Flash Attention.

config.json

@@ -41,5 +41,6 @@
   "share_decoder_input_output_embed": false,
   "torch_dtype": "float32",
   "transformers_version": "4.34.0.dev0",
-  "use_cache": true
+  "use_cache": true,
+  "attn_implementation": "eager"
 }

configuration_indictrans.py

@@ -118,6 +118,7 @@ class IndicTransConfig(PretrainedConfig):
         pad_token_id=1,
         bos_token_id=0,
         eos_token_id=2,
+        attn_implementation="eager",
         **kwargs,
     ):
         self.encoder_vocab_size = encoder_vocab_size
@@ -146,7 +147,8 @@ class IndicTransConfig(PretrainedConfig):
         self.num_hidden_layers = encoder_layers
         self.scale_embedding = scale_embedding
         self.share_decoder_input_output_embed = share_decoder_input_output_embed
-
+        self.attn_implementation = attn_implementation
+        
         super().__init__(
             pad_token_id=pad_token_id,
             bos_token_id=bos_token_id,
@@ -304,4 +306,4 @@ class IndicTransOnnxConfig(OnnxSeq2SeqConfigWithPast):
                 )
         return common_inputs
 
-    generate_dummy_inputs = _generate_dummy_inputs_for_default_and_seq2seq_lm
+    generate_dummy_inputs = _generate_dummy_inputs_for_default_and_seq2seq_lm

modeling_indictrans.py

@@ -23,15 +23,28 @@ import torch.nn as nn
 from torch.nn import functional as F
 
 from transformers.activations import ACT2FN
+
+from transformers.modeling_attn_mask_utils import (
+    _prepare_4d_attention_mask,
+    _prepare_4d_attention_mask_for_sdpa,
+    _prepare_4d_causal_attention_mask,
+    _prepare_4d_causal_attention_mask_for_sdpa,
+)
+
 from transformers.integrations.deepspeed import is_deepspeed_zero3_enabled
 from transformers.modeling_outputs import (
     BaseModelOutput,
     BaseModelOutputWithPastAndCrossAttentions,
     Seq2SeqLMOutput,
-    Seq2SeqModelOutput,
+    Seq2SeqModelOutput
 )
 
-from transformers.utils import logging
+from transformers.utils import (
+    logging,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+)
+ 
 from transformers.modeling_utils import PreTrainedModel
 
 from .configuration_indictrans import IndicTransConfig
@@ -39,10 +52,25 @@ from .configuration_indictrans import IndicTransConfig
 
 logger = logging.get_logger(__name__)
 
-_CONFIG_FOR_DOC = "IndicTransConfig"
-
 INDICTRANS_PRETRAINED_MODEL_ARCHIVE_LIST = [""]
 
+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
+
+
+# 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.bart.modeling_bart.shift_tokens_right
 def shift_tokens_right(
@@ -63,54 +91,6 @@ def shift_tokens_right(
     return shifted_input_ids
 
 
-# Copied from transformers.models.bart.modeling_bart._make_causal_mask
-def _make_causal_mask(
-    input_ids_shape: torch.Size,
-    dtype: torch.dtype,
-    device: torch.device,
-    past_key_values_length: int = 0,
-):
-    """
-    Make causal mask used for bi-directional self-attention.
-    """
-    bsz, tgt_len = input_ids_shape
-    mask = torch.full((tgt_len, tgt_len), torch.finfo(dtype).min, device=device)
-    mask_cond = torch.arange(mask.size(-1), device=device)
-    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
-    mask = mask.to(dtype)
-
-    if past_key_values_length > 0:
-        mask = torch.cat(
-            [
-                torch.zeros(
-                    tgt_len, past_key_values_length, dtype=dtype, device=device
-                ),
-                mask,
-            ],
-            dim=-1,
-        )
-    return mask[None, None, :, :].expand(
-        bsz, 1, tgt_len, tgt_len + past_key_values_length
-    )
-
-
-# Copied from transformers.models.bart.modeling_bart._expand_mask
-def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
-    """
-    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
-    """
-    bsz, src_len = mask.size()
-    tgt_len = tgt_len if tgt_len is not None else src_len
-
-    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
-
-    inverted_mask = 1.0 - expanded_mask
-
-    return inverted_mask.masked_fill(
-        inverted_mask.to(torch.bool), torch.finfo(dtype).min
-    )
-
-
 def create_position_ids_from_input_ids(
     input_ids, padding_idx, past_key_values_length=0
 ):
@@ -247,12 +227,15 @@ class IndicTransAttention(nn.Module):
         dropout: float = 0.0,
         is_decoder: bool = False,
         bias: bool = True,
+        is_causal: bool = False,
+        config: Optional[IndicTransConfig] = None,
     ):
         super().__init__()
         self.embed_dim = embed_dim
         self.num_heads = num_heads
         self.dropout = dropout
         self.head_dim = embed_dim // num_heads
+        self.config = config
 
         if (self.head_dim * num_heads) != self.embed_dim:
             raise ValueError(
@@ -261,6 +244,7 @@ class IndicTransAttention(nn.Module):
             )
         self.scaling = self.head_dim**-0.5
         self.is_decoder = is_decoder
+        self.is_causal = is_causal
 
         self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
         self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
@@ -402,17 +386,345 @@ class IndicTransAttention(nn.Module):
         attn_output = self.out_proj(attn_output)
 
         return attn_output, attn_weights_reshaped, past_key_value
+    
+
+class IndicTransFlashAttention2(IndicTransAttention):
+    """
+    IndicTrans flash attention module. This module inherits from `IndicTransAttention` 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]]]:
+        # IndicTransFlashAttention2 attention does not support output_attentions
+        if output_attentions:
+            raise ValueError("IndicTransFlashAttention2 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 IndicTransSdpaAttention(IndicTransAttention):
+    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(
+                "IndicTransModel is using IndicTransSdpaAttention, 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 = F.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
+
+
+INDICTRANS_ATTENTION_CLASSES = {
+    "eager": IndicTransAttention,
+    "sdpa": IndicTransSdpaAttention,
+    "flash_attention_2": IndicTransFlashAttention2,
+}
 
 # Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->IndicTrans
 class IndicTransEncoderLayer(nn.Module):
     def __init__(self, config: IndicTransConfig):
         super().__init__()
         self.embed_dim = config.encoder_embed_dim
-        self.self_attn = IndicTransAttention(
+        self.self_attn = INDICTRANS_ATTENTION_CLASSES[config._attn_implementation](
             embed_dim=self.embed_dim,
             num_heads=config.encoder_attention_heads,
             dropout=config.attention_dropout,
+            config=config,
         )
         self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
         self.dropout = config.dropout
@@ -490,22 +802,25 @@ class IndicTransDecoderLayer(nn.Module):
         super().__init__()
         self.embed_dim = config.decoder_embed_dim
 
-        self.self_attn = IndicTransAttention(
+        self.self_attn = INDICTRANS_ATTENTION_CLASSES[config._attn_implementation](
             embed_dim=self.embed_dim,
             num_heads=config.decoder_attention_heads,
             dropout=config.attention_dropout,
             is_decoder=True,
+            is_causal=True,
+            config=config,
         )
         self.dropout = config.dropout
         self.activation_fn = ACT2FN[config.activation_function]
         self.activation_dropout = config.activation_dropout
 
         self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
-        self.encoder_attn = IndicTransAttention(
+        self.encoder_attn = INDICTRANS_ATTENTION_CLASSES[config._attn_implementation](
             self.embed_dim,
             config.decoder_attention_heads,
             dropout=config.attention_dropout,
             is_decoder=True,
+            config=config,
         )
         self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
         self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
@@ -693,6 +1008,9 @@ class IndicTransEncoder(IndicTransPreTrainedModel):
             nn.LayerNorm(embed_dim) if config.layernorm_embedding else None
         )
 
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+        self._use_sdpa = config._attn_implementation == "sdpa"
+
         self.gradient_checkpointing = False
         # Initialize weights and apply final processing
         self.post_init()
@@ -782,10 +1100,18 @@ class IndicTransEncoder(IndicTransPreTrainedModel):
             hidden_states = self.layernorm_embedding(hidden_states)
         hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
 
-        # expand attention_mask
         if attention_mask is not None:
-            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
-            attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)
+            if self._use_flash_attention_2:
+                attention_mask = attention_mask if 0 in attention_mask else None
+            elif self._use_sdpa and head_mask is None and not output_attentions:
+                # output_attentions=True & head_mask can not be supported when using SDPA, fall back to
+                # the manual implementation that requires a 4D causal mask in all cases.
+                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+                attention_mask = _prepare_4d_attention_mask_for_sdpa(attention_mask, inputs_embeds.dtype)
+            else:
+                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+                attention_mask = _prepare_4d_attention_mask(attention_mask, inputs_embeds.dtype)
+
 
         encoder_states = () if output_hidden_states else None
         all_attentions = () if output_attentions else None
@@ -909,6 +1235,9 @@ class IndicTransDecoder(IndicTransPreTrainedModel):
             nn.LayerNorm(embed_dim) if config.layernorm_embedding else None
         )
 
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+        self._use_sdpa = config._attn_implementation == "sdpa"
+
         self.gradient_checkpointing = False
         # Initialize weights and apply final processing
         self.post_init()
@@ -1031,29 +1360,43 @@ class IndicTransDecoder(IndicTransPreTrainedModel):
         if inputs_embeds is None:
             inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
 
-        # create causal mask
-        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
-        combined_attention_mask = None
-        if input_shape[-1] > 1:
-            combined_attention_mask = _make_causal_mask(
+
+        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
+        elif self._use_sdpa and not output_attentions and cross_attn_head_mask is None:
+            # output_attentions=True & cross_attn_head_mask can not be supported when using SDPA, and we fall back on
+            # the manual implementation that requires a 4D causal mask in all cases.
+            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
+                attention_mask,
                 input_shape,
-                inputs_embeds.dtype,
-                device=inputs_embeds.device,
-                past_key_values_length=past_key_values_length,
+                inputs_embeds,
+                past_key_values_length,
             )
-
-        if attention_mask is not None and combined_attention_mask is not None:
-            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
-            combined_attention_mask = combined_attention_mask + _expand_mask(
-                attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
+        else:
+            # 4d mask is passed through the layers
+            attention_mask = _prepare_4d_causal_attention_mask(
+                attention_mask, input_shape, inputs_embeds, past_key_values_length
             )
 
         # expand encoder attention mask
         if encoder_hidden_states is not None and encoder_attention_mask is not None:
-            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
-            encoder_attention_mask = _expand_mask(
-                encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
-            )
+            if self._use_flash_attention_2:
+                encoder_attention_mask = encoder_attention_mask if 0 in encoder_attention_mask else None
+            elif self._use_sdpa and cross_attn_head_mask is None and not output_attentions:
+                # output_attentions=True & cross_attn_head_mask can not be supported when using SDPA, and we fall back on
+                # the manual implementation that requires a 4D causal mask in all cases.
+                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+                encoder_attention_mask = _prepare_4d_attention_mask_for_sdpa(
+                    encoder_attention_mask,
+                    inputs_embeds.dtype,
+                    tgt_len=input_shape[-1],
+                )
+            else:
+                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+                encoder_attention_mask = _prepare_4d_attention_mask(
+                    encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
+                )
 
         # embed positions
         positions = self.embed_positions(
@@ -1124,7 +1467,7 @@ class IndicTransDecoder(IndicTransPreTrainedModel):
                     layer_outputs = torch.utils.checkpoint.checkpoint(
                         create_custom_forward(decoder_layer),
                         hidden_states,
-                        combined_attention_mask,
+                        attention_mask,
                         encoder_hidden_states,
                         encoder_attention_mask,
                         head_mask[idx] if head_mask is not None else None,
@@ -1136,7 +1479,7 @@ class IndicTransDecoder(IndicTransPreTrainedModel):
                 else:
                     layer_outputs = decoder_layer(
                         hidden_states,
-                        attention_mask=combined_attention_mask,
+                        attention_mask=attention_mask,
                         encoder_hidden_states=encoder_hidden_states,
                         encoder_attention_mask=encoder_attention_mask,
                         layer_head_mask=(