delete file
Browse files- btlm_3b_convert.py +0 -242
btlm_3b_convert.py
DELETED
@@ -1,242 +0,0 @@
|
|
1 |
-
# -*- coding: utf-8 -*-
|
2 |
-
"""btlm_3b_convert.ipynb
|
3 |
-
|
4 |
-
Automatically generated by Colaboratory.
|
5 |
-
|
6 |
-
Original file is located at
|
7 |
-
https://colab.research.google.com/drive/1kuWDvfRlc0BplrityIIwpdf404mljhRK
|
8 |
-
"""
|
9 |
-
|
10 |
-
!pip install -U transformers
|
11 |
-
!pip install -U accelerate #git+https://github.com/huggingface/accelerate.git
|
12 |
-
!pip install -U bitsandbytes #git+ https://github.com/timdettmers/bitsandbytes.git
|
13 |
-
|
14 |
-
from transformers import AutoTokenizer, AutoModelForCausalLM
|
15 |
-
|
16 |
-
# Load the tokenizer and model
|
17 |
-
tokenizer = AutoTokenizer.from_pretrained("cerebras/btlm-3b-8k-base")
|
18 |
-
model = AutoModelForCausalLM.from_pretrained(
|
19 |
-
"cerebras/btlm-3b-8k-base",
|
20 |
-
trust_remote_code=True,
|
21 |
-
torch_dtype="auto",
|
22 |
-
load_in_8bit=True,
|
23 |
-
offload_folder="offload",
|
24 |
-
)
|
25 |
-
|
26 |
-
# Set the prompt for generating text
|
27 |
-
prompt = "Albert Einstein was known for "
|
28 |
-
|
29 |
-
# Tokenize the prompt and convert to PyTorch tensors
|
30 |
-
inputs = tokenizer(prompt, return_tensors="pt")
|
31 |
-
|
32 |
-
# Generate text using the model
|
33 |
-
outputs = model.generate(
|
34 |
-
**inputs,
|
35 |
-
num_beams=5,
|
36 |
-
max_new_tokens=50,
|
37 |
-
early_stopping=True,
|
38 |
-
no_repeat_ngram_size=2
|
39 |
-
)
|
40 |
-
|
41 |
-
# Convert the generated token IDs back to text
|
42 |
-
generated_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)
|
43 |
-
|
44 |
-
# Print the generated text
|
45 |
-
print(generated_text[0])
|
46 |
-
|
47 |
-
print(model)
|
48 |
-
|
49 |
-
list_vars = model.state_dict()
|
50 |
-
for name in list_vars.keys():
|
51 |
-
print(name, "=>", list_vars[name])
|
52 |
-
|
53 |
-
import sys
|
54 |
-
import os
|
55 |
-
import struct
|
56 |
-
import json
|
57 |
-
|
58 |
-
import torch
|
59 |
-
from transformers import AutoConfig
|
60 |
-
|
61 |
-
config = AutoConfig.from_pretrained("cerebras/btlm-3b-8k-base", trust_remote_code=True)
|
62 |
-
hparams = config.to_dict()
|
63 |
-
fname_out = "btlm-3b.ggml.bin"
|
64 |
-
|
65 |
-
print(json.dumps(hparams, indent=4, sort_keys=True))
|
66 |
-
|
67 |
-
import re
|
68 |
-
import numpy as np
|
69 |
-
import tensorflow as tf
|
70 |
-
|
71 |
-
bfloat16 = tf.bfloat16.as_numpy_dtype
|
72 |
-
|
73 |
-
|
74 |
-
fout = open(fname_out, "wb")
|
75 |
-
|
76 |
-
fout.write(struct.pack("i", 0x67676D6C))
|
77 |
-
fout.write(struct.pack("i", hparams["vocab_size"]))
|
78 |
-
fout.write(struct.pack("i", hparams["n_positions"]))
|
79 |
-
fout.write(struct.pack("i", hparams["n_embd"]))
|
80 |
-
fout.write(struct.pack("i", hparams["n_head"]))
|
81 |
-
fout.write(struct.pack("i", hparams["n_layer"]))
|
82 |
-
fout.write(struct.pack("i", hparams["n_inner"]))
|
83 |
-
fout.write(struct.pack("i", 1))
|
84 |
-
|
85 |
-
for i in range(hparams["vocab_size"]):
|
86 |
-
text = tokenizer.decode([i]).encode('utf-8')
