File size: 11,436 Bytes
1e5dcf2 |
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 |
"""
Muon Optimizer for BitTransformerLM Extensions
==============================================
Implementation of the Muon optimizer with orthogonal momentum updates.
Based on "Muon: Momentum Orthogonalized by Newton's method" research.
Key features:
- Orthogonal momentum updates
- Better convergence properties than Adam/AdamW
- Memory efficient implementation
- Compatible with BitTransformerLM's training infrastructure
"""
import math
import torch
from torch.optim.optimizer import Optimizer
from typing import Any, Dict, List, Optional, Tuple, Union
import warnings
class Muon(Optimizer):
"""
Muon optimizer with orthogonal momentum updates.
This implementation provides momentum updates that are orthogonalized using
Newton's method, leading to more stable training dynamics.
Args:
params: Iterable of parameters to optimize
lr: Learning rate (default: 1e-3)
momentum: Momentum factor (default: 0.95)
nesterov: Enable Nesterov momentum (default: False)
backend: Backend for orthogonalization ('newtonschulz' or 'svd')
update_period: Period for updating orthogonalization (default: 1)
rank_deficiency_threshold: Threshold for rank deficiency detection
eps: Small constant for numerical stability (default: 1e-8)
weight_decay: Weight decay coefficient (default: 0.0)
"""
def __init__(
self,
params,
lr: float = 1e-3,
momentum: float = 0.95,
nesterov: bool = False,
backend: str = "newtonschulz",
update_period: int = 1,
rank_deficiency_threshold: float = 1e-6,
eps: float = 1e-8,
weight_decay: float = 0.0,
):
if not 0.0 <= lr:
raise ValueError(f"Invalid learning rate: {lr}")
if not 0.0 <= momentum <= 1.0:
raise ValueError(f"Invalid momentum value: {momentum}")
if not 0.0 <= weight_decay:
raise ValueError(f"Invalid weight_decay value: {weight_decay}")
if backend not in ["newtonschulz", "svd"]:
raise ValueError(f"Invalid backend: {backend}")
defaults = dict(
lr=lr,
momentum=momentum,
nesterov=nesterov,
backend=backend,
update_period=update_period,
rank_deficiency_threshold=rank_deficiency_threshold,
eps=eps,
weight_decay=weight_decay,
)
super().__init__(params, defaults)
def _orthogonalize_newtonschulz(self, matrix: torch.Tensor, num_iterations: int = 5) -> torch.Tensor:
"""Orthogonalize matrix using Newton-Schulz iteration."""
# Handle different shapes
original_shape = matrix.shape
if matrix.dim() > 2:
matrix = matrix.view(-1, matrix.shape[-1])
if matrix.shape[0] >= matrix.shape[1]:
# Tall matrix - orthogonalize columns
X = matrix.clone()
for _ in range(num_iterations):
A = X.T @ X
X = X @ (1.5 * torch.eye(A.shape[0], device=A.device, dtype=A.dtype) - 0.5 * A)
else:
# Wide matrix - orthogonalize rows
X = matrix.clone()
for _ in range(num_iterations):
A = X @ X.T
X = (1.5 * torch.eye(A.shape[0], device=A.device, dtype=A.dtype) - 0.5 * A) @ X
return X.view(original_shape)
def _orthogonalize_svd(self, matrix: torch.Tensor) -> torch.Tensor:
"""Orthogonalize matrix using SVD decomposition."""
original_shape = matrix.shape
if matrix.dim() > 2:
matrix = matrix.view(-1, matrix.shape[-1])
try:
U, _, Vt = torch.linalg.svd(matrix, full_matrices=False)
orthogonal = U @ Vt
return orthogonal.view(original_shape)
except torch._C._LinAlgError:
# Fallback to Newton-Schulz if SVD fails
warnings.warn("SVD failed, falling back to Newton-Schulz")
return self._orthogonalize_newtonschulz(matrix)
@torch.no_grad()
def step(self, closure=None):
"""Perform a single optimization step."""
