Merge pull request meta-llama#900 from flu0r1ne/main

Fix key-value caching for seqlen != 1 (Issue meta-llama#899)

llama/model.py

@@ -289,12 +289,12 @@ def forward(
         values = self.cache_v[:bsz, : start_pos + seqlen]
 
         # repeat k/v heads if n_kv_heads < n_heads
-        keys = repeat_kv(keys, self.n_rep)  # (bs, seqlen, n_local_heads, head_dim)
-        values = repeat_kv(values, self.n_rep)  # (bs, seqlen, n_local_heads, head_dim)
+        keys = repeat_kv(keys, self.n_rep)  # (bs, cache_len + seqlen, n_local_heads, head_dim)
+        values = repeat_kv(values, self.n_rep)  # (bs, cache_len + seqlen, n_local_heads, head_dim)
 
         xq = xq.transpose(1, 2)  # (bs, n_local_heads, seqlen, head_dim)
-        keys = keys.transpose(1, 2)
-        values = values.transpose(1, 2)
+        keys = keys.transpose(1, 2) # (bs, n_local_heads, cache_len + seqlen, head_dim)
+        values = values.transpose(1, 2) # (bs, n_local_heads, cache_len + seqlen, head_dim)
         scores = torch.matmul(xq, keys.transpose(2, 3)) / math.sqrt(self.head_dim)
         if mask is not None:
             scores = scores + mask  # (bs, n_local_heads, seqlen, cache_len + seqlen)
@@ -474,9 +474,19 @@ def forward(self, tokens: torch.Tensor, start_pos: int):
         mask = None
         if seqlen > 1:
             mask = torch.full(
-                (1, 1, seqlen, seqlen), float("-inf"), device=tokens.device
+                (seqlen, seqlen), float("-inf"), device=tokens.device
             )
-            mask = torch.triu(mask, diagonal=start_pos + 1).type_as(h)
+
+            mask = torch.triu(mask, diagonal=1)
+
+            # When performing key-value caching, we compute the attention scores
+            # only for the new sequence. Thus, the matrix of scores is of size
+            # (seqlen, cache_len + seqlen), and the only masked entries are (i, j) for
+            # j > cache_len + i, since row i corresponds to token cache_len + i.
+            mask = torch.hstack([
+                torch.zeros((seqlen, start_pos), device=tokens.device),
+                mask
+            ]).type_as(h)
 
         for layer in self.layers:
             h = layer(h, start_pos, freqs_cis, mask)