Spaces:
Running
Running
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,394 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gc
|
2 |
+
import gradio as gr
|
3 |
+
import torch
|
4 |
+
from huggingface_hub import hf_hub_download, HfApi, login, list_repo_files
|
5 |
+
from safetensors import safe_open
|
6 |
+
from safetensors.torch import save_file, load_file
|
7 |
+
import os
|
8 |
+
import shutil
|
9 |
+
import json
|
10 |
+
|
11 |
+
api = HfApi()
|
12 |
+
|
13 |
+
def info_fn(text):
|
14 |
+
gr.Info(text)
|
15 |
+
|
16 |
+
def warning_fn(text):
|
17 |
+
gr.Warning(text)
|
18 |
+
|
19 |
+
def load_lora_state(lora_model_name):
|
20 |
+
"""Download and load LoRA adapter weights"""
|
21 |
+
temp_lora_dir = "/tmp/lora_adapter"
|
22 |
+
os.makedirs(temp_lora_dir, exist_ok=True)
|
23 |
+
|
24 |
+
# Download adapter config
|
25 |
+
config_path = hf_hub_download(
|
26 |
+
repo_id=lora_model_name,
|
27 |
+
filename="adapter_config.json",
|
28 |
+
local_dir=temp_lora_dir,
|
29 |
+
local_dir_use_symlinks=False
|
30 |
+
)
|
31 |
+
|
32 |
+
with open(config_path, 'r') as f:
|
33 |
+
lora_config = json.load(f)
|
34 |
+
|
35 |
+
scale = lora_config['lora_alpha'] / lora_config['r']
|
36 |
+
|
37 |
+
# Download adapter weights
|
38 |
+
try:
|
39 |
+
adapter_path = hf_hub_download(
|
40 |
+
repo_id=lora_model_name,
|
41 |
+
filename="adapter_model.safetensors",
|
42 |
+
local_dir=temp_lora_dir,
|
43 |
+
local_dir_use_symlinks=False
|
44 |
+
)
|
45 |
+
lora_state = load_file(adapter_path, device='cpu')
|
46 |
+
except:
|
47 |
+
adapter_path = hf_hub_download(
|
48 |
+
repo_id=lora_model_name,
|
49 |
+
filename="adapter_model.bin",
|
50 |
+
local_dir=temp_lora_dir,
|
51 |
+
local_dir_use_symlinks=False
|
52 |
+
)
|
53 |
+
lora_state = torch.load(adapter_path, map_location='cpu')
|
54 |
+
|
55 |
+
return lora_state, scale, temp_lora_dir
|
56 |
+
|
57 |
+
def find_lora_weights(lora_state, key):
|
58 |
+
"""Find corresponding LoRA A and B weights for a given key"""
|
59 |
+
lora_A = None
|
60 |
+
lora_B = None
|
61 |
+
|
62 |
+
# Remove .weight suffix and handle potential prefixes
|
63 |
+
clean_key = key.replace('.weight', '')
|
64 |
+
|
65 |
+
for lora_key, lora_weight in lora_state.items():
|
66 |
+
if clean_key in lora_key or clean_key.replace('language_model.', '') in lora_key:
|
67 |
+
if 'lora_A' in lora_key:
|
68 |
+
lora_A = lora_weight
|
69 |
+
elif 'lora_B' in lora_key:
|
70 |
+
lora_B = lora_weight
|
71 |
+
|
72 |
+
# Both should be None or both should have values
|
73 |
+
if (lora_A is None) != (lora_B is None):
|
74 |
+
return None, None
|
75 |
+
|
76 |
+
return lora_A, lora_B
|
77 |
+
|
78 |
+
def download_and_upload_non_model_files(base_model_name, output_repo_name):
|
79 |
+
"""Download and upload non-model files (config, tokenizer, etc.)"""
