Adding Flash Attention 2 Support for GPT2 (#29226)

* First commit to add flash attention 2 for GPT-2

* more improvements

* Make GPT2 pass tests and fixed Decison Transformers copies

* Fixed missing arg

* fix copies

* Added expected speedup

* Update src/transformers/models/gpt2/modeling_gpt2.py

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

* Update src/transformers/models/gpt2/modeling_gpt2.py

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

* Update src/transformers/models/gpt2/modeling_gpt2.py

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

* Added test

* Fixed attn attribute

* Update docs/source/en/model_doc/gpt2.md

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

* Update docs/source/en/model_doc/gpt2.md

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

* Update Decision transformer attentions

* More updates

* Passing tests

* Fix copies

* Fix copies part 2

* Decision transformer updates

* Update src/transformers/models/gpt2/modeling_gpt2.py

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

* Fix copies

* Decision transformer not supporting flash attn

* Addressed comments

* Addressed comments

* Addressed comments

---------

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/gpt2.md

@@ -60,6 +60,73 @@ This model was contributed by [thomwolf](https://huggingface.co/thomwolf). The o
 - Enabling the *scale_attn_by_inverse_layer_idx* and *reorder_and_upcast_attn* flags will apply the training stability
   improvements from [Mistral](https://github.com/stanford-crfm/mistral/) (for PyTorch only).
 
+## Usage example
+
+The `generate()` method can be used to generate text using GPT2 model.
+
+```python
+>>> from transformers import AutoModelForCausalLM, AutoTokenizer
+
+>>> model = AutoModelForCausalLM.from_pretrained("gpt2")
+>>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
+
+>>> prompt = "GPT2 is a model developed by OpenAI."
+
+>>> input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+
+>>> gen_tokens = model.generate(
+...     input_ids,
+...     do_sample=True,
+...     temperature=0.9,
+...     max_length=100,
+... )
+>>> gen_text = tokenizer.batch_decode(gen_tokens)[0]
+```
+
+## Using Flash Attention 2
+
+Flash Attention 2 is a faster, optimized version of the attention scores computation which relies on `cuda` kernels.
+
+### 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
+>>> import torch
+>>> from transformers import AutoModelForCausalLM, AutoTokenizer
+>>> device = "cuda" # the device to load the model onto
+
+>>> model = AutoModelForCausalLM.from_pretrained("gpt2", torch_dtype=torch.float16, attn_implementation="flash_attention_2")
+>>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
+
+>>> 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]
+```
+
+
+### Expected speedups
+
+Below is an expected speedup diagram that compares pure inference time between the native implementation in transformers using `gpt2` checkpoint and the Flash Attention 2 version of the model using a sequence length of 512.
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/EduardoPacheco/documentation-images/resolve/main/gpt2_flash_attention_2_speedup.jpg">
+</div>
+
 ## Resources
 
 A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with GPT2. 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

@@ -42,6 +42,7 @@ FlashAttention-2 is currently supported for the following architectures:
 * [Cohere](https://huggingface.co/docs/transformers/model_doc/cohere#transformers.CohereModel)
 * [DistilBert](https://huggingface.co/docs/transformers/model_doc/distilbert#transformers.DistilBertModel)
 * [Gemma](https://huggingface.co/docs/transformers/model_doc/gemma#transformers.GemmaModel)
+* [GPT2](https://huggingface.co/docs/transformers/model_doc/gpt2)
 * [GPTBigCode](https://huggingface.co/docs/transformers/model_doc/gpt_bigcode#transformers.GPTBigCodeModel)
 * [GPTNeo](https://huggingface.co/docs/transformers/model_doc/gpt_neo#transformers.GPTNeoModel)
 * [GPTNeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox#transformers.GPTNeoXModel)

src/transformers/models/decision_transformer/modeling_decision_transformer.py

@@ -108,7 +108,7 @@ def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
 class DecisionTransformerGPT2Attention(nn.Module):
     def __init__(self, config, is_cross_attention=False, layer_idx=None):
         super().__init__()
-
+        self.config = config
         max_positions = config.max_position_embeddings
         self.register_buffer(
             "bias",
@@ -146,6 +146,7 @@ class DecisionTransformerGPT2Attention(nn.Module):
 
         self.attn_dropout = nn.Dropout(config.attn_pdrop)
         self.resid_dropout = nn.Dropout(config.resid_pdrop)
+        self.is_causal = True
 
         self.pruned_heads = set()
 
@@ -346,6 +347,7 @@ class DecisionTransformerGPT2MLP(nn.Module):
 
