model: add merge code
Browse files
merge.py
ADDED
|
@@ -0,0 +1,123 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import argparse
|
| 3 |
+
from safetensors import safe_open
|
| 4 |
+
from safetensors.torch import save_file
|
| 5 |
+
import json
|
| 6 |
+
from tqdm import tqdm
|
| 7 |
+
|
| 8 |
+
def get_tensor_locations(input_dir):
|
| 9 |
+
tensor_locations = {}
|
| 10 |
+
for i in tqdm(range(1, 52), desc="Scanning input files"): # 51 splits
|
| 11 |
+
file_path = os.path.join(input_dir, f"model-{i:05d}-of-00051.safetensors")
|
| 12 |
+
with safe_open(file_path, framework="pt", device="cpu") as f:
|
| 13 |
+
for key in f.keys():
|
| 14 |
+
tensor_locations[key] = i
|
| 15 |
+
return tensor_locations
|
| 16 |
+
|
| 17 |
+
def create_merge_plan(tensor_locations, layer_config):
|
| 18 |
+
merge_plan = []
|
| 19 |
+
new_layer_idx = 0
|
| 20 |
+
new_file_idx = 1
|
| 21 |
+
|
| 22 |
+
# Special handling for specific weights
|
| 23 |
+
special_weights = {
|
| 24 |
+
"model.embed_tokens.weight": 1,
|
| 25 |
+
"lm_head.weight": 48,
|
| 26 |
+
"model.norm.weight": 48
|
| 27 |
+
}
|
| 28 |
+
|
| 29 |
+
for slice_config in layer_config:
|
| 30 |
+
start, end = slice_config['layer_range']
|
| 31 |
+
for i in range(start, end):
|
| 32 |
+
layer_tensors = []
|
| 33 |
+
for key in tensor_locations.keys():
|
| 34 |
+
if key.startswith(f"model.layers.{i}."):
|
| 35 |
+
new_key = key.replace(f"model.layers.{i}", f"model.layers.{new_layer_idx}")
|
| 36 |
+
layer_tensors.append({
|
| 37 |
+
'old_key': key,
|
| 38 |
+
'new_key': new_key,
|
| 39 |
+
'original_file_index': tensor_locations[key],
|
| 40 |
+
'new_file_index': new_file_idx
|
| 41 |
+
})
|
| 42 |
+
if layer_tensors:
|
| 43 |
+
merge_plan.extend(layer_tensors)
|
| 44 |
+
new_file_idx += 1
|
| 45 |
+
new_layer_idx += 1
|
| 46 |
+
|
| 47 |
+
# Add special weights to their original locations
|
| 48 |
+
for key, file_index in special_weights.items():
|
| 49 |
+
merge_plan.append({
|
| 50 |
+
'old_key': key,
|
| 51 |
+
'new_key': key,
|
| 52 |
+
'original_file_index': file_index,
|
| 53 |
+
'new_file_index': file_index
|
| 54 |
+
})
|
| 55 |
+
|
| 56 |
+
# Add any remaining non-layer tensors to the first file
|
| 57 |
+
for key, file_index in tensor_locations.items():
|
| 58 |
+
if not key.startswith("model.layers.") and key not in special_weights:
|
| 59 |
+
merge_plan.append({
|
| 60 |
+
'old_key': key,
|
| 61 |
+
'new_key': key,
|
| 62 |
+
'original_file_index': file_index,
|
| 63 |
+
'new_file_index': 1
|
| 64 |
+
})
|
| 65 |
+
|
| 66 |
+
return merge_plan
|
| 67 |
+
|
| 68 |
+
def merge_layers(input_dir, output_dir, merge_plan):
|
| 69 |
+
output_tensors = {}
|
| 70 |
+
current_new_file_index = 1
|
| 71 |
+
max_file_index = max(item['new_file_index'] for item in merge_plan)
|
| 72 |
+
|
| 73 |
+
with tqdm(total=len(merge_plan), desc="Merging layers") as pbar:
|
| 74 |
+
for file_index in range(1, max_file_index + 1):
|
| 75 |
+
for item in merge_plan:
|
| 76 |
+
if item['new_file_index'] == file_index:
|
| 77 |
+
input_file = os.path.join(input_dir, f"model-{item['original_file_index']:05d}-of-00051.safetensors")
|
| 78 |
+
with safe_open(input_file, framework="pt", device="cpu") as f:
|
| 79 |
+
tensor = f.get_tensor(item['old_key'])
|
| 80 |
+
output_tensors[item['new_key']] = tensor
|
| 81 |
+
pbar.update(1)
|
| 82 |
+
|
| 83 |
+
if output_tensors:
|
| 84 |
+
output_file = os.path.join(output_dir, f"model-{file_index:05d}-of-{max_file_index:05d}.safetensors")
|
| 85 |
+
save_file(output_tensors, output_file)
|
| 86 |
+
output_tensors = {}
|
| 87 |
+
|
| 88 |
+
print(f"Merged model saved to {output_dir}")
|
| 89 |
+
|
| 90 |
+
def main():
|
| 91 |
+
parser = argparse.ArgumentParser(description="Merge and split Mistral model")
|
| 92 |
+
parser.add_argument("input_dir", help="Directory containing input safetensors files")
|
| 93 |
+
parser.add_argument("output_dir", help="Directory for output safetensors files")
|
| 94 |
+
parser.add_argument("--dry-run", action="store_true", help="Perform a dry run and output merge plan")
|
| 95 |
+
args = parser.parse_args()
|
| 96 |
+
|
| 97 |
+
layer_config = [
|
| 98 |
+
{'layer_range': [0, 20]},
|
| 99 |
+
{'layer_range': [10, 30]},
|
| 100 |
+
{'layer_range': [20, 40]},
|
| 101 |
+
{'layer_range': [30, 50]},
|
| 102 |
+
{'layer_range': [40, 60]},
|
| 103 |
+
{'layer_range': [50, 70]},
|
| 104 |
+
{'layer_range': [60, 80]},
|
| 105 |
+
{'layer_range': [70, 87]}
|
| 106 |
+
]
|
| 107 |
+
|
| 108 |
+
tensor_locations = get_tensor_locations(args.input_dir)
|
| 109 |
+
merge_plan = create_merge_plan(tensor_locations, layer_config)
|
| 110 |
+
|
| 111 |
+
if args.dry_run:
|
| 112 |
+
print("Merge plan:")
|
| 113 |
+
print(json.dumps(merge_plan, indent=2))
|
| 114 |
+
with open("merge_plan_large.json", "w") as f:
|
| 115 |
+
json.dump(merge_plan, f, indent=2)
|
| 116 |
+
print("Merge plan saved to merge_plan.json")
|
| 117 |
+
else:
|
| 118 |
+
os.makedirs(args.output_dir, exist_ok=True)
|
| 119 |
+
merge_layers(args.input_dir, args.output_dir, merge_plan)
|
| 120 |
+
print(f"Merged model saved to {args.output_dir}")
|
| 121 |
+
|
| 122 |
+
if __name__ == "__main__":
|
| 123 |
+
main()
|