-
Notifications
You must be signed in to change notification settings - Fork 0
/
convert-pth-to-ggml.py
177 lines (147 loc) · 5.29 KB
/
convert-pth-to-ggml.py
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
# Convert a LLaMA model checkpoint to a ggml compatible file
#
# Load the model using Torch
# Iterate over all variables and write them to a binary file.
#
# For each variable, write the following:
# - Number of dimensions (int)
# - Name length (int)
# - Dimensions (int[n_dims])
# - Name (char[name_length])
# - Data (float[n_dims])
#
# By default, the bigger matrices are converted to 16-bit floats.
# This can be disabled by adding the "use-f32" CLI argument.
#
# At the start of the ggml file we write the model parameters
# and vocabulary.
#
import sys
import json
import struct
import numpy as np
import torch
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
dir_model = sys.argv[1]
fname_hparams = sys.argv[1] + "/params.json"
fname_tokenizer = sys.argv[1] + "/../tokenizer.model"
def get_n_parts(dim):
if dim == 4096:
return 1
elif dim == 5120:
return 2
elif dim == 6656:
return 4
elif dim == 8192:
return 8
else:
print("Invalid dim: " + str(dim))
sys.exit(1)
# possible data types
# ftype == 0 -> float32
# ftype == 1 -> float16
#
# map from ftype to string
ftype_str = ["f32", "f16"]
ftype = 1
if len(sys.argv) > 2:
ftype = int(sys.argv[2])
if ftype < 0 or ftype > 1:
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:
hparams = json.load(f)
tokenizer = SentencePieceProcessor(fname_tokenizer)
hparams.update({"vocab_size": tokenizer.vocab_size()})
n_parts = get_n_parts(hparams["dim"])
print(hparams)
print('n_parts = ', n_parts)
for p in range(n_parts):
print('Processing part ', p)
#fname_model = sys.argv[1] + "/consolidated.00.pth"
fname_model = sys.argv[1] + "/consolidated.0" + str(p) + ".pth"
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")
fout = open(fname_out, "wb")
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))
# 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:
# 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 k, v in model.items():
name = k
shape = v.shape
# skip layers.X.attention.inner_attention.rope.freqs
if name[-5:] == "freqs":
continue
print("Processing variable: " + name + " with shape: ", shape, " and type: ", v.dtype)
#data = tf.train.load_variable(dir_model, name).squeeze()
data = v.numpy().squeeze()
n_dims = len(data.shape);
# for efficiency - transpose some matrices
# "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()
dshape = data.shape
# default type is fp16
ftype_cur = 1
if ftype == 0 or n_dims == 1:
print(" Converting to float32")
data = data.astype(np.float32)
ftype_cur = 0
# header
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("")