File size: 12,360 Bytes
f8519d1 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 |
import math
from dataclasses import dataclass
from typing import Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from pscan import pscan
"""
This file closely follows the mamba_simple.py from the official Mamba implementation, and the mamba-minimal by @johnma2006.
The major differences are :
-the convolution is done with torch.nn.Conv1d
-the selective scan is done in PyTorch
A sequential version of the selective scan is also available for comparison.
- A Mamba model is composed of several layers, which are ResidualBlock.
- A ResidualBlock is composed of a MambaBlock, a normalization, and a residual connection : ResidualBlock(x) = mamba(norm(x)) + x
- This leaves us with the MambaBlock : its input x is (B, L, D) and its outputs y is also (B, L, D) (B=batch size, L=seq len, D=model dim).
First, we expand x into (B, L, 2*ED) (where E is usually 2) and split it into x and z, each (B, L, ED).
Then, we apply the short 1d conv to x, followed by an activation function (silu), then the SSM.
We then multiply it by silu(z).
See Figure 3 of the paper (page 8) for a visual representation of a MambaBlock.
"""
@dataclass
class MambaConfig:
d_model: int # D
n_layers: int
dt_rank: Union[int, str] = 'auto'
d_state: int = 16 # N in paper/comments
expand_factor: int = 2 # E in paper/comments
d_conv: int = 4
dt_min: float = 0.001
dt_max: float = 0.1
dt_init: str = "random" # "random" or "constant"
dt_scale: float = 1.0
dt_init_floor = 1e-4
bias: bool = False
conv_bias: bool = True
pscan: bool = True # use parallel scan mode or sequential mode when training
def __post_init__(self):
self.d_inner = self.expand_factor * self.d_model # E*D = ED in comments
if self.dt_rank == 'auto':
self.dt_rank = math.ceil(self.d_model / 16)
class Mamba(nn.Module):
def __init__(self, config: MambaConfig):
super().__init__()
self.config = config
self.layers = nn.ModuleList([ResidualBlock(config) for _ in range(config.n_layers)])
#self.norm_f = RMSNorm(config.d_model)
def forward(self, x):
# x : (B, L, D)
# y : (B, L, D)
for layer in self.layers:
x = layer(x)
#x = self.norm_f(x)
return x
def step(self, x, caches):
# x : (B, L, D)
# caches : [cache(layer) for all layers], cache : (h, inputs)
# y : (B, L, D)
# caches : [cache(layer) for all layers], cache : (h, inputs)
for i, layer in enumerate(self.layers):
x, caches[i] = layer.step(x, caches[i])
return x, caches
class ResidualBlock(nn.Module):
def __init__(self, config: MambaConfig):
super().__init__()
self.mixer = MambaBlock(config)
self.norm = RMSNorm(config.d_model)
def forward(self, x):
# x : (B, L, D)
# output : (B, L, D)
output = self.mixer(self.norm(x)) + x
return output
def step(self, x, cache):
# x : (B, D)
# cache : (h, inputs)
# h : (B, ED, N)
# inputs: (B, ED, d_conv-1)
# output : (B, D)
# cache : (h, inputs)
output, cache = self.mixer.step(self.norm(x), cache)
output = output + x
return output, cache
class MambaBlock(nn.Module):
def __init__(self, config: MambaConfig):
super().__init__()
self.config = config
# projects block input from D to 2*ED (two branches)
self.in_proj = nn.Linear(config.d_model, 2 * config.d_inner, bias=config.bias)
self.conv1d = nn.Conv1d(in_channels=config.d_inner, out_channels=config.d_inner,
kernel_size=config.d_conv, bias=config.conv_bias,
groups=config.d_inner,
padding=config.d_conv - 1)
nn.init.kaiming_normal_(self.conv1d.weight, mode='fan_out', nonlinearity='leaky_relu')
# projects x to input-dependent Δ, B, C
self.x_proj = nn.Linear(config.d_inner, config.dt_rank + 2 * config.d_state, bias=False)
# projects Δ from dt_rank to d_inner
self.dt_proj = nn.Linear(config.dt_rank, config.d_inner, bias=True)
# dt initialization
# dt weights
dt_init_std = config.dt_rank**-0.5 * config.dt_scale
if config.dt_init == "constant":
nn.init.constant_(self.dt_proj.weight, dt_init_std)
elif config.dt_init == "random":
nn.init.uniform_(self.dt_proj.weight, -dt_init_std, dt_init_std)
else:
raise NotImplementedError
# dt bias
dt = torch.exp(
torch.rand(config.d_inner) * (math.log(config.dt_max) - math.log(config.dt_min)) + math.log(config.dt_min)
).clamp(min=config.dt_init_floor)
inv_dt = dt + torch.log(-torch.expm1(-dt)) # inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
with torch.no_grad():
self.dt_proj.bias.copy_(inv_dt)
#self.dt_proj.bias._no_reinit = True # initialization would set all Linear.bias to zero, need to mark this one as _no_reinit
# todo : explain why removed
# S4D real initialization
A = torch.arange(1, config.d_state + 1, dtype=torch.float32).repeat(config.d_inner, 1)
self.A_log = nn.Parameter(torch.log(A)) # why store A in log ? to keep A < 0 (cf -torch.exp(...)) ? for gradient stability ?
