Static attention: do not specialize on input sequence length (pytorch#13373)

Summary:

Hardcode embedding/head dims which are fixed and use -1 for the sequence length dimension when reshaping, this allows us to quantize the graph once and export it to different sequence lengths.

Differential Revision: D80181012

Author: sxu

examples/models/llama/static_attention.py

@@ -759,7 +759,7 @@ def forward(
 
         bsz, seq_len, dim = x.shape
         if self.use_conv2d:
-            x = x.reshape(bsz, seq_len, 1, dim).transpose(1, 3)
+            x = x.reshape(bsz, -1, 1, dim).transpose(1, 3)
 
         new_qs = [wq(x) for wq in self.wqs]
         new_ks = [wk(x) for wk in self.wks]
@@ -768,9 +768,7 @@ def forward(
         if self.use_conv2d:
 
             def from_conv2ds(ts):
-                return [
-                    t.reshape(bsz, self.head_dim, seq_len).transpose(1, 2) for t in ts
-                ]
+                return [t.reshape(bsz, self.head_dim, -1).transpose(1, 2) for t in ts]
 
             new_qs = from_conv2ds(new_qs)
             new_ks = from_conv2ds(new_ks)
@@ -800,9 +798,11 @@ def from_conv2ds(ts):
 
         if self.use_conv2d:
             y = (
-                self.wo(y.reshape(bsz, seq_len, 1, -1).transpose(1, 3))
+                self.wo(
+                    y.reshape(bsz, -1, 1, self.n_heads * self.head_dim).transpose(1, 3)
+                )
                 .transpose(1, 3)
-                .reshape(bsz, seq_len, -1)
+                .reshape(bsz, -1, self.dim)
             )
         else:
             y = self.wo(y)