 # Copied from transformers.models.gpt2.modeling_gpt2.GPT2Block with GPT2->DecisionTransformerGPT2
 class DecisionTransformerGPT2Block(nn.Module):
+    # Ignore copy
     def __init__(self, config, layer_idx=None):
         super().__init__()
         hidden_size = config.hidden_size
@@ -497,7 +499,6 @@ class DecisionTransformerGPT2Model(DecisionTransformerGPT2PreTrainedModel):
     def set_input_embeddings(self, new_embeddings):
         self.wte = new_embeddings
 
-    # Copied from transformers.models.gpt2.modeling_gpt2.GPT2Model.forward
     def forward(
         self,
         input_ids: Optional[torch.LongTensor] = None,
@@ -548,7 +549,7 @@ class DecisionTransformerGPT2Model(DecisionTransformerGPT2PreTrainedModel):
             position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
             position_ids = position_ids.unsqueeze(0)
 
-        # GPT2Attention mask.
+        # Attention mask.
         if attention_mask is not None:
             if batch_size <= 0:
                 raise ValueError("batch_size has to be defined and > 0")

src/transformers/models/gpt2/modeling_gpt2.py

@@ -22,6 +22,7 @@ from dataclasses import dataclass
 from typing import Optional, Tuple, Union
 
 import torch
+import torch.nn.functional as F
 import torch.utils.checkpoint
 from torch import nn
 from torch.cuda.amp import autocast
@@ -42,6 +43,8 @@ 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,
     logging,
     replace_return_docstrings,
 )
@@ -49,6 +52,11 @@ from ...utils.model_parallel_utils import assert_device_map, get_device_map
 from .configuration_gpt2 import GPT2Config
 
 
+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
+
+
 logger = logging.get_logger(__name__)
 
 _CHECKPOINT_FOR_DOC = "openai-community/gpt2"
@@ -58,6 +66,19 @@ _CONFIG_FOR_DOC = "GPT2Config"
 from ..deprecated._archive_maps import GPT2_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,
+    )
+
+
 def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
     """Load tf checkpoints in a pytorch model"""
     try:
@@ -117,7 +138,7 @@ def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
 class GPT2Attention(nn.Module):
     def __init__(self, config, is_cross_attention=False, layer_idx=None):
         super().__init__()
-
+        self.config = config
         max_positions = config.max_position_embeddings
         self.register_buffer(
             "bias",
@@ -155,6 +176,7 @@ class GPT2Attention(nn.Module):
 
         self.attn_dropout = nn.Dropout(config.attn_pdrop)
         self.resid_dropout = nn.Dropout(config.resid_pdrop)
+        self.is_causal = True
 
         self.pruned_heads = set()
 
@@ -335,6 +357,210 @@ class GPT2Attention(nn.Module):
         return outputs  # a, present, (attentions)
 
 
+class GPT2FlashAttention2(GPT2Attention):
+    """
+    GPT2 flash attention module. This module inherits from `GPT2Attention` 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 __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 forward(
+        self,
+        hidden_states: Optional[Tuple[torch.FloatTensor]],
+        layer_past: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
+        bsz, _, _ = hidden_states.size()
+        if encoder_hidden_states is not None:
+            if not hasattr(self, "q_attn"):
+                raise ValueError(
+                    "If class is used as cross attention, the weights `q_attn` have to be defined. "
+                    "Please make sure to instantiate class with `GPT2Attention(..., is_cross_attention=True)`."
+                )
+
+            query = self.q_attn(hidden_states)
+            key, value = self.c_attn(encoder_hidden_states).split(self.split_size, dim=2)
+            attention_mask = encoder_attention_mask
+        else:
+            query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
+
+        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)
+
+        present = None
+        if use_cache is True:
+            present = (key, value)
+
+        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.attn_dropout.p 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:
+            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.c_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
+        )
+
+        attn_weights_reshaped = attn_output.reshape(bsz, query_length, self.num_heads * self.head_dim)
+        attn_output = self.c_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 (`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 GPT2MLP(nn.Module):
     def __init__(self, intermediate_size, config):
         super().__init__()
@@ -352,18 +578,25 @@ class GPT2MLP(nn.Module):
         return hidden_states
 
