Update A.py
Browse files
A.py
CHANGED
|
@@ -34,7 +34,7 @@ def resize_tensor_shapes(tensor1, tensor2):
|
|
| 34 |
|
| 35 |
return tensor1_resized, tensor2_resized
|
| 36 |
|
| 37 |
-
def merge_checkpoints(ckpt1, ckpt2, blend_ratio=0.
|
| 38 |
print(f"Merging checkpoints with blend ratio: {blend_ratio}")
|
| 39 |
merged = {}
|
| 40 |
all_keys = set(ckpt1.keys()).union(set(ckpt2.keys()))
|
|
@@ -49,7 +49,7 @@ def merge_checkpoints(ckpt1, ckpt2, blend_ratio=0.75):
|
|
| 49 |
else:
|
| 50 |
merged[key] = t2
|
| 51 |
|
| 52 |
-
# Control the final size to be
|
| 53 |
control_output_size(merged, target_size_gb=26)
|
| 54 |
|
| 55 |
return merged
|
|
@@ -64,14 +64,20 @@ def control_output_size(merged, target_size_gb):
|
|
| 64 |
excess_size = current_size_bytes - target_size_bytes
|
| 65 |
print(f"Current size exceeds target by {excess_size / (1024**2):.2f} MB. Adjusting...")
|
| 66 |
|
| 67 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
| 68 |
for key in merged.keys():
|
| 69 |
tensor = merged[key]
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
|
|
|
|
|
|
| 75 |
|
| 76 |
def cleanup_files(*file_paths):
|
| 77 |
for file_path in file_paths:
|
|
@@ -83,8 +89,8 @@ if __name__ == "__main__":
|
|
| 83 |
try:
|
| 84 |
model1_path = "mangledMergeFlux_v0Bfloat16Dev.safetensors"
|
| 85 |
model2_path = "output_checkpoint.safetensors"
|
| 86 |
-
blend_ratio = 0.
|
| 87 |
-
output_file = "
|
| 88 |
|
| 89 |
# Loading models
|
| 90 |
model1 = load_model(model1_path)
|
|
|
|
| 34 |
|
| 35 |
return tensor1_resized, tensor2_resized
|
| 36 |
|
| 37 |
+
def merge_checkpoints(ckpt1, ckpt2, blend_ratio=0.6):
|
| 38 |
print(f"Merging checkpoints with blend ratio: {blend_ratio}")
|
| 39 |
merged = {}
|
| 40 |
all_keys = set(ckpt1.keys()).union(set(ckpt2.keys()))
|
|
|
|
| 49 |
else:
|
| 50 |
merged[key] = t2
|
| 51 |
|
| 52 |
+
# Control the final size to be strictly 26 GB
|
| 53 |
control_output_size(merged, target_size_gb=26)
|
| 54 |
|
| 55 |
return merged
|
|
|
|
| 64 |
excess_size = current_size_bytes - target_size_bytes
|
| 65 |
print(f"Current size exceeds target by {excess_size / (1024**2):.2f} MB. Adjusting...")
|
| 66 |
|
| 67 |
+
# Calculate the total number of elements to reduce
|
| 68 |
+
elements_to_reduce = excess_size // 4 # Assuming 4 bytes per float32 tensor
|
| 69 |
+
total_elements = sum(tensor.numel() for tensor in merged.values())
|
| 70 |
+
|
| 71 |
+
# Distribute the reduction uniformly across all tensors
|
| 72 |
for key in merged.keys():
|
| 73 |
tensor = merged[key]
|
| 74 |
+
num_elements = tensor.numel()
|
| 75 |
+
# Calculate how much to reduce from this tensor
|
| 76 |
+
reduction = min(elements_to_reduce, num_elements)
|
| 77 |
+
merged[key] = tensor.flatten()[:num_elements - reduction].view(tensor.shape)
|
| 78 |
+
elements_to_reduce -= reduction
|
| 79 |
+
if elements_to_reduce <= 0:
|
| 80 |
+
break
|
| 81 |
|
| 82 |
def cleanup_files(*file_paths):
|
| 83 |
for file_path in file_paths:
|
|
|
|
| 89 |
try:
|
| 90 |
model1_path = "mangledMergeFlux_v0Bfloat16Dev.safetensors"
|
| 91 |
model2_path = "output_checkpoint.safetensors"
|
| 92 |
+
blend_ratio = 0.6 # Set to 60%
|
| 93 |
+
output_file = "output_checkpoint.safetensors"
|
| 94 |
|
| 95 |
# Loading models
|
| 96 |
model1 = load_model(model1_path)
|