feat: add flash_attn and dropout support

configuration_stablelm_epoch.py

@@ -64,6 +64,8 @@ class StableLMEpochConfig(PretrainedConfig):
             (not used by all models). Only relevant if `config.is_decoder=True`.
         tie_word_embeddings(`bool`, *optional*, defaults to `False`):
             Whether to tie weight embeddings
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
     """
     model_type = "stablelm_epoch"
     keys_to_ignore_at_inference = ["past_key_values"]
@@ -86,6 +88,7 @@ class StableLMEpochConfig(PretrainedConfig):
         bos_token_id=0,
         eos_token_id=2,
         tie_word_embeddings=False,
+        attention_dropout: float = 0.0,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -102,6 +105,7 @@ class StableLMEpochConfig(PretrainedConfig):
         self.norm_eps = norm_eps
         self.use_cache = use_cache
         self.tie_word_embeddings = tie_word_embeddings
+        self.attention_dropout = attention_dropout
         super().__init__(
             bos_token_id=bos_token_id,
             eos_token_id=eos_token_id,

modeling_stablelm_epoch.py

@@ -19,23 +19,48 @@
 """ PyTorch StableLM Epoch model. """
 from typing import Optional, Tuple, Union
 import math
+import warnings
 
 import torch
+import torch.nn.functional as F
 import torch.utils.checkpoint
 from torch import nn
 from torch.nn import CrossEntropyLoss
+
+from transformers.cache_utils import Cache
 from transformers.modeling_outputs import (
     BaseModelOutputWithPast,
     CausalLMOutputWithPast,
 )
 from transformers.modeling_utils import PreTrainedModel
-from transformers.utils import logging
+from transformers.utils import logging, is_flash_attn_greater_or_equal_2_10
+
 from .configuration_stablelm_epoch import StableLMEpochConfig
 
+try:
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input
+except:
+    flash_attn_func, flash_attn_varlen_func = None, None
+    index_first_axis, pad_input, unpad_input = None, None, None
+
 
 logger = logging.get_logger(__name__)
 
 
+# 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.bart.modeling_bart._make_causal_mask
 def _make_causal_mask(
     input_ids_shape: torch.Size,
@@ -165,6 +190,8 @@ class Attention(nn.Module):
         self.num_key_value_heads = config.num_key_value_heads
         self.num_key_value_groups = self.num_heads // self.num_key_value_heads
         self.max_position_embeddings = config.max_position_embeddings
+        self.is_causal = True
+        self.attention_dropout = config.attention_dropout
 
         if (self.head_dim * self.num_heads) != self.hidden_size:
             raise ValueError(
@@ -248,6 +275,7 @@ class Attention(nn.Module):
 
         # Upcast attention to fp32
         attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
         attn_output = torch.matmul(attn_weights, value_states)
 
         if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
@@ -269,10 +297,202 @@ class Attention(nn.Module):
         return attn_output, attn_weights, past_key_value
 
 
+class FlashAttention2(Attention):
+    """
+    Reference: https://github.com/huggingface/transformers/blob/5d36025ca13d05151b7a0c761e90d429c4644a30/src/transformers/models/llama/modeling_llama.py#L456
+    """
+
+    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.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # FlashAttention2 attention does not support output_attentions
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+
+            # overwrite attention_mask with padding_mask
+            attention_mask = kwargs.pop("padding_mask")
+
+        output_attentions = False
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        query_rot = query_states[..., : self.rotary_ndims]
+        query_pass = query_states[..., self.rotary_ndims :]
+        key_rot = key_states[..., : self.rotary_ndims]
+        key_pass = key_states[..., self.rotary_ndims :]
+
+        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)
+        query_states, key_states = apply_rotary_pos_emb(query_rot, key_rot, cos, sin, position_ids)
+
+        # [batch_size, num_heads, seq_len, head_dim]
+        query_states = torch.cat((query_states, query_pass), dim=-1)
+        key_states = torch.cat((key_states, 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
+
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        dropout_rate = self.attention_dropout if self.training else 0.0
+
+        attn_output = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate
+        )
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    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 FlashAttention2 __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
+
+    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),
+        )
+
+
+ATTENTION_CLASSES = {
+    "eager": Attention,
+    "flash_attention_2": FlashAttention2,
+}
+
+
 class DecoderLayer(nn.Module):
     def __init__(self, config: StableLMEpochConfig):
         super().__init__()
-        self.self_attn = Attention(config)
+        self.self_attn = ATTENTION_CLASSES[config._attn_implementation](config=config)
         self.mlp = MLP(config)
         self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.norm_eps)
         self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.norm_eps)
@@ -328,6 +548,7 @@ class StableLMEpochPreTrainedModel(PreTrainedModel):
     supports_gradient_checkpointing = True
     _no_split_modules = ["DecoderLayer"]
     _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
 
     def _init_weights(self, module: nn.Module):
         """Initialize the weights"""
@@ -355,6 +576,7 @@ class StableLMEpochModel(StableLMEpochPreTrainedModel):
         self.layers = nn.ModuleList([DecoderLayer(config) for _ in range(config.num_hidden_layers)])
         self.norm = nn.LayerNorm(config.hidden_size, eps=config.norm_eps)
 
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
         self.gradient_checkpointing = False
         # Initialize weights and apply final processing
         self.post_init()
@@ -428,10 +650,6 @@ class StableLMEpochModel(StableLMEpochPreTrainedModel):
         seq_length_with_past = seq_length
         past_key_values_length = 0
 
-        if past_key_values is not None:
-            past_key_values_length = past_key_values[0][0].shape[2]
-            seq_length_with_past = seq_length_with_past + past_key_values_length
-
         if position_ids is None:
             device = input_ids.device if input_ids is not None else inputs_embeds.device
             position_ids = torch.arange(
@@ -447,18 +665,22 @@ class StableLMEpochModel(StableLMEpochPreTrainedModel):
         if inputs_embeds is None:
             inputs_embeds = self.embed_tokens(input_ids)
         # Embed positions
-        if attention_mask is None:
-            attention_mask = torch.ones(
-                (batch_size, seq_length_with_past),
-                dtype=torch.bool,
-                device=inputs_embeds.device,
+        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
+        else:
+            if attention_mask is None:
+                attention_mask = torch.ones(
+                    (batch_size, seq_length_with_past),
+                    dtype=torch.bool,
+                    device=inputs_embeds.device,
+                )
+            attention_mask = self._prepare_decoder_attention_mask(
+                attention_mask,
+                (batch_size, seq_length),
+                inputs_embeds,
+                past_key_values_length,
             )
-        attention_mask = self._prepare_decoder_attention_mask(
-            attention_mask,
-            (batch_size, seq_length),
-            inputs_embeds,
-            past_key_values_length,
-        )
 
         hidden_states = inputs_embeds
 
@@ -643,8 +865,17 @@ class StableLMEpochForCausalLM(StableLMEpochPreTrainedModel):
         **kwargs,
     ):
         # Trim decoder_input_ids if past is used
-        if past_key_values and past_key_values[0] is not None:
-            input_ids = input_ids[:, -1:]
+        if past_key_values is not None:
+            past_length = past_key_values[0][0].shape[2]
+
+            # Some generation methods already pass only the last input ID
+            if input_ids.shape[1] > past_length:
+                remove_prefix_length = past_length
+            else:
+                # Default to old behavior: keep only final ID
+                remove_prefix_length = input_ids.shape[1] - 1
+
+            input_ids = input_ids[:, remove_prefix_length:]
 
         position_ids = kwargs.get("position_ids", None)
         if attention_mask is not None and position_ids is None: