Spaces:
Sleeping
Sleeping
Upload folder using huggingface_hub
Browse files- run_transformers_training.py +825 -622
run_transformers_training.py
CHANGED
@@ -184,227 +184,291 @@ def load_configs(base_path):
|
|
184 |
raise
|
185 |
|
186 |
def parse_args():
|
187 |
-
|
188 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
189 |
return parser.parse_args()
|
190 |
|
191 |
def load_model_and_tokenizer(config):
|
192 |
-
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
193 |
try:
|
194 |
-
if
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
|
|
205 |
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
|
219 |
-
|
220 |
-
|
221 |
-
|
222 |
-
|
223 |
-
|
224 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
225 |
else:
|
226 |
-
#
|
227 |
-
|
228 |
-
|
229 |
-
|
230 |
-
# Check for flash attention as the last dependency check
|
231 |
-
use_flash_attention = config.get("use_flash_attention", True)
|
232 |
-
if use_flash_attention and not find_spec("flash_attn"):
|
233 |
-
logger.warning("flash-attn not found. Will continue without flash attention.")
|
234 |
-
logger.warning("To use flash attention, install with: pip install flash-attn --no-build-isolation")
|
235 |
-
use_flash_attention = False
|
236 |
-
|
237 |
-
# Set device map based on config or default to "auto"
|
238 |
-
device_map = config.get("hardware", {}).get("hardware_setup", {}).get("device_map", "auto")
|
239 |
-
|
240 |
-
# Calculate max memory settings if multiple GPUs are available
|
241 |
-
max_memory = None
|
242 |
-
if gpu_count > 1:
|
243 |
-
memory_per_gpu = config.get("hardware", {}).get("specs", {}).get("vram_per_gpu", 24)
|
244 |
-
max_memory = {i: f"{int(memory_per_gpu * 0.85)}GiB" for i in range(gpu_count)}
|
245 |
-
max_memory["cpu"] = "64GiB" # Allow CPU offloading if needed
|
246 |
-
|
247 |
-
# Load model with proper error handling for out-of-memory
|
248 |
-
try:
|
249 |
-
# Improved memory settings for multi-GPU setup
|
250 |
-
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
|
251 |
|
252 |
-
|
253 |
-
|
254 |
-
|
255 |
-
|
256 |
-
|
257 |
-
|
258 |
-
|
259 |
)
|
260 |
-
except RuntimeError as e:
|
261 |
-
if "CUDA out of memory" in str(e):
|
262 |
-
logger.error("Out of GPU memory. Consider using a smaller batch size or gradient accumulation steps.")
|
263 |
-
raise
|
264 |
-
else:
|
265 |
-
# Try again with CPU placement to see if it's a memory issue
|
266 |
-
logger.warning(f"Error loading model on default device: {str(e)}")
|
267 |
-
logger.warning("Attempting to load with device_map='cpu' and no specific dtype")
|
268 |
-
model, tokenizer = FastLanguageModel.from_pretrained(
|
269 |
-
model_name=model_name,
|
270 |
-
max_seq_length=config.get("max_seq_length", 2048) or config.get("tokenizer", {}).get("max_seq_length", 2048),
|
271 |
-
dtype=None,
|
272 |
-
device_map={"": "cpu"},
|
273 |
-
)
|
274 |
-
logger.warning("Model loaded on CPU. Training will be very slow.")
|
275 |
-
|
276 |
-
# Ensure model and optimizer init is on the same device
|
277 |
-
logger.info(f"Model device map: {model.hf_device_map if hasattr(model, 'hf_device_map') else 'Not available'}")
|
278 |
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
# Apply optimizations
|
290 |
-
model = FastLanguageModel.get_peft_model(
|
291 |
-
model,
|
292 |
-
r=unsloth_config.get("r", 32),
|
293 |
-
target_modules=unsloth_config.get("target_modules",
|
294 |
-
["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]),
|
295 |
-
lora_alpha=unsloth_config.get("alpha", 16),
|
296 |
-
lora_dropout=lora_dropout, # Using the value from config or default
|
297 |
-
bias="none",
|
298 |
-
use_gradient_checkpointing=config.get("gradient_checkpointing", True) or config.get("training", {}).get("gradient_checkpointing", True),
|
299 |
-
random_state=config.get("seed", 42),
|
300 |
-
)
|
301 |
-
logger.info("Unsloth optimizations applied successfully")
|
302 |
-
|
303 |
-
# Set up tokenizer settings
|
304 |
-
chat_template = config.get("chat_template") or config.get("tokenizer", {}).get("chat_template")
|
305 |
-
if chat_template:
|
306 |
-
try:
|
307 |
-
# Get the correct chat template for phi models
|
308 |
-
template = get_chat_template("phi")
|
309 |
-
# Correctly apply the template to the tokenizer (it's a string)
|
310 |
-
if isinstance(template, str):
|
311 |
-
tokenizer.chat_template = template
|
312 |
-
logger.info("Set phi chat template (string)")
|
313 |
else:
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
320 |
|
321 |
-
#
|
322 |
-
|
323 |
-
|
324 |
-
|
|
|
|
|
|
|
|
|
325 |
|
326 |
return model, tokenizer
|
327 |
-
|
328 |
except Exception as e:
|
329 |
-
|
330 |
-
|
331 |
-
|
332 |
|
333 |
-
def load_dataset_with_mapping(
|
334 |
-
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
335 |
try:
|
336 |
-
# Load dataset
|
337 |
-
|
338 |
-
dataset_split = dataset_config.get("dataset", {}).get("split", "train")
|
339 |
|
340 |
-
if
|
341 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
342 |
|
343 |
-
|
344 |
|
345 |
-
|
346 |
-
|
|
|
|
|
|
|
|
|
|
|
347 |
|
348 |
-
#
|
349 |
-
if
|
350 |
-
|
351 |
-
|
352 |
-
|
353 |
-
|
354 |
-
|
355 |
-
|
356 |
-
|
357 |
-
|
358 |
-
|
359 |
-
|
360 |
-
|
361 |
-
|
362 |
-
|
363 |
-
|
364 |
-
|
365 |
-
|
366 |
-
|
367 |
-
|
368 |
-
|
369 |
-
|
370 |
-
|
371 |
-
|
372 |
-
|
373 |
-
|
374 |
-
|
375 |
-
|
376 |
-
|
377 |
-
|
378 |
-
|
379 |
-
|
380 |
-
logger.info("Validating conversation structure...")
|
381 |
-
for i in range(min(5, len(dataset))):
|
382 |
-
conv = dataset[i].get("conversations")
|
383 |
-
if conv is None:
|
384 |
-
logger.warning(f"Example {i} has None as 'conversations' value")
|
385 |
-
elif not isinstance(conv, list):
|
386 |
-
logger.warning(f"Example {i} has non-list 'conversations': {type(conv)}")
|
387 |
-
elif len(conv) == 0:
|
388 |
-
logger.warning(f"Example {i} has empty conversations list")
|
389 |
else:
|
390 |
-
|
391 |
-
first_entry = conv[0]
|
392 |
-
if isinstance(first_entry, dict) and "content" in first_entry:
|
393 |
-
logger.info(f"Content field example: {str(first_entry['content'])[:50]}...")
|
394 |
-
else:
|
395 |
-
logger.warning(f"Example {i} missing 'content' key in conversation")
|
396 |
-
|
397 |
-
except Exception as dataset_error:
|
398 |
-
logger.error(f"Failed to load dataset {dataset_name}: {str(dataset_error)}")
|
399 |
-
logger.error("Make sure the dataset exists and you have proper access permissions")
|
400 |
-
logger.error("This could be due to authentication issues with your HF_TOKEN")
|
401 |
-
raise
|
402 |
|
403 |
return dataset
|
404 |
|
405 |
except Exception as e:
|
406 |
-
|
407 |
-
|
|
|
408 |
|
409 |
def format_phi_chat(messages, dataset_config):
|
410 |
"""Format messages according to phi-4's chat template and dataset config.
