292 lines
11 KiB
Python
292 lines
11 KiB
Python
# Ported from ostris/ai-toolkit (MIT License)
|
|
# https://github.com/ostris/ai-toolkit/blob/main/toolkit/optimizers/automagic.py
|
|
# Copyright (c) ostris. Used under MIT License.
|
|
from typing import List
|
|
import torch
|
|
from library.optimizers.optimizer_utils import Auto8bitTensor, copy_stochastic, stochastic_grad_accummulation
|
|
|
|
try:
|
|
from optimum.quanto import QBytesTensor
|
|
_QUANTO_AVAILABLE = True
|
|
except ImportError:
|
|
QBytesTensor = None
|
|
_QUANTO_AVAILABLE = False
|
|
|
|
|
|
class Automagic(torch.optim.Optimizer):
|
|
def __init__(
|
|
self,
|
|
params,
|
|
lr=1e-6,
|
|
min_lr=1e-7,
|
|
max_lr=1e-3,
|
|
lr_bump=1e-6,
|
|
eps=(1e-30, 1e-3),
|
|
clip_threshold=1.0,
|
|
beta2=0.999,
|
|
weight_decay=0.0,
|
|
do_paramiter_swapping=False,
|
|
paramiter_swapping_factor=0.1,
|
|
):
|
|
self.lr = lr
|
|
if self.lr > 1e-3:
|
|
print(f"Warning! Start lr is very high: {self.lr}. Forcing to 1e-6.")
|
|
self.lr = 1e-6
|
|
self.min_lr = min_lr
|
|
self.max_lr = max_lr
|
|
self.lr_bump = lr_bump
|
|
|
|
defaults = {
|
|
"lr": lr,
|
|
"eps": eps,
|
|
"clip_threshold": clip_threshold,
|
|
"beta2": beta2,
|
|
"weight_decay": weight_decay,
|
|
}
|
|
super().__init__(params, defaults)
|
|
|
|
self.base_lrs: List[float] = [lr for group in self.param_groups]
|
|
self.is_stochastic_rounding_accumulation = False
|
|
|
|
for group in self.param_groups:
|
|
for param in group['params']:
|
|
if param.requires_grad and param.dtype != torch.float32:
|
|
self.is_stochastic_rounding_accumulation = True
|
|
param.register_post_accumulate_grad_hook(stochastic_grad_accummulation)
|
|
|
|
self.do_paramiter_swapping = do_paramiter_swapping
|
|
self.paramiter_swapping_factor = paramiter_swapping_factor
|
|
self._total_paramiter_size = sum(
|
|
torch.numel(p) for group in self.param_groups for p in group['params']
|
|
)
|
|
print(f"Automagic: total training parameters: {self._total_paramiter_size:,}")
|
|
|
|
if self.do_paramiter_swapping:
|
|
self.enable_paramiter_swapping(self.paramiter_swapping_factor)
|
|
|
|
def enable_paramiter_swapping(self, paramiter_swapping_factor=0.1):
|
|
self.do_paramiter_swapping = True
|
|
self.paramiter_swapping_factor = paramiter_swapping_factor
|
|
self.swap_paramiters()
|
|
|
|
def swap_paramiters(self):
|
|
import random
|
|
all_params = []
|
|
for group in self.param_groups:
|
|
for param in group['params']:
|
|
param.requires_grad_(False)
|
|
param.grad = None
|
|
all_params.append(param)
|
|
random.shuffle(all_params)
|
|
|
|
target = int(self._total_paramiter_size * self.paramiter_swapping_factor)
|
|
total = 0
|
|
for param in all_params:
|
|
total += torch.numel(param)
|
|
if total >= target:
|
|
break
|
|
else:
|
|
param.requires_grad_(True)
|
|
|
|
@staticmethod
|
|
def _get_lr(param_group, param_state):
|
|
return param_state.get("avg_lr", 0.0)
|
|
|
|
def _get_group_lr(self, group):
|
|
lrs = [self._get_lr(group, self.state[p]) for p in group["params"]]
|
|
return sum(lrs) / len(lrs) if lrs else self.lr
|
|
|
|
@staticmethod
|
|
def _rms(tensor):
|
|
return tensor.norm(2) / (tensor.numel() ** 0.5)
|
|
|
|
@staticmethod
|
|
def _approx_sq_grad(exp_avg_sq_row, exp_avg_sq_col):
|
|
r_factor = (exp_avg_sq_row / exp_avg_sq_row.mean(dim=-1, keepdim=True)).rsqrt_().unsqueeze(-1)
|
|
c_factor = exp_avg_sq_col.unsqueeze(-2).rsqrt()
|
|
return torch.mul(r_factor, c_factor)
|
|
|
|
def step_hook(self):
|
|
if not self.is_stochastic_rounding_accumulation:
|
|
return
|
|
for group in self.param_groups:
|
|
for param in group['params']:
|
|
if param.requires_grad and hasattr(param, "_accum_grad"):
|
|
param.grad = param._accum_grad
|
|
del param._accum_grad
|
|
|
|
def get_learning_rates(self):
|
|
lrs = [self._get_group_lr(group) for group in self.param_groups]
|
|
return lrs if lrs else self.base_lrs
|
|
|
|
def get_avg_learning_rate(self):
|
|
lrs = self.get_learning_rates()
|
|
return sum(lrs) / len(lrs)
|
|
|
|
@torch.no_grad()
|
|
def step(self, closure=None):
|
|
self.step_hook()
|
|
loss = None
|
|
if closure is not None:
|
|
loss = closure()
|
|
|
|
for group in self.param_groups:
|
|
for p in group["params"]:
|
|
if p.grad is None or not p.requires_grad:
|
|
continue
|
|
|
|
grad = p.grad
|
|
if grad.dtype != torch.float32:
|
|
grad = grad.to(torch.float32)
|
|
if grad.is_sparse:
|
|
raise RuntimeError("Automagic does not support sparse gradients.")
|
|
|
|
state = self.state[p]
|
|
grad_shape = grad.shape
|
|
factored = len(grad_shape) >= 2
|
|
|
|
if len(state) == 0:
|
|
self.initialize_state(p)
|
|
else:
|
|
if factored:
|
|
if "exp_avg_sq_row" not in state or "exp_avg_sq_col" not in state:
|
|
state["exp_avg_sq_row"] = torch.zeros(p.shape[:-1]).to(grad)
|
|
state["exp_avg_sq_col"] = torch.zeros(p.shape[:-2] + p.shape[-1:]).to(grad)
|
|
else:
|
|
state["exp_avg_sq_row"] = state["exp_avg_sq_row"].to(grad)
|
|
state["exp_avg_sq_col"] = state["exp_avg_sq_col"].to(grad)
|
|
else:
|
|
if "exp_avg_sq" not in state:
|
|
state["exp_avg_sq"] = torch.zeros_like(grad)
|
|
else:
|
|
state["exp_avg_sq"] = state["exp_avg_sq"].to(grad)
|
|
|
|
p_data_fp32 = p
|
|
if _QUANTO_AVAILABLE and isinstance(p_data_fp32, QBytesTensor):
|
|
p_data_fp32 = p_data_fp32.dequantize()
|
|
if p.dtype != torch.float32:
|
|
p_data_fp32 = p_data_fp32.clone().float()
|
|
|
|
if "step" not in state:
|
|
state["step"] = 0
|
|
state["step"] += 1
|
|
state["RMS"] = self._rms(p_data_fp32)
|
|
|
|
beta2 = group["beta2"]
|
|
eps = group["eps"]
|
|
if isinstance(eps, (tuple, list)):
|
|
eps = eps[0]
|
|
|
|
update = (grad ** 2) + eps
|
|
if factored:
|
|
exp_avg_sq_row = state["exp_avg_sq_row"]
|
|
exp_avg_sq_col = state["exp_avg_sq_col"]
|
|
exp_avg_sq_row.mul_(beta2).add_(update.mean(dim=-1), alpha=(1.0 - beta2))
|
|
exp_avg_sq_col.mul_(beta2).add_(update.mean(dim=-2), alpha=(1.0 - beta2))
|
|
update = self._approx_sq_grad(exp_avg_sq_row, exp_avg_sq_col)
|
|
update.mul_(grad)
|
|
else:
|
|
exp_avg_sq = state["exp_avg_sq"]
|
|
exp_avg_sq.mul_(beta2).add_(update, alpha=(1.0 - beta2))
|
|
update = exp_avg_sq.rsqrt().mul_(grad)
|
|
|
|
update.div_((self._rms(update) / group["clip_threshold"]).clamp_(min=1.0))
|
|
|
|
if 'last_polarity' not in state or 'lr_mask' not in state:
|
|
self.initialize_state(p)
|
|
|
|
