dummy data

mosaicml · Jan 17, 2024 · b227bcf · b227bcf
1 parent 3d8cda8
commit b227bcf
Showing 1 changed file with 8 additions and 8 deletions.
diff --git a/llmfoundry/models/mpt/modeling_mpt.py b/llmfoundry/models/mpt/modeling_mpt.py
@@ -232,11 +232,13 @@ def gen_flash_attn_padding_info(
         attention_mask_in_length: Optional[torch.Tensor] = None,
         attention_mask: Optional[torch.Tensor] = None):
     flash_attn_padding_info = {}
+    dummy_data = torch.ones((bsz, past_key_len + S),
+                            dtype=torch.bool,
+                            device=device)
     if attention_mask_in_length is None:
         key_padding_mask = attention_mask
         if key_padding_mask is None:
-            key_padding_mask = torch.ones((bsz, past_key_len + S),
-                                          dtype=torch.bool, device=device)
+            key_padding_mask = dummy_data
         query_padding_mask = key_padding_mask[:, -S:]
         unpadding_function = bert_padding.unpad_input
     else:
@@ -245,13 +247,11 @@ def gen_flash_attn_padding_info(
         unpadding_function = bert_padding.unpad_input_for_concatenated_sequences
 
     _, indices_q, cu_seqlens_q, max_seqlen_q = unpadding_function(
-        torch.zeros(bsz, S, 1, device=device), query_padding_mask)
+        dummy_data[:, :S, None], query_padding_mask)
     _, indices_k, cu_seqlens_k, max_seqlen_k = unpadding_function(
-        torch.zeros(bsz, past_key_len + S, 1, device=device),
-        key_padding_mask)
-    _, indices_v, _, _ = unpadding_function(
-        torch.zeros(bsz, past_key_len + S, 1, device=device),
-        key_padding_mask)
+        dummy_data[:, :, None], key_padding_mask)
+    _, indices_v, _, _ = unpadding_function(dummy_data[:, :, None],
+                                            key_padding_mask)
 
     flash_attn_padding_info['indices_q'] = indices_q
     flash_attn_padding_info['indices_k'] = indices_k