Cleaner attention interfaces

src/transformers/integrations/flash_attention.py

@@ -29,18 +29,13 @@ def flash_attention_forward(
     key = key.transpose(1, 2)
     value = value.transpose(1, 2)
 
-    if query.dtype == torch.float32:
-        query = query.to(torch.float16)
-        key = key.to(torch.float16)
-        value = value.to(torch.float16)
-
     attn_output = _flash_attention_forward(
         query,
         key,
         value,
         attention_mask,
-        seq_len,
-        module.is_causal,
+        query_length=seq_len,
+        is_causal=module.is_causal,
         dropout=dropout,
         softmax_scale=scaling,
         sliding_window=sliding_window,

src/transformers/integrations/flex_attention.py

@@ -2,10 +2,10 @@
 
 import torch
 
-from ..utils import is_torch_greater_or_equal
+from ..utils import is_torch_flex_attn_available
 
 
-if is_torch_greater_or_equal("2.5"):
+if is_torch_flex_attn_available():
     from torch.nn.attention.flex_attention import flex_attention
 
 
@@ -37,8 +37,12 @@ def causal_mod(score, b, h, q_idx, kv_idx):
         score_mod=causal_mod,
         enable_gqa=True,
         scale=scaling,
+        # Last time checked on PyTorch == 2.5.1: Flex Attention always computes the lse regardless.
+        # For simplification, we thus always return it as no additional computations are introduced.
         return_lse=True,
     )
+    # lse is returned in float32
+    attention_weights = attention_weights.to(value.dtype)
     attn_output = attn_output.transpose(1, 2).contiguous()
 
     return attn_output, attention_weights

src/transformers/integrations/sdpa_attention.py

@@ -1,6 +1,20 @@
 from typing import Optional, Tuple
 
 import torch
+from packaging import version
+
+from ..utils import get_torch_version
+
+
+def sdpa_needs_contiguous_inputs():
+    """Check if the currently installed version of torch sdpa is bugged with non-contiguous inputs.
+    Reference: https://github.com/pytorch/pytorch/issues/112577
+    """
+    torch_version = version.parse(get_torch_version())
+    return version.parse("2.1.0") <= torch_version < version.parse("2.2.0")
+
+
+_needs_contiguous_inputs = sdpa_needs_contiguous_inputs()
 
 
 def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
@@ -34,10 +48,15 @@ def sdpa_attention_forward(
     if attention_mask is not None:
         causal_mask = causal_mask[:, :, :, : key.shape[-2]]
 
-    query = query.contiguous()
-    key = key.contiguous()
-    value = value.contiguous()
+    # SDPA with memory-efficient backend is bugged with non-contiguous inputs and custom attn_mask for some torch versions
+    # Reference: https://github.com/pytorch/pytorch/issues/112577.
+    if _needs_contiguous_inputs:
+        query = query.contiguous()
+        key = key.contiguous()
+        value = value.contiguous()
 
+    # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+    # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
     if is_causal is None:
         is_causal = causal_mask is None and query.shape[2] > 1
 

src/transformers/modeling_utils.py

@@ -1473,12 +1473,7 @@ def _autoset_attn_implementation(
                     ' We recommend to just use `attn_implementation="flash_attention_2"` when loading the model.'
                 )
 
-            if not isinstance(config._attn_implementation, dict) and config._attn_implementation not in [
-                "eager",
-                "sdpa",
-                "flash_attention_2",
-                "flex_attention",
-            ]:
+            if not isinstance(config._attn_implementation, dict) and config._attn_implementation not in ["eager"] + list(ALL_ATTENTION_FUNCTIONS.keys()):
                 message = f'Specified `attn_implementation="{config._attn_implementation}"` is not supported. The only possible arguments are `attn_implementation="eager"` (manual attention implementation)'
                 if cls._supports_flash_attn_2:
                     message += ', `"attn_implementation=flash_attention_2"` (implementation using flash attention 2)'
@@ -1540,6 +1535,8 @@ def _autoset_attn_implementation(
                     "Using the `SDPA` attention implementation on multi-gpu setup with ROCM may lead to performance issues due to the FA backend. Disabling it to use alternative backends."
                 )
                 torch.backends.cuda.enable_flash_sdp(False)
+        elif requested_attn_implementation in ALL_ATTENTION_FUNCTIONS.keys():
+            config._attn_implementation = requested_attn_implementation
         elif isinstance(requested_attn_implementation, dict):
             config._attn_implementation = None
         else: