improvement(tools): optimize with argparse
This commit is contained in:
tpoisonooo
parent
904d2a8d6a
commit
3c7cb413fb
1 changed files with 129 additions and 112 deletions
|
|
@ -22,19 +22,27 @@ import json
|
||||||
import struct
|
import struct
|
||||||
import numpy as np
|
import numpy as np
|
||||||
import torch
|
import torch
|
||||||
|
import argparse
|
||||||
|
import os
|
||||||
|
|
||||||
from sentencepiece import SentencePieceProcessor
|
from sentencepiece import SentencePieceProcessor
|
||||||
|
|
||||||
if len(sys.argv) < 3:
|
|
||||||
print("Usage: convert-ckpt-to-ggml.py dir-model ftype\n")
|
|
||||||
print(" ftype == 0 -> float32")
|
|
||||||
print(" ftype == 1 -> float16")
|
|
||||||
sys.exit(1)
|
|
||||||
|
|
||||||
# output in the same directory as the model
|
def parse_args():
|
||||||
dir_model = sys.argv[1]
|
parser = argparse.ArgumentParser(
|
||||||
|
description='Convert ckpt models to ggml models.')
|
||||||
|
parser.add_argument('dir_model',
|
||||||
|
type=str,
|
||||||
|
help='Directory path of the checkpoint model')
|
||||||
|
parser.add_argument('ftype',
|
||||||
|
type=str,
|
||||||
|
choices=['f32', 'f16'],
|
||||||
|
help='Data type of the converted tensor, f32 or f16')
|
||||||
|
parser.add_argument('out_dir',
|
||||||
|
type=str,
|
||||||
|
help='Directory path for storing ggml model')
|
||||||
|
return parser.parse_args()
|
||||||
|
|
||||||
fname_hparams = sys.argv[1] + "/params.json"
|
|
||||||
fname_tokenizer = sys.argv[1] + "/../tokenizer.model"
|
|
||||||
|
|
||||||
def get_n_parts(dim):
|
def get_n_parts(dim):
|
||||||
if dim == 4096:
|
if dim == 4096:
|
||||||
|
|
@ -49,129 +57,138 @@ def get_n_parts(dim):
|
||||||
print("Invalid dim: " + str(dim))
|
print("Invalid dim: " + str(dim))
|
||||||
sys.exit(1)
|
sys.exit(1)
|
||||||
|
|
||||||
# possible data types
|
|
||||||
# ftype == 0 -> float32
|
|
||||||
# ftype == 1 -> float16
|
|
||||||
#
|
|
||||||
# map from ftype to string
|
|
||||||
ftype_str = ["f32", "f16"]
|
|
||||||
|
|
||||||
ftype = 1
|
def main():
|
||||||
if len(sys.argv) > 2:
|
args = parse_args()
|
||||||
ftype = int(sys.argv[2])
|
dir_model = args.dir_model
|
||||||
if ftype < 0 or ftype > 1:
|
out_dir = args.out_dir
|
||||||
print("Invalid ftype: " + str(ftype))
|
|
||||||
sys.exit(1)
|
|
||||||
fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
|
|
||||||
|
|
||||||
with open(fname_hparams, "r") as f:
|
if not os.path.exists(out_dir):
|
||||||
hparams = json.load(f)
|
os.mkdir(out_dir)
|
||||||
|
|
||||||
tokenizer = SentencePieceProcessor(fname_tokenizer)
|
ftype = args.ftype
|
||||||
|
ftype_int = {'f32': 0, 'f16': 1}
|
||||||
|
fname_hparams = os.path.join(dir_model, 'params.json')
|
||||||
|
fname_tokenizer = os.path.join(dir_model, '..', 'tokenizer.model')
|
||||||
|
|
||||||
hparams.update({"vocab_size": tokenizer.vocab_size()})
|
with open(fname_hparams, "r") as f:
|
||||||
|
hparams = json.load(f)
|
||||||
|
|
||||||
n_parts = get_n_parts(hparams["dim"])
|
tokenizer = SentencePieceProcessor(fname_tokenizer)
|
||||||
|
|
||||||
print(hparams)
|
hparams.update({"vocab_size": tokenizer.vocab_size()})
|
||||||
print('n_parts = ', n_parts)
|
|
||||||
|
|
||||||
for p in range(n_parts):
|
n_parts = get_n_parts(hparams["dim"])
|
||||||
print('Processing part ', p)
|
|
||||||
|
|
||||||
#fname_model = sys.argv[1] + "/consolidated.00.pth"
|
print(hparams)
|
||||||
fname_model = sys.argv[1] + "/consolidated.0" + str(p) + ".pth"
|
print('n_parts = ', n_parts)
|
||||||
fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
|
|
||||||
if (p > 0):
|
|
||||||
fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin" + "." + str(p)
|
|
||||||
|
|
||||||
model = torch.load(fname_model, map_location="cpu")
|
for p in range(n_parts):
|
||||||
|
print('Processing part ', p)
|
||||||
|
|
||||||
fout = open(fname_out, "wb")
|
#fname_model = sys.argv[1] + "/consolidated.00.pth"
|
||||||
|
fname_model = os.path.join(dir_model, "consolidated.0{}.pth".format(p))
|
||||||
fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
|
if p > 0:
|
||||||
fout.write(struct.pack("i", hparams["vocab_size"]))
|
fname_out = os.path.join(out_dir,
|
||||||
fout.write(struct.pack("i", hparams["dim"]))
|
"ggml-model-{}.bin.{}".format(ftype, p))
|
||||||
fout.write(struct.pack("i", hparams["multiple_of"]))
|
|
||||||
fout.write(struct.pack("i", hparams["n_heads"]))
|
|
||||||
fout.write(struct.pack("i", hparams["n_layers"]))
|
|
||||||
fout.write(struct.pack("i", hparams["dim"] // hparams["n_heads"])) # rot (obsolete)
|
|
||||||
fout.write(struct.pack("i", ftype))
