Add SDPA support for M2M100 (#33309)

* Add SDPA support for M2M100

* [run_slow] m2m_100, nllb

docs/source/en/model_doc/m2m_100.md

@@ -163,3 +163,21 @@ Below is an expected speedup diagram that compares pure inference time between t
 <div style="text-align: center">
 <img src="https://huggingface.co/datasets/visheratin/documentation-images/resolve/main/nllb-speedup.webp">
 </div>
+
+## Using Scaled Dot Product Attention (SDPA)
+PyTorch includes a native scaled dot-product attention (SDPA) operator as part of `torch.nn.functional`. This function
+encompasses several implementations that can be applied depending on the inputs and the hardware in use. See the
+[official documentation](https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html)
+or the [GPU Inference](https://huggingface.co/docs/transformers/main/en/perf_infer_gpu_one#pytorch-scaled-dot-product-attention)
+page for more information.
+
+SDPA is used by default for `torch>=2.1.1` when an implementation is available, but you may also set
+`attn_implementation="sdpa"` in `from_pretrained()` to explicitly request SDPA to be used.
+
+```python
+from transformers import M2M100ForConditionalGeneration
+model = M2M100ForConditionalGeneration.from_pretrained("facebook/m2m100_418M", torch_dtype=torch.float16, attn_implementation="sdpa")
+...
+```
+
+For the best speedups, we recommend loading the model in half-precision (e.g. `torch.float16` or `torch.bfloat16`).

docs/source/en/model_doc/nllb.md

@@ -188,3 +188,21 @@ Below is an expected speedup diagram that compares pure inference time between t
 <div style="text-align: center">
 <img src="https://huggingface.co/datasets/visheratin/documentation-images/resolve/main/nllb-speedup.webp">
 </div>
+
+## Using Scaled Dot Product Attention (SDPA)
+PyTorch includes a native scaled dot-product attention (SDPA) operator as part of `torch.nn.functional`. This function
+encompasses several implementations that can be applied depending on the inputs and the hardware in use. See the
+[official documentation](https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html)
+or the [GPU Inference](https://huggingface.co/docs/transformers/main/en/perf_infer_gpu_one#pytorch-scaled-dot-product-attention)
+page for more information.
+
+SDPA is used by default for `torch>=2.1.1` when an implementation is available, but you may also set
+`attn_implementation="sdpa"` in `from_pretrained()` to explicitly request SDPA to be used.
+
+```python
+from transformers import AutoModelForSeq2SeqLM
+model = AutoModelForSeq2SeqLM.from_pretrained("facebook/nllb-200-distilled-600M", torch_dtype=torch.float16, attn_implementation="sdpa")
+...
+```
+
+For the best speedups, we recommend loading the model in half-precision (e.g. `torch.float16` or `torch.bfloat16`).

docs/source/en/perf_infer_gpu_one.md

@@ -233,11 +233,13 @@ For now, Transformers supports SDPA inference and training for the following arc
 * [Jamba](https://huggingface.co/docs/transformers/model_doc/jamba#transformers.JambaModel)
 * [Llama](https://huggingface.co/docs/transformers/model_doc/llama#transformers.LlamaModel)
 * [LLaVA-Onevision](https://huggingface.co/docs/transformers/model_doc/llava_onevision)
+* [M2M100](https://huggingface.co/docs/transformers/model_doc/m2m_100#transformers.M2M100Model)
 * [Mimi](https://huggingface.co/docs/transformers/model_doc/mimi)
 * [Mistral](https://huggingface.co/docs/transformers/model_doc/mistral#transformers.MistralModel)
 * [Mixtral](https://huggingface.co/docs/transformers/model_doc/mixtral#transformers.MixtralModel)
 * [Musicgen](https://huggingface.co/docs/transformers/model_doc/musicgen#transformers.MusicgenModel)
 * [MusicGen Melody](https://huggingface.co/docs/transformers/model_doc/musicgen_melody#transformers.MusicgenMelodyModel)
+* [NLLB](https://huggingface.co/docs/transformers/model_doc/nllb)
 * [OLMo](https://huggingface.co/docs/transformers/model_doc/olmo#transformers.OlmoModel)
 * [OLMoE](https://huggingface.co/docs/transformers/model_doc/olmoe#transformers.OlmoeModel)
 * [PaliGemma](https://huggingface.co/docs/transformers/model_doc/paligemma#transformers.PaliGemmaForConditionalGeneration)

