92 lines
4.2 KiB
Python
Vendored
92 lines
4.2 KiB
Python
Vendored
from collections import OrderedDict
|
|
import os
|
|
import sys
|
|
from typing import Dict
|
|
import typing
|
|
import torch
|
|
import bitsandbytes as bnb
|
|
from argparse import ArgumentParser
|
|
|
|
parser = ArgumentParser()
|
|
parser.add_argument("--type", default="pissa", type=str)
|
|
parser.add_argument("--base_model", default="", type=str)
|
|
parser.add_argument("--lora_init", default="none", type=str)
|
|
parser.add_argument("--lora_checkpoint", default="", type=str)
|
|
parser.add_argument("--output", default="", type=str)
|
|
parser.add_argument("--quant", default="none", type=str)
|
|
parser.add_argument("--device", default="cuda", type=str)
|
|
parser.add_argument("--lora_alpha", default=16, type=int)
|
|
args = parser.parse_args()
|
|
device= args.device
|
|
base_model = args.base_model
|
|
init_lora= args.lora_init
|
|
lora= args.lora_checkpoint
|
|
output= args.output
|
|
quant= args.quant
|
|
lora_alpha = args.lora_alpha
|
|
|
|
with torch.no_grad():
|
|
w: Dict[str, torch.Tensor] = torch.load(base_model, map_location='cpu')
|
|
# merge LoRA-only slim checkpoint into the main weights
|
|
w_lora: Dict[str, torch.Tensor] = torch.load(lora, map_location='cpu')
|
|
|
|
if args.type=='pissa':
|
|
w_init_lora: Dict[str, torch.Tensor] = torch.load(init_lora, map_location='cpu')
|
|
for k in w_lora.keys():
|
|
w[k] = w_lora[k]
|
|
output_w: typing.OrderedDict[str, torch.Tensor] = OrderedDict()
|
|
# merge LoRA weights
|
|
keys = list(w.keys())
|
|
for k in keys:
|
|
if k.endswith('.weight'):
|
|
prefix = k[:-len('.weight')]
|
|
lora_A = prefix + '.lora_A'
|
|
lora_B = prefix + '.lora_B'
|
|
init_lora_A = prefix + '.init_lora_A'
|
|
init_lora_B = prefix + '.init_lora_B'
|
|
if lora_A in keys:
|
|
assert lora_B in keys
|
|
print(f'merging {lora_A} and {lora_B} into {k}')
|
|
assert w[lora_B].shape[1] == w[lora_A].shape[0]
|
|
lora_r = w[lora_B].shape[1]
|
|
w[k] = w[k].to(device=device)
|
|
w[lora_A] = w[lora_A].to(device=device)
|
|
w[lora_B] = w[lora_B].to(device=device)
|
|
|
|
if args.type=='pissa':
|
|
w_init_lora[init_lora_A] = w_init_lora[init_lora_A].to(device=device)
|
|
w_init_lora[init_lora_B] = w_init_lora[init_lora_B].to(device=device)
|
|
if quant=='4bit':
|
|
qw,qs = bnb.functional.quantize_4bit(w[k]- w_init_lora[init_lora_B] @ w_init_lora[init_lora_A])
|
|
w[k] = (bnb.functional.dequantize_4bit(qw,quant_state=qs)).to(dtype=torch.bfloat16)
|
|
elif quant == 'nf4':
|
|
qw,qs = bnb.functional.quantize_nf4(w[k]- w_init_lora[init_lora_B] @ w_init_lora[init_lora_A])
|
|
w[k] = (bnb.functional.dequantize_nf4(qw,quant_state=qs)).to(dtype=torch.bfloat16)
|
|
elif quant == 'fp4':
|
|
qw,qs = bnb.functional.quantize_fp4(w[k]- w_init_lora[init_lora_B] @ w_init_lora[init_lora_A])
|
|
w[k] = (bnb.functional.dequantize_fp4(qw,quant_state=qs)).to(dtype=torch.bfloat16)
|
|
else:
|
|
w[k] = (w[k]- w_init_lora[init_lora_B] @ w_init_lora[init_lora_A]).to(dtype=torch.bfloat16)
|
|
w[k] += w[lora_B] @ w[lora_A]
|
|
else:
|
|
if quant=='4bit':
|
|
qw,qs = bnb.functional.quantize_4bit(w[k])
|
|
w[k] = (bnb.functional.dequantize_4bit(qw,quant_state=qs)).to(dtype=torch.bfloat16)
|
|
elif quant=='nf4':
|
|
qw,qs = bnb.functional.quantize_nf4(w[k])
|
|
w[k] = (bnb.functional.dequantize_nf4(qw,quant_state=qs)).to(dtype=torch.bfloat16)
|
|
elif quant=='fp4':
|
|
qw,qs = bnb.functional.quantize_fp4(w[k])
|
|
w[k] = (bnb.functional.dequantize_fp4(qw,quant_state=qs)).to(dtype=torch.bfloat16)
|
|
w[k] += w[lora_B] @ w[lora_A] * (lora_alpha / lora_r)
|
|
output_w[k] = w[k].to(device='cpu', copy=True)
|
|
del w[k]
|
|
del w[lora_A]
|
|
del w[lora_B]
|
|
continue
|
|
|
|
if 'lora' not in k:
|
|
print(f'retaining {k}')
|
|
output_w[k] = w[k].clone()
|
|
del w[k]
|
|
torch.save(output_w, output) |