[`FA2`] Add flash attention for `GPT-Neo` (#26486)

* added flash attention for gpt-neo

* small change
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* readme updated

* .

* changes

* removed padding_mask

* Update src/transformers/models/gpt_neo/modeling_gpt_neo.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

---------
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

docs/source/en/model_doc/gpt_neo.md

@@ -54,6 +54,46 @@ The `generate()` method can be used to generate text using GPT Neo model.
 >>> gen_text = tokenizer.batch_decode(gen_tokens)[0]
 ```
 
+## Combining GPT-Neo and Flash Attention 2
+
+First, make sure to install the latest version of Flash Attention 2 to include the sliding window attention feature.
+
+```bash
+pip install -U flash-attn --no-build-isolation
+```
+
+Make also sure that you have a hardware that is compatible with Flash-Attention 2. Read more about it in the official documentation of flash-attn repository. Make also sure to load your model in half-precision (e.g. `torch.float16``)
+
+To load and run a model using Flash Attention 2, refer to the snippet below:
+
+```python
+>>> import torch
+>>> from transformers import AutoModelForCausalLM, AutoTokenizer
+>>> device = "cuda" # the device to load the model onto
+
+>>> model = AutoModelForCausalLM.from_pretrained("EleutherAI/gpt-neo-2.7B", torch_dtype=torch.float16, use_flash_attention_2=True)
+>>> tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neo-2.7B")
+
+>>> prompt = "def hello_world():"
+
+>>> model_inputs = tokenizer([prompt], return_tensors="pt").to(device)
+>>> model.to(device)
+
+>>> generated_ids = model.generate(**model_inputs, max_new_tokens=100, do_sample=True)
+>>> tokenizer.batch_decode(generated_ids)[0]
+"def hello_world():\n    >>> run_script("hello.py")\n    >>> exit(0)\n<|endoftext|>"
+```
+
+### Expected speedups
+
+Below is an expected speedup diagram that compares pure inference time between the native implementation in transformers using `EleutherAI/gpt-neo-2.7B` checkpoint and the Flash Attention 2 version of the model.
+Note that for GPT-Neo it is not possible to train / run on very long context as the max [position embeddings](https://huggingface.co/EleutherAI/gpt-neo-2.7B/blob/main/config.json#L58 ) is limited to 2048 - but this is applicable to all gpt-neo models and not specific to FA-2
+
+<div style="text-align: center">
+<img src="https://user-images.githubusercontent.com/49240599/272241893-b1c66b75-3a48-4265-bc47-688448568b3d.png">
+</div>
+
+
 ## Resources
 
 - [Text classification task guide](../tasks/sequence_classification)

src/transformers/models/gpt_neo/modeling_gpt_neo.py

@@ -19,11 +19,13 @@ import os
 from typing import Optional, Tuple, Union
 
 import torch
+import torch.nn.functional as F
 import torch.utils.checkpoint
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
     BaseModelOutputWithPastAndCrossAttentions,
@@ -34,10 +36,28 @@ from ...modeling_outputs import (
     TokenClassifierOutput,
 )
 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_torch_fx_available,
+    logging,
+)
 from .configuration_gpt_neo import GPTNeoConfig
 
 
+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
+
+
+# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
+# It means that the function will not be traced through and simply appear as a node in the graph.
+if is_torch_fx_available():
+    _prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask)
+
+
 logger = logging.get_logger(__name__)
 
 _CONFIG_FOR_DOC = "GPTNeoConfig"
@@ -50,6 +70,19 @@ GPT_NEO_PRETRAINED_MODEL_ARCHIVE_LIST = [
 _CHECKPOINT_FOR_DOC = "EleutherAI/gpt-neo-1.3B"
 
 
+# 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.torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
 def load_tf_weights_in_gpt_neo(model, config, gpt_neo_checkpoint_path):
     """Load tf checkpoints in a pytorch model"""
     try:
@@ -133,6 +166,7 @@ def load_tf_weights_in_gpt_neo(model, config, gpt_neo_checkpoint_path):
 class GPTNeoSelfAttention(nn.Module):
     def __init__(self, config, attention_type):
         super().__init__()
+        self.config = config
 
         max_positions = config.max_position_embeddings
         bias = torch.tril(torch.ones((max_positions, max_positions), dtype=bool)).view(
@@ -150,6 +184,7 @@ class GPTNeoSelfAttention(nn.Module):
 
         self.attn_dropout = nn.Dropout(float(config.attention_dropout))
         self.resid_dropout = nn.Dropout(float(config.resid_dropout))
+        self.is_causal = True
 
         self.embed_dim = config.hidden_size
         self.num_heads = config.num_heads
@@ -253,6 +288,185 @@ class GPTNeoSelfAttention(nn.Module):
         return outputs  # a, present, (attentions)
 
 
+class GPTNeoFlashAttention2(GPTNeoSelfAttention):
+    """
+    GPTNeo flash attention module. This module inherits from `GPTNeoSelfAttention` 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.
+    """
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        layer_past=None,
+        head_mask=None,
+        use_cache=False,
+        output_attentions=False,
+    ):
+        bsz, _, _ = hidden_states.size()
+
+        query = self.q_proj(hidden_states)
+        key = self.k_proj(hidden_states)
+        value = self.v_proj(hidden_states)
+
+        query = self._split_heads(query, self.num_heads, self.head_dim)
+        key = self._split_heads(key, self.num_heads, self.head_dim)
+        value = self._split_heads(value, self.num_heads, self.head_dim)
+
+        if layer_past is not None:
+            past_key = layer_past[0]
+            past_value = layer_past[1]
+            key = torch.cat((past_key, key), dim=-2)