|
87 |
-
fout.write(struct.pack("i", len(text)))
|
88 |
-
fout.write(text)
|
89 |
-
|
90 |
-
|
91 |
-
# for name in list_vars.keys():
|
92 |
-
# print(name, "=>", list_vars[name])
|
93 |
-
|
94 |
-
|
95 |
-
for name in list_vars.keys():
|
96 |
-
if name[-14:] == ".weight_format":
|
97 |
-
print("FOUND " + name)
|
98 |
-
continue
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
print("Processing variable: " + name)
|
104 |
-
data = list_vars[name].squeeze().cpu().type(dtype=torch.float16).numpy()
|
105 |
-
print(" with shape: ", data.shape)
|
106 |
-
|
107 |
-
# rename headers to keep compatibility
|
108 |
-
if name == "transformer.ln_f.weight":
|
109 |
-
name = "model/ln_f/g"
|
110 |
-
elif name == "transformer.ln_f.bias":
|
111 |
-
name = "model/ln_f/b"
|
112 |
-
elif name == "transformer.wte.weight":
|
113 |
-
name = "model/wte"
|
114 |
-
elif name == "transformer.wpe.weight":
|
115 |
-
name = "model/wpe"
|
116 |
-
elif name == "lm_head.weight":
|
117 |
-
name = "model/lm_head"
|
118 |
-
elif name == "transformer.relative_pe.slopes":
|
119 |
-
name = "model/relative_pe/slopes"
|
120 |
-
elif re.match(r"transformer.h\.\d+\.ln_1\.weight", name):
|
121 |
-
i = re.findall("\d+", name)[0]
|
122 |
-
name = f"model/h{i}/ln_1/g"
|
123 |
-
elif re.match(r"transformer.h\.\d+\.ln_1\.bias", name):
|
124 |
-
i = re.findall("\d+", name)[0]
|
125 |
-
name = f"model/h{i}/ln_1/b"
|
126 |
-
elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.weight", name):
|
127 |
-
i = re.findall("\d+", name)[0]
|
128 |
-
name = f"model/h{i}/attn/c_attn/w"
|
129 |
-
elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.bias", name):
|
130 |
-
i = re.findall("\d+", name)[0]
|
131 |
-
name = f"model/h{i}/attn/c_attn/b"
|
132 |
-
elif re.match(r"transformer.h\.\d+\.attn\.c_proj\.weight", name):
|
133 |
-
i = re.findall("\d+", name)[0]
|
134 |
-
name = f"model/h{i}/attn/c_proj/w"
|
135 |
-
elif re.match(r"transformer.h.\d+.attn.c_proj.bias", name):
|
136 |
-
i = re.findall("\d+", name)[0]
|
137 |
-
name = f"model/h{i}/attn/c_proj/b"
|
138 |
-
elif re.match(r"transformer.h.\d+.ln_2.weight", name):
|
139 |
-
i = re.findall("\d+", name)[0]
|
140 |
-
name = f"model/h{i}/ln_2/g"
|
141 |
-
elif re.match(r"transformer.h.\d+.ln_2.bias", name):
|
142 |
-
i = re.findall("\d+", name)[0]
|
143 |
-
name = f"model/h{i}/ln_2/b"
|
144 |
-
elif re.match(r"transformer.h.\d+.mlp.c_fc.weight", name):
|
145 |
-
i = re.findall("\d+", name)[0]
|
146 |
-
name = f"model/h{i}/mlp/c_fc/w"
|
147 |
-
elif re.match(r"transformer.h.\d+.mlp.c_fc.bias", name):
|
148 |
-
i = re.findall("\d+", name)[0]
|
149 |
-
name = f"model/h{i}/mlp/c_fc/b"
|
150 |
-
elif re.match(r"transformer.h.\d+.mlp.c_proj.weight", name):
|
151 |
-
i = re.findall("\d+", name)[0]
|
152 |
-
name = f"model/h{i}/mlp/c_proj/w"
|
153 |
-
elif re.match(r"transformer.h.\d+.mlp.c_proj.bias", name):
|
154 |
-
i = re.findall("\d+", name)[0]
|
155 |
-
name = f"model/h{i}/mlp/c_proj/b"
|
156 |
-
# NEW
|
157 |
-
elif re.match(r"transformer.h.\d+.attn.c_proj.SCB", name):
|
158 |
-
