File size: 4,540 Bytes
1c60c10
 
f2a2029
 
1c60c10
f2a2029
 
 
 
37293dc
f2a2029
37293dc
f2a2029
 
e8717d3
f2a2029
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
37293dc
 
 
 
 
f2a2029
 
 
 
 
 
 
1c60c10
 
f2a2029
 
 
37293dc
 
 
 
f2a2029
 
37293dc
 
 
 
 
 
f2a2029
 
 
37293dc
 
 
 
f2a2029
 
 
 
37293dc
 
 
 
 
f2a2029
 
 
 
 
37293dc
 
 
 
 
1c60c10
f2a2029
 
 
 
 
1c60c10
f2a2029
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
"""Flash attention monkey patch for llama model"""

# copied from https://github.com/lm-sys/FastChat/blob/main/fastchat/train/llama_flash_attn_monkey_patch.py

from typing import Optional, Tuple

import torch
import transformers
from einops import rearrange
from flash_attn.bert_padding import pad_input, unpad_input
from flash_attn.flash_attn_interface import flash_attn_unpadded_qkvpacked_func
from transformers.models.llama.modeling_llama import apply_rotary_pos_emb


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_unpadded_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]

        # pylint: disable=invalid-name
        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_unpadded_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,
):  # pylint: disable=unused-argument
    # [bsz, seq_len]
    return attention_mask


def replace_llama_attn_with_flash_attn():
    transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (  # pylint: disable=protected-access
        _prepare_decoder_attention_mask
    )
    transformers.models.llama.modeling_llama.LlamaAttention.forward = forward