|
80 |
+
temp_config_dir = "/tmp/config_files"
|
81 |
+
os.makedirs(temp_config_dir, exist_ok=True)
|
82 |
+
|
83 |
+
try:
|
84 |
+
# List all files in the repository
|
85 |
+
files = list_repo_files(repo_id=base_model_name)
|
86 |
+
|
87 |
+
# Filter non-model files
|
88 |
+
non_model_files = [
|
89 |
+
f for f in files
|
90 |
+
if not (f.startswith('model') and f.endswith('.safetensors'))
|
91 |
+
]
|
92 |
+
|
93 |
+
# Download and upload each non-model file
|
94 |
+
for filename in non_model_files:
|
95 |
+
if filename.endswith(('.gguf', '.bin')) and 'model' in filename:
|
96 |
+
continue # Skip other model formats
|
97 |
+
|
98 |
+
try:
|
99 |
+
file_path = hf_hub_download(
|
100 |
+
repo_id=base_model_name,
|
101 |
+
filename=filename,
|
102 |
+
local_dir=temp_config_dir,
|
103 |
+
local_dir_use_symlinks=False
|
104 |
+
)
|
105 |
+
|
106 |
+
# Upload to output repo
|
107 |
+
api.upload_file(
|
108 |
+
path_or_fileobj=file_path,
|
109 |
+
path_in_repo=filename,
|
110 |
+
repo_id=output_repo_name,
|
111 |
+
repo_type="model"
|
112 |
+
)
|
113 |
+
|
114 |
+
except Exception as e:
|
115 |
+
info_fn(f"Skipping {filename}: {e}")
|
116 |
+
|
117 |
+
finally:
|
118 |
+
shutil.rmtree(temp_config_dir, ignore_errors=True)
|
119 |
+
|
120 |
+
def merge_lora_efficient(hf_token, base_model_name, lora_model_name, output_repo_name,
|
121 |
+
lora_scale, lm_head_scale, multiplicative_lora, progress=gr.Progress()):
|
122 |
+
temp_lora_dir = None
|
123 |
+
try:
|
124 |
+
login(hf_token)
|
125 |
+
|
126 |
+
progress(0.1, desc="Loading LoRA adapter...")
|
127 |
+
info_fn("Loading LoRA adapter...")
|
128 |
+
|
129 |
+
# Load LoRA state (this downloads the adapter)
|
130 |
+
lora_state, base_scale, temp_lora_dir = load_lora_state(lora_model_name)
|
131 |
+
|
132 |
+
# Apply LoRA scale multiplier
|
133 |
+
scale = base_scale * lora_scale
|
134 |
+
info_fn(f"Using LoRA scale: {scale} (base: {base_scale}, multiplier: {lora_scale})")
|
135 |
+
|
136 |
+
progress(0.2, desc="Creating output repository...")
|
137 |
+
|
138 |
+
# Create repository
|
139 |
+
try:
|
140 |
+
repo_url = api.create_repo(repo_id=output_repo_name, exist_ok=True)
|
141 |
+
info_fn(f"Repository created/updated: {repo_url}")
|
142 |
+
except Exception as e:
|
143 |
+
warning_fn(f"Repository might already exist: {e}")
|
144 |
+
|
145 |
+
progress(0.3, desc="Uploading configuration files...")
|
146 |
+
info_fn("Uploading configuration files...")
|
147 |
+
|
148 |
+
# Download and upload non-model files
|
149 |
+
download_and_upload_non_model_files(base_model_name, output_repo_name)
|
150 |
+
|
151 |
+
progress(0.4, desc="Finding model shards...")
|
152 |
+
info_fn("Finding model shards...")
|
153 |
+
|
154 |
+
# Get list of all safetensors files
|
155 |
+
all_files = list_repo_files(repo_id=base_model_name)
|
156 |
+
shard_files = [f for f in all_files if f.startswith('model') and f.endswith('.safetensors')]
|
157 |
+
|
158 |
+
if not shard_files:
|
159 |
+
raise FileNotFoundError("No model safetensors files found in the repository")
|
160 |
+
|
161 |
+
info_fn(f"Found {len(shard_files)} model shards to process")
|
162 |
+
|
163 |
+
merged_tensors = 0
|
164 |
+
scaled_lm_heads = 0
|
165 |
+
total_shards = len(shard_files)
|
166 |
+
|
167 |
+
# Process each shard individually
|
168 |
+
for i, shard_filename in enumerate(shard_files):
|
169 |
+
progress(0.4 + (i / total_shards) * 0.5,
|
170 |
+
desc=f"Processing {shard_filename} ({i+1}/{total_shards})")
|
171 |
+
info_fn(f"Processing shard {i+1}/{total_shards}: {shard_filename}")
|
172 |
+
|
173 |
+
# Create temporary directory for this shard only
|
174 |
+
temp_shard_dir = f"/tmp/shard_{i}"
|
175 |
+
os.makedirs(temp_shard_dir, exist_ok=True)
|
176 |
+
|
177 |
+
try:
|
178 |
+
# Download the current shard
|
179 |
+
shard_path = hf_hub_download(
|
180 |
+
repo_id=base_model_name,
|
181 |
+
filename=shard_filename,
|
182 |
+
local_dir=temp_shard_dir,
|
183 |
+
local_dir_use_symlinks=False
|
184 |
+
)
|
185 |
+
|
186 |
+
# Process the shard
|
187 |
+
tensors = {}
|
188 |
+
shard_merged_count = 0
|
189 |
+
shard_lm_head_count = 0
|
190 |
+
|
191 |
+
with safe_open(shard_path, framework='pt', device='cpu') as f:
|
192 |
+
# Get metadata if available
|
193 |
+
metadata = f.metadata() if hasattr(f, 'metadata') else {}
|
194 |
+
|
195 |
+
for key in f.keys():
|
196 |
+
tensor = f.get_tensor(key)
|
197 |
+
|
198 |
+
# Apply lm_head scaling if applicable
|
199 |
+
if key.endswith('lm_head.weight') and lm_head_scale != 1.0:
|
200 |
+
info_fn(f"Scaling {key} by {lm_head_scale}")
|
201 |
+
original_dtype = tensor.dtype
|
202 |
+
tensor = tensor.to(torch.float32)
|
203 |
+
tensor = tensor * lm_head_scale
|
204 |
+
tensor = tensor.to(original_dtype)
|
205 |
+
shard_lm_head_count += 1
|
206 |
+
scaled_lm_heads += 1
|
207 |
+
|
208 |
+
# Try to find corresponding LoRA weights
|
209 |
+
lora_A, lora_B = find_lora_weights(lora_state, key)
|
210 |
+
|
211 |
+
if lora_A is not None and lora_B is not None:
|
212 |
+
lora_type = "Multiplicative" if multiplicative_lora else "Additive"
|
213 |
+
info_fn(f"Merging {lora_type} LoRA weights for {key}")
|
214 |
+
shard_merged_count += 1
|
215 |
+
merged_tensors += 1
|
216 |
+
|
217 |
+
# Convert to float32 for computation
|
218 |
+
original_dtype = tensor.dtype
|
219 |
+
tensor_f32 = tensor.to(torch.float32)
|
220 |
+
lora_A_f32 = lora_A.to(torch.float32)
|
221 |
+
lora_B_f32 = lora_B.to(torch.float32)
|
222 |
+
|
223 |
+
if multiplicative_lora:
|
224 |
+
# Apply Multiplicative-LoRA: W = W + scale * B @ A @ W
|
225 |
+
tensor_f32 += scale * lora_B_f32 @ lora_A_f32 @ tensor_f32
|
226 |
+
else:
|
227 |
+
# Apply standard LoRA: W = W + scale * B @ A
|
228 |
+
tensor_f32 += scale * lora_B_f32 @ lora_A_f32
|
229 |
+
|
230 |
+
# Convert back to original dtype
|
231 |
+
tensor = tensor_f32.to(original_dtype)
|
232 |
+
|
233 |
+
# Clean up intermediate tensors
|
234 |
+
del tensor_f32, lora_A_f32, lora_B_f32
|
235 |
+
if torch.cuda.is_available():
|
236 |
+
torch.cuda.empty_cache()
|
237 |
+
|
238 |
+
tensors[key] = tensor
|
239 |
+
|
240 |
+
# Save processed shard to temporary file
|
241 |
+
output_shard_path = os.path.join(temp_shard_dir, f"processed_{shard_filename}")
|
242 |
+
save_file(tensors, output_shard_path, metadata=metadata)
|
243 |
+
|
244 |
+
info_fn(f"Shard {shard_filename}:\n- Merged {shard_merged_count} tensors\n- Scaled {shard_lm_head_count} lm_head tensors")
|
245 |
+
|
246 |
+
# Upload the processed shard
|
247 |
+
api.upload_file(
|
248 |
+
path_or_fileobj=output_shard_path,
|
249 |
+
path_in_repo=shard_filename,
|
250 |
+
repo_id=output_repo_name,
|
251 |
+
repo_type="model"
|
252 |
+
)
|
253 |
+
|
254 |
+
# Clean up this shard's data
|
255 |
+
del tensors
|
256 |
+
gc.collect()
|
257 |
+
|
258 |
+
finally:
|
259 |
+
# Always clean up the temporary shard directory
|
260 |
+
shutil.rmtree(temp_shard_dir, ignore_errors=True)
|
261 |
+
|
262 |
+
progress(1.0, desc="Upload completed!")
|
263 |
+
|
264 |
+
success_msg = f"β Successfully merged and uploaded model!\nModel URL: https://huggingface.co/{output_repo_name}\nProcessed {total_shards} shards\nMerged {merged_tensors} layers with LoRA weights\nScaled {scaled_lm_heads} lm_head layers"
|
265 |
+
info_fn("Merge completed successfully!")
|
266 |
+
|
267 |
+
return success_msg
|
268 |
+
|
269 |
+
except Exception as e:
|
270 |
+
error_msg = f"β Error during merge: {str(e)}"
|
271 |
+
warning_fn(error_msg)
|
272 |
+
return error_msg
|
273 |
+
|
274 |
+
finally:
|
275 |
+
# Cleanup LoRA directory
|
276 |
+
if temp_lora_dir and os.path.exists(temp_lora_dir):
|
277 |
+
shutil.rmtree(temp_lora_dir, ignore_errors=True)
|
278 |
+
gc.collect()
|
279 |
+
|
280 |
+
INTRODUCTION_TEXT = """
|
281 |
+
## Memory-Efficient LoRA Merge
|
282 |
+
|
283 |
+
This tool merges LoRA (Low-Rank Adaptation) adapters with base models using a memory-efficient approach that processes model files individually, significantly reducing memory requirements compared to traditional methods.
|
284 |
+
|
285 |
+
### Key Features
|
286 |
+
- **Minimal Memory Usage**: Processes one model shard at a time instead of loading the entire model
|
287 |
+
- **Streaming Processing**: Downloads β Processes β Uploads β Deletes each shard sequentially
|
288 |
+
- **Automatic Cleanup**: Temporary files are automatically removed after processing
|
289 |
+
- **Progress Tracking**: Real-time status updates throughout the merge process
|
290 |
+
- **Advanced Options**: Configurable LoRA scaling, LM head scaling, and multiplicative LoRA support
|
291 |
+
|
292 |
+
### How It Works
|
293 |
+
LoRA enables efficient fine-tuning by adding small adapter weights rather than modifying the entire model. This tool applies the LoRA transformation with configurable scaling:
|
294 |
+
|
295 |
+
- **Standard Additive-LoRA**: `W_new = W + scale Γ B^T @ A`
|
296 |
+
- **Multiplicative LoRA**: `W_new = W + scale Γ B^T @ A @ W`
|
297 |
+
|
298 |
+
Additionally, the model's default temperature behavior can be adjusted by scaling the `lm_head.weight` tensor:
|
299 |
+
|
300 |
+
- **Up-scaling**: Makes the model's outputs more peaked, requiring lower temperature settings for the same output distribution
|
301 |
+
- **Down-scaling**: Makes the model's outputs flatter, requiring higher temperature settings for the same output distribution
|
302 |
+
- **Examples**:
|
303 |
+
- Scaling `lm_head.weight` by `1.25` makes the new model with `temperature = 1.0` act like the old model with `temperature = 0.8`
|
304 |
+
- Scaling `lm_head.weight` by `0.667` makes the new model with `temperature = 1.0` act like the old model with `temperature = 1.5`
|
305 |
+
|
306 |
+
### Memory Efficiency
|
307 |
+
- **Traditional approach**: Loads entire model (~15GB+ for 7B parameter models)
|
308 |
+
- **This approach**: Peak usage determined by largest shard size, not total model size
|
309 |
+
- **Result**: Enables merging of much larger models on limited hardware
|
310 |
+
|
311 |
+
### Example Usage
|
312 |
