Ali-Forootani
commited on
Commit
•
062d453
1
Parent(s):
ed27633
Update README.md
Browse files
README.md
CHANGED
@@ -37,8 +37,8 @@ Once it's installed, we can import the necessary libraries and log in to W&B (op
|
|
37 |
|
38 |
"""
|
39 |
wandb
|
40 |
-
https://wandb.ai/
|
41 |
-
you need wb_token
|
42 |
"""
|
43 |
|
44 |
import gc
|
@@ -233,7 +233,180 @@ Training the model on these 1,000 samples and 20 epochs took about 22 hours on a
|
|
233 |
|
234 |
|
235 |
|
236 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
237 |
|
238 |
<!-- Provide a longer summary of what this model is. -->
|
239 |
|
|
|
37 |
|
38 |
"""
|
39 |
wandb
|
40 |
+
https://wandb.ai/wandb_account
|
41 |
+
you need wb_token as well
|
42 |
"""
|
43 |
|
44 |
import gc
|
|
|
233 |
|
234 |
|
235 |
|
236 |
+
|
237 |
+
|
238 |
+
## Test the model
|
239 |
+
|
240 |
+
# -*- coding: utf-8 -*-
|
241 |
+
"""
|
242 |
+
Created on Wed Jul 3 15:57:22 2024
|
243 |
+
|
244 |
+
@author: Ali forootani
|
245 |
+
"""
|
246 |
+
|
247 |
+
"""
|
248 |
+
!pip install -U transformers datasets accelerate peft trl bitsandbytes wandb
|
249 |
+
!pip install -qqq flash-attn
|
250 |
+
!pip install -qU transformers accelerate
|
251 |
+
"""
|
252 |
+
|
253 |
+
"""
|
254 |
+
wandb
|
255 |
+
https://wandb.ai/your_account
|
256 |
+
dde689e74d3f9146d2d116b098016f5e0d9cc202
|
257 |
+
"""
|
258 |
+
|
259 |
+
|
260 |
+
```python
|
261 |
+
import gc
|
262 |
+
import os
|
263 |
+
|
264 |
+
import torch
|
265 |
+
import wandb
|
266 |
+
from datasets import load_dataset
|
267 |
+
|
268 |
+
|
269 |
+
|
270 |
+
# Directly insert your Weights & Biases API key here
|
271 |
+
wb_token = 'your_wb_token'
|
272 |
+
wandb.login(key=wb_token)
|
273 |
+
|
274 |
+
|
275 |
+
from peft import LoraConfig, PeftModel, prepare_model_for_kbit_training
|
276 |
+
|
277 |
+
from transformers import (
|
278 |
+
AutoModelForCausalLM,
|
279 |
+
AutoTokenizer,
|
280 |
+
BitsAndBytesConfig,
|
281 |
+
TrainingArguments,
|
282 |
+
pipeline,)
|
283 |
+
|
284 |
+
from trl import ORPOConfig, ORPOTrainer, setup_chat_format
|
285 |
+
|
286 |
+
|
287 |
+
|
288 |
+
|
289 |
+
"""
|
290 |
+
https://huggingface.co/blog/mlabonne/orpo-llama-3
|
291 |
+
|
292 |
+
mlabonne/orpo-dpo-mix-40k
|
293 |
+
|
294 |
+
https://huggingface.co/datasets/mlabonne/orpo-dpo-mix-40k/tree/main
|
295 |
+
"""
|
296 |
+
|
297 |
+
if torch.cuda.get_device_capability()[0] >= 128:
|
298 |
+
|
299 |
+
attn_implementation = "flash_attention_2"
|
300 |
+
torch_dtype = torch.bfloat16
|
301 |
+
else:
|
302 |
+
attn_implementation = "eager"
|
303 |
+
torch_dtype = torch.float16
|
304 |
+
|
305 |
+
|
306 |
+
##################################
|
307 |
+
|
308 |
+
import sys
|
309 |
+
import os
|
310 |
+
|
311 |
+
cwd = os.getcwd()
|
312 |
+
# sys.path.append(cwd + '/my_directory')
|
313 |
+
sys.path.append(cwd)
|
314 |
+
|
315 |
+
|
316 |
+
def setting_directory(depth):
|
317 |
+
current_dir = os.path.abspath(os.getcwd())
|
318 |
+
root_dir = current_dir
|
319 |
+
for i in range(depth):
|
320 |
+
root_dir = os.path.abspath(os.path.join(root_dir, os.pardir))
|
321 |
+
sys.path.append(os.path.dirname(root_dir))
|
322 |
+
return root_dir
|
323 |
+
|
324 |
+
|
325 |
+
model_path = "/data/bio-eng-llm/llm_repo/mlabonne/OrpoLlama-3-8B"
|
326 |
+
|
327 |
+
|
328 |
+
###################################
|
329 |
+
###################################
|
330 |
+
|
331 |
+
"""
|
332 |
+
# Model
|
333 |
+
base_model = "meta-llama/Meta-Llama-3-8B"
|
334 |
+
new_model = "OrpoLlama-3-8B"
|
335 |
+
"""
|
336 |
+
|
337 |
+
|
338 |
+
# QLoRA config
|
339 |
+
bnb_config = BitsAndBytesConfig(
|
340 |
+
load_in_4bit=True,
|
341 |
+
bnb_4bit_quant_type="nf4",
|
342 |
+
bnb_4bit_compute_dtype= torch_dtype,
|
343 |
+
bnb_4bit_use_double_quant=True,
|
344 |
+
)
|
345 |
+
|
346 |
+
# LoRA config
|
347 |
+
peft_config = LoraConfig(
|
348 |
+
r=16,
|
349 |
+
lora_alpha=32,
|
350 |
+
lora_dropout=0.05,
|
351 |
+
bias="none",
|
352 |
+
task_type="CAUSAL_LM",
|
353 |
+
target_modules=['up_proj', 'down_proj', 'gate_proj', 'k_proj', 'q_proj', 'v_proj', 'o_proj']
|
354 |
+
)
|
355 |
+
|
356 |
+
|
357 |
+
# Reload tokenizer and model
|
358 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path)
|
359 |
+
model = AutoModelForCausalLM.from_pretrained(
|
360 |
+
model_path,
|
361 |
+
low_cpu_mem_usage=True,
|
362 |
+
return_dict=True,
|
363 |
+
torch_dtype=torch.float16,
|
364 |
+
device_map="auto",
|
365 |
+
)
|
366 |
+
model, tokenizer = setup_chat_format(model, tokenizer)
|
367 |
+
|
368 |
+
|
369 |
+
|
370 |
+
root_dir = setting_directory(0)
|
371 |
+
epochs = 20
|
372 |
+
|
373 |
+
new_model_path = root_dir + f"models/fine_tuned_models/OrpoLlama-3-8B_{epochs}e_qa_qa"
|
374 |
+
|
375 |
+
|
376 |
+
### Merge adapter with base model
|
377 |
+
model = PeftModel.from_pretrained(model, new_model_path)
|
378 |
+
model = model.merge_and_unload()
|
379 |
+
|
380 |
+
print("#############################")
|
381 |
+
print("#############################")
|
382 |
+
print(model)
|
383 |
+
|
384 |
+
|
385 |
+
|
386 |
+
|
387 |
+
# Pushing the model into the Huggingface hub
|
388 |
+
|
389 |
+
from huggingface_hub import HfApi, login
|
390 |
+
|
391 |
+
#########################################################
|
392 |
+
#########################################################
|
393 |
+
#########################################################
|
394 |
+
######## Repo token
|
395 |
+
# Login to Hugging Face
|
396 |
+
login(token="your_huggingface_token")
|
397 |
+
|
398 |
+
# Define your Hugging Face repository name
|
399 |
+
repo_name = "your_name/OrpoLlama-3-8B_fine_tune_trl"
|
400 |
+
|
401 |
+
|
402 |
+
|
403 |
+
# Push the model and tokenizer 2
|
404 |
+
model.push_to_hub(repo_name, use_auth_token=True)
|
405 |
+
tokenizer.push_to_hub(repo_name, use_auth_token=True)
|
406 |
+
```
|
407 |
+
|
408 |
+
|
409 |
+
|
410 |
|
411 |
<!-- Provide a longer summary of what this model is. -->
|
412 |
|