File size: 9,681 Bytes
72c0672 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 | # Copyright (c) Meta Platforms, Inc. and affiliates.
from dataclasses import dataclass, field
from enum import Enum
from typing import Optional, Tuple
import torch
from torch import nn
from torch.nn import functional as F
from lingua.transformer import FeedForward, InitStdFactor, RMSNorm
from lingua.probe import log_stats
from apps.fastRNN.component.rnn_common import conv1d, scan
@dataclass
class BaseHawkArgs:
dim: int = 512
n_layers: int = 8
n_heads: int = 1
multiple_of: int = 256
ffn_dim_multiplier: Optional[float] = None
lru_dim_multiplier: Optional[float] = None
conv_size: Optional[int] = None
norm_eps: float = 1e-5
init_base_std: Optional[float] = None
init_std_factor: str = "disabled"
_MAX_SQRT_GRADIENT: float = 1000.0
class SqrtBoundDerivative(torch.autograd.Function):
"""Computes a square root with a gradient clipped at `_MAX_SQRT_GRADIENT`."""
@staticmethod
def forward(ctx, x: torch.Tensor) -> torch.Tensor:
"""The forward pass, which is a normal `sqrt`."""
ctx.save_for_backward(x)
return torch.sqrt(x)
@staticmethod
def backward(ctx, grad_output: torch.Tensor) -> torch.Tensor:
"""The backward pass, which clips the `sqrt` gradient."""
(x,) = ctx.saved_tensors
clipped_x_times_4 = torch.clip(4.0 * x, min=1 / (_MAX_SQRT_GRADIENT**2))
return grad_output / torch.sqrt(clipped_x_times_4)
def sqrt_bounded_derivative(x: torch.Tensor) -> torch.Tensor:
return SqrtBoundDerivative.apply(x)
class RGLRU(nn.Module):
def __init__(
self,
dim: int,
n_heads: int,
head_dim: int,
conv_size: Optional[int] = None,
):
super().__init__()
assert dim % n_heads == 0, f"dim {dim} must be divisible by n_heads {n_heads}"
self.dim = dim
self.head_dim = head_dim
self.n_heads = n_heads
assert (
head_dim * n_heads == dim
), f"dim {dim} must be equal to n_heads {n_heads} * head_dim {head_dim}"
self.c = 8.0
self.conv_size = conv_size
if conv_size is not None:
assert (dim % 8 == 0) and (
conv_size in [2, 3, 4]
), f"Causal conv1d only supports conv_size in [2, 3, 4] and hidden_dim/head_dim % 8 == 0, got {dim} and {conv_size}"
self.conv_dim = self.dim
self.conv_weight = nn.Parameter(torch.empty((self.conv_dim, conv_size)))
self.register_parameter("a", nn.Parameter(torch.empty((head_dim))))
self.input_gate = nn.Linear(n_heads * head_dim, dim, bias=False)
self.a_gate = nn.Linear(n_heads * head_dim, dim, bias=False)
def forward(
self, x: torch.Tensor, tok_idx: torch.Tensor, cu_seqlens: torch.Tensor, impl: str = "parallel"
) -> torch.Tensor:
bsz, seqlen, _ = x.shape
if self.conv_size is not None:
conv1d_w = log_stats(self.conv_weight, "conv1d.w")
x = conv1d(
x=x.transpose(1, 2),
conv_weight=conv1d_w,
tok_idx=tok_idx,
cu_seqlens=cu_seqlens,
impl=impl,
cache=self.cache.conv_cache if hasattr(self, "cache") else None,
).transpose(1, 2)
gate_x = F.sigmoid(self.input_gate(x.view_as(x)))
gate_a = F.sigmoid(self.a_gate(x.view_as(x)))
gate_x = gate_x.transpose(1, 2).reshape(
bsz * self.n_heads, self.head_dim, seqlen
)
gate_a = gate_a.transpose(1, 2).reshape(
bsz * self.n_heads, self.head_dim, seqlen
)
a = (
F.softplus(self.a)
.unsqueeze(0)
.unsqueeze(-1)
.expand(bsz * self.n_heads, self.head_dim, seqlen)
)
log_a = -self.c * gate_a * a
a = log_a.exp()
multiplier = sqrt_bounded_derivative(1.0 - (2.0 * log_a).exp())
x = x.transpose(1, 2).reshape(bsz * self.n_heads, self.head_dim, seqlen)
h = scan(
a=a.contiguous(),
b=(multiplier * gate_x * x).contiguous(),
cu_seqlens=cu_seqlens,
impl=impl,
cache=self.cache.state_cache if hasattr(self, "cache") else None,
)
h = h.view(bsz, self.dim, seqlen).transpose(1, 2)
h = log_stats(h, "hidden_state")
return h
def reset_parameters(self, init_std, factor):
in_init_std = init_std or (self.dim ** (-0.5))
in_init_std = in_init_std / factor
for w in [self.input_gate, self.a_gate]:
nn.init.trunc_normal_(