 
+GPT2_ATTENTION_CLASSES = {
+    "eager": GPT2Attention,
+    "flash_attention_2": GPT2FlashAttention2,
+}
+
+
 class GPT2Block(nn.Module):
     def __init__(self, config, layer_idx=None):
         super().__init__()
         hidden_size = config.hidden_size
         inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
+        attention_class = GPT2_ATTENTION_CLASSES[config._attn_implementation]
 
         self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
-        self.attn = GPT2Attention(config, layer_idx=layer_idx)
+        self.attn = attention_class(config=config, layer_idx=layer_idx)
         self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
 
         if config.add_cross_attention:
-            self.crossattention = GPT2Attention(config, is_cross_attention=True, layer_idx=layer_idx)
+            self.crossattention = attention_class(config=config, is_cross_attention=True, layer_idx=layer_idx)
             self.ln_cross_attn = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
 
         self.mlp = GPT2MLP(inner_dim, config)
@@ -443,6 +676,7 @@ class GPT2PreTrainedModel(PreTrainedModel):
     supports_gradient_checkpointing = True
     _no_split_modules = ["GPT2Block"]
     _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
 
     def __init__(self, *inputs, **kwargs):
         super().__init__(*inputs, **kwargs)
@@ -673,6 +907,7 @@ class GPT2Model(GPT2PreTrainedModel):
         self.model_parallel = False
         self.device_map = None
         self.gradient_checkpointing = False
+        self._attn_implementation = config._attn_implementation
 
         # Initialize weights and apply final processing
         self.post_init()
@@ -790,25 +1025,26 @@ class GPT2Model(GPT2PreTrainedModel):
             position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
             position_ids = position_ids.unsqueeze(0)
 
-        # GPT2Attention mask.
+        # 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.
+            if self._attn_implementation == "flash_attention_2":
-            # Sizes are [batch_size, 1, 1, to_seq_length]
+                attention_mask = attention_mask if 0 in attention_mask else None
-            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+            else:
-            # this attention mask is more simple than the triangular masking of causal attention
+                # We create a 3D attention mask from a 2D tensor mask.
-            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+                # Sizes are [batch_size, 1, 1, to_seq_length]
-            attention_mask = attention_mask[:, None, None, :]
+                # 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
+                # 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
+                # 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.
+                # 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
+                # Since we are adding it to the raw scores before the softmax, this is
-            # effectively the same as removing these entirely.
+                # effectively the same as removing these entirely.
-            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+                attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
-            attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
+                attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
 
         # If a 2D or 3D attention mask is provided for the cross-attention
         # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
@@ -817,7 +1053,8 @@ class GPT2Model(GPT2PreTrainedModel):
             encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
             if encoder_attention_mask is None:
                 encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
-            encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+            if self._attn_implementation != "flash_attention_2":
+                encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask)
         else:
             encoder_attention_mask = None
 

tests/models/gpt2/test_modeling_gpt2.py

@@ -19,8 +19,17 @@ import gc
 import math
 import unittest
 
+import pytest
+
 from transformers import GPT2Config, is_torch_available
-from transformers.testing_utils import backend_empty_cache, require_torch, slow, torch_device
+from transformers.testing_utils import (
+    backend_empty_cache,
+    require_flash_attn,
+    require_torch,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
 
 from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
@@ -858,3 +867,40 @@ class GPT2ModelLanguageGenerationTest(unittest.TestCase):
                 "but said in a statement to The Associated Press that"
             ],
         )
+
+    @require_flash_attn
+    @require_torch_gpu
+    @pytest.mark.flash_attn_test
+    @slow
+    def test_flash_attn_2_generate_padding_left(self):
+        """
+        Overwritting the common test as the test is flaky on tiny models
+        """
+        model = GPT2LMHeadModel.from_pretrained("gpt2", torch_dtype=torch.float16).to(0)
+
+        tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+
+        texts = ["hi", "Hello this is a very long sentence"]
+
+        tokenizer.padding_side = "left"
+        tokenizer.pad_token = tokenizer.eos_token
+
+        inputs = tokenizer(texts, return_tensors="pt", padding=True).to(0)
+
+        output_native = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        output_native = tokenizer.batch_decode(output_native)
+
+        model = GPT2LMHeadModel.from_pretrained(
+            "gpt2", device_map={"": 0}, attn_implementation="flash_attention_2", torch_dtype=torch.float16
+        )
+
+        output_fa_2 = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        output_fa_2 = tokenizer.batch_decode(output_fa_2)
+
+        expected_output = [
+            "<|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|>hi, who was born in the city of Kolkata, was a member of the Kolkata",
+            "Hello this is a very long sentence. I'm sorry. I'm sorry. I'm sorry. I'm sorry. I'm sorry",
+        ]
+
+        self.assertListEqual(output_native, output_fa_2)
+        self.assertListEqual(output_native, expected_output)