-
Notifications
You must be signed in to change notification settings - Fork 32
/
Copy pathexport.py
199 lines (143 loc) · 10.6 KB
/
export.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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
# Export safetensor model to lm.rs format
import torch
import struct
import argparse
import json
import re
import gc
import numpy as np
from contextlib import ExitStack
from safetensors import safe_open
from utils.io import write_tensors_by_group
if __name__ == "__main__":
model_types = ["GEMMA", "LLAMA", "PHI"]
parser = argparse.ArgumentParser(description="Export safetensors model to lm.rs format.")
parser.add_argument('--files', type=str, nargs='+', required=True, help='a list of safetensor file paths')
parser.add_argument('--config', type=str, required=True, help='path of the config file of the model')
parser.add_argument('--save-path', type=str, required=True, help='path of the output model file')
parser.add_argument('--quantize', action='store_true', default=False, help='use quantization')
parser.add_argument('--quantize-type', type=int, default=1, help='type of quantization - 1 for Q8_0, 2 for Q4_0')
parser.add_argument('--group-size', type=int, default=128, help='groups to use in quantization')
parser.add_argument('--type', type=str, required=True, choices=model_types, help='model type')
parser.add_argument('--vision-config', type=str, required=False, help='path to the vision model config')
version = 4
args = parser.parse_args()
if args.vision_config and args.type != "PHI":
print("Error: --multimodal can only be used when --type is PHI.")
sys.exit(1)
if args.quantize:
assert args.quantize_type == 1 or args.quantize_type == 2
quantize_type = args.quantize_type if args.quantize else 0 # 0 for no quantization, 1 for Q8_0, 2 for Q4_0
group_size = args.group_size
ew = []
model_type = model_types.index(args.type)
multimodal = 1 if args.vision_config else 0
with open(args.config, 'r') as file:
cfg = json.load(file)
out_file = open(f"{args.save_path}.lmrs", 'wb', buffering=False)
#lmrs
out_file.write(struct.pack('I', 0x73726d6c))
out_file.write(struct.pack('I', version))
head_dim = cfg["head_dim"] if "head_dim" in cfg else cfg["hidden_size"] // cfg["num_attention_heads"]
header = struct.pack('IIIIIIIIff', cfg["hidden_size"], cfg["intermediate_size"], cfg["num_hidden_layers"], cfg["num_attention_heads"], head_dim,
cfg["num_key_value_heads"], cfg["vocab_size"], cfg["max_position_embeddings"], cfg["rms_norm_eps"], cfg["rope_theta"])
out_file.write(header)
types = struct.pack('BB', quantize_type, model_type)
out_file.write(types)
with ExitStack() as stack:
files = [stack.enter_context(safe_open(file_path, framework="pt", device="cpu")) for file_path in args.files]
if args.quantize:
dim = files[0].get_tensor("model.embed_tokens.weight").shape[1]
while dim % group_size != 0:
group_size //= 2
print(f"BACKOFF: reducing group size to {group_size} to fit hidden_dim")
out_file.write(struct.pack('I', group_size))
out_file.write(struct.pack('B', multimodal))
pad = 256 - out_file.tell()
assert pad >= 0
out_file.write(b'\0' * pad)
# Embedding table / cls layer weights
ew.extend(write_tensors_by_group(files, "model.embed_tokens.weight", out_file, m_type="", quantize_type=quantize_type))
# Attention weights
write_tensors_by_group(files, "input_layernorm", out_file)
if args.type == "PHI":
ew.extend(write_tensors_by_group(files, "self_attn.qkv_proj", out_file, quantize_type=quantize_type, splits=3, split_idx=0))
ew.extend(write_tensors_by_group(files, "self_attn.qkv_proj", out_file, quantize_type=quantize_type, splits=3, split_idx=1))
ew.extend(write_tensors_by_group(files, "self_attn.qkv_proj", out_file, quantize_type=quantize_type, splits=3, split_idx=2))
else:
ew.extend(write_tensors_by_group(files, "self_attn.q_proj", out_file, quantize_type=quantize_type))
ew.extend(write_tensors_by_group(files, "self_attn.k_proj", out_file, quantize_type=quantize_type))
ew.extend(write_tensors_by_group(files, "self_attn.v_proj", out_file, quantize_type=quantize_type))
ew.extend(write_tensors_by_group(files, "self_attn.o_proj", out_file, quantize_type=quantize_type))
# FFN weights
write_tensors_by_group(files, "post_attention_layernorm", out_file)
if args.type == "GEMMA":
write_tensors_by_group(files, "pre_feedforward_layernorm", out_file)
if args.type == "PHI":
ew.extend(write_tensors_by_group(files, "mlp.gate_up_proj", out_file, quantize_type=quantize_type, splits=2, split_idx=0))
ew.extend(write_tensors_by_group(files, "mlp.down_proj", out_file, quantize_type=quantize_type))
ew.extend(write_tensors_by_group(files, "mlp.gate_up_proj", out_file, quantize_type=quantize_type, splits=2, split_idx=1))
else:
ew.extend(write_tensors_by_group(files, "mlp.gate_proj", out_file, quantize_type=quantize_type))
ew.extend(write_tensors_by_group(files, "mlp.down_proj", out_file, quantize_type=quantize_type))
ew.extend(write_tensors_by_group(files, "mlp.up_proj", out_file, quantize_type=quantize_type))
if args.type == "GEMMA":
write_tensors_by_group(files, "post_feedforward_layernorm", out_file)
# Final norm weights
write_tensors_by_group(files, "model.norm.weight", out_file, m_type="")
if args.type == "PHI":
ew.extend(write_tensors_by_group(files, "lm_head.weight", out_file, m_type="", quantize_type=quantize_type))
if args.vision_config:
with open(args.vision_config, 'r') as file:
vision_cfg = json.load(file)
prev_pos = out_file.tell()
vision_head_dim = vision_cfg["vision_config"]["hidden_size"] // vision_cfg["vision_config"]["num_attention_heads"]
vision_header = struct.pack('IIIIIfII', vision_cfg["vision_config"]["hidden_size"], vision_cfg["vision_config"]["intermediate_size"],
vision_cfg["vision_config"]["num_hidden_layers"], vision_cfg["vision_config"]["num_attention_heads"], vision_head_dim, vision_cfg["vision_config"]["layer_norm_eps"],
vision_cfg["vision_config"]["patch_size"], vision_cfg["vision_config"]["image_size"])
out_file.write(vision_header)
out_file.write(struct.pack('B', quantize_type))
out_file.write(struct.pack('I', group_size))
pad = 128 - (out_file.tell() - prev_pos)
assert pad >= 0
out_file.write(b'\0' * pad)
# CLIP encoder
ew.extend(write_tensors_by_group(files, "class_embedding", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "patch_embedding.weight", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "position_embedding.weight", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "layer_norm1.weight", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "layer_norm1.bias", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "layer_norm2.weight", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "layer_norm2.bias", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "self_attn.q_proj.weight", out_file, m_type="model.vision_embed_tokens", quantize_type=quantize_type))
ew.extend(write_tensors_by_group(files, "self_attn.q_proj.bias", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "self_attn.k_proj.weight", out_file, m_type="model.vision_embed_tokens", quantize_type=quantize_type))
ew.extend(write_tensors_by_group(files, "self_attn.k_proj.bias", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "self_attn.v_proj.weight", out_file, m_type="model.vision_embed_tokens", quantize_type=quantize_type))
ew.extend(write_tensors_by_group(files, "self_attn.v_proj.bias", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "self_attn.out_proj.weight", out_file, m_type="model.vision_embed_tokens", quantize_type=quantize_type))
ew.extend(write_tensors_by_group(files, "self_attn.out_proj.bias", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "mlp.fc1.weight", out_file, m_type="model.vision_embed_tokens", quantize_type=quantize_type))
ew.extend(write_tensors_by_group(files, "mlp.fc1.bias", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "mlp.fc2.weight", out_file, m_type="model.vision_embed_tokens", quantize_type=quantize_type))
ew.extend(write_tensors_by_group(files, "mlp.fc2.bias", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "pre_layrnorm.weight", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "pre_layrnorm.bias", out_file, m_type="model.vision_embed_tokens"))
# Processor
prev_pos = out_file.tell()
processor_header = struct.pack('II', vision_cfg["vision_config"]["intermediate_size"], cfg["hidden_size"])
out_file.write(processor_header)
out_file.write(struct.pack('B', quantize_type))
out_file.write(struct.pack('I', group_size))
pad = 128 - (out_file.tell() - prev_pos)
assert pad >= 0
out_file.write(b'\0' * pad)
ew.extend(write_tensors_by_group(files, "glb_GN", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "sub_GN", out_file, m_type="model.vision_embed_tokens"))
ew.extend(write_tensors_by_group(files, "img_projection", out_file, m_type="weight", quantize_type=quantize_type))
ew.extend(write_tensors_by_group(files, "img_projection", out_file, m_type="bias"))
if args.quantize:
ew.sort(reverse=True)
print(f"Max quantization group error across all weights: {ew[0]}. Mean: {sum(ew)/len(ew)}.")
print(f"Successfully converted {args.type} model to lmrs format.")
out_file.close()