Update README.md
Browse files
README.md
CHANGED
@@ -12,4 +12,53 @@ The original Llama 3 8b (base) special token weights are zero, which might cause
|
|
12 |
<|end_header_id|>
|
13 |
```
|
14 |
|
15 |
-
We set the weights of these tokens in `embed` and `lm_head` to be the mean of all other tokens.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
<|end_header_id|>
|
13 |
```
|
14 |
|
15 |
+
We set the weights of these tokens in `embed` and `lm_head` to be the mean of all other tokens.
|
16 |
+
|
17 |
+
Code for making this model:
|
18 |
+
|
19 |
+
```python
|
20 |
+
import argparse
|
21 |
+
|
22 |
+
import transformers
|
23 |
+
import torch
|
24 |
+
|
25 |
+
|
26 |
+
def init_eot_embedding_llama3(model_path, output_dir, special_tokens=["<|eot_id|>", "<|start_header_id|>", "<|end_header_id|>"], mean_cutoff=128000, dtype=torch.bfloat16):
|
27 |
+
tokenizer = transformers.AutoTokenizer.from_pretrained(model_path)
|
28 |
+
model = transformers.AutoModelForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, torch_dtype=dtype)
|
29 |
+
|
30 |
+
assert model.model.embed_tokens.weight.shape[0] >= mean_cutoff
|
31 |
+
assert model.lm_head.weight.shape[0] >= mean_cutoff
|
32 |
+
|
33 |
+
with torch.no_grad():
|
34 |
+
for token in special_tokens:
|
35 |
+
token_id = tokenizer.convert_tokens_to_ids(token)
|
36 |
+
|
37 |
+
print (f"Token {token} ID {token_id}")
|
38 |
+
|
39 |
+
model.model.embed_tokens.weight[token_id] = torch.mean(model.model.embed_tokens.weight[:mean_cutoff].to(torch.float32), dim=0).to(dtype)
|
40 |
+
model.lm_head.weight[token_id] = torch.mean(model.lm_head.weight[:mean_cutoff].to(torch.float32), dim=0).to(dtype)
|
41 |
+
|
42 |
+
# Save
|
43 |
+
tokenizer.save_pretrained(output_dir)
|
44 |
+
model.save_pretrained(output_dir)
|
45 |
+
|
46 |
+
|
47 |
+
def main():
|
48 |
+
parser = argparse.ArgumentParser()
|
49 |
+
parser.add_argument(
|
50 |
+
"--model-path",
|
51 |
+
help="Location of model, or HuggingFace repo ID",
|
52 |
+
)
|
53 |
+
parser.add_argument(
|
54 |
+
"--output-dir",
|
55 |
+
help="Location to write resulting model and tokenizer",
|
56 |
+
)
|
57 |
+
|
58 |
+
init_eot_embedding_llama3(**vars(parser.parse_args()))
|
59 |
+
|
60 |
+
|
61 |
+
if __name__ == "__main__":
|
62 |
+
main()
|
63 |
+
|
64 |
+
```
|