model/llama_flash_attention.py

import math
from typing import List, Optional, Tuple
import logging

import torch
from torch import nn

import transformers
from transformers.models.llama.modeling_llama import apply_rotary_pos_emb

from einops import rearrange
from flash_attn.flash_attn_interface import (  # pip3 install "flash-attn>=2.0"
    flash_attn_varlen_qkvpacked_func,
)
from flash_attn.bert_padding import unpad_input, pad_input
from transformers.models.llama.modeling_llama import _make_causal_mask, _expand_mask

def forward(
    self,
    hidden_states: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = 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]]]:
    """Input shape: Batch x Time x Channel

    attention_mask: [bsz, q_len]
    """
    bsz, q_len, _ = hidden_states.size()

    query_states = (
        self.q_proj(hidden_states)
        .view(bsz, q_len, self.num_heads, self.head_dim)
        .transpose(1, 2)
    )
    key_states = (
        self.k_proj(hidden_states)
        .view(bsz, q_len, self.num_heads, self.head_dim)
        .transpose(1, 2)
    )
    value_states = (
        self.v_proj(hidden_states)
        .view(bsz, q_len, self.num_heads, self.head_dim)
        .transpose(1, 2)
    )
    # [bsz, q_len, nh, hd]
    # [bsz, nh, q_len, hd]

    kv_seq_len = key_states.shape[-2]
    assert past_key_value is None, "past_key_value is not supported"

    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
    query_states, key_states = apply_rotary_pos_emb(
        query_states, key_states, cos, sin, position_ids
    )
    # [bsz, nh, t, hd]
    assert not output_attentions, "output_attentions is not supported"
    assert not use_cache, "use_cache is not supported"

    # Flash attention codes from
    # https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/flash_attention.py

    # transform the data into the format required by flash attention
    qkv = torch.stack(
        [query_states, key_states, value_states], dim=2
    )  # [bsz, nh, 3, q_len, hd]
    qkv = qkv.transpose(1, 3)  # [bsz, q_len, 3, nh, hd]
    # We have disabled _prepare_decoder_attention_mask in LlamaModel
    # the attention_mask should be the same as the key_padding_mask
    key_padding_mask = attention_mask

    if key_padding_mask is None:
        qkv = rearrange(qkv, "b s ... -> (b s) ...")
        max_s = q_len
        cu_q_lens = torch.arange(
            0, (bsz + 1) * q_len, step=q_len, dtype=torch.int32, device=qkv.device
        )
        output = flash_attn_varlen_qkvpacked_func(
            qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
        )
        output = rearrange(output, "(b s) ... -> b s ...", b=bsz)
    else:
        nheads = qkv.shape[-2]
        x = rearrange(qkv, "b s three h d -> b s (three h d)")
        x_unpad, indices, cu_q_lens, max_s = unpad_input(x, key_padding_mask)
        x_unpad = rearrange(
            x_unpad, "nnz (three h d) -> nnz three h d", three=3, h=nheads
        )
        output_unpad = flash_attn_varlen_qkvpacked_func(
            x_unpad, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
        )
        output = rearrange(
            pad_input(
                rearrange(output_unpad, "nnz h d -> nnz (h d)"), indices, bsz, q_len
            ),
            "b s (h d) -> b s h d",
            h=nheads,
        )
    return self.o_proj(rearrange(output, "b s h d -> b s (h d)")), None, None


# Disable the transformation of the attention mask in LlamaModel as the flash attention
# requires the attention mask to be the same as the key_padding_mask
def _prepare_decoder_attention_mask(
    self, attention_mask, input_shape, inputs_embeds, past_key_values_length
):
    # [bsz, seq_len]
    return attention_mask



    
def _original_prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
    # create causal mask
    # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
    combined_attention_mask = None
    if input_shape[-1] > 1:
        combined_attention_mask = _make_causal_mask(
            input_shape,
            inputs_embeds.dtype,
            device=inputs_embeds.device,
            past_key_values_length=past_key_values_length,
        )

    if attention_mask is not None:
        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
        expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
            inputs_embeds.device
        )
        combined_attention_mask = (
            expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
        )

    return combined_attention_mask


def original_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]]]:
    bsz, q_len, _ = hidden_states.size()

    query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
    key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
    value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).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)
    query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
    # [bsz, nh, t, hd]

    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

    attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)

    if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
        raise ValueError(
            f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
            f" {attn_weights.size()}"
        )

    if attention_mask is not None:
        if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
            raise ValueError(
                f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
            )
        attn_weights = attn_weights + attention_mask
        attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))

    # upcast attention to fp32
    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
    attn_output = torch.matmul(attn_weights, value_states)

    if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
        raise ValueError(
            f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
            f" {attn_output.size()}"
        )

    attn_output = attn_output.transpose(1, 2)
    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)

    attn_output = self.o_proj(attn_output)

    if not output_attentions:
        attn_weights = None

    return attn_output, attn_weights, past_key_value


def replace_llama_attn_with_flash_attn():
    cuda_major, cuda_minor = torch.cuda.get_device_capability()
    if cuda_major < 8:
        logging.warning(
            "Flash attention is only supported on A100 or H100 GPU during training due to head dim > 64 backward."
            "ref: https://github.com/HazyResearch/flash-attention/issues/190#issuecomment-1523359593"
        )
    transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (
        _prepare_decoder_attention_mask
    )
    transformers.models.llama.modeling_llama.LlamaAttention.forward = forward


def replace_flash_attn_with_original_attn():
    cuda_major, cuda_minor = torch.cuda.get_device_capability()
    if cuda_major < 8:
        logging.warning(
            "Flash attention is only supported on A100 or H100 GPU during training due to head dim > 64 backward."
            "ref: https://github.com/HazyResearch/flash-attention/issues/190#issuecomment-1523359593"
        )
    transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = _original_prepare_decoder_attention_mask
    transformers.models.llama.modeling_llama.LlamaAttention.forward = original_forward