|
|
||||||
|
|
||||||
# Is this correct??
|
|
||||||
for i in range(tokenizer.vocab_size()):
|
|
||||||
if tokenizer.is_unknown(i):
|
|
||||||
# "<unk>" token (translated as ??)
|
|
||||||
text = " \u2047 ".encode("utf-8")
|
|
||||||
fout.write(struct.pack("i", len(text)))
|
|
||||||
fout.write(text)
|
|
||||||
elif tokenizer.is_control(i):
|
|
||||||
# "<s>"/"</s>" tokens
|
|
||||||
fout.write(struct.pack("i", 0))
|
|
||||||
elif tokenizer.is_byte(i):
|
|
||||||
# "<U+XX>" tokens (which may be invalid UTF-8)
|
|
||||||
piece = tokenizer.id_to_piece(i)
|
|
||||||
if len(piece) != 6:
|
|
||||||
print("Invalid token: " + piece)
|
|
||||||
sys.exit(1)
|
|
||||||
byte_value = int(piece[3:-1], 16)
|
|
||||||
fout.write(struct.pack("i", 1))
|
|
||||||
fout.write(struct.pack("B", byte_value))
|
|
||||||
else:
|
else:
|
||||||
# normal token. Uses U+2581 (LOWER ONE EIGHTH BLOCK) to represent spaces.
|
fname_out = os.path.join(out_dir,
|
||||||
text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode("utf-8")
|
"ggml-model-{}.bin".format(ftype))
|
||||||
fout.write(struct.pack("i", len(text)))
|
|
||||||
fout.write(text)
|
|
||||||
|
|
||||||
for k, v in model.items():
|
model = torch.load(fname_model, map_location="cpu")
|
||||||
name = k
|
|
||||||
shape = v.shape
|
|
||||||
|
|
||||||
# skip layers.X.attention.inner_attention.rope.freqs
|
fout = open(fname_out, "wb")
|
||||||
if name[-5:] == "freqs":
|
|
||||||
continue
|
|
||||||
|
|
||||||
print("Processing variable: " + name + " with shape: ", shape, " and type: ", v.dtype)
|
fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
|
||||||
|
fout.write(struct.pack("i", hparams["vocab_size"]))
|
||||||
|
fout.write(struct.pack("i", hparams["dim"]))
|
||||||
|
fout.write(struct.pack("i", hparams["multiple_of"]))
|
||||||
|
fout.write(struct.pack("i", hparams["n_heads"]))
|
||||||
|
fout.write(struct.pack("i", hparams["n_layers"]))
|
||||||
|
fout.write(struct.pack("i", hparams["dim"] //
|
||||||
|
hparams["n_heads"])) # rot (obsolete)
|
||||||
|
fout.write(struct.pack("i", ftype_int[ftype]))
|
||||||
|
|
||||||
#data = tf.train.load_variable(dir_model, name).squeeze()
|
# Is this correct??
|
||||||
data = v.numpy().squeeze()
|
for i in range(tokenizer.vocab_size()):
|
||||||
n_dims = len(data.shape);
|
if tokenizer.is_unknown(i):
|
||||||
|
# "<unk>" token (translated as ??)
|
||||||
|
text = " \u2047 ".encode("utf-8")
|
||||||
|
fout.write(struct.pack("i", len(text)))
|
||||||
|
fout.write(text)
|
||||||
|
elif tokenizer.is_control(i):
|
||||||
|
# "<s>"/"</s>" tokens
|
||||||
|
fout.write(struct.pack("i", 0))
|
||||||
|
elif tokenizer.is_byte(i):
|
||||||
|
# "<U+XX>" tokens (which may be invalid UTF-8)
|
||||||
|
piece = tokenizer.id_to_piece(i)
|
||||||
|
if len(piece) != 6:
|
||||||
|
print("Invalid token: " + piece)
|
||||||
|
sys.exit(1)
|
||||||
|
byte_value = int(piece[3:-1], 16)
|
||||||
|
fout.write(struct.pack("i", 1))
|
||||||
|
fout.write(struct.pack("B", byte_value))
|
||||||
|
else:
