # Only store the expanded embeddings if auto_expand=True.
+            # In multithreading environments, mutating the weights of a module
+            # may cause trouble. Set auto_expand=False if this happens.
+            weights = SinusoidalPositionalEmbedding.get_embedding(
                 max_pos, self.embedding_dim, self.padding_idx
             ).to(self.weights)
+            if self.auto_expand:
+                self.weights = weights
 
         if incremental_state is not None:
             # positions is the same for every token when decoding a single step
             pos = timestep.view(-1)[0] + 1 if timestep is not None else seq_len
             if self.onnx_trace:
                 return (
-                    self.weights.index_select(index=self.padding_idx + pos, dim=0)
+                    weights.index_select(index=self.padding_idx + pos, dim=0)
                     .unsqueeze(1)
                     .repeat(bsz, 1, 1)
                 )
-            return self.weights[self.padding_idx + pos, :].expand(bsz, 1, -1)
+            return weights[self.padding_idx + pos, :].expand(bsz, 1, -1)
 
         positions = utils.make_positions(
             input, self.padding_idx, onnx_trace=self.onnx_trace
         )
         if self.onnx_trace:
-            flat_embeddings = self.weights.detach().index_select(0, positions.view(-1))
+            flat_embeddings = weights.detach().index_select(0, positions.view(-1))
             embedding_shape = torch.cat(
                 (bsz.view(1), seq_len.view(1), torch.tensor([-1], dtype=torch.long))
             )
@@ -105,7 +118,5 @@ def forward(
             )
             return embeddings
         return (
-            self.weights.index_select(0, positions.view(-1))
-            .view(bsz, seq_len, -1)
-            .detach()
+            weights.index_select(0, positions.view(-1)).view(bsz, seq_len, -1).detach()
         )