Upload prune.py
Browse filesUpload pruning script used to create Jamba-Small-v1
prune.py
ADDED
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import re
|
3 |
+
|
4 |
+
import torch
|
5 |
+
from modeling_jamba import JambaForCausalLM
|
6 |
+
|
7 |
+
model = JambaForCausalLM.from_pretrained("ai21labs/Jamba-v0.1", device_map="cpu", torch_dtype=torch.bfloat16)
|
8 |
+
|
9 |
+
def prune_and_copy_additional_layers(original_state_dict):
|
10 |
+
layer_mapping = {
|
11 |
+
0: 0,
|
12 |
+
1: 1,
|
13 |
+
2: 2,
|
14 |
+
3: 2,
|
15 |
+
4: 4,
|
16 |
+
5: 5,
|
17 |
+
6: 30,
|
18 |
+
7: 31
|
19 |
+
}
|
20 |
+
|
21 |
+
new_state_dict = {}
|
22 |
+
|
23 |
+
# Copy specified layers from the original state dict
|
24 |
+
for new_idx, orig_idx in layer_mapping.items():
|
25 |
+
prefix = f"model.layers.{orig_idx}"
|
26 |
+
for key, value in original_state_dict.items():
|
27 |
+
if key.startswith(prefix):
|
28 |
+
new_key = key.replace(f"layers.{orig_idx}", f"layers.{new_idx}")
|
29 |
+
new_state_dict[new_key] = value
|
30 |
+
|
31 |
+
global_keys = ['model.embed_tokens.weight', 'model.final_layernorm.weight', 'lm_head.weight']
|
32 |
+
for key in global_keys:
|
33 |
+
new_state_dict[key] = original_state_dict[key]
|
34 |
+
|
35 |
+
return new_state_dict
|
36 |
+
|
37 |
+
pruned_state_dict = prune_and_copy_additional_layers(model.state_dict())
|
38 |
+
|
39 |
+
print("Saving weights...")
|
40 |
+
torch.save(pruned_state_dict, '/scratch/nbrown9/jamba-small-v2.bin')
|
41 |
+
print("Done!")
|