w.weight, std=in_init_std, a=-3 * in_init_std, b=3 * in_init_std
)
min_rad, max_rad = 0.9, 0.999
self.a.data.uniform_(min_rad**2 + 1e-8, max_rad**2 + 1e-8)
self.a.data.log_().mul_(0.5)
if self.conv_size is not None:
conv_std = init_std or (self.conv_size ** (-0.5))
nn.init.trunc_normal_(
self.conv_weight, std=conv_std, a=-3 * conv_std, b=3 * conv_std
)
class RGLRUBlock(nn.Module):
def __init__(
self,
dim: int,
hidden_dim: int,
n_heads: int,
multiple_of: int,
lru_dim_multiplier: Optional[float],
conv_size: Optional[int] = None,
):
super().__init__()
if lru_dim_multiplier is not None:
hidden_dim = int(lru_dim_multiplier * hidden_dim)
hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
assert (
hidden_dim % n_heads == 0
), f"Hidden dim must be divisible by n_heads: {hidden_dim} % {n_heads} != 0"
self.dim = dim
self.hidden_dim = hidden_dim
self.wy = nn.Linear(
dim,
hidden_dim,
bias=False,
)
self.wx = nn.Linear(
dim,
hidden_dim,
bias=False,
)
self.rglru = RGLRU(
dim=hidden_dim,
n_heads=n_heads,
head_dim=hidden_dim // n_heads,
conv_size=conv_size,
)
self.wo = nn.Linear(
hidden_dim,
dim,
bias=False,
)
def forward(
self, x: torch.Tensor, tok_idx: torch.Tensor, cu_seqlens: torch.Tensor, impl: str = "parallel"
) -> torch.Tensor:
h = self.rglru(self.wx(x), tok_idx=tok_idx, cu_seqlens=cu_seqlens, impl=impl)
h = h * F.silu(self.wy(x))
y = x + self.wo(h)
return y
def init_weights(self, init_std: Optional[float], factor: InitStdFactor):
self.rglru.reset_parameters(init_std, factor)
in_init_std = init_std or (self.dim ** (-0.5))
out_init_std = init_std or (self.hidden_dim ** (-0.5))
in_init_std = in_init_std / factor
out_init_std = out_init_std / factor
for w in [self.wy, self.wx]:
nn.init.trunc_normal_(
w.weight, std=in_init_std, a=-3 * in_init_std, b=3 * in_init_std
)
nn.init.trunc_normal_(
self.wo.weight, std=out_init_std, a=-3 * out_init_std, b=3 * out_init_std
)
class HawkBlock(nn.Module):
def __init__(self, args: BaseHawkArgs):
super().__init__()
self.rlgru_block = RGLRUBlock(
dim=args.dim,
hidden_dim=int(4 / 3 * args.dim),
n_heads=args.n_heads,
conv_size=args.conv_size,
multiple_of=args.multiple_of,
lru_dim_multiplier=args.lru_dim_multiplier,
)
self.feed_forward = FeedForward(
dim=args.dim,
hidden_dim=4 * args.dim,
multiple_of=args.multiple_of,
ffn_dim_multiplier=args.ffn_dim_multiplier,
)
self.rlgru_norm = RMSNorm(args.dim, eps=args.norm_eps)
self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps)
def forward(
self, x: torch.Tensor, tok_idx: torch.Tensor, cu_seqlens: torch.Tensor, impl: str = "parallel"
) -> torch.Tensor:
x = x + self.rlgru_block(self.rlgru_norm(x), tok_idx=tok_idx, cu_seqlens=cu_seqlens, impl=impl)
x = x + self.feed_forward(self.ffn_norm(x))
return x
def init_weights(self, init_std: Optional[float], factor: InitStdFactor):
self.rlgru_block.init_weights(init_std, factor)
self.rlgru_norm.reset_parameters()
self.feed_forward.reset_parameters()
self.ffn_norm.reset_parameters()
class BaseHawk(nn.Module):
def __init__(self, args: BaseHawkArgs):
super().__init__()
self.dim = args.dim
self.init_base_std = args.init_base_std
self.init_std_factor = InitStdFactor(args.init_std_factor)
self.layers = nn.ModuleList()
for _ in range(args.n_layers):
self.layers.append(HawkBlock(args))
def forward(
self, h: torch.Tensor, tok_idx: torch.Tensor, cu_seqlens: torch.Tensor, impl: str = "parallel"
) -> torch.Tensor:
for i, layer in enumerate(self.layers):
h = layer(h, tok_idx=tok_idx, cu_seqlens=cu_seqlens, impl=impl)
return h
def reset_parameters(self):
pass
def init_weights(self):
self.reset_parameters()
for depth, layer in enumerate(self.layers):
factor = {
InitStdFactor.CURRENT_DEPTH: (2 * (depth + 1)) ** 0.5,
InitStdFactor.GLOBAL_DEPTH: (2 * (len(self.layers) + 1)) ** 0.5,
InitStdFactor.DIM_RATIO: self.dim / 4096,
InitStdFactor.DISABLED: 1.0,
}[self.init_std_factor]
layer.init_weights(self.init_base_std, factor)
|