[`FA-2`] Add Flash Attention to `Phi` (#27661)

* add FA and modify doc file

* test_flash_attn_2_generate_padding_right test overwritten

* comment

* modify persimmon modeling file

* added speedup graph

* more changes

docs/source/en/model_doc/phi.md

@@ -76,7 +76,7 @@ The original code for Phi-1 and Phi-1.5 can be found [here](https://huggingface.
 ```python
 >>> from transformers import PhiForCausalLM, AutoTokenizer
 
->>> # define the model and tokenzier.
+>>> # define the model and tokenizer.
 >>> model = PhiForCausalLM.from_pretrained("susnato/phi-1_5_dev")
 >>> tokenizer = AutoTokenizer.from_pretrained("susnato/phi-1_5_dev")
 
@@ -94,6 +94,46 @@ The original code for Phi-1 and Phi-1.5 can be found [here](https://huggingface.
 ```
 
 
+## Combining Phi 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 PhiForCausalLM, AutoTokenizer
+
+>>> # define the model and tokenizer and push the model and tokens to the GPU.
+>>> model = PhiForCausalLM.from_pretrained("susnato/phi-1_5_dev", torch_dtype=torch.float16, use_flash_attention_2=True).to("cuda")
+>>> tokenizer = AutoTokenizer.from_pretrained("susnato/phi-1_5_dev")
+
+>>> # feel free to change the prompt to your liking.
+>>> prompt = "If I were an AI that had just achieved"
+
+>>> # apply the tokenizer.
+>>> tokens = tokenizer(prompt, return_tensors="pt").to("cuda")
+
+>>> # use the model to generate new tokens.
+>>> generated_output = model.generate(**tokens, use_cache=True, max_new_tokens=10)
+
+>>> tokenizer.batch_decode(generated_output)[0]
+'If I were an AI that had just achieved a breakthrough in machine learning, I would be thrilled'
+```
+
+### Expected speedups
+Below is an expected speedup diagram that compares pure inference time between the native implementation in transformers using `susnato/phi-1_dev` checkpoint and the Flash Attention 2 version of the model using a sequence length of 2048.
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/ybelkada/documentation-images/resolve/main/phi_1_speedup_plot.jpg">
+</div>
+
+
 ## PhiConfig
 
 [[autodoc]] PhiConfig

src/transformers/models/persimmon/modeling_persimmon.py

@@ -187,6 +187,7 @@ class PersimmonAttention(nn.Module):
         self.max_position_embeddings = config.max_position_embeddings
         self.rope_theta = config.rope_theta
         self.partial_rotary_factor = config.partial_rotary_factor
+        self.is_causal = True
 
         if (self.head_dim * self.num_heads) != self.hidden_size:
             raise ValueError(

src/transformers/models/phi/modeling_phi.py

@@ -20,6 +20,7 @@ import math
 from typing import List, 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
@@ -37,12 +38,19 @@ from ...utils import (
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
     logging,
     replace_return_docstrings,
 )
 from .configuration_phi import PhiConfig
 
 
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+
 logger = logging.get_logger(__name__)
 
 _CHECKPOINT_FOR_DOC = "susnato/phi-1_dev"
@@ -55,6 +63,19 @@ PHI_PRETRAINED_MODEL_ARCHIVE_LIST = [
 ]
 
 
+# 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,
+    )
+
+
 # Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Phi
 class PhiRotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
@@ -205,6 +226,7 @@ class PhiAttention(nn.Module):
         self.max_position_embeddings = config.max_position_embeddings
         self.rope_theta = config.rope_theta
         self.partial_rotary_factor = config.partial_rotary_factor
+        self.is_causal = True
 
         if (self.head_dim * self.num_heads) != self.hidden_size:
             raise ValueError(
@@ -361,10 +383,233 @@ class PhiAttention(nn.Module):
         return attn_output, attn_weights, past_key_value
 