i = re.findall("\d+", name)[0]
|
159 |
-
name = f"model/h{i}/attn/c_proj/scb"
|
160 |
-
elif re.match(r"transformer.h.\d+.attn.c_attn.SCB", name):
|
161 |
-
i = re.findall("\d+", name)[0]
|
162 |
-
name = f"model/h{i}/attn/c_attn/scb"
|
163 |
-
elif re.match(r"transformer.h.\d+.mlp.c_fc.SCB", name):
|
164 |
-
i = re.findall("\d+", name)[0]
|
165 |
-
name = f"model/h{i}/mlp/c_fc/scb"
|
166 |
-
elif re.match(r"transformer.h.\d+.mlp.c_fc2.weight", name):
|
167 |
-
i = re.findall("\d+", name)[0]
|
168 |
-
name = f"model/h{i}/mlp/c_fc2/weight"
|
169 |
-
elif re.match(r"transformer.h.\d+.mlp.c_fc2.bias", name):
|
170 |
-
i = re.findall("\d+", name)[0]
|
171 |
-
name = f"model/h{i}/mlp/c_fc2/bias"
|
172 |
-
elif re.match(r"transformer.h.\d+.mlp.c_fc2.SCB", name):
|
173 |
-
i = re.findall("\d+", name)[0]
|
174 |
-
name = f"model/h{i}/mlp/c_fc2/scb"
|
175 |
-
elif re.match(r"transformer.h.\d+.mlp.c_proj.SCB", name):
|
176 |
-
i = re.findall("\d+", name)[0]
|
177 |
-
name = f"model/h{i}/mlp/c_proj/scb"
|
178 |
-
|
179 |
-
else:
|
180 |
-
print("Unrecognized variable name. %s", name)
|
181 |
-
|
182 |
-
|
183 |
-
# we don't need these
|
184 |
-
if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
|
185 |
-
print(" Skipping variable: " + name)
|
186 |
-
continue
|
187 |
-
|
188 |
-
n_dims = len(data.shape);
|
189 |
-
|
190 |
-
# ftype == 0 -> float32, ftype == 1 -> float16
|
191 |
-
ftype = 0;
|
192 |
-
if True: #use_b16
|
193 |
-
if (name == "model/wte" or name == "model/lm_head" or name[-2:] == "/g" or name[-2:] == "/w") and n_dims == 2:
|
194 |
-
print(" Converting to float16")
|
195 |
-
data = data.astype(np.float16)
|
196 |
-
ftype = 1
|
197 |
-
else:
|
198 |
-
print(" Converting to float32")
|
199 |
-
data = data.astype(np.float32)
|
200 |
-
ftype = 0
|
201 |
-
|
202 |
-
# for efficiency - transpose the projection matrices
|
203 |
-
# "model/h.*/attn/c_attn/w"
|
204 |
-
# "model/h.*/attn/c_proj/w"
|
205 |
-
# "model/h.*/mlp/c_fc/w"
|
206 |
-
# "model/h.*/mlp/c_proj/w"
|
207 |
-
if name[-14:] == "/attn/c_attn/w" or \
|
208 |
-
name[-14:] == "/attn/c_proj/w" or \
|
209 |
-
name[-11:] == "/mlp/c_fc/w" or \
|
210 |
-
name[-13:] == "/mlp/c_proj/w":
|
211 |
-
print(" Transposing")
|
212 |
-
data = data.transpose()
|
213 |
-
|
214 |
-
# header
|
215 |
-
str = name.encode('utf-8')
|
216 |
-
fout.write(struct.pack("iii", n_dims, len(str), ftype))
|
217 |
-
for i in range(n_dims):
|
218 |
-
fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
|
219 |
-
fout.write(str);
|
220 |
-
|
221 |
-
# data
|
222 |
-
data.tofile(fout)
|
223 |
-
|
224 |
-
fout.close()
|
225 |
-
|
226 |
-
print("Done. Output file: " + fname_out)
|
227 |
-
print("")
|
228 |
-
|
229 |
-
|
230 |
-
# write_binary()
|
231 |
-
|
232 |
-
from huggingface_hub import login, HfApi
|
233 |
-
|
234 |
-
login()
|
235 |
-
|
236 |
-
api = HfApi()
|
237 |
-
|
238 |
-
api.upload_folder(
|
239 |
-
folder_path="/content/btlm-3b-ggml",
|
240 |
-
repo_id="bornjre/btlm-3b-ggml",
|
241 |
-
repo_type="model",
|
242 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|