|
@@ -528,110 +592,292 @@ class SimpleDataCollator:
|
|
528 |
# Return empty batch if no valid examples
|
529 |
return {k: [] for k in batch}
|
530 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
531 |
class LoggingCallback(TrainerCallback):
|
532 |
def __init__(self, model=None, dataset=None):
|
533 |
super().__init__()
|
534 |
self.training_started = time.time()
|
535 |
self.last_log_time = time.time()
|
536 |
-
self.
|
|
|
|
|
537 |
self.model = model
|
538 |
self.dataset = dataset
|
539 |
|
540 |
def on_train_begin(self, args, state, control, **kwargs):
|
541 |
-
|
542 |
-
|
543 |
-
|
544 |
-
|
545 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
546 |
|
547 |
-
|
548 |
-
|
549 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
550 |
|
551 |
-
|
552 |
-
|
553 |
-
|
554 |
-
log_info(f"Training plan: {len(self.dataset or [])} examples over {args.num_train_epochs} epochs ≈ {total_steps} steps")
|
555 |
-
log_info(f"Batch size: {args.per_device_train_batch_size} × {args.gradient_accumulation_steps} steps × {NUM_GPUS} GPUs = {total_batch_size} total")
|
556 |
|
557 |
-
|
558 |
-
|
559 |
-
|
560 |
-
|
561 |
-
|
562 |
-
|
563 |
-
|
564 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
565 |
|
566 |
def check_dependencies():
|
567 |
-
"""
|
568 |
-
|
569 |
-
|
570 |
-
|
571 |
-
# Define required packages with versions
|
572 |
required_packages = {
|
573 |
-
"unsloth": ">=2024.3",
|
574 |
-
"transformers": ">=4.38.0",
|
575 |
-
"peft": ">=0.9.0",
|
576 |
-
"accelerate": ">=0.27.0"
|
577 |
}
|
578 |
|
579 |
-
#
|
580 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
581 |
try:
|
|
|
582 |
if package == "unsloth" and not unsloth_available:
|
583 |
-
missing_packages.append(f"{package}{
|
|
|
|
|
584 |
elif package == "peft" and not peft_available:
|
585 |
-
missing_packages.append(f"{package}{
|
586 |
-
|
587 |
-
|
588 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
589 |
except ImportError:
|
590 |
-
missing_packages.append(f"{package}{
|
|
|
591 |
|
592 |
-
# Check
|
593 |
-
|
594 |
-
|
595 |
-
|
596 |
-
|
597 |
-
|
598 |
-
|
599 |
-
|
600 |
-
|
601 |
-
if transformers_idx < unsloth_idx:
|
602 |
-
order_issues.append("For optimal performance, unsloth should be imported before transformers")
|
603 |
-
except ValueError:
|
604 |
-
pass
|
605 |
-
except Exception as e:
|
606 |
-
logger.warning(f"Could not check module import order: {str(e)}")
|
607 |
-
|
608 |
-
# Check optional dependencies
|
609 |
-
optional_packages = {
|
610 |
-
"flash_attn": "Flash attention support",
|
611 |
-
"bitsandbytes": "4-bit quantization support"
|
612 |
-
}
|
613 |
|
614 |
-
|
615 |
-
|
616 |
-
|
617 |
-
|
618 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
619 |
|
620 |
# Report missing required packages
|
621 |
if missing_packages:
|
622 |
-
|
623 |
for pkg in missing_packages:
|
624 |
-
|
625 |
-
|
626 |
-
|
627 |
return False
|
628 |
|
629 |
-
|
630 |
-
for issue in order_issues:
|
631 |
-
logger.warning(issue)
|
632 |
-
|
633 |
return True
|
634 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
635 |
def update_huggingface_space():
|
636 |
"""Update the Hugging Face Space with the current code."""
|
637 |
log_info("Updating Hugging Face Space...")
