-
Notifications
You must be signed in to change notification settings - Fork 23
/
flash_attention.py
178 lines (117 loc) · 6.56 KB
/
flash_attention.py
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
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
import math
import jax
from functools import partial
from jax import nn
from jax import custom_vjp
from jax import numpy as jnp, lax, jit
from jax.numpy import einsum
from einops import rearrange
# constants
EPSILON = 1e-10
MASK_VALUE = -1e10
Q_CHUNK_SIZE = 1024
K_CHUNK_SIZE = 1024
# flash attention
def _query_chunk_flash_attention(chunk_idx, q, k, v, key_mask):
q_len, batch, heads, dim, k_len, v_dim = *q.shape, k.shape[0], v.shape[-1]
scale = 1 / jnp.sqrt(dim)
q_scaled = q * scale
def chunk_scanner(carries, _):
chunk_idx, out, row_sum, row_max = carries
k_chunk_sizes = min(K_CHUNK_SIZE, k_len)
k_chunk = lax.dynamic_slice(k, (chunk_idx, 0, 0, 0), slice_sizes=(k_chunk_sizes, batch, heads, dim))
v_chunk = lax.dynamic_slice(v, (chunk_idx, 0, 0, 0), slice_sizes=(k_chunk_sizes, batch, heads, v_dim))
key_mask_chunk = lax.dynamic_slice(key_mask, (chunk_idx, 0), slice_sizes=(k_chunk_sizes, batch))
attn_weights = einsum('i ... d, j ... d -> i ... j', q_scaled, k_chunk)
key_mask_chunk = rearrange(key_mask_chunk, 'j b -> 1 b 1 j')
attn_weights = jnp.where(key_mask_chunk, attn_weights, MASK_VALUE)
block_row_max = jnp.max(attn_weights, axis = -1, keepdims = True)
exp_weights = jnp.exp(attn_weights - block_row_max)
exp_weights = jnp.where(key_mask_chunk, exp_weights, 0.)
block_row_sum = jnp.sum(exp_weights, axis = -1, keepdims = True) + EPSILON
exp_values = einsum('i ... j, j ... d -> i ... d', exp_weights, v_chunk)
new_row_max = jnp.maximum(block_row_max, row_max)
exp_row_max_diff = jnp.exp(row_max - new_row_max)
exp_block_row_max_diff = jnp.exp(block_row_max - new_row_max)
new_row_sum = exp_row_max_diff * row_sum + exp_block_row_max_diff * block_row_sum
out = (row_sum / new_row_sum) * exp_row_max_diff * out + \
(exp_block_row_max_diff / new_row_sum) * exp_values
return (chunk_idx + k_chunk_sizes, out, new_row_sum, new_row_max), None
out = jnp.zeros((q_len, batch, heads, dim))
row_sum = jnp.zeros((q_len, batch, heads, 1))
row_max = jnp.ones((q_len, batch, heads, 1)) * -1e6
(_, out, row_sum, row_max), _ = lax.scan(chunk_scanner, init = (0, out, row_sum, row_max), xs = None, length = math.ceil(k_len / K_CHUNK_SIZE))
row_sum = rearrange(row_sum, 'n ... 1 -> n ...')
row_max = rearrange(row_max, 'n ... 1 -> n ...')
lse = jnp.log(row_sum) + row_max
return out, lse
def _flash_attention(q, k, v, key_mask):
batch, heads, q_len, dim, v_dim = *q.shape, v.shape[-1]
def chunk_scanner(chunk_idx, _):
chunk_sizes = min(Q_CHUNK_SIZE, q_len)
q_chunk = lax.dynamic_slice(q, (chunk_idx, 0, 0, 0), slice_sizes = (chunk_sizes, batch, heads, dim))
return (chunk_idx + chunk_sizes, _query_chunk_flash_attention(chunk_idx, q_chunk, k, v, key_mask))
q, k, v = map(lambda t: rearrange(t, 'b h n d -> n b h d'), (q, k, v))
key_mask = rearrange(key_mask, 'b j -> j b')
_, (out, lse) = lax.scan(chunk_scanner, init = 0, xs = None, length = math.ceil(q_len / Q_CHUNK_SIZE))
out = rearrange(out, 'c n b h d -> b h (c n) d')
lse = rearrange(lse, 'c n b h -> b h (c n)')
return out, lse
@custom_vjp
@jit
def flash_attention(q, k, v, key_mask):
out, _ = _flash_attention(q, k, v, key_mask)
return out
@jit
def flash_attention_forward(q, k, v, key_mask):
out, lse = _flash_attention(q, k, v, key_mask)
return out, (q, k, v, key_mask, out, lse)
def _query_chunk_flash_attention_backward(q, k, v, key_mask, o, do, lse):
q_len, batch, heads, dim, k_len, v_dim = *q.shape, v.shape[0], v.shape[-1]
scale = 1 / jnp.sqrt(dim)
q_scaled = q * scale
def chunk_scanner(carries, _):
chunk_idx, dq = carries
k_chunk_sizes = min(K_CHUNK_SIZE, k_len)
k_chunk = lax.dynamic_slice(k, (chunk_idx, batch, heads, 0), slice_sizes=(k_chunk_sizes, batch, heads, dim))
v_chunk = lax.dynamic_slice(v, (chunk_idx, batch, heads, 0), slice_sizes=(k_chunk_sizes, batch, heads, v_dim))
key_mask_chunk = lax.dynamic_slice(key_mask, (chunk_idx, batch), slice_sizes=(k_chunk_sizes, batch))
attn_weights = einsum('i ... d, j ... d -> i ... j', q_scaled, k_chunk)
p = jnp.exp(attn_weights - lse)
key_mask_chunk = rearrange(key_mask_chunk, 'j b -> 1 b 1 j')
p = jnp.where(key_mask_chunk, p, 0.)
dv_chunk = einsum('i ... j, i ... d -> j ... d', p, do)
dp = einsum('i ... d, j ... d -> i ... j', do, v_chunk)
D = jnp.sum(do * o, axis = -1, keepdims = True)
ds = p * scale * (dp - D)
dq_chunk = einsum('i ... j, j ... d -> i ... d', ds, k_chunk)
dk_chunk = einsum('i ... j, i ... d -> j ... d', ds, q)
return (chunk_idx + k_chunk_sizes, dq + dq_chunk), (dk_chunk, dv_chunk)
dq = jnp.zeros_like(q)
(_, dq), (dk, dv) = lax.scan(chunk_scanner, init = (0, dq), xs = None, length = math.ceil(k_len / K_CHUNK_SIZE))
dk = rearrange(dk, 'c n ... -> (c n) ...')
dv = rearrange(dv, 'c n ... -> (c n) ...')
return dq, dk, dv
@jit
def flash_attention_backward(res, do):
q, k, v, key_mask, o, lse = res
batch, heads, q_len, dim = q.shape
lse = rearrange(lse, 'b h n -> n b h 1')
q, k, v, o, do = map(lambda t: rearrange(t, 'b h n d -> n b h d'), (q, k, v, o, do))
key_mask = rearrange(key_mask, 'b j -> j b')
dk = jnp.zeros_like(k)
dv = jnp.zeros_like(v)
def chunk_scanner(carries, _):
chunk_idx, dk, dv = carries
chunk_sizes = min(Q_CHUNK_SIZE, q_len)
q_chunk = lax.dynamic_slice(q, (chunk_idx, batch, heads, 0), slice_sizes = (chunk_sizes, batch, heads, q.shape[-1]))
lse_chunk = lax.dynamic_slice(lse, (chunk_idx, batch, heads, 0), slice_sizes = (chunk_sizes, batch, heads, 1))
o_chunk = lax.dynamic_slice(o, (chunk_idx, batch, heads, 0), slice_sizes = (chunk_sizes, batch, heads, o.shape[-1]))
do_chunk = lax.dynamic_slice(do, (chunk_idx, batch, heads, 0), slice_sizes = (chunk_sizes, batch, heads, do.shape[-1]))
dq_chunk, dk_chunk, dv_chunk = _query_chunk_flash_attention_backward(q_chunk, k, v, key_mask, o_chunk, do_chunk, lse_chunk)
return (chunk_idx + chunk_sizes, dk + dk_chunk, dv + dv_chunk), dq_chunk
(_, dk, dv), dq = lax.scan(chunk_scanner, init = (0, dk, dv), xs = None, length = math.ceil(q_len / Q_CHUNK_SIZE))
dq = rearrange(dq, 'c n b h d -> b h (c n) d')
dk, dv = map(lambda t: rearrange(t, 'n b h d -> b h n d'), (dk, dv))
return dq, dk, dv, None
flash_attention.defvjp(flash_attention_forward, flash_attention_backward)