loss = None
if closure is not None:
with torch.enable_grad():
loss = closure()
for group in self.param_groups:
for p in group["params"]:
if p.grad is None:
continue
grad = p.grad
if grad.dtype in {torch.float16, torch.bfloat16}:
grad = grad.float()
state = self.state[p]
# State initialization
if len(state) == 0:
state["step"] = 0
state["momentum_buffer"] = torch.zeros_like(p, memory_format=torch.preserve_format)
momentum_buffer = state["momentum_buffer"]
state["step"] += 1
# Weight decay
if group["weight_decay"] != 0:
grad = grad.add(p, alpha=group["weight_decay"])
# Apply momentum
momentum_buffer.mul_(group["momentum"]).add_(grad)
# Orthogonalize momentum every update_period steps
if state["step"] % group["update_period"] == 0 and momentum_buffer.numel() > 1:
# Only orthogonalize if we have sufficient dimensions
if momentum_buffer.dim() >= 2 and min(momentum_buffer.shape[-2:]) > 1:
if group["backend"] == "newtonschulz":
orthogonal_momentum = self._orthogonalize_newtonschulz(momentum_buffer)
else:
orthogonal_momentum = self._orthogonalize_svd(momentum_buffer)
# Check for rank deficiency
rank_ratio = torch.linalg.matrix_norm(orthogonal_momentum) / torch.linalg.matrix_norm(momentum_buffer)
if rank_ratio < group["rank_deficiency_threshold"]:
warnings.warn("Detected rank deficiency in momentum buffer")
else:
momentum_buffer.copy_(orthogonal_momentum)
# Apply Nesterov acceleration if enabled
if group["nesterov"]:
update = grad.add(momentum_buffer, alpha=group["momentum"])
else:
update = momentum_buffer
# Apply update
p.add_(update, alpha=-group["lr"])
return loss
def configure_muon_optimizer(
model: torch.nn.Module,
lr: float = 1e-3,
momentum: float = 0.95,
weight_decay: float = 0.01,
total_steps: Optional[int] = None,
warmup_ratio: float = 0.1,
nesterov: bool = False,
backend: str = "newtonschulz",
**muon_kwargs
) -> Tuple[Muon, Optional[torch.optim.lr_scheduler._LRScheduler]]:
"""
Configure Muon optimizer with OneCycle learning rate schedule.
This function provides a drop-in replacement for BitTransformerLM's
configure_optimizer function, using Muon instead of AdamW.
Args:
model: PyTorch model to optimize
lr: Peak learning rate
momentum: Momentum factor for Muon
weight_decay: Weight decay coefficient
total_steps: Total training steps for OneCycle schedule
warmup_ratio: Fraction of steps for warmup
nesterov: Enable Nesterov momentum
backend: Orthogonalization backend
**muon_kwargs: Additional arguments for Muon optimizer
Returns:
Tuple of (optimizer, scheduler)
"""
# Filter parameters that need weight decay
decay_params = []
no_decay_params = []
for name, param in model.named_parameters():
if not param.requires_grad:
continue
# Apply weight decay to weights but not biases/norms
if param.dim() >= 2:
decay_params.append(param)
else:
no_decay_params.append(param)
param_groups = [
{"params": decay_params, "weight_decay": weight_decay},
{"params": no_decay_params, "weight_decay": 0.0},
]
optimizer = Muon(
param_groups,
lr=lr,
momentum=momentum,
nesterov=nesterov,
backend=backend,
**muon_kwargs
)
scheduler = None
if total_steps is not None and total_steps > 0:
scheduler = torch.optim.lr_scheduler.OneCycleLR(
optimizer,
max_lr=lr,
total_steps=total_steps,
pct_start=warmup_ratio,
anneal_strategy='cos',
cycle_momentum=False, # Muon handles momentum internally
div_factor=25.0,
final_div_factor=1e4,
)
return optimizer, scheduler
def create_muon_training_config(
lr: float = 1e-3,
momentum: float = 0.95,
weight_decay: float = 0.01,
backend: str = "newtonschulz",
nesterov: bool = False,
**kwargs
) -> Dict[str, Any]:
"""
Create a training configuration dictionary for Muon optimizer.
This can be used with BitTransformerLM's training scripts by passing
the config to the training loop.
Args:
lr: Learning rate
momentum: Momentum factor
weight_decay: Weight decay coefficient
backend: Orthogonalization backend
nesterov: Enable Nesterov momentum
**kwargs: Additional configuration options
Returns:
Dictionary containing training configuration
"""
config = {
"optimizer_type": "muon",
"optimizer_config": {
"lr": lr,
"momentum": momentum,
"weight_decay": weight_decay,
"backend": backend,
"nesterov": nesterov,
**kwargs
},
"scheduler_type": "onecycle",
}
return config
# Example usage and integration helpers
def integrate_with_bittransformerlm():
"""
Example of how to integrate Muon optimizer with BitTransformerLM training.
Usage:
from BTLM_Extensions.muon_optimizer import configure_muon_optimizer
# Replace the standard optimizer configuration
optimizer, scheduler = configure_muon_optimizer(
model, lr=1e-3, momentum=0.95, total_steps=1000
)
# Use in training loop
train_loop(model, data, optimizer=optimizer, scheduler=scheduler)
"""
pass
if __name__ == "__main__":
# Simple test of the optimizer
import torch.nn as nn
model = nn.Sequential(
nn.Linear(10, 20),
nn.ReLU(),
nn.Linear(20, 1)
)
optimizer, scheduler = configure_muon_optimizer(model, lr=1e-3, total_steps=100)
# Simple training step
x = torch.randn(32, 10)
y = torch.randn(32, 1)
pred = model(x)
loss = nn.functional.mse_loss(pred, y)
loss.backward()
optimizer.step()
if scheduler:
scheduler.step()
print("Muon optimizer test completed successfully!")
print(f"Loss: {loss.item():.4f}") |