|
@@ -709,381 +955,338 @@ def validate_huggingface_credentials():
|
|
709 |
logger.warning(f"Error validating Hugging Face credentials: {str(e)}")
|
710 |
return False
|
711 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
712 |
def main():
|
|
|
|
|
|
|
|
|
|
|
|
|
713 |
# Set up logging
|
714 |
logger.info("Starting training process")
|
715 |
|
716 |
try:
|
717 |
-
#
|
718 |
-
if not check_dependencies():
|
719 |
-
logger.error("Aborting due to missing critical dependencies")
|
720 |
-
return 1
|
721 |
-
|
722 |
-
# Parse arguments
|
723 |
args = parse_args()
|
724 |
|
725 |
-
#
|
726 |
-
|
727 |
|
728 |
-
#
|
729 |
-
validate_huggingface_credentials()
|
730 |
-
|
731 |
-
# Load configuration
|
732 |
try:
|
733 |
-
|
734 |
-
hardware_config = transformers_config.get("hardware", {})
|
735 |
-
dataset_config = transformers_config.get("dataset", {})
|
736 |
-
logger.info("Configuration loaded successfully")
|
737 |
except Exception as e:
|
738 |
-
logger.error(f"Error
|
739 |
return 1
|
740 |
|
741 |
-
#
|
742 |
-
|
743 |
-
|
744 |
-
|
745 |
-
|
746 |
-
|
747 |
-
log_info("Running in non-distributed mode (single process)")
|
748 |
-
|
749 |
-
# Set random seed for reproducibility
|
750 |
-
seed = transformers_config.get("seed", 42)
|
751 |
-
set_seed(seed)
|
752 |
-
logger.info(f"Set random seed to {seed}")
|
753 |
-
|
754 |
-
# Load model and tokenizer using the consolidated config
|
755 |
-
model, tokenizer = load_model_and_tokenizer(transformers_config)
|
756 |
-
|
757 |
-
# Empty CUDA cache to ensure clean state
|
758 |
-
if CUDA_AVAILABLE:
|
759 |
-
torch.cuda.empty_cache()
|
760 |
-
log_info("Cleared CUDA cache")
|
761 |
|
762 |
-
#
|
763 |
-
|
764 |
-
|
765 |
-
|
|
|
|
|
766 |
|
767 |
-
|
768 |
-
|
769 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
770 |
|
|
|
771 |
try:
|
772 |
-
|
773 |
-
|
774 |
-
|
775 |
-
|
776 |
-
if dataset is None:
|
777 |
-
logger.error("Dataset is None! Cannot proceed with training.")
|
778 |
-
return 1
|
779 |
-
|
780 |
-
if not hasattr(dataset, '__len__') or len(dataset) == 0:
|
781 |
-
logger.error("Dataset is empty! Cannot proceed with training.")
|
782 |
-
return 1
|
783 |
-
|
784 |
-
log_info(f"Dataset loaded with {len(dataset)} examples")
|
785 |
-
|
786 |
-
# Minimal validation before proceeding
|
787 |
-
if dataset is None or len(dataset) == 0:
|
788 |
-
logger.error("Dataset is empty or None! Cannot proceed with training.")
