memory/modeling_flash_attention_utils.py

# coding=utf-8
# Copyright 2024 The Fairseq Authors and the HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import inspect
import os
from typing import Optional, Tuple

import torch
import torch.nn.functional as F

from transformers.utils import is_flash_attn_2_available


if is_flash_attn_2_available():
    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
    from flash_attn import flash_attn_func, flash_attn_varlen_func

    _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)


def _get_unpad_data(attention_mask: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, int]:
    """
    Retrieves indexing data required to repad unpadded (ragged) tensors.

    Arguments:
        attention_mask (`torch.Tensor`):
            Boolean or int tensor of shape (batch_size, sequence_length), 1 means valid and 0 means not valid.

    Return:
        indices (`torch.Tensor):
            The indices of non-masked tokens from the flattened input sequence.
        cu_seqlens (`torch.Tensor`):
            The cumulative sequence lengths, used to index into ragged (unpadded) tensors. `cu_seqlens` shape is (batch_size + 1,).
        max_seqlen_in_batch (`int`):
            Maximum sequence length in batch.
    """
    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 _upad_input(
    query_layer: torch.Tensor,
    key_layer: torch.Tensor,
    value_layer: torch.Tensor,
    attention_mask: torch.Tensor,
    query_length: int,
):
    """
    Unpads query, key, and values tensors, using a single dimension for all tokens even though they belong to different batches.

    This function is used instead of `flash_attn.bert_padding.unpad_input` in order to avoid the recomputation of the same intermediary
    tensors for query, key, value tensors.

    Arguments:
        query_layer (`torch.Tensor`):
            Query state with padding. Shape: (batch_size, query_length, num_heads, head_dim).
        key_layer (`torch.Tensor`):
            Key state with padding. Shape: (batch_size, kv_seq_len, num_key_value_heads, head_dim).
        value_layer (`torch.Tensor`):
            Value state with padding. Shape: (batch_size, kv_seq_len, num_key_value_heads, head_dim).
        attention_mask (`torch.Tensor`):
            Boolean or int tensor of shape (batch_size, sequence_length), 1 means valid and 0 means not valid.
        query_length (`int`):
            Target length.

    Return:
        query_layer (`torch.Tensor):
            Query state without padding. Shape: (total_target_length, num_heads, head_dim).
        key_layer (`torch.Tensor`):
            Key state with padding. Shape: (total_source_length, num_key_value_heads, head_dim).
        value_layer (`torch.Tensor`):
            Value state with padding. Shape: (total_source_length, num_key_value_heads, head_dim).
        indices_q (`torch.Tensor`):
            The indices of non-masked tokens from the flattened input target sequence.
        (cu_seqlens_q, cu_seqlens_k) (`Tuple[int]`):
            The cumulative sequence lengths for the target (query) and source (key, value), used to index into ragged (unpadded) tensors. `cu_seqlens` shape is (batch_size + 1,).
        (max_seqlen_in_batch_q, max_seqlen_in_batch_k) (`Tuple[int]`):
            Maximum sequence length in batch (`max_seqlen_in_batch_q` for the target sequence i.e. query, `max_seqlen_in_batch_k` for the source sequence i.e. key/value).
    """
    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, -1, 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),
    )


def prepare_fa2_from_position_ids(query, key, value, position_ids):
    """
    This function returns necessary arguments to call `flash_attn_varlen_func`.
    All three query, key, value states will be flattened.
    Cummulative lengths of each examples in the batch will be extracted from position_ids.

    NOTE: ideally cummulative lengths should be prepared at the data collator stage

    Arguments:
        query (`torch.Tensor`):
            Query state with padding. Shape: (batch_size, query_length, num_heads, head_dim).
        key (`torch.Tensor`):
            Key state with padding. Shape: (batch_size, kv_seq_len, num_key_value_heads, head_dim).
        value (`torch.Tensor`):
            Value state with padding. Shape: (batch_size, kv_seq_len, num_key_value_heads, head_dim).
        position_ids (`torch.Tensor`):
            Boolean or int tensor of shape (batch_size, sequence_length), 1 means valid and 0 means not valid.

