Cleaner attention interfaces (#35342)

* cleaner attention interfaces

* correctly set the _attn_implementation when adding other functions to it

* update

* Update modeling_utils.py

* CIs

src/transformers/integrations/flash_attention.py

@@ -29,23 +29,34 @@ def flash_attention_forward(
     key = key.transpose(1, 2)
     value = value.transpose(1, 2)
 
+    # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+    # therefore the input hidden states gets silently casted in float32. Hence, we need
+    # cast them back in the correct dtype just to be sure everything works as expected.
+    # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+    # in fp32. (usually our RMSNorm modules handle it correctly)
+    target_dtype = None
     if query.dtype == torch.float32:
-        query = query.to(torch.float16)
-        key = key.to(torch.float16)
-        value = value.to(torch.float16)
+        if torch.is_autocast_enabled():
+            target_dtype = torch.get_autocast_gpu_dtype()
+        # Handle the case where the model is quantized
+        elif hasattr(module.config, "_pre_quantization_dtype"):
+            target_dtype = module.config._pre_quantization_dtype
+        else:
+            target_dtype = next(layer for layer in module.modules() if isinstance(layer, torch.nn.Linear)).weight.dtype
 
     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,
         softcap=softcap,
         use_top_left_mask=_use_top_left_mask,
+        target_dtype=target_dtype,
         **kwargs,
     )
 

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

@@ -34,10 +34,14 @@ def sdpa_attention_forward(
     if attention_mask is not None:
         causal_mask = causal_mask[:, :, :, : key.shape[-2]]
 
+    # 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.
     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

@@ -1474,11 +1474,8 @@ def _autoset_attn_implementation(
                 )
 
             if not isinstance(config._attn_implementation, dict) and config._attn_implementation not in [
-                "eager",
-                "sdpa",
-                "flash_attention_2",
-                "flex_attention",
-            ]:
+                "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 +1537,8 @@ 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 list(ALL_ATTENTION_FUNCTIONS.keys()):
+            config._attn_implementation = requested_attn_implementation
         elif isinstance(requested_attn_implementation, dict):
             config._attn_implementation = None
         else: