Add sdpa and fa2 the Wav2vec2 family. (#30121)

* add sdpa to wav2vec.
Co-authored-by: kamilakesbi <kamil@huggingface.co>
Co-authored-by: jp1924 <jp42maru@gmail.com>

* add fa2 to wav2vec2

* add tests

* fix attention_mask compatibility with fa2

* minor dtype fix

* replace fa2 slow test

* fix fa2 slow test

* apply code review + add fa2 batch test

* add sdpa and fa2 to hubert

* sdpa and fa2 to data2vec_audio

* sdpa and fa2 to Sew

* sdpa to unispeech + unispeech sat

* small fix

* attention mask in tests
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* add_speedup_benchmark_to_doc

---------
Co-authored-by: kamil@huggingface.co <kamil.akesbi@gmail.com>
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

docs/source/en/model_doc/hubert.md

@@ -44,6 +44,42 @@ This model was contributed by [patrickvonplaten](https://huggingface.co/patrickv
 - Hubert model was fine-tuned using connectionist temporal classification (CTC) so the model output has to be decoded
   using [`Wav2Vec2CTCTokenizer`].
 
+
+## Using Flash Attention 2
+
+Flash Attention 2 is an faster, optimized version of the model.
+
+### Installation 
+
+First, check whether your hardware is compatible with Flash Attention 2. The latest list of compatible hardware can be found in the [official documentation](https://github.com/Dao-AILab/flash-attention#installation-and-features). If your hardware is not compatible with Flash Attention 2, you can still benefit from attention kernel optimisations through Better Transformer support covered [above](https://huggingface.co/docs/transformers/main/en/model_doc/bark#using-better-transformer).
+
+Next, [install](https://github.com/Dao-AILab/flash-attention#installation-and-features) the latest version of Flash Attention 2:
+
+```bash
+pip install -U flash-attn --no-build-isolation
+```
+
+### Usage
+
+Below is an expected speedup diagram comparing the pure inference time between the native implementation in transformers of `facebook/hubert-large-ls960-ft`, the flash-attention-2 and the sdpa (scale-dot-product-attention) version. We show the average speedup obtained on the `librispeech_asr` `clean` validation split: 
+
+```python
+>>> from transformers import Wav2Vec2Model
+
+model = Wav2Vec2Model.from_pretrained("facebook/hubert-large-ls960-ft", torch_dtype=torch.float16, attn_implementation="flash_attention_2").to(device)
+...
+```
+
+### Expected speedups
+
+Below is an expected speedup diagram comparing the pure inference time between the native implementation in transformers of the `facebook/hubert-large-ls960-ft` model and the flash-attention-2 and sdpa (scale-dot-product-attention) versions. . We show the average speedup obtained on the `librispeech_asr` `clean` validation split: 
+
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/kamilakesbi/transformers_image_doc/resolve/main/data/Hubert_speedup.png">
+</div>
+
+
 ## Resources
 
 - [Audio classification task guide](../tasks/audio_classification)

docs/source/en/model_doc/wav2vec2.md

@@ -39,6 +39,42 @@ This model was contributed by [patrickvonplaten](https://huggingface.co/patrickv
 - Wav2Vec2 model was trained using connectionist temporal classification (CTC) so the model output has to be decoded
   using [`Wav2Vec2CTCTokenizer`].
 
+## Using Flash Attention 2
+
+Flash Attention 2 is an faster, optimized version of the model.
+
+### Installation 
+
+First, check whether your hardware is compatible with Flash Attention 2. The latest list of compatible hardware can be found in the [official documentation](https://github.com/Dao-AILab/flash-attention#installation-and-features). If your hardware is not compatible with Flash Attention 2, you can still benefit from attention kernel optimisations through Better Transformer support covered [above](https://huggingface.co/docs/transformers/main/en/model_doc/bark#using-better-transformer).
+
+Next, [install](https://github.com/Dao-AILab/flash-attention#installation-and-features) the latest version of Flash Attention 2:
+
+```bash
+pip install -U flash-attn --no-build-isolation
+```
+
+### Usage
+
+To load a model using Flash Attention 2, we can pass the argument `attn_implementation="flash_attention_2"` to [`.from_pretrained`](https://huggingface.co/docs/transformers/main/en/main_classes/model#transformers.PreTrainedModel.from_pretrained). We'll also load the model in half-precision (e.g. `torch.float16`), since it results in almost no degradation to audio quality but significantly lower memory usage and faster inference:
+
+```python
+>>> from transformers import Wav2Vec2Model
+
+model = Wav2Vec2Model.from_pretrained("facebook/wav2vec2-large-960h-lv60-self", torch_dtype=torch.float16, attn_implementation="flash_attention_2").to(device)
+...
+```
+
+### Expected speedups
+
+Below is an expected speedup diagram comparing the pure inference time between the native implementation in transformers of the `facebook/wav2vec2-large-960h-lv60-self` model and the flash-attention-2 and sdpa (scale-dot-product-attention) versions. . We show the average speedup obtained on the `librispeech_asr` `clean` validation split: 
+
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/kamilakesbi/transformers_image_doc/resolve/main/data/Wav2Vec2_speedup.png">
+</div>
+
+
+
 ## Resources
 
 A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with Wav2Vec2. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.

docs/source/en/perf_infer_gpu_one.md

@@ -70,6 +70,12 @@ FlashAttention-2 is currently supported for the following architectures:
 * [Qwen2](https://huggingface.co/docs/transformers/model_doc/qwen2#transformers.Qwen2Model)
 * [Qwen2MoE](https://huggingface.co/docs/transformers/model_doc/qwen2_moe#transformers.Qwen2MoeModel)
 * [Whisper](https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperModel)
+* [Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2#transformers.Wav2Vec2Model)
+* [Hubert](https://huggingface.co/docs/transformers/model_doc/hubert#transformers.HubertModel)
+* [data2vec_audio](https://huggingface.co/docs/transformers/main/en/model_doc/data2vec#transformers.Data2VecAudioModel)
+* [Sew](https://huggingface.co/docs/transformers/main/en/model_doc/sew#transformers.SEWModel)
+* [UniSpeech](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/unispeech#transformers.UniSpeechModel)
+* [unispeech_sat](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/unispeech-sat#transformers.UniSpeechSatModel)
 
 You can request to add FlashAttention-2 support for another model by opening a GitHub Issue or Pull Request.
 
@@ -203,6 +209,13 @@ For now, Transformers supports SDPA inference and training for the following arc
 * [Qwen2MoE](https://huggingface.co/docs/transformers/model_doc/qwen2_moe#transformers.Qwen2MoeModel)
 * [Musicgen](https://huggingface.co/docs/transformers/model_doc/musicgen#transformers.MusicgenModel)
 * [MusicGen Melody](https://huggingface.co/docs/transformers/model_doc/musicgen_melody#transformers.MusicgenMelodyModel)
+* [wav2vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2#transformers.Wav2Vec2Model)
+* [Hubert](https://huggingface.co/docs/transformers/model_doc/hubert#transformers.HubertModel)
+* [data2vec_audio](https://huggingface.co/docs/transformers/main/en/model_doc/data2vec#transformers.Data2VecAudioModel)
+* [Sew](https://huggingface.co/docs/transformers/main/en/model_doc/sew#transformers.SEWModel)
+* [UniSpeech](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/unispeech#transformers.UniSpeechModel)
+* [unispeech_sat](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/unispeech-sat#transformers.UniSpeechSatModel)
+
 
 <Tip>
 

src/transformers/models/data2vec/modeling_data2vec_audio.py

@@ -20,6 +20,7 @@ from typing import Optional, Tuple, Union
 
 import numpy as np
 import torch
+import torch.nn.functional as F
 import torch.utils.checkpoint
 from torch import nn
 from torch.nn import CrossEntropyLoss
@@ -39,12 +40,18 @@ from ...utils import (
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
     is_peft_available,
     logging,
 )
 from .configuration_data2vec_audio import Data2VecAudioConfig
 
 
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
 logger = logging.get_logger(__name__)
 
 
@@ -65,6 +72,19 @@ _CTC_EXPECTED_LOSS = 66.95
 from ..deprecated._archive_maps import DATA2VEC_AUDIO_PRETRAINED_MODEL_ARCHIVE_LIST  # noqa: F401, E402
 
 
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
 # Copied from transformers.models.wav2vec2.modeling_wav2vec2._compute_mask_indices
 def _compute_mask_indices(
     shape: Tuple[int, int],
@@ -478,6 +498,335 @@ class Data2VecAudioAttention(nn.Module):
         return attn_output, attn_weights_reshaped, past_key_value
 
 
+# Copied from transformers.models.bart.modeling_bart.BartFlashAttention2 with Bart->Data2VecAudio
+class Data2VecAudioFlashAttention2(Data2VecAudioAttention):
+    """
+    Data2VecAudio flash attention module. This module inherits from `Data2VecAudioAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def _reshape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # Data2VecAudioFlashAttention2 attention does not support output_attentions
+        if output_attentions:
+            raise ValueError("Data2VecAudioFlashAttention2 attention does not support output_attentions")
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, q_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self._reshape(self.q_proj(hidden_states), -1, bsz)
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0].transpose(1, 2)
+            value_states = past_key_value[1].transpose(1, 2)
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._reshape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0].transpose(1, 2), key_states], dim=1)
+            value_states = torch.cat([past_key_value[1].transpose(1, 2), value_states], dim=1)
+        else:
+            # self_attention
+            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states.transpose(1, 2), value_states.transpose(1, 2))
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-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. (LlamaRMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=self.dropout
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, -1)
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._flash_attention_forward
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
+
+        return attn_output
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._upad_input
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+class Data2VecAudioSdpaAttention(Data2VecAudioAttention):
+    # Copied from transformers.models.bart.modeling_bart.BartSdpaAttention.forward with Bart->Data2VecAudio
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+        if output_attentions or layer_head_mask is not None:
+            # TODO: Improve this warning with e.g. `model.config._attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "Data2VecAudioModel is using Data2VecAudioSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True` or `layer_head_mask` not None. Falling back to the manual attention"
+                ' implementation, but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states,
+                key_value_states=key_value_states,
+                past_key_value=past_key_value,
+                attention_mask=attention_mask,
+                layer_head_mask=layer_head_mask,
+                output_attentions=output_attentions,
+            )
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, tgt_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states)
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        query_states = self._shape(query_states, tgt_len, bsz)
+
+        # NOTE: SDPA with memory-efficient backend is currently (torch==2.1.2) bugged when using non-contiguous inputs and a custom attn_mask,
+        # but we are fine here as `_shape` do call `.contiguous()`. Reference: https://github.com/pytorch/pytorch/issues/112577
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.dropout if self.training else 0.0,
+            # The tgt_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case tgt_len == 1.
+            is_causal=self.is_causal and attention_mask is None and tgt_len > 1,
+        )
+
+        if attn_output.size() != (bsz, self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2)
+
+        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+        # partitioned across GPUs when using tensor-parallelism.
+        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+DATA2VEC2AUDIO_ATTENTION_CLASSES = {
+    "eager": Data2VecAudioAttention,
+    "sdpa": Data2VecAudioSdpaAttention,
+    "flash_attention_2": Data2VecAudioFlashAttention2,
+}
+
+
 # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeedForward with Wav2Vec2->Data2VecAudio
 class Data2VecAudioFeedForward(nn.Module):
     def __init__(self, config):
@@ -503,16 +852,17 @@ class Data2VecAudioFeedForward(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayer with Wav2Vec2->Data2VecAudio
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayer with Wav2Vec2->Data2VecAudio, WAV2VEC2->DATA2VEC2AUDIO
 class Data2VecAudioEncoderLayer(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.attention = Data2VecAudioAttention(
+        self.attention = DATA2VEC2AUDIO_ATTENTION_CLASSES[config._attn_implementation](
             embed_dim=config.hidden_size,
             num_heads=config.num_attention_heads,
             dropout=config.attention_dropout,
             is_decoder=False,
         )
+
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.feed_forward = Data2VecAudioFeedForward(config)
@@ -548,6 +898,7 @@ class Data2VecAudioEncoder(nn.Module):
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layers = nn.ModuleList([Data2VecAudioEncoderLayer(config) for _ in range(config.num_hidden_layers)])
         self.gradient_checkpointing = False
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
 
     def forward(
         self,
@@ -564,7 +915,10 @@ class Data2VecAudioEncoder(nn.Module):
             # make sure padded tokens output 0
             expand_attention_mask = attention_mask.unsqueeze(-1).repeat(1, 1, hidden_states.shape[2])
             hidden_states[~expand_attention_mask] = 0
-
+            if self._use_flash_attention_2:
+                # 2d mask is passed through the layers
+                attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+            else:
                 # extend attention_mask
                 attention_mask = 1.0 - attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)
                 attention_mask = attention_mask * torch.finfo(hidden_states.dtype).min
@@ -681,6 +1035,8 @@ class Data2VecAudioPreTrainedModel(PreTrainedModel):
     base_model_prefix = "data2vec_audio"
     main_input_name = "input_values"
     supports_gradient_checkpointing = True
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
 
     def _init_weights(self, module):
         """Initialize the weights"""

src/transformers/models/hubert/modeling_hubert.py

@@ -19,6 +19,7 @@ from typing import Optional, Tuple, Union
 
 import numpy as np
 import torch
+import torch.nn.functional as F
 import torch.utils.checkpoint
 from torch import nn
 from torch.nn import CrossEntropyLoss
@@ -31,12 +32,19 @@ from ...utils import (
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
     logging,
     replace_return_docstrings,
 )
 from .configuration_hubert import HubertConfig
 
 
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+
 logger = logging.get_logger(__name__)
 
 _HIDDEN_STATES_START_POSITION = 1
@@ -61,6 +69,19 @@ _SEQ_CLASS_EXPECTED_LOSS = 8.53
 from ..deprecated._archive_maps import HUBERT_PRETRAINED_MODEL_ARCHIVE_LIST  # noqa: F401, E402
 
 
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
 # Copied from transformers.models.wav2vec2.modeling_wav2vec2._compute_mask_indices
 def _compute_mask_indices(
     shape: Tuple[int, int],
@@ -541,6 +562,335 @@ class HubertAttention(nn.Module):
         return attn_output, attn_weights_reshaped, past_key_value
 
 
+# Copied from transformers.models.bart.modeling_bart.BartFlashAttention2 with Bart->Hubert
+class HubertFlashAttention2(HubertAttention):
+    """
+    Hubert flash attention module. This module inherits from `HubertAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def _reshape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # HubertFlashAttention2 attention does not support output_attentions
+        if output_attentions:
+            raise ValueError("HubertFlashAttention2 attention does not support output_attentions")
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, q_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self._reshape(self.q_proj(hidden_states), -1, bsz)
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0].transpose(1, 2)
+            value_states = past_key_value[1].transpose(1, 2)
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._reshape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0].transpose(1, 2), key_states], dim=1)
+            value_states = torch.cat([past_key_value[1].transpose(1, 2), value_states], dim=1)
+        else:
+            # self_attention
+            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states.transpose(1, 2), value_states.transpose(1, 2))
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-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. (LlamaRMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=self.dropout
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, -1)
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._flash_attention_forward
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
+
+        return attn_output
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._upad_input
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+class HubertSdpaAttention(HubertAttention):
+    # Copied from transformers.models.bart.modeling_bart.BartSdpaAttention.forward with Bart->Hubert
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+        if output_attentions or layer_head_mask is not None:
+            # TODO: Improve this warning with e.g. `model.config._attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "HubertModel is using HubertSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True` or `layer_head_mask` not None. Falling back to the manual attention"
+                ' implementation, but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states,
+                key_value_states=key_value_states,
+                past_key_value=past_key_value,
+                attention_mask=attention_mask,
+                layer_head_mask=layer_head_mask,
+                output_attentions=output_attentions,
+            )
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, tgt_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states)
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        query_states = self._shape(query_states, tgt_len, bsz)
+
+        # NOTE: SDPA with memory-efficient backend is currently (torch==2.1.2) bugged when using non-contiguous inputs and a custom attn_mask,
+        # but we are fine here as `_shape` do call `.contiguous()`. Reference: https://github.com/pytorch/pytorch/issues/112577
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.dropout if self.training else 0.0,
+            # The tgt_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case tgt_len == 1.
+            is_causal=self.is_causal and attention_mask is None and tgt_len > 1,
+        )
+
+        if attn_output.size() != (bsz, self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2)
+
+        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+        # partitioned across GPUs when using tensor-parallelism.
+        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+HUBERT_ATTENTION_CLASSES = {
+    "eager": HubertAttention,
+    "sdpa": HubertSdpaAttention,
+    "flash_attention_2": HubertFlashAttention2,
+}
+
+
 # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeedForward with Wav2Vec2->Hubert
 class HubertFeedForward(nn.Module):
     def __init__(self, config):
@@ -566,16 +916,17 @@ class HubertFeedForward(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayer with Wav2Vec2->Hubert
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayer with Wav2Vec2->Hubert, WAV2VEC2->HUBERT
 class HubertEncoderLayer(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.attention = HubertAttention(
+        self.attention = HUBERT_ATTENTION_CLASSES[config._attn_implementation](
             embed_dim=config.hidden_size,
             num_heads=config.num_attention_heads,
             dropout=config.attention_dropout,
             is_decoder=False,
         )
+
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.feed_forward = HubertFeedForward(config)
@@ -627,11 +978,11 @@ class HubertAttnAdapterLayer(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayerStableLayerNorm with Wav2Vec2->Hubert
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayerStableLayerNorm with Wav2Vec2->Hubert, WAV2VEC2->HUBERT
 class HubertEncoderLayerStableLayerNorm(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.attention = HubertAttention(
+        self.attention = HUBERT_ATTENTION_CLASSES[config._attn_implementation](
             embed_dim=config.hidden_size,
             num_heads=config.num_attention_heads,
             dropout=config.attention_dropout,
@@ -683,6 +1034,7 @@ class HubertEncoder(nn.Module):
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layers = nn.ModuleList([HubertEncoderLayer(config) for _ in range(config.num_hidden_layers)])
         self.gradient_checkpointing = False
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
 
     def forward(
         self,
@@ -699,7 +1051,10 @@ class HubertEncoder(nn.Module):
             # make sure padded tokens output 0
             expand_attention_mask = attention_mask.unsqueeze(-1).repeat(1, 1, hidden_states.shape[2])
             hidden_states[~expand_attention_mask] = 0
-
+            if self._use_flash_attention_2:
+                # 2d mask is passed through the layers
+                attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+            else:
                 # extend attention_mask
                 attention_mask = 1.0 - attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)
                 attention_mask = attention_mask * torch.finfo(hidden_states.dtype).min
@@ -767,6 +1122,7 @@ class HubertEncoderStableLayerNorm(nn.Module):
             [HubertEncoderLayerStableLayerNorm(config) for _ in range(config.num_hidden_layers)]
         )
         self.gradient_checkpointing = False
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
 
     def forward(
         self,
@@ -783,7 +1139,10 @@ class HubertEncoderStableLayerNorm(nn.Module):
             # make sure padded tokens are not attended to
             expand_attention_mask = attention_mask.unsqueeze(-1).repeat(1, 1, hidden_states.shape[2])
             hidden_states[~expand_attention_mask] = 0
-
+            if self._use_flash_attention_2:
+                # 2d mask is passed through the layers
+                attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+            else:
                 # extend attention_mask
                 attention_mask = 1.0 - attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)
                 attention_mask = attention_mask * torch.finfo(hidden_states.dtype).min
@@ -851,6 +1210,8 @@ class HubertPreTrainedModel(PreTrainedModel):
     base_model_prefix = "hubert"
     main_input_name = "input_values"
     supports_gradient_checkpointing = True
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
 
     def _init_weights(self, module):
         """Initialize the weights"""

src/transformers/models/sew/modeling_sew.py

@@ -20,6 +20,7 @@ from typing import Optional, Tuple, Union
 
 import numpy as np
 import torch
+import torch.nn.functional as F
 import torch.utils.checkpoint
 from torch import nn
 from torch.nn import CrossEntropyLoss
@@ -28,10 +29,22 @@ from ...activations import ACT2FN
 from ...integrations.deepspeed import is_deepspeed_zero3_enabled
 from ...modeling_outputs import BaseModelOutput, CausalLMOutput, SequenceClassifierOutput
 from ...modeling_utils import PreTrainedModel
-from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from ...utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+)
 from .configuration_sew import SEWConfig
 
 
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+
 logger = logging.get_logger(__name__)
 
 
@@ -59,6 +72,19 @@ _SEQ_CLASS_EXPECTED_LOSS = 9.52
 from ..deprecated._archive_maps import SEW_PRETRAINED_MODEL_ARCHIVE_LIST  # noqa: F401, E402
 
 
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
 # Copied from transformers.models.wav2vec2.modeling_wav2vec2._compute_mask_indices
 def _compute_mask_indices(
     shape: Tuple[int, int],
@@ -536,6 +562,335 @@ class SEWAttention(nn.Module):
         return attn_output, attn_weights_reshaped, past_key_value
 
 
+# Copied from transformers.models.bart.modeling_bart.BartFlashAttention2 with Bart->SEW
+class SEWFlashAttention2(SEWAttention):
+    """
+    SEW flash attention module. This module inherits from `SEWAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def _reshape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # SEWFlashAttention2 attention does not support output_attentions
+        if output_attentions:
+            raise ValueError("SEWFlashAttention2 attention does not support output_attentions")
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, q_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self._reshape(self.q_proj(hidden_states), -1, bsz)
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0].transpose(1, 2)
+            value_states = past_key_value[1].transpose(1, 2)
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._reshape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0].transpose(1, 2), key_states], dim=1)
+            value_states = torch.cat([past_key_value[1].transpose(1, 2), value_states], dim=1)
+        else:
+            # self_attention
+            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states.transpose(1, 2), value_states.transpose(1, 2))
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-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. (LlamaRMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=self.dropout
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, -1)
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._flash_attention_forward
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
+
+        return attn_output
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._upad_input
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+class SEWSdpaAttention(SEWAttention):
+    # Copied from transformers.models.bart.modeling_bart.BartSdpaAttention.forward with Bart->SEW
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+        if output_attentions or layer_head_mask is not None:
+            # TODO: Improve this warning with e.g. `model.config._attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "SEWModel is using SEWSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True` or `layer_head_mask` not None. Falling back to the manual attention"
+                ' implementation, but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states,
+                key_value_states=key_value_states,
+                past_key_value=past_key_value,
+                attention_mask=attention_mask,
+                layer_head_mask=layer_head_mask,
+                output_attentions=output_attentions,
+            )
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, tgt_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states)
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        query_states = self._shape(query_states, tgt_len, bsz)
+
+        # NOTE: SDPA with memory-efficient backend is currently (torch==2.1.2) bugged when using non-contiguous inputs and a custom attn_mask,
+        # but we are fine here as `_shape` do call `.contiguous()`. Reference: https://github.com/pytorch/pytorch/issues/112577
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.dropout if self.training else 0.0,
+            # The tgt_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case tgt_len == 1.
+            is_causal=self.is_causal and attention_mask is None and tgt_len > 1,
+        )
+
+        if attn_output.size() != (bsz, self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2)
+
+        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+        # partitioned across GPUs when using tensor-parallelism.
+        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+SEW_ATTENTION_CLASSES = {
+    "eager": SEWAttention,
+    "sdpa": SEWSdpaAttention,
+    "flash_attention_2": SEWFlashAttention2,
+}
+
+
 # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeedForward with Wav2Vec2->SEW
 class SEWFeedForward(nn.Module):
     def __init__(self, config):
@@ -561,16 +916,17 @@ class SEWFeedForward(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayer with Wav2Vec2->SEW
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayer with Wav2Vec2->SEW, WAV2VEC2->SEW
 class SEWEncoderLayer(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.attention = SEWAttention(
+        self.attention = SEW_ATTENTION_CLASSES[config._attn_implementation](
             embed_dim=config.hidden_size,
             num_heads=config.num_attention_heads,
             dropout=config.attention_dropout,
             is_decoder=False,
         )
+
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.feed_forward = SEWFeedForward(config)
@@ -607,6 +963,7 @@ class SEWEncoder(nn.Module):
         self.layers = nn.ModuleList([SEWEncoderLayer(config) for _ in range(config.num_hidden_layers)])
         self.upsample = SEWUpsampling(config)
         self.gradient_checkpointing = False
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
 
     def forward(
         self,
@@ -620,6 +977,12 @@ class SEWEncoder(nn.Module):
         all_self_attentions = () if output_attentions else None
 
         if attention_mask is not None:
+            if self._use_flash_attention_2:
+                # make sure padded tokens output 0
+                hidden_states[~attention_mask] = 0.0
+                # 2d mask is passed through the layers
+                attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+            else:
                 # make sure padded tokens output 0
                 hidden_states[~attention_mask] = 0.0
 
@@ -710,6 +1073,8 @@ class SEWPreTrainedModel(PreTrainedModel):
     base_model_prefix = "sew"
     main_input_name = "input_values"
     supports_gradient_checkpointing = True
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
 
     def _init_weights(self, module):
         """Initialize the weights"""

src/transformers/models/unispeech/modeling_unispeech.py

@@ -21,6 +21,7 @@ from typing import Optional, Tuple, Union
 
 import numpy as np
 import torch
+import torch.nn.functional as F
 import torch.utils.checkpoint
 from torch import nn
 from torch.nn import CrossEntropyLoss
@@ -34,12 +35,19 @@ from ...utils import (
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
     logging,
     replace_return_docstrings,
 )
 from .configuration_unispeech import UniSpeechConfig
 
 
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+
 logger = logging.get_logger(__name__)
 
 
@@ -60,6 +68,19 @@ _CTC_EXPECTED_LOSS = 17.17
 from ..deprecated._archive_maps import UNISPEECH_PRETRAINED_MODEL_ARCHIVE_LIST  # noqa: F401, E402
 
 
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
 @dataclass
 class UniSpeechForPreTrainingOutput(ModelOutput):
     """
@@ -577,6 +598,335 @@ class UniSpeechAttention(nn.Module):
         return attn_output, attn_weights_reshaped, past_key_value
 
 
+# Copied from transformers.models.bart.modeling_bart.BartFlashAttention2 with Bart->UniSpeech
+class UniSpeechFlashAttention2(UniSpeechAttention):
+    """
+    UniSpeech flash attention module. This module inherits from `UniSpeechAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def _reshape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # UniSpeechFlashAttention2 attention does not support output_attentions
+        if output_attentions:
+            raise ValueError("UniSpeechFlashAttention2 attention does not support output_attentions")
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, q_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self._reshape(self.q_proj(hidden_states), -1, bsz)
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0].transpose(1, 2)
+            value_states = past_key_value[1].transpose(1, 2)
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._reshape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0].transpose(1, 2), key_states], dim=1)
+            value_states = torch.cat([past_key_value[1].transpose(1, 2), value_states], dim=1)
+        else:
+            # self_attention
+            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states.transpose(1, 2), value_states.transpose(1, 2))
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-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. (LlamaRMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=self.dropout
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, -1)
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._flash_attention_forward
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
+
+        return attn_output
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._upad_input
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+class UniSpeechSdpaAttention(UniSpeechAttention):
+    # Copied from transformers.models.bart.modeling_bart.BartSdpaAttention.forward with Bart->UniSpeech
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+        if output_attentions or layer_head_mask is not None:
+            # TODO: Improve this warning with e.g. `model.config._attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "UniSpeechModel is using UniSpeechSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True` or `layer_head_mask` not None. Falling back to the manual attention"
+                ' implementation, but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states,
+                key_value_states=key_value_states,
+                past_key_value=past_key_value,
+                attention_mask=attention_mask,
+                layer_head_mask=layer_head_mask,
+                output_attentions=output_attentions,
+            )
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, tgt_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states)
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        query_states = self._shape(query_states, tgt_len, bsz)
+
+        # NOTE: SDPA with memory-efficient backend is currently (torch==2.1.2) bugged when using non-contiguous inputs and a custom attn_mask,
+        # but we are fine here as `_shape` do call `.contiguous()`. Reference: https://github.com/pytorch/pytorch/issues/112577
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.dropout if self.training else 0.0,
+            # The tgt_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case tgt_len == 1.
+            is_causal=self.is_causal and attention_mask is None and tgt_len > 1,
+        )
+
+        if attn_output.size() != (bsz, self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2)
+
+        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+        # partitioned across GPUs when using tensor-parallelism.
+        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+UNISPEECH_ATTENTION_CLASSES = {
+    "eager": UniSpeechAttention,
+    "sdpa": UniSpeechSdpaAttention,
+    "flash_attention_2": UniSpeechFlashAttention2,
+}
+
+
 # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeedForward with Wav2Vec2->UniSpeech
 class UniSpeechFeedForward(nn.Module):
     def __init__(self, config):
@@ -602,16 +952,17 @@ class UniSpeechFeedForward(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayer with Wav2Vec2->UniSpeech
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayer with Wav2Vec2->UniSpeech, WAV2VEC2->UNISPEECH
 class UniSpeechEncoderLayer(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.attention = UniSpeechAttention(
+        self.attention = UNISPEECH_ATTENTION_CLASSES[config._attn_implementation](
             embed_dim=config.hidden_size,
             num_heads=config.num_attention_heads,
             dropout=config.attention_dropout,
             is_decoder=False,
         )
+
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.feed_forward = UniSpeechFeedForward(config)
@@ -663,11 +1014,11 @@ class UniSpeechAttnAdapterLayer(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayerStableLayerNorm with Wav2Vec2->UniSpeech
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayerStableLayerNorm with Wav2Vec2->UniSpeech, WAV2VEC2->UNISPEECH
 class UniSpeechEncoderLayerStableLayerNorm(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.attention = UniSpeechAttention(
+        self.attention = UNISPEECH_ATTENTION_CLASSES[config._attn_implementation](
             embed_dim=config.hidden_size,
             num_heads=config.num_attention_heads,
             dropout=config.attention_dropout,
@@ -719,6 +1070,7 @@ class UniSpeechEncoder(nn.Module):
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layers = nn.ModuleList([UniSpeechEncoderLayer(config) for _ in range(config.num_hidden_layers)])
         self.gradient_checkpointing = False
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
 
     def forward(
         self,
@@ -735,7 +1087,10 @@ class UniSpeechEncoder(nn.Module):
             # make sure padded tokens output 0
             expand_attention_mask = attention_mask.unsqueeze(-1).repeat(1, 1, hidden_states.shape[2])
             hidden_states[~expand_attention_mask] = 0
-
+            if self._use_flash_attention_2:
+                # 2d mask is passed through the layers
+                attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+            else:
                 # extend attention_mask
                 attention_mask = 1.0 - attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)
                 attention_mask = attention_mask * torch.finfo(hidden_states.dtype).min
@@ -803,6 +1158,7 @@ class UniSpeechEncoderStableLayerNorm(nn.Module):
             [UniSpeechEncoderLayerStableLayerNorm(config) for _ in range(config.num_hidden_layers)]
         )
         self.gradient_checkpointing = False
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
 
     def forward(
         self,
@@ -819,7 +1175,10 @@ class UniSpeechEncoderStableLayerNorm(nn.Module):
             # make sure padded tokens are not attended to
             expand_attention_mask = attention_mask.unsqueeze(-1).repeat(1, 1, hidden_states.shape[2])
             hidden_states[~expand_attention_mask] = 0
-
+            if self._use_flash_attention_2:
+                # 2d mask is passed through the layers
+                attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+            else:
                 # extend attention_mask
                 attention_mask = 1.0 - attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)
                 attention_mask = attention_mask * torch.finfo(hidden_states.dtype).min
@@ -957,6 +1316,8 @@ class UniSpeechPreTrainedModel(PreTrainedModel):
     base_model_prefix = "unispeech"
     main_input_name = "input_values"
     supports_gradient_checkpointing = True
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
 
     def _init_weights(self, module):
         """Initialize the weights"""

src/transformers/models/unispeech_sat/modeling_unispeech_sat.py

@@ -21,6 +21,7 @@ from typing import Optional, Tuple, Union
 
 import numpy as np
 import torch
+import torch.nn.functional as F
 import torch.utils.checkpoint
 from torch import nn
 from torch.nn import CrossEntropyLoss
@@ -41,6 +42,8 @@ from ...utils import (
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
     is_peft_available,
     logging,
     replace_return_docstrings,
@@ -48,6 +51,11 @@ from ...utils import (
 from .configuration_unispeech_sat import UniSpeechSatConfig
 
 
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+
 logger = logging.get_logger(__name__)
 
 
@@ -76,6 +84,19 @@ _XVECTOR_EXPECTED_OUTPUT = 0.97
 from ..deprecated._archive_maps import UNISPEECH_SAT_PRETRAINED_MODEL_ARCHIVE_LIST  # noqa: F401, E402
 
 
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
 @dataclass
 class UniSpeechSatForPreTrainingOutput(ModelOutput):
     """
@@ -594,6 +615,335 @@ class UniSpeechSatAttention(nn.Module):
         return attn_output, attn_weights_reshaped, past_key_value
 
 
+# Copied from transformers.models.bart.modeling_bart.BartFlashAttention2 with Bart->UniSpeechSat
+class UniSpeechSatFlashAttention2(UniSpeechSatAttention):
+    """
+    UniSpeechSat flash attention module. This module inherits from `UniSpeechSatAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def _reshape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # UniSpeechSatFlashAttention2 attention does not support output_attentions
+        if output_attentions:
+            raise ValueError("UniSpeechSatFlashAttention2 attention does not support output_attentions")
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, q_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self._reshape(self.q_proj(hidden_states), -1, bsz)
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0].transpose(1, 2)
+            value_states = past_key_value[1].transpose(1, 2)
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._reshape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0].transpose(1, 2), key_states], dim=1)
+            value_states = torch.cat([past_key_value[1].transpose(1, 2), value_states], dim=1)
+        else:
+            # self_attention
+            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states.transpose(1, 2), value_states.transpose(1, 2))
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-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. (LlamaRMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=self.dropout
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, -1)
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._flash_attention_forward
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
+
+        return attn_output
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._upad_input
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+class UniSpeechSatSdpaAttention(UniSpeechSatAttention):
+    # Copied from transformers.models.bart.modeling_bart.BartSdpaAttention.forward with Bart->UniSpeechSat
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+        if output_attentions or layer_head_mask is not None:
+            # TODO: Improve this warning with e.g. `model.config._attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "UniSpeechSatModel is using UniSpeechSatSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True` or `layer_head_mask` not None. Falling back to the manual attention"
+                ' implementation, but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states,
+                key_value_states=key_value_states,
+                past_key_value=past_key_value,
+                attention_mask=attention_mask,
+                layer_head_mask=layer_head_mask,
+                output_attentions=output_attentions,
+            )
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, tgt_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states)
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        query_states = self._shape(query_states, tgt_len, bsz)
+
+        # NOTE: SDPA with memory-efficient backend is currently (torch==2.1.2) bugged when using non-contiguous inputs and a custom attn_mask,
+        # but we are fine here as `_shape` do call `.contiguous()`. Reference: https://github.com/pytorch/pytorch/issues/112577
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.dropout if self.training else 0.0,
+            # The tgt_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case tgt_len == 1.
+            is_causal=self.is_causal and attention_mask is None and tgt_len > 1,
+        )
+
+        if attn_output.size() != (bsz, self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2)
+
+        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+        # partitioned across GPUs when using tensor-parallelism.
+        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+UNISPEECHSAT_ATTENTION_CLASSES = {
+    "eager": UniSpeechSatAttention,
+    "sdpa": UniSpeechSatSdpaAttention,
+    "flash_attention_2": UniSpeechSatFlashAttention2,
+}
+
+
 # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeedForward with Wav2Vec2->UniSpeechSat
 class UniSpeechSatFeedForward(nn.Module):
     def __init__(self, config):
@@ -619,16 +969,17 @@ class UniSpeechSatFeedForward(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayer with Wav2Vec2->UniSpeechSat
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayer with Wav2Vec2->UniSpeechSat, WAV2VEC2->UNISPEECHSAT
 class UniSpeechSatEncoderLayer(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.attention = UniSpeechSatAttention(
+        self.attention = UNISPEECHSAT_ATTENTION_CLASSES[config._attn_implementation](
             embed_dim=config.hidden_size,
             num_heads=config.num_attention_heads,
             dropout=config.attention_dropout,
             is_decoder=False,
         )
+
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.feed_forward = UniSpeechSatFeedForward(config)
@@ -680,11 +1031,11 @@ class UniSpeechSatAttnAdapterLayer(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayerStableLayerNorm with Wav2Vec2->UniSpeechSat
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayerStableLayerNorm with Wav2Vec2->UniSpeechSat, WAV2VEC2->UNISPEECHSAT
 class UniSpeechSatEncoderLayerStableLayerNorm(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.attention = UniSpeechSatAttention(
+        self.attention = UNISPEECHSAT_ATTENTION_CLASSES[config._attn_implementation](
             embed_dim=config.hidden_size,
             num_heads=config.num_attention_heads,
             dropout=config.attention_dropout,
@@ -736,6 +1087,7 @@ class UniSpeechSatEncoder(nn.Module):
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layers = nn.ModuleList([UniSpeechSatEncoderLayer(config) for _ in range(config.num_hidden_layers)])
         self.gradient_checkpointing = False
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
 
     def forward(
         self,
@@ -752,7 +1104,10 @@ class UniSpeechSatEncoder(nn.Module):
             # make sure padded tokens output 0
             expand_attention_mask = attention_mask.unsqueeze(-1).repeat(1, 1, hidden_states.shape[2])
             hidden_states[~expand_attention_mask] = 0
-
+            if self._use_flash_attention_2:
+                # 2d mask is passed through the layers
+                attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+            else:
                 # extend attention_mask
                 attention_mask = 1.0 - attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)
                 attention_mask = attention_mask * torch.finfo(hidden_states.dtype).min
@@ -820,6 +1175,7 @@ class UniSpeechSatEncoderStableLayerNorm(nn.Module):
             [UniSpeechSatEncoderLayerStableLayerNorm(config) for _ in range(config.num_hidden_layers)]
         )
         self.gradient_checkpointing = False
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
 
     def forward(
         self,
@@ -836,7 +1192,10 @@ class UniSpeechSatEncoderStableLayerNorm(nn.Module):
             # make sure padded tokens are not attended to
             expand_attention_mask = attention_mask.unsqueeze(-1).repeat(1, 1, hidden_states.shape[2])
             hidden_states[~expand_attention_mask] = 0
-
+            if self._use_flash_attention_2:
+                # 2d mask is passed through the layers
+                attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+            else:
                 # extend attention_mask
                 attention_mask = 1.0 - attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)
                 attention_mask = attention_mask * torch.finfo(hidden_states.dtype).min
@@ -974,6 +1333,8 @@ class UniSpeechSatPreTrainedModel(PreTrainedModel):
     base_model_prefix = "unispeech_sat"
     main_input_name = "input_values"
     supports_gradient_checkpointing = True
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
 
     def _init_weights(self, module):
         """Initialize the weights"""

src/transformers/models/wav2vec2/modeling_wav2vec2.py

@@ -21,6 +21,7 @@ from typing import Optional, Tuple, Union
 
 import numpy as np
 import torch
+import torch.nn.functional as F
 import torch.utils.checkpoint
 from torch import nn
 from torch.nn import CrossEntropyLoss
@@ -44,6 +45,8 @@ from ...utils import (
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
     cached_file,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
     is_peft_available,
     is_safetensors_available,
     logging,
@@ -59,6 +62,10 @@ if is_safetensors_available():
     from safetensors.torch import load_file as safe_load_file
 
 
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
 logger = logging.get_logger(__name__)
 
 
@@ -92,6 +99,19 @@ _XVECTOR_EXPECTED_OUTPUT = 0.98
 from ..deprecated._archive_maps import WAV_2_VEC_2_PRETRAINED_MODEL_ARCHIVE_LIST  # noqa: F401, E402
 
 
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
 @dataclass
 class Wav2Vec2ForPreTrainingOutput(ModelOutput):
     """
@@ -642,6 +662,336 @@ class Wav2Vec2Attention(nn.Module):
         return attn_output, attn_weights_reshaped, past_key_value
 
 
+# Copied from transformers.models.bart.modeling_bart.BartFlashAttention2 with Bart->Wav2Vec2
+class Wav2Vec2FlashAttention2(Wav2Vec2Attention):
+    """
+    Wav2Vec2 flash attention module. This module inherits from `Wav2Vec2Attention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def _reshape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # Wav2Vec2FlashAttention2 attention does not support output_attentions
+        if output_attentions:
+            raise ValueError("Wav2Vec2FlashAttention2 attention does not support output_attentions")
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, q_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self._reshape(self.q_proj(hidden_states), -1, bsz)
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0].transpose(1, 2)
+            value_states = past_key_value[1].transpose(1, 2)
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._reshape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0].transpose(1, 2), key_states], dim=1)
+            value_states = torch.cat([past_key_value[1].transpose(1, 2), value_states], dim=1)
+        else:
+            # self_attention
+            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states.transpose(1, 2), value_states.transpose(1, 2))
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-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. (LlamaRMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=self.dropout
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, -1)
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._flash_attention_forward
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
+
+        return attn_output
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._upad_input
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+class Wav2Vec2SdpaAttention(Wav2Vec2Attention):
+    # Copied from transformers.models.bart.modeling_bart.BartSdpaAttention.forward with Bart->Wav2Vec2
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+        if output_attentions or layer_head_mask is not None:
+            # TODO: Improve this warning with e.g. `model.config._attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "Wav2Vec2Model is using Wav2Vec2SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True` or `layer_head_mask` not None. Falling back to the manual attention"
+                ' implementation, but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states,
+                key_value_states=key_value_states,
+                past_key_value=past_key_value,
+                attention_mask=attention_mask,
+                layer_head_mask=layer_head_mask,
+                output_attentions=output_attentions,
+            )
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, tgt_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states)
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        query_states = self._shape(query_states, tgt_len, bsz)
+
+        # NOTE: SDPA with memory-efficient backend is currently (torch==2.1.2) bugged when using non-contiguous inputs and a custom attn_mask,
+        # but we are fine here as `_shape` do call `.contiguous()`. Reference: https://github.com/pytorch/pytorch/issues/112577
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.dropout if self.training else 0.0,
+            # The tgt_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case tgt_len == 1.
+            is_causal=self.is_causal and attention_mask is None and tgt_len > 1,
+        )
+
+        if attn_output.size() != (bsz, self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2)
+
+        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+        # partitioned across GPUs when using tensor-parallelism.
+        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+WAV2VEC2_ATTENTION_CLASSES = {
+    "eager": Wav2Vec2Attention,
+    "sdpa": Wav2Vec2SdpaAttention,
+    "flash_attention_2": Wav2Vec2FlashAttention2,
+}
+
+
 class Wav2Vec2FeedForward(nn.Module):
     def __init__(self, config):
         super().__init__()
@@ -669,12 +1019,13 @@ class Wav2Vec2FeedForward(nn.Module):
 class Wav2Vec2EncoderLayer(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.attention = Wav2Vec2Attention(
+        self.attention = WAV2VEC2_ATTENTION_CLASSES[config._attn_implementation](
             embed_dim=config.hidden_size,
             num_heads=config.num_attention_heads,
             dropout=config.attention_dropout,
             is_decoder=False,
         )
+
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.feed_forward = Wav2Vec2FeedForward(config)
@@ -703,7 +1054,7 @@ class Wav2Vec2EncoderLayer(nn.Module):
 class Wav2Vec2EncoderLayerStableLayerNorm(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.attention = Wav2Vec2Attention(
+        self.attention = WAV2VEC2_ATTENTION_CLASSES[config._attn_implementation](
             embed_dim=config.hidden_size,
             num_heads=config.num_attention_heads,
             dropout=config.attention_dropout,
@@ -754,6 +1105,7 @@ class Wav2Vec2Encoder(nn.Module):
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layers = nn.ModuleList([Wav2Vec2EncoderLayer(config) for _ in range(config.num_hidden_layers)])
         self.gradient_checkpointing = False
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
 
     def forward(
         self,
@@ -770,7 +1122,10 @@ class Wav2Vec2Encoder(nn.Module):
             # make sure padded tokens output 0
             expand_attention_mask = attention_mask.unsqueeze(-1).repeat(1, 1, hidden_states.shape[2])
             hidden_states[~expand_attention_mask] = 0
-
+            if self._use_flash_attention_2:
+                # 2d mask is passed through the layers
+                attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+            else:
                 # extend attention_mask
                 attention_mask = 1.0 - attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)
                 attention_mask = attention_mask * torch.finfo(hidden_states.dtype).min
@@ -837,6 +1192,7 @@ class Wav2Vec2EncoderStableLayerNorm(nn.Module):
             [Wav2Vec2EncoderLayerStableLayerNorm(config) for _ in range(config.num_hidden_layers)]
         )
         self.gradient_checkpointing = False
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
 
     def forward(
         self,
@@ -853,7 +1209,10 @@ class Wav2Vec2EncoderStableLayerNorm(nn.Module):
             # make sure padded tokens are not attended to
             expand_attention_mask = attention_mask.unsqueeze(-1).repeat(1, 1, hidden_states.shape[2])
             hidden_states[~expand_attention_mask] = 0
-
+            if self._use_flash_attention_2:
+                # 2d mask is passed through the layers
+                attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+            else:
                 # extend attention_mask
                 attention_mask = 1.0 - attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)
                 attention_mask = attention_mask * torch.finfo(hidden_states.dtype).min
@@ -1071,6 +1430,8 @@ class Wav2Vec2PreTrainedModel(PreTrainedModel):
     base_model_prefix = "wav2vec2"
     main_input_name = "input_values"
     supports_gradient_checkpointing = True
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
 
     def _init_weights(self, module):
         """Initialize the weights"""
@@ -1742,6 +2103,8 @@ class Wav2Vec2ForPreTraining(Wav2Vec2PreTrainedModel):
         quantized_features, codevector_perplexity = self.quantizer(
             extract_features, mask_time_indices=mask_time_indices
         )
+
+        quantized_features = quantized_features.to(self.project_q.weight.dtype)
         quantized_features = self.project_q(quantized_features)
 
         loss = contrastive_loss = diversity_loss = None

src/transformers/models/wav2vec2_conformer/modeling_wav2vec2_conformer.py

@@ -1515,6 +1515,8 @@ class Wav2Vec2ConformerForPreTraining(Wav2Vec2ConformerPreTrainedModel):
         quantized_features, codevector_perplexity = self.quantizer(
             extract_features, mask_time_indices=mask_time_indices
         )
+
+        quantized_features = quantized_features.to(self.project_q.weight.dtype)
         quantized_features = self.project_q(quantized_features)
 
         loss = contrastive_loss = diversity_loss = None

tests/models/wav2vec2/test_modeling_wav2vec2.py

@@ -25,6 +25,7 @@ import unittest
 
 import numpy as np
 from datasets import load_dataset
+from pytest import mark
 
 from transformers import Wav2Vec2Config, is_torch_available
 from transformers.testing_utils import (
@@ -33,9 +34,11 @@ from transformers.testing_utils import (
     is_pt_flax_cross_test,
     is_pyctcdecode_available,
     is_torchaudio_available,
+    require_flash_attn,
     require_pyctcdecode,
     require_soundfile,
     require_torch,
+    require_torch_gpu,
     require_torchaudio,
     run_test_in_subprocess,
     slow,
@@ -1995,3 +1998,52 @@ class Wav2Vec2ModelIntegrationTest(unittest.TestCase):
 
         for lang in LANG_MAP.keys():
             assert run_model(lang) == TRANSCRIPTIONS[lang]
+
+    @require_flash_attn
+    @require_torch_gpu
+    @mark.flash_attn_test
+    def test_inference_ctc_fa2(self):
+        model_fa = Wav2Vec2ForCTC.from_pretrained(
+            "facebook/wav2vec2-base-960h", attn_implementation="flash_attention_2", torch_dtype=torch.bfloat16
+        )
+        model_fa.to(torch_device)
+        processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h", do_lower_case=True)
+        input_speech = self._load_datasamples(1)
+
+        input_values = processor(input_speech, return_tensors="pt").input_values.to(torch_device)
+
+        with torch.no_grad():
+            logits = model_fa(input_values.to(torch.bfloat16)).logits
+
+        predicted_ids = torch.argmax(logits, dim=-1)
+        predicted_trans = processor.batch_decode(predicted_ids)
+
+        EXPECTED_TRANSCRIPTIONS = ["a man said to the universe sir i exist"]
+        self.assertListEqual(predicted_trans, EXPECTED_TRANSCRIPTIONS)
+
+    @require_flash_attn
+    @require_torch_gpu
+    @mark.flash_attn_test
+    def test_inference_ctc_fa2_batched(self):
+        model_fa = Wav2Vec2ForCTC.from_pretrained(
+            "facebook/wav2vec2-base-960h", attn_implementation="flash_attention_2", torch_dtype=torch.bfloat16
+        )
+        model_fa.to(torch_device)
+        processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h", do_lower_case=True)
+
+        input_speech = self._load_datasamples(2)
+
+        inputs = processor(input_speech, return_tensors="pt", padding=True, return_attention_mask=True)
+        inputs = inputs.to(torch_device)
+
+        with torch.no_grad():
+            logits = model_fa(inputs.input_values.to(torch.bfloat16), attention_mask=inputs.attention_mask).logits
+
+        predicted_ids = torch.argmax(logits, dim=-1)
+        predicted_trans = processor.batch_decode(predicted_ids)
+
+        EXPECTED_TRANSCRIPTIONS = [
+            "a man said to the universe sir i exist",
+            "sweat covered brion's body trickling into the tight lowing cloth that was the only garment he wore",
+        ]
+        self.assertListEqual(predicted_trans, EXPECTED_TRANSCRIPTIONS)