    Return:
        query (`torch.Tensor):
            Query state without padding. Shape: (total_target_length, num_heads, head_dim).
        key (`torch.Tensor`):
            Key state with padding. Shape: (total_source_length, num_key_value_heads, head_dim).
        value (`torch.Tensor`):
            Value state with padding. Shape: (total_source_length, num_key_value_heads, head_dim).
        indices_q (`torch.Tensor`):
            The indices of non-masked tokens from the flattened input target sequence.
        (cu_seqlens_q, cu_seqlens_k) (`Tuple[int]`):
            The cumulative sequence lengths for the target (query) and source (key, value), used to index into ragged (unpadded) tensors. `cu_seqlens` shape is (batch_size + 1,).
        (max_seqlen_in_batch_q, max_seqlen_in_batch_k) (`Tuple[int]`):
            Maximum sequence length in batch (`max_seqlen_in_batch_q` for the target sequence i.e. query, `max_seqlen_in_batch_k` for the source sequence i.e. key/value).
    """
    query = query.view(-1, query.size(-2), query.size(-1))
    key = key.view(-1, key.size(-2), key.size(-1))
    value = value.view(-1, value.size(-2), value.size(-1))
    position_ids = position_ids.flatten()
    indices_q = torch.arange(position_ids.size(0), device=position_ids.device, dtype=torch.int32)

    cu_seq_lens = torch.cat(
        (
            indices_q[position_ids == 0],
            torch.tensor(position_ids.size(), device=position_ids.device, dtype=torch.int32),
        )
    )

    max_length = position_ids.max() + 1

    return (query, key, value, indices_q, (cu_seq_lens, cu_seq_lens), (max_length, max_length))


def _flash_attention_forward(
    query_states: torch.Tensor,
    key_states: torch.Tensor,
    value_states: torch.Tensor,
    attention_mask: torch.Tensor,
    query_length: int,
    is_causal: bool,
    dropout: float = 0.0,
    position_ids: Optional[torch.Tensor] = None,
    softmax_scale: Optional[float] = None,
    sliding_window: Optional[int] = None,
    use_top_left_mask: bool = False,
    softcap: Optional[float] = None,
    deterministic: bool = 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)
        use_top_left_mask (`bool`, defaults to `False`):
            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.
        softcap (`float`, *optional*):
            Softcap for the attention logits, used e.g. in gemma2.
        deterministic (`bool`, *optional*):
            Determines if the deterministic option introduced in flash_attn>=2.4.1 is enabled.
    """
    if not use_top_left_mask:
        causal = 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 transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__.
        causal = is_causal and query_length != 1

    # Assuming 4D tensors, key_states.shape[1] is the key/value sequence length (source length).
    use_sliding_windows = (
        _flash_supports_window_size and sliding_window is not None and key_states.shape[1] > sliding_window
    )
    flash_kwargs = {"window_size": (sliding_window, sliding_window)} if use_sliding_windows else {}

    if deterministic is None:
        deterministic = os.environ.get("FLASH_ATTENTION_DETERMINISTIC", "0") == "1"
    flash_kwargs["deterministic"] = deterministic

    if softcap is not None:
        flash_kwargs["softcap"] = softcap

    # 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 = _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,
            **flash_kwargs,
        )
        attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)

    # if position_ids is provided and check not all examples (row) contain only 1 sequence, and is in pre-fill/training stage
    # then use `flash_attn_varlen_func` to prevent cross-example attention and also allow padding free approach
    elif (
        position_ids is not None and not (position_ids[:, -1] == position_ids.size(1) - 1).all() and query_length != 1
    ):
        batch_size = query_states.size(0)
        query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = prepare_fa2_from_position_ids(
            query_states, key_states, value_states, position_ids
        )

        cu_seqlens_q, cu_seqlens_k = cu_seq_lens
        max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens

        attn_output = 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,
            **flash_kwargs,
        )

        attn_output = attn_output.view(batch_size, -1, attn_output.size(-2), attn_output.size(-1))

    else:
        attn_output = flash_attn_func(
            query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal, **flash_kwargs
        )

    return attn_output