|
||||||
|
# normal token. Uses U+2581 (LOWER ONE EIGHTH BLOCK) to represent spaces.
|
||||||
|
text = tokenizer.id_to_piece(i).replace("\u2581",
|
||||||
|
" ").encode("utf-8")
|
||||||
|
fout.write(struct.pack("i", len(text)))
|
||||||
|
fout.write(text)
|
||||||
|
|
||||||
# for efficiency - transpose some matrices
|
for k, v in model.items():
|
||||||
# "model/h.*/attn/c_attn/w"
|
name = k
|
||||||
# "model/h.*/attn/c_proj/w"
|
shape = v.shape
|
||||||
# "model/h.*/mlp/c_fc/w"
|
|
||||||
# "model/h.*/mlp/c_proj/w"
|
|
||||||
#if name[-14:] == "/attn/c_attn/w" or \
|
|
||||||
# name[-14:] == "/attn/c_proj/w" or \
|
|
||||||
# name[-11:] == "/mlp/c_fc/w" or \
|
|
||||||
# name[-13:] == "/mlp/c_proj/w":
|
|
||||||
# print(" Transposing")
|
|
||||||
# data = data.transpose()
|
|
||||||
|
|
||||||
dshape = data.shape
|
# skip layers.X.attention.inner_attention.rope.freqs
|
||||||
|
if name[-5:] == "freqs":
|
||||||
|
continue
|
||||||
|
|
||||||
# default type is fp16
|
print("Processing variable: " + name + " with shape: ", shape,
|
||||||
ftype_cur = 1
|
" and type: ", v.dtype)
|
||||||
if ftype == 0 or n_dims == 1:
|
|
||||||
print(" Converting to float32")
|
|
||||||
data = data.astype(np.float32)
|
|
||||||
ftype_cur = 0
|
|
||||||
|
|
||||||
# header
|
#data = tf.train.load_variable(dir_model, name).squeeze()
|
||||||
sname = name.encode('utf-8')
|
data = v.numpy().squeeze()
|
||||||
fout.write(struct.pack("iii", n_dims, len(sname), ftype_cur))
|
n_dims = len(data.shape)
|
||||||
for i in range(n_dims):
|
|
||||||
fout.write(struct.pack("i", dshape[n_dims - 1 - i]))
|
|
||||||
fout.write(sname);
|
|
||||||
|
|
||||||
# data
|
# for efficiency - transpose some matrices
|
||||||
data.tofile(fout)
|
# "model/h.*/attn/c_attn/w"
|
||||||
|
# "model/h.*/attn/c_proj/w"
|
||||||
|
# "model/h.*/mlp/c_fc/w"
|
||||||
|
# "model/h.*/mlp/c_proj/w"
|
||||||
|
#if name[-14:] == "/attn/c_attn/w" or \
|
||||||
|
# name[-14:] == "/attn/c_proj/w" or \
|
||||||
|
# name[-11:] == "/mlp/c_fc/w" or \
|
||||||
|
# name[-13:] == "/mlp/c_proj/w":
|
||||||
|
# print(" Transposing")
|
||||||
|
# data = data.transpose()
|
||||||
|
|
||||||
# I hope this deallocates the memory ..
|
dshape = data.shape
|
||||||
model = None
|
|
||||||
|
|
||||||
fout.close()
|
# default type is fp16
|
||||||
|
ftype_cur = 1
|
||||||
|
if ftype == 'f32' or n_dims == 1:
|
||||||
|
print(" Converting to float32")
|
||||||
|
data = data.astype(np.float32)
|
||||||
|
ftype_cur = 0
|
||||||
|
|
||||||
print("Done. Output file: " + fname_out + ", (part ", p, ")")
|
# header
|
||||||
print("")
|
sname = name.encode('utf-8')
|
||||||
|
fout.write(struct.pack("iii", n_dims, len(sname), ftype_cur))
|
||||||
|
for i in range(n_dims):
|
||||||
|
fout.write(struct.pack("i", dshape[n_dims - 1 - i]))
|
||||||
|
fout.write(sname)
|
||||||
|
|
||||||
|
# data
|
||||||
|
data.tofile(fout)
|
||||||
|
|
||||||
|
# I hope this deallocates the memory ..
|
||||||
|
model = None
|
||||||
|
|
||||||
|
fout.close()
|
||||||
|
|
||||||
|
print("Done. Output file: " + fname_out + ", (part ", p, ")")
|
||||||
|
print("")
|
||||||
|
|
||||||
|
|
||||||
|
if __name__ == '__main__':
|
||||||
|
main()
|
||||||
|
|
|
||||||
Loading…
Add table
Add a link
Reference in a new issue