Spaces:
Running
Running
File size: 6,981 Bytes
891b88f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 |
import math
from abc import ABC, abstractmethod
from itertools import product
from typing import Any, Optional
import attr
import numpy as np
import torch
@attr.s
class AttentionMask(ABC):
query_context_size: int = attr.ib(validator=lambda i, a, x: x >= 1) # type: ignore
key_context_size: int = attr.ib(validator=lambda i, a, x: x >= 1) # type: ignore
block_size: int = attr.ib(validator=lambda i, a, x: x >= 1) # type: ignore
n_head: int = attr.ib(validator=lambda i, a, x: x >= 1) # type: ignore
is_head_specific: bool = attr.ib(default=False)
n_query_pad: int = attr.ib(default=0)
n_key_pad: int = attr.ib(default=0)
def __attrs_post_init__(self) -> None:
if self.query_context_size % self.block_size != 0:
raise ValueError()
if self.key_context_size % self.block_size != 0:
raise ValueError()
if self.n_query_pad >= self.query_context_size:
raise ValueError()
if self.n_key_pad >= self.key_context_size:
raise ValueError()
self.n_query_block = self.query_context_size // self.block_size
self.n_key_block = self.key_context_size // self.block_size
self.first_pad_query_block_idx = self.n_query_block - int(
math.ceil(self.n_query_pad / self.block_size)
)
self.first_pad_key_block_idx = self.n_key_block - int(
math.ceil(self.n_key_pad / self.block_size)
)
def _make_global_layout(self) -> None:
if not self.is_head_specific:
m = np.ones([self.n_query_block, self.n_key_block], dtype=np.bool)
r = product(*[range(n) for n in m.shape])
for qb, kb in r:
m[qb, kb] = np.any(self.block_layout(None, 0, qb, kb, 0))
else:
m = np.ones([self.n_head, self.n_query_block, self.n_key_block], dtype=np.bool)
r = product(*[range(n) for n in m.shape])
for h, qb, kb in r:
m[h, qb, kb] = np.any(self.block_layout(None, h, qb, kb, 0))
self.global_layout = m
@abstractmethod
def _block_layout(
self, blk_shape: Any, head_idx: int, query_idx: int, key_idx: int, blk_idx: int
) -> np.ndarray:
raise NotImplementedError()
def block_layout(
self, blk_shape: Any, head_idx: int, query_idx: int, key_idx: int, blk_idx: int
) -> np.ndarray:
"""
`query_idx`, `key_idx` are block-level, zero-based indices.
"""
m = np.ones([self.block_size, self.block_size], dtype=np.bool)
if query_idx >= self.first_pad_query_block_idx:
n_pad = min(
self.block_size,
(query_idx + 1) * self.block_size - (self.query_context_size - self.n_query_pad),
)
assert n_pad > 0
m[self.block_size - n_pad :] = False
if key_idx >= self.first_pad_key_block_idx:
n_pad = min(
self.block_size,
(key_idx + 1) * self.block_size - (self.key_context_size - self.n_key_pad),
)
assert n_pad > 0
m[:, self.block_size - n_pad :] = False
return m & self._block_layout(blk_shape, head_idx, query_idx, key_idx, blk_idx)
@attr.s
class DenseAttentionMask(AttentionMask):
def __attrs_post_init__(self) -> None:
super().__attrs_post_init__()
self.global_layout = np.ones([self.n_query_block, self.n_key_block], dtype=np.bool)
n_zero_query_blocks = self.n_query_pad // self.block_size
n_zero_key_blocks = self.n_key_pad // self.block_size
self.global_layout[self.n_query_block - n_zero_query_blocks :] = False
self.global_layout[:, self.n_key_block - n_zero_key_blocks :] = False
def _block_layout(
self, blk_shape: Any, head_idx: int, query_idx: int, key_idx: int, blk_idx: int
) -> np.ndarray:
return np.ones([self.block_size, self.block_size], dtype=np.bool)
@attr.s
class DenseCausalAttentionMask(AttentionMask):
def __attrs_post_init__(self) -> None:
super().__attrs_post_init__()
self.global_layout = np.tril(np.ones([self.n_query_block, self.n_key_block], dtype=np.bool))
n_zero_query_blocks = self.n_query_pad // self.block_size
n_zero_key_blocks = self.n_key_pad // self.block_size
self.global_layout[self.n_query_block - n_zero_query_blocks :] = False
self.global_layout[:, self.n_key_block - n_zero_key_blocks :] = False
def _block_layout(
self, blk_shape: Any, head_idx: int, query_idx: int, key_idx: int, blk_idx: int
) -> np.ndarray:
if query_idx > key_idx:
return np.ones(2 * [self.block_size], dtype=np.bool)
elif query_idx < key_idx:
return np.zeros(2 * [self.block_size], dtype=np.bool)
else:
return np.tril(np.ones(2 * [self.block_size], dtype=np.bool))
@attr.s(eq=False, repr=False)
class AttentionInfo:
n_heads: int = attr.ib()
ctx_blks_q: int = attr.ib()
ctx_blks_k: int = attr.ib()
block_size: int = attr.ib()
pytorch_attn_bias: Optional[torch.Tensor] = attr.ib()
def to_attention_info(d: AttentionMask) -> AttentionInfo:
return AttentionInfo(
n_heads=d.n_head,
ctx_blks_q=d.n_query_block,
ctx_blks_k=d.n_key_block,
block_size=d.block_size,
pytorch_attn_bias=None,
)
def make_full_layout(d: AttentionMask) -> np.ndarray:
"""
Returns the `context_size x context_size` layout matrix described by `d`. If the layout is dependent on the index of
the attention head, a `attention_head x context_size x context_size` layout matrix is returned instead.
"""
if not d.is_head_specific:
u = np.reshape(d.global_layout, [d.n_query_block, d.n_key_block, 1, 1])
r = product(range(d.n_query_block), range(d.n_key_block))
v = np.array([d.block_layout(None, 0, i, j, 0) for i, j in r])
v = np.reshape(v, [d.n_query_block, d.n_key_block, d.block_size, d.block_size])
w = u * v
w = np.transpose(w, [0, 2, 1, 3])
w = np.reshape(w, [d.query_context_size, d.key_context_size])
return w
else:
if len(d.global_layout.shape) == 2:
u = np.reshape(d.global_layout, [1, d.n_query_block, d.n_key_block, 1, 1])
u = np.tile(u, [d.n_head, 1, 1, 1, 1])
elif len(d.global_layout.shape) == 3:
u = np.reshape(d.global_layout, [d.n_head, d.n_query_block, d.n_key_block, 1, 1])
else:
raise RuntimeError()
s = product(range(d.n_head), range(d.n_query_block), range(d.n_key_block))
v = np.array([d.block_layout(None, i, j, k, 0) for i, j, k in s])
v = np.reshape(v, [d.n_head, d.n_query_block, d.n_key_block, d.block_size, d.block_size])
w = u * v
w = np.transpose(w, [0, 1, 3, 2, 4])
w = np.reshape(w, [d.n_head, d.query_context_size, d.key_context_size])
return w
|