File size: 2,138 Bytes
4318522 d96f432 4318522 d96f432 4318522 d96f432 4318522 | 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 | import torch
from budget_rl import train as inherited
from hinted_rl.train import weight_hash
@torch.no_grad()
def initialize(model, config):
if config.conditioning in ('native', 'none'):
return dict(recipe='Preserve pretrained weights exactly before task SFT.',
initialized_weights=weight_hash(model), residual_projection_scale=1., qk_rms=None)
preserve_attention = getattr(config, 'preserve_pretrained_attention', False)
attention = ([block.attn.c_attn.weight[:2 * block.attn.n_embd].detach().clone()
for block in model.h] if preserve_attention else [])
receipt = inherited.initialize(model, config)
if preserve_attention:
for block, original in zip(model.h, attention, strict=True):
block.attn.c_attn.weight[:2 * block.attn.n_embd].copy_(original)
receipt.update(recipe='Rescale residual output projections; preserve pretrained Q/K exactly before SFT and learn them without per-step normalization.',
qk_rms=None, preserve_pretrained_attention=True, initialized_weights=weight_hash(model))
return receipt
def optimizer_for(model, config, phase):
if config.optimizer == 'relative':
return inherited.optimizer_for(model, config, phase)
if config.optimizer != 'adamw':
raise ValueError('Unknown optimizer: ' + config.optimizer)
head_lr = config.sft_lr if phase == 'sft' else config.rl_lr
hidden_lr = config.sft_hidden_lr if phase == 'sft' else config.rl_hidden_lr
groups = [dict(params=[parameter], lr=head_lr if parameter is model.lm_head.weight else hidden_lr,
task_base_lr=head_lr if parameter is model.lm_head.weight else hidden_lr,
parameter_name=name) for name, parameter in model.named_parameters()]
return torch.optim.AdamW(groups, betas=(.9, .95), weight_decay=0., eps=1e-8)
def optimizer_step(model, optimizer, config):
if config.conditioning in ('native', 'none') or getattr(config, 'preserve_pretrained_attention', False):
optimizer.step()
else:
inherited.optimizer_step(model, optimizer, config)
|