bug fix

llm2vec/models/bidirectional_llama.py

@@ -32,6 +32,14 @@ def is_transformers_attn_greater_or_equal_4_38():
         "4.38.0"
     )
 
+def is_transformers_attn_greater_or_equal_4_40():
+    if not _is_package_available("transformers"):
+        return False
+
+    return version.parse(importlib.metadata.version("transformers")) >= version.parse(
+        "4.40.0"
+    )
+
 
 class ModifiedLlamaAttention(LlamaAttention):
     def __init__(self, *args, **kwargs):
@@ -99,12 +107,20 @@ def __init__(self, config: LlamaConfig):
         # Initialize weights and apply final processing
         self.post_init()
 
-    def _update_causal_mask(self, attention_mask, input_tensor, cache_position):
+    def _update_causal_mask(self, attention_mask, input_tensor, cache_position, past_seen_tokens=None):
         if self.config._attn_implementation == "flash_attention_2":
             if attention_mask is not None and 0.0 in attention_mask:
                 return attention_mask
             return None
 
+        if is_transformers_attn_greater_or_equal_4_40() and self.config._attn_implementation == "sdpa":
+            # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument,
+            # in order to dispatch on Flash Attention 2.
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask, inputs_embeds=input_tensor, past_key_values_length=past_seen_tokens
+            ):
+                return None
+
         dtype, device = input_tensor.dtype, input_tensor.device
         min_dtype = torch.finfo(dtype).min
         sequence_length = input_tensor.shape[1]
@@ -116,7 +132,7 @@ def _update_causal_mask(self, attention_mask, input_tensor, cache_position):
             target_length = (
                 attention_mask.shape[-1]
                 if isinstance(attention_mask, torch.Tensor)
-                else cache_position[-1] + 1
+                else (cache_position[-1] + 1 if not is_transformers_attn_greater_or_equal_4_40() else past_seen_tokens + sequence_length + 1)
             )
 
         causal_mask = torch.zeros(
@@ -164,19 +180,9 @@ def _update_causal_mask(self, attention_mask, input_tensor, cache_position):
             and attention_mask is not None
             and attention_mask.device.type == "cuda"
         ):
-            # TODO: For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400).
-            is_tracing = (
-                torch.jit.is_tracing()
-                or isinstance(input_tensor, torch.fx.Proxy)
-                or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
+            causal_mask = AttentionMaskConverter._unmask_unattended(
+                causal_mask, min_dtype
             )
-            if not is_tracing and torch.any(attention_mask != 1):
-                # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
-                # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
-                # Details: https://github.com/pytorch/pytorch/issues/110213
-                causal_mask = AttentionMaskConverter._unmask_unattended(
-                    causal_mask, min_dtype
-                )
 
         return causal_mask
 

llm2vec/version.py

@@ -1 +1 @@
-__version__ = "0.1.0"
+__version__ = "0.1.5"