+
- **Base Model:** `microsoft/DialoGPT-medium`
|
313 |
+
- **LoRA Adapter:** `username/my-trained-lora`
|
314 |
+
- **Output Name:** `username/dialogpt-merged`
|
315 |
+
|
316 |
+
### Attribution
|
317 |
+
This tool builds upon excellent work from the community:
|
318 |
+
|
319 |
+
- **Base implementation:** [Weyaxi/merge-lora](https://huggingface.co/spaces/Weyaxi/merge-lora)
|
320 |
+
- **Memory-efficient method:** [qlora-pipe](https://github.com/tdrussell/qlora-pipe/blob/main/tools/merge_lora.py) by tdrussell
|
321 |
+
"""
|
322 |
+
|
323 |
+
with gr.Blocks(title="Memory-Efficient LoRA Merge", theme=gr.themes.Soft()) as demo:
|
324 |
+
gr.Markdown(INTRODUCTION_TEXT)
|
325 |
+
|
326 |
+
with gr.Row():
|
327 |
+
with gr.Column(scale=1):
|
328 |
+
gr.Markdown("### Configuration")
|
329 |
+
hf_token = gr.Textbox(
|
330 |
+
label="Hugging Face Token",
|
331 |
+
placeholder="hf_...",
|
332 |
+
type="password",
|
333 |
+
info="Token with write access to create repositories"
|
334 |
+
)
|
335 |
+
base_model_name = gr.Textbox(
|
336 |
+
label="Base Model Repository",
|
337 |
+
placeholder="microsoft/DialoGPT-medium",
|
338 |
+
info="The original model to merge LoRA into"
|
339 |
+
)
|
340 |
+
lora_model_name = gr.Textbox(
|
341 |
+
label="LoRA Adapter Repository",
|
342 |
+
placeholder="username/my-lora-adapter",
|
343 |
+
info="Repository containing adapter_model.safetensors"
|
344 |
+
)
|
345 |
+
output_repo_name = gr.Textbox(
|
346 |
+
label="Output Repository Name",
|
347 |
+
placeholder="username/my-merged-model",
|
348 |
+
info="Name for the new merged model repository"
|
349 |
+
)
|
350 |
+
|
351 |
+
gr.Markdown("### Advanced Options")
|
352 |
+
lora_scale = gr.Number(
|
353 |
+
label="LoRA Scale",
|
354 |
+
value=1.0,
|
355 |
+
minimum=0.0,
|
356 |
+
maximum=10.0,
|
357 |
+
step=0.1,
|
358 |
+
info="Multiplier for LoRA strength (1.0 = default)"
|
359 |
+
)
|
360 |
+
lm_head_scale = gr.Number(
|
361 |
+
label="LM Head Scale",
|
362 |
+
value=1.0,
|
363 |
+
minimum=0.1,
|
364 |
+
maximum=5.0,
|
365 |
+
step=0.05,
|
366 |
+
info="Multiplier for lm_head weights (1.0 = default)"
|
367 |
+
)
|
368 |
+
multiplicative_lora = gr.Checkbox(
|
369 |
+
label="Multiplicative LoRA",
|
370 |
+
value=False,
|
371 |
+
info="Apply a \"multiplicative-LoRA\" instead of a standard \"additive-LoRA\""
|
372 |
+
)
|
373 |
+
|
374 |
+
with gr.Column(scale=1):
|
375 |
+
gr.Markdown("### Status")
|
376 |
+
output_text = gr.Textbox(
|
377 |
+
label="Merge Progress & Results",
|
378 |
+
lines=20,
|
379 |
+
interactive=False,
|
380 |
+
show_copy_button=True
|
381 |
+
)
|
382 |
+
|
383 |
+
with gr.Row():
|
384 |
+
submit_btn = gr.Button("Start LoRA Merge", variant="primary", size="lg")
|
385 |
+
|
386 |
+
submit_btn.click(
|
387 |
+
fn=merge_lora_efficient,
|
388 |
+
inputs=[hf_token, base_model_name, lora_model_name, output_repo_name,
|
389 |
+
lora_scale, lm_head_scale, multiplicative_lora],
|
390 |
+
outputs=output_text
|
391 |
+
)
|
392 |
+
|
393 |
+
demo.queue()
|
394 |
+
demo.launch(show_error=True)
|