 
+class PhiFlashAttention2(PhiAttention):
+    """
+    Phi flash attention module. This module inherits from `PhiAttention` 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 forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # PhiFlashAttention2 attention does not support output_attentions
+
+        output_attentions = False
+
+        bsz, q_len, _ = hidden_states.size()
+
+        # [batch_size, seq_length, 3 x hidden_size]
+        fused_qkv = self.query_key_value(hidden_states)
+
+        # 3 x [batch_size, seq_length, num_heads, head_dim]
+        (query_states, key_states, value_states) = self._split_heads(fused_qkv)
+
+        if self.qk_layernorm:
+            query_states = self.q_layernorm(query_states)
+            key_states = self.k_layernorm(key_states)
+
+        # [batch_size, num_heads, seq_length, head_dim] -> [batch_size, seq_length, num_heads, head_dim]
+        query_states = query_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+        key_states = key_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]
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+
+        # Partial rotary embedding
+        query_rot, query_pass = (
+            query_states[..., : self.rotary_emb.dim],
+            query_states[..., self.rotary_emb.dim :],
+        )
+        key_rot, key_pass = (
+            key_states[..., : self.rotary_emb.dim],
+            key_states[..., self.rotary_emb.dim :],
+        )
+        # [batch_size, seq_length, num_heads, head_dim // config.partial_rotary_factor]
+        query_rot, key_rot = apply_rotary_pos_emb(query_rot, key_rot, cos, sin, position_ids)
+
+        # [batch_size, seq_length, num_heads, head_dim]
+        query_states = torch.cat((query_rot, query_pass), dim=-1)
+        key_states = torch.cat((key_rot, key_pass), dim=-1)
+
+        if past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        tgt_len = key_states.shape[2]
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        query_states = query_states.transpose(1, 2).view(bsz, q_len, self.num_heads, self.head_dim)
+        key_states = key_states.transpose(1, 2).view(bsz, tgt_len, self.num_heads, self.head_dim)
+        value_states = value_states.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.
+
+        if query_states.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_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=attn_dropout, softmax_scale=1.0
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim)
+        attn_output = self.dense(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 (`int`, *optional*):
+                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 PhiDecoderLayer(nn.Module):
     def __init__(self, config: PhiConfig):
         super().__init__()
-        self.self_attn = PhiAttention(config=config)
+        self.self_attn = (
+            PhiAttention(config=config)
+            if not getattr(config, "_flash_attn_2_enabled", False)
+            else PhiFlashAttention2(config=config)
+        )
         self.mlp = PhiMLP(config)
         self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.resid_dropout = nn.Dropout(config.resid_pdrop)
@@ -450,6 +695,7 @@ class PhiPreTrainedModel(PreTrainedModel):
     base_model_prefix = "model"
     supports_gradient_checkpointing = True
     _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
 
     def _init_weights(self, module):
         std = self.config.initializer_range
@@ -609,11 +855,12 @@ class PhiModel(PhiPreTrainedModel):
 
         inputs_embeds = self.embed_dropout(inputs_embeds)
 
-        # embed positions
-        if attention_mask is None:
-            attention_mask = torch.ones(
-                (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
-            )
+        # 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, (batch_size, seq_length), inputs_embeds, past_key_values_length
             )

tests/models/phi/test_modeling_phi.py

@@ -18,8 +18,17 @@
 
 import unittest
 
+import pytest
+
 from transformers import PhiConfig, is_torch_available
-from transformers.testing_utils import require_torch, slow, torch_device
+from transformers.testing_utils import (
+    require_bitsandbytes,
+    require_flash_attn,
+    require_torch,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
 
 from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
@@ -31,6 +40,7 @@ if is_torch_available():
     import torch
 
     from transformers import (
+        AutoTokenizer,
         PhiForCausalLM,
         PhiForSequenceClassification,
         PhiForTokenClassification,
@@ -350,6 +360,43 @@ class PhiModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin,
         result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
         self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
 
+    @require_flash_attn
+    @require_torch_gpu
+    @require_bitsandbytes
+    @pytest.mark.flash_attn_test
+    @slow
+    # Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_flash_attn_2_generate_padding_right with LlamaForCausalLM->PhiForCausalLM,LlamaTokenizer->AutoTokenizer,meta-llama/Llama-2-7b-hf->susnato/phi-1_5_dev
+    def test_flash_attn_2_generate_padding_right(self):
+        """
+        Overwritting the common test as the test is flaky on tiny models
+        """
+        model = PhiForCausalLM.from_pretrained(
+            "susnato/phi-1_5_dev",
+            load_in_4bit=True,
+            device_map={"": 0},
+        )
+
+        tokenizer = AutoTokenizer.from_pretrained("susnato/phi-1_5_dev")
+
+        texts = ["hi", "Hello this is a very long sentence"]
+
+        tokenizer.padding_side = "right"
+        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 = PhiForCausalLM.from_pretrained(
+            "susnato/phi-1_5_dev", load_in_4bit=True, device_map={"": 0}, use_flash_attention_2=True
+        )
+
+        output_fa_2 = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        output_fa_2 = tokenizer.batch_decode(output_fa_2)
+
+        self.assertListEqual(output_native, output_fa_2)
+
 
 @slow
 @require_torch