+            value = torch.cat((past_value, value), dim=-2)
+
+        if use_cache is True:
+            present = (key, value)
+        else:
+            present = None
+
+        query_length = query.shape[2]
+        tgt_len = key.shape[2]
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        query = query.transpose(1, 2).view(bsz, query_length, self.num_heads, self.head_dim)
+        key = key.transpose(1, 2).view(bsz, tgt_len, self.num_heads, self.head_dim)
+        value = value.transpose(1, 2).view(bsz, tgt_len, self.num_heads, self.head_dim)
+
+        attn_dropout = self.config.attention_dropout if self.training else 0.0
+
+        # 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)
+
+        if query.dtype == torch.float32:
+            # Handle the case where the model is quantized
+            if 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 = query.to(target_dtype)
+            key = key.to(target_dtype)
+            value = value.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query, key, value, attention_mask, query_length, dropout=attn_dropout, softmax_scale=1.0
+        )
+
+        attn_weights_reshaped = attn_output.reshape(bsz, query_length, self.num_heads * self.head_dim)
+        attn_output = self.out_proj(attn_weights_reshaped)
+        attn_output = self.resid_dropout(attn_output)
+
+        outputs = (attn_output, present)
+        if output_attentions:
+            outputs += (attn_weights_reshaped,)
+
+        return outputs
+
+    # 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 (`int`, *optional*):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        # 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=self.is_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=self.is_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 GPTNeoAttention(nn.Module):
     def __init__(self, config, layer_id=0):
         super().__init__()
@@ -261,7 +475,11 @@ class GPTNeoAttention(nn.Module):
         self.attention_type = self.attention_layers[layer_id]
 
         if self.attention_type in ["global", "local"]:
-            self.attention = GPTNeoSelfAttention(config, self.attention_type)
+            self.attention = (
+                GPTNeoSelfAttention(config, self.attention_type)
+                if not getattr(config, "_flash_attn_2_enabled", False)
+                else GPTNeoFlashAttention2(config, self.attention_type)
+            )
         else:
             raise NotImplementedError(
                 "Only attn layer types 'global' and 'local' exist, but got `config.attention_layers`: "
@@ -364,6 +582,7 @@ class GPTNeoPreTrainedModel(PreTrainedModel):
     supports_gradient_checkpointing = True
     _no_split_modules = ["GPTNeoBlock"]
     _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
 
     def __init__(self, *inputs, **kwargs):
         super().__init__(*inputs, **kwargs)
@@ -524,10 +743,8 @@ class GPTNeoModel(GPTNeoPreTrainedModel):
             self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
             input_shape = input_ids.size()
             input_ids = input_ids.view(-1, input_shape[-1])
-            batch_size = input_ids.shape[0]
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]
-            batch_size = inputs_embeds.shape[0]
         else:
             raise ValueError("You have to specify either input_ids or inputs_embeds")
 
@@ -546,26 +763,6 @@ class GPTNeoModel(GPTNeoPreTrainedModel):
             position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
             position_ids = position_ids.unsqueeze(0)
 
-        # Attention mask.
-        if attention_mask is not None:
-            if batch_size <= 0:
-                raise ValueError("batch_size has to be defined and > 0")
-            attention_mask = attention_mask.view(batch_size, -1)
-            # We create a 3D attention mask from a 2D tensor mask.
-            # Sizes are [batch_size, 1, 1, to_seq_length]
-            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
-            # this attention mask is more simple than the triangular masking of causal attention
-            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
-            attention_mask = attention_mask[:, None, None, :]
-
-            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
-            # masked positions, this operation will create a tensor which is 0.0 for
-            # positions we want to attend and the dtype's smallest value for masked positions.
-            # Since we are adding it to the raw scores before the softmax, this is
-            # effectively the same as removing these entirely.
-            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
-            attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
-
         # Prepare head mask if needed
         # 1.0 in head_mask indicate we keep the head
         # attention_probs has shape bsz x num_heads x N x N
@@ -577,6 +774,14 @@ class GPTNeoModel(GPTNeoPreTrainedModel):
         position_embeds = self.wpe(position_ids)
         hidden_states = inputs_embeds + position_embeds
 
+        # Attention mask.
+        if getattr(self.config, "_flash_attn_2_enabled", False):
+            # 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:
+            # 4d mask is passed through the layers
+            attention_mask = _prepare_4d_causal_attention_mask(attention_mask, input_shape, inputs_embeds, past_length)
+
         if token_type_ids is not None:
             token_type_embeds = self.wte(token_type_ids)
             hidden_states = hidden_states + token_type_embeds