self.D = nn.Parameter(torch.ones(config.d_inner))
# projects block output from ED back to D
self.out_proj = nn.Linear(config.d_inner, config.d_model, bias=config.bias)
def forward(self, x):
# x : (B, L, D)
# y : (B, L, D)
_, L, _ = x.shape
xz = self.in_proj(x) # (B, L, 2*ED)
x, z = xz.chunk(2, dim=-1) # (B, L, ED), (B, L, ED)
# x branch
x = x.transpose(1, 2) # (B, ED, L)
x = self.conv1d(x)[:, :, :L] # depthwise convolution over time, with a short filter
x = x.transpose(1, 2) # (B, L, ED)
x = F.silu(x)
y = self.ssm(x)
# z branch
z = F.silu(z)
output = y * z
output = self.out_proj(output) # (B, L, D)
return output
def ssm(self, x):
# x : (B, L, ED)
# y : (B, L, ED)
A = -torch.exp(self.A_log.float()) # (ED, N)
D = self.D.float()
# TODO remove .float()
deltaBC = self.x_proj(x) # (B, L, dt_rank+2*N)
delta, B, C = torch.split(deltaBC, [self.config.dt_rank, self.config.d_state, self.config.d_state], dim=-1) # (B, L, dt_rank), (B, L, N), (B, L, N)
delta = F.softplus(self.dt_proj(delta)) # (B, L, ED)
if self.config.pscan:
y = self.selective_scan(x, delta, A, B, C, D)
else:
y = self.selective_scan_seq(x, delta, A, B, C, D)
return y
def selective_scan(self, x, delta, A, B, C, D):
# x : (B, L, ED)
# Δ : (B, L, ED)
# A : (ED, N)
# B : (B, L, N)
# C : (B, L, N)
# D : (ED)
# y : (B, L, ED)
deltaA = torch.exp(delta.unsqueeze(-1) * A) # (B, L, ED, N)
deltaB = delta.unsqueeze(-1) * B.unsqueeze(2) # (B, L, ED, N)
BX = deltaB * (x.unsqueeze(-1)) # (B, L, ED, N)
hs = pscan(deltaA, BX)
y = (hs @ C.unsqueeze(-1)).squeeze(3) # (B, L, ED, N) @ (B, L, N, 1) -> (B, L, ED, 1)
y = y + D * x
return y
def selective_scan_seq(self, x, delta, A, B, C, D):
# x : (B, L, ED)
# Δ : (B, L, ED)
# A : (ED, N)
# B : (B, L, N)
# C : (B, L, N)
# D : (ED)
# y : (B, L, ED)
_, L, _ = x.shape
deltaA = torch.exp(delta.unsqueeze(-1) * A) # (B, L, ED, N)
deltaB = delta.unsqueeze(-1) * B.unsqueeze(2) # (B, L, ED, N)
BX = deltaB * (x.unsqueeze(-1)) # (B, L, ED, N)
h = torch.zeros(x.size(0), self.config.d_inner, self.config.d_state, device=deltaA.device) # (B, ED, N)
hs = []
for t in range(0, L):
h = deltaA[:, t] * h + BX[:, t]
hs.append(h)
hs = torch.stack(hs, dim=1) # (B, L, ED, N)
y = (hs @ C.unsqueeze(-1)).squeeze(3) # (B, L, ED, N) @ (B, L, N, 1) -> (B, L, ED, 1)
y = y + D * x
return y
# -------------------------- inference -------------------------- #
"""
Concerning auto-regressive inference
The cool part of using Mamba : inference is constant wrt to sequence length
We just have to keep in cache, for each layer, two things :
- the hidden state h (which is (B, ED, N)), as you typically would when doing inference with a RNN
- the last d_conv-1 inputs of the layer, to be able to compute the 1D conv which is a convolution over the time dimension
(d_conv is fixed so this doesn't incur a growing cache as we progress on generating the sequence)
(and d_conv is usually very small, like 4, so we just have to "remember" the last 3 inputs)
Concretely, these two quantities are put inside a cache tuple, and are named h and inputs respectively.