|
789 |
-
return 1
|
790 |
-
|
791 |
-
# Create data collator
|
792 |
-
data_collator = SimpleDataCollator(tokenizer, dataset_config)
|
793 |
-
|
794 |
-
# Verify precision settings - ensure only one of bf16/fp16 is set, with bf16 taking precedence
|
795 |
-
# First check hardware config, then transformers config
|
796 |
-
use_bf16 = False
|
797 |
-
use_fp16 = False
|
798 |
-
|
799 |
-
# Check hardware config first
|
800 |
-
hardware_precision = hardware_config.get("training_optimizations", {}).get("mixed_precision", "")
|
801 |
-
if hardware_precision.lower() == "bf16":
|
802 |
-
use_bf16 = True
|
803 |
-
log_info("Using BF16 precision from hardware config")
|
804 |
-
elif hardware_precision.lower() == "fp16":
|
805 |
-
use_fp16 = True
|
806 |
-
log_info("Using FP16 precision from hardware config")
|
807 |
-
else:
|
808 |
-
# Fall back to transformers config
|
809 |
-
use_bf16 = transformers_config.get("bf16", False) or transformers_config.get("torch_dtype", "") == "bfloat16"
|
810 |
-
use_fp16 = transformers_config.get("fp16", False) and not use_bf16 # Only use fp16 if bf16 is not set
|
811 |
-
log_info(f"Using precision: {'bf16' if use_bf16 else 'fp16' if use_fp16 else 'full precision'}")
|
812 |
-
|
813 |
-
# Get per device batch size - from transformers config, but possibly overridden by hardware config
|
814 |
-
per_device_batch_size = transformers_config.get("training", {}).get("per_device_train_batch_size", 16)
|
815 |
-
gradient_accumulation_steps = transformers_config.get("training", {}).get("gradient_accumulation_steps", 3)
|
816 |
-
|
817 |
-
# Get multi-GPU strategy from hardware config (default to data_parallel)
|
818 |
-
multi_gpu_strategy = hardware_config.get("training_optimizations", {}).get("multi_gpu_strategy", "data_parallel")
|
819 |
-
logger.info(f"Multi-GPU strategy: {multi_gpu_strategy}")
|
820 |
-
|
821 |
-
# For multi-GPU setup, adjust for better balance
|
822 |
-
if CUDA_AVAILABLE and NUM_GPUS > 1:
|
823 |
-
log_info(f"Multi-GPU setup: Adjusting for {NUM_GPUS} GPUs")
|
824 |
-
|
825 |
-
# Set up FSDP for multi-GPU training if specified and in distributed mode
|
826 |
-
fsdp_config = None
|
827 |
-
if multi_gpu_strategy == "fsdp" and is_distributed and NUM_GPUS > 1:
|
828 |
-
try:
|
829 |
-
from torch.distributed.fsdp import (
|
830 |
-
FullyShardedDataParallel as FSDP,
|
831 |
-
MixedPrecision,
|
832 |
-
BackwardPrefetch,
|
833 |
-
ShardingStrategy,
|
834 |
-
CPUOffload,
|
835 |
-
)
|
836 |
-
from torch.distributed.fsdp.wrap import (
|
837 |
-
transformer_auto_wrap_policy,
|
838 |
-
enable_wrap,
|
839 |
-
wrap,
|
840 |
-
)
|
841 |
-
|
842 |
-
log_info("Using FSDP for distributed training")
|
843 |
-
|
844 |
-
# Configure FSDP
|
845 |
-
fsdp_config = {
|
846 |
-
"fsdp_transformer_layer_cls_to_wrap": ["LlamaDecoderLayer"],
|
847 |
-
"fsdp_offload_params": False,
|
848 |
-
"fsdp_backward_prefetch": "BACKWARD_PRE",
|
849 |
-
"fsdp_min_num_params": 1e6,
|
850 |
-
"fsdp_sharding_strategy": 1, # FULL_SHARD
|
851 |
-
}
|
852 |
-
|
853 |
-
if use_bf16 or use_fp16:
|
854 |
-
precision_type = "bf16" if use_bf16 else "fp16"
|
855 |
-
fsdp_config["fsdp_state_dict_type"] = "FULL_STATE_DICT"
|
856 |
-
log_info(f"FSDP using mixed precision: {precision_type}")
|
857 |
-
except ImportError:
|
858 |
-
log_info("FSDP imports failed, falling back to standard DDP")
|
859 |
-
fsdp_config = None
|
860 |
-
elif multi_gpu_strategy == "fsdp" and not is_distributed:
|
861 |
-
log_info("FSDP disabled: requires distributed environment (use torchrun or accelerate)")
|
862 |
-