src/transformers/models/m2m_100/modeling_m2m_100.py

@@ -24,7 +24,12 @@ from torch.nn import CrossEntropyLoss
 from ...activations import ACT2FN
 from ...generation import GenerationMixin
 from ...integrations.deepspeed import is_deepspeed_zero3_enabled
-from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
+from ...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 ...modeling_outputs import (
     BaseModelOutput,
     BaseModelOutputWithPastAndCrossAttentions,
@@ -428,6 +433,113 @@ class M2M100FlashAttention2(M2M100Attention):
         return attn_output, None, past_key_value
 
 
+# Copied from transformers.models.bart.modeling_bart.BartSdpaAttention with Bart->M2M100
+class M2M100SdpaAttention(M2M100Attention):
+    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(
+                "M2M100Model is using M2M100SdpaAttention, 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)
+
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+        # 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 = True if self.is_causal and attention_mask is None and tgt_len > 1 else False
+
+        # 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,
+            is_causal=is_causal,
+        )
+
+        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
+
+
 # Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->M2M100, MBART->M2M100
 class M2M100EncoderLayer(nn.Module):
     def __init__(self, config: M2M100Config):
@@ -502,6 +614,7 @@ class M2M100EncoderLayer(nn.Module):
 M2M100_ATTENTION_CLASSES = {
     "eager": M2M100Attention,
     "flash_attention_2": M2M100FlashAttention2,
+    "sdpa": M2M100SdpaAttention,
 }
 
 
@@ -632,6 +745,7 @@ class M2M100PreTrainedModel(PreTrainedModel):
     supports_gradient_checkpointing = True
     _no_split_modules = ["M2M100EncoderLayer", "M2M100DecoderLayer"]
     _supports_flash_attn_2 = True
+    _supports_sdpa = True
 
     def _init_weights(self, module):
         std = self.config.init_std
@@ -805,6 +919,7 @@ class M2M100Encoder(M2M100PreTrainedModel):
         self.layers = nn.ModuleList([M2M100EncoderLayer(config) for _ in range(config.encoder_layers)])
         self.layer_norm = nn.LayerNorm(config.d_model)
         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
@@ -887,6 +1002,11 @@ class M2M100Encoder(M2M100PreTrainedModel):
         if attention_mask is not None:
             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)
@@ -981,6 +1101,7 @@ class M2M100Decoder(M2M100PreTrainedModel):
         )
         self.layers = nn.ModuleList([M2M100DecoderLayer(config) for _ in range(config.decoder_layers)])
         self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+        self._use_sdpa = config._attn_implementation == "sdpa"
         self.layer_norm = nn.LayerNorm(config.d_model)
 
         self.gradient_checkpointing = False
@@ -1094,6 +1215,15 @@ class M2M100Decoder(M2M100PreTrainedModel):
         if self._use_flash_attention_2:
             # 2d mask is passed through the layers
             combined_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.
+            combined_attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
+                attention_mask,
+                input_shape,
+                inputs_embeds,
+                past_key_values_length,
+            )
         else:
             # 4d mask is passed through the layers
             combined_attention_mask = _prepare_4d_causal_attention_mask(
@@ -1104,6 +1234,15 @@ class M2M100Decoder(M2M100PreTrainedModel):
         if encoder_hidden_states is not None and encoder_attention_mask is not None:
             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(