h is (B, ED, N), and inputs is (B, ED, d_conv-1)
The MambaBlock.step() receives this cache, and, along with outputing the output, alos outputs the updated cache for the next call.
The cache object is initialized as follows : (None, torch.zeros()).
When h is None, the selective scan function detects it and start with h=0.
The torch.zeros() isn't a problem (it's same as just feeding the input, because the conv1d is padded)
As we need one such cache variable per layer, we store a caches object, which is simply a list of cache object. (See mamba_lm.py)
"""
def step(self, x, cache):
# x : (B, D)
# cache : (h, inputs)
# h : (B, ED, N)
# inputs : (B, ED, d_conv-1)
# y : (B, D)
# cache : (h, inputs)
h, inputs = cache
xz = self.in_proj(x) # (B, 2*ED)
x, z = xz.chunk(2, dim=1) # (B, ED), (B, ED)
# x branch
x_cache = x.unsqueeze(2)
x = self.conv1d(torch.cat([inputs, x_cache], dim=2))[:, :, self.config.d_conv-1] # (B, ED)
x = F.silu(x)
y, h = self.ssm_step(x, h)
# z branch
z = F.silu(z)
output = y * z
output = self.out_proj(output) # (B, D)
# prepare cache for next call
inputs = torch.cat([inputs[:, :, 1:], x_cache], dim=2) # (B, ED, d_conv-1)
cache = (h, inputs)
return output, cache
def ssm_step(self, x, h):
# x : (B, ED)
# h : (B, ED, N)
# y : (B, ED)
# h : (B, ED, N)
A = -torch.exp(self.A_log.float()) # (ED, N) # todo : ne pas le faire tout le temps, puisque c'est indépendant de la timestep
D = self.D.float()
# TODO remove .float()
deltaBC = self.x_proj(x) # (B, dt_rank+2*N)
delta, B, C = torch.split(deltaBC, [self.config.dt_rank, self.config.d_state, self.config.d_state], dim=-1) # (B, dt_rank), (B, N), (B, N)
delta = F.softplus(self.dt_proj(delta)) # (B, ED)
deltaA = torch.exp(delta.unsqueeze(-1) * A) # (B, ED, N)
deltaB = delta.unsqueeze(-1) * B.unsqueeze(1) # (B, ED, N)
BX = deltaB * (x.unsqueeze(-1)) # (B, ED, N)
if h is None:
h = torch.zeros(x.size(0), self.config.d_inner, self.config.d_state, device=deltaA.device) # (B, ED, N)
h = deltaA * h + BX # (B, ED, N)
y = (h @ C.unsqueeze(-1)).squeeze(2) # (B, ED, N) @ (B, N, 1) -> (B, ED, 1)
y = y + D * x
# todo : pq h.squeeze(1) ??
return y, h.squeeze(1)
# taken straight from https://github.com/johnma2006/mamba-minimal/blob/master/model.py
class RMSNorm(nn.Module):
def __init__(self, d_model: int, eps: float = 1e-5):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(d_model))
def forward(self, x):
output = x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) * self.weight
return output
|