last_polarity = state['last_polarity']
|
|
current_polarity = (update > 0).to(torch.bool)
|
|
sign_agreement = torch.where(last_polarity == current_polarity, 1, -1)
|
|
state['last_polarity'] = current_polarity
|
|
|
|
lr_mask = state['lr_mask'].to(torch.float32)
|
|
new_lr = torch.where(
|
|
sign_agreement > 0,
|
|
lr_mask + self.lr_bump,
|
|
lr_mask - self.lr_bump,
|
|
)
|
|
new_lr = torch.clamp(new_lr, min=self.min_lr, max=self.max_lr)
|
|
update.mul_(new_lr)
|
|
|
|
state['lr_mask'] = Auto8bitTensor(new_lr)
|
|
state['avg_lr'] = torch.mean(new_lr)
|
|
|
|
if group["weight_decay"] != 0:
|
|
weight_decay_update = p_data_fp32 * (-group["weight_decay"]) * new_lr
|
|
p_data_fp32.add_(weight_decay_update)
|
|
|
|
p_data_fp32.add_(-update)
|
|
|
|
if p.dtype != torch.float32:
|
|
copy_stochastic(p, p_data_fp32)
|
|
|
|
return loss
|
|
|
|
def initialize_state(self, p):
|
|
state = self.state[p]
|
|
state["step"] = 0
|
|
if 'lr_mask' not in state:
|
|
state['lr_mask'] = Auto8bitTensor(
|
|
torch.ones(p.shape).to(p.device, dtype=torch.float32) * self.lr
|
|
)
|
|
state['avg_lr'] = torch.mean(state['lr_mask'].to(torch.float32))
|
|
if 'last_polarity' not in state:
|
|
state['last_polarity'] = torch.zeros(p.shape, dtype=torch.bool, device=p.device)
|
|
factored = len(p.shape) >= 2
|
|
if factored:
|
|
state["exp_avg_sq_row"] = torch.zeros(p.shape[:-1]).to(p)
|
|
state["exp_avg_sq_col"] = torch.zeros(p.shape[:-2] + p.shape[-1:]).to(p)
|
|
else:
|
|
state["exp_avg_sq"] = torch.zeros_like(p)
|
|
state["RMS"] = 0
|
|
|
|
def state_dict(self, *args, **kwargs):
|
|
orig = super().state_dict(*args, **kwargs)
|
|
new_state = {}
|
|
for p, state in orig['state'].items():
|
|
s = {k: v for k, v in state.items() if k != 'lr_mask'}
|
|
if 'lr_mask' in state:
|
|
s['lr_mask'] = state['lr_mask'].state_dict()
|
|
new_state[p] = s
|
|
orig['state'] = new_state
|
|
return orig
|
|
|
|
def load_state_dict(self, state_dict, strict=True):
|
|
is_valid = any(
|
|
'lr_mask' in pstate
|
|
for pstate in state_dict.get('state', {}).values()
|
|
if isinstance(pstate, dict)
|
|
)
|
|
if not is_valid:
|
|
return
|
|
|
|
stripped = {
|
|
'state': {pid: {k: v for k, v in ps.items() if k != 'lr_mask'}
|
|
for pid, ps in state_dict['state'].items()},
|
|
'param_groups': state_dict['param_groups'],
|
|
}
|
|
super().load_state_dict(stripped)
|
|
|
|
current_params = [p for g in self.param_groups for p in g['params'] if p.requires_grad]
|
|
saved_ids = list(state_dict['state'].keys())
|
|
for i, param in enumerate(current_params):
|
|
if i >= len(saved_ids):
|
|
break
|
|
saved = state_dict['state'][saved_ids[i]]
|
|
if 'lr_mask' not in saved:
|
|
continue
|
|
if param not in self.state:
|
|
self.initialize_state(param)
|
|
try:
|
|
lm = saved['lr_mask']
|
|
if 'quantized' in lm and lm['quantized'].shape == param.shape:
|
|
self.state[param]['lr_mask'] = Auto8bitTensor(lm)
|
|
else:
|
|
raise ValueError("shape mismatch")
|
|
except Exception as e:
|
|
print(f"Automagic: reinit lr_mask for param {i}: {e}")
|
|
self.state[param]['lr_mask'] = Auto8bitTensor(
|
|
torch.ones(param.shape).to(param.device, dtype=torch.float32) * self.lr
|
|
)
|