log_info("Using DataParallel for multi-GPU training instead")
|
863 |
-
else:
|
864 |
-
log_info(f"Using {multi_gpu_strategy} for multi-GPU training")
|
865 |
-
|
866 |
-
# Get system settings from hardware config
|
867 |
-
dataloader_workers = hardware_config.get("system_settings", {}).get("dataloader_num_workers", 2)
|
868 |
-
pin_memory = hardware_config.get("system_settings", {}).get("dataloader_pin_memory", True)
|
869 |
-
|
870 |
-
# Set up training arguments
|
871 |
-
log_info("Setting up training arguments")
|
872 |
-
|
873 |
-
# Handle FSDP configuration
|
874 |
-
fsdp_config = transformers_config.get("distributed_training", {}).get("fsdp_config", {})
|
875 |
-
fsdp_enabled = fsdp_config.get("enabled", False)
|
876 |
-
|
877 |
-
# Only set FSDP args if explicitly enabled
|
878 |
-
fsdp_args = None
|
879 |
-
if fsdp_enabled and is_distributed and NUM_GPUS > 1:
|
880 |
-
fsdp_args = {
|
881 |
-
"fsdp": ["full_shard", "auto_wrap"],
|
882 |
-
"fsdp_transformer_layer_cls_to_wrap": "LlamaDecoderLayer",
|
883 |
-
"fsdp_offload_params": fsdp_config.get("offload_params", False),
|
884 |
-
"fsdp_state_dict_type": "FULL_STATE_DICT",
|
885 |
-
"fsdp_sharding_strategy": 1, # FULL_SHARD
|
886 |
-
}
|
887 |
-
log_info("FSDP configuration enabled")
|
888 |
-
else:
|
889 |
-
log_info("FSDP disabled, using standard data parallel")
|
890 |
-
|
891 |
-
# Check if we're running in a Space
|
892 |
-
is_space = bool(os.environ.get("SPACE_ID"))
|
893 |
-
|
894 |
-
# Create training arguments
|
895 |
-
training_args = TrainingArguments(
|
896 |
-
output_dir=transformers_config.get("output_dir", "./results") or transformers_config.get("checkpointing", {}).get("output_dir", "./results"),
|
897 |
-
num_train_epochs=transformers_config.get("training", {}).get("num_train_epochs", 3),
|
898 |
-
per_device_train_batch_size=per_device_batch_size,
|
899 |
-
gradient_accumulation_steps=gradient_accumulation_steps,
|
900 |
-
learning_rate=transformers_config.get("training", {}).get("learning_rate", 2e-5),
|
901 |
-
weight_decay=transformers_config.get("training", {}).get("weight_decay", 0.01),
|
902 |
-
warmup_ratio=transformers_config.get("training", {}).get("warmup_ratio", 0.05),
|
903 |
-
lr_scheduler_type=transformers_config.get("training", {}).get("lr_scheduler_type", "cosine"),
|
904 |
-
logging_steps=transformers_config.get("training", {}).get("logging_steps", 10),
|
905 |
-
save_strategy=transformers_config.get("checkpointing", {}).get("save_strategy", "steps"),
|
906 |
-
save_steps=transformers_config.get("checkpointing", {}).get("save_steps", 100),
|
907 |
-
save_total_limit=transformers_config.get("checkpointing", {}).get("save_total_limit", 3),
|
908 |
-
fp16=use_fp16,
|
909 |
-
bf16=use_bf16,
|
910 |
-
max_grad_norm=transformers_config.get("training", {}).get("max_grad_norm", 1.0),
|
911 |
-
push_to_hub=transformers_config.get("huggingface_hub", {}).get("push_to_hub", False),
|
912 |
-
hub_model_id=transformers_config.get("huggingface_hub", {}).get("hub_model_id", None),
|
913 |
-
hub_token=None if is_space else os.environ.get("HF_TOKEN", None),
|
914 |
-
report_to="tensorboard",
|
915 |
-
remove_unused_columns=False, # Keep all columns
|
916 |
-
gradient_checkpointing=transformers_config.get("training", {}).get("gradient_checkpointing", True),
|
917 |
-
dataloader_pin_memory=pin_memory,
|
918 |
-
optim=transformers_config.get("training", {}).get("optim", "adamw_torch"),
|
919 |
-
ddp_find_unused_parameters=False, # Improve distributed training efficiency
|
920 |
-
dataloader_drop_last=False, # Process all examples
|
921 |
-
dataloader_num_workers=dataloader_workers,
|
922 |
-
no_cuda=False if CUDA_AVAILABLE else True, # Use CUDA if available
|
923 |
-
**({} if fsdp_args is None else fsdp_args) # Only include FSDP args if configured
|
924 |
-
)
|
925 |
-
|
926 |
-
log_info("Training arguments created successfully")
|
927 |
-
|
928 |
-
# Validate dataset before creating sampler
|
929 |
-
if dataset is None:
|
930 |
-
raise ValueError("Dataset is None - cannot create sampler")
|
931 |
-
|
932 |
-
# Create sequential sampler to maintain original dataset order
|
933 |
-
sequential_sampler = torch.utils.data.SequentialSampler(dataset)
|
934 |
-
log_info("Sequential sampler created")
|
935 |
-
|
936 |
-
# Initialize trainer first
|
937 |
-
log_info("Initializing Trainer")
|
938 |
-
trainer = Trainer(
|
939 |
-
model=model,
|
940 |
-
args=training_args,
|
941 |
-
train_dataset=dataset,
|
942 |
-
data_collator=data_collator,
|
943 |
-
callbacks=[LoggingCallback(model=model, dataset=dataset)],
|
944 |
-
)
|
945 |
-
|
946 |
-
# Then override the get_train_dataloader method
|
947 |
-
def custom_get_train_dataloader():
|
948 |
-
"""Custom dataloader that preserves original dataset order"""
|
949 |
-
log_info("Creating sequential dataloader to maintain original dataset order")
|
950 |
-
|
951 |
-
# Safety check - make sure dataset exists and is not None
|
952 |
-
if dataset is None:
|
953 |
-
raise ValueError("Dataset is None - cannot create dataloader")
|
954 |
-
|
955 |
-
# Make sure dataset is not empty
|
956 |
-
if len(dataset) == 0:
|
957 |
-
raise ValueError("Dataset is empty - cannot create dataloader")
|
958 |
-
|
959 |
-
# Create a simple sequential sampler
|
960 |
-
sequential_sampler = torch.utils.data.SequentialSampler(dataset)
|
961 |
-
|
962 |
-
# Verification of sequence preservation flags - simplified
|
963 |
-
data_loading_config = dataset_config.get("data_loading", {})
|
964 |
-
shuffle_enabled = data_loading_config.get("shuffle", False)
|
965 |
-
|
966 |
-
if shuffle_enabled:
|
967 |
-
log_info("WARNING: Shuffle is enabled in configuration! This will be overridden to preserve order.")
|
968 |
-
# We enforce sequential processing regardless of config
|
969 |
-
|
970 |
-
# Log our approach clearly
|
971 |
-
log_info("Using SequentialSampler to guarantee dataset order is preserved based on prompt_number")
|
972 |
-
|
973 |
-
# Verify column order and check for 'conversations' field
|
974 |
-
expected_order = ["prompt_number", "article_id", "conversations"]
|
975 |
-
if hasattr(dataset, 'column_names'):
|
976 |
-
actual_order = dataset.column_names
|
977 |
-
|
978 |
-
# Verify all required fields exist
|
979 |
-
missing_fields = [field for field in ["conversations"] if field not in actual_order]
|
980 |
-
if missing_fields:
|
981 |
-
raise ValueError(f"Dataset missing critical fields: {missing_fields}")
|
982 |
-
|
983 |
-
if actual_order == expected_order:
|
984 |
-
log_info(f"Confirmed dataset columns are in expected order: {', '.join(expected_order)}")
|
985 |
-
else:
|
986 |
-
log_info(f"Note: Dataset columns ({', '.join(actual_order)}) are not in expected order ({', '.join(expected_order)})")
|
987 |
-
log_info("This is handled correctly by field-based access, but noting for clarity")
|
988 |
-
|
989 |
-
log_info("Dataset is pre-processed with prompt_number field indicating the correct sequence")
|
990 |
-
|
991 |
-
# Validate a few samples before proceeding
|
992 |
-
for i in range(min(3, len(dataset))):
|
993 |
-
sample = dataset[i]
|
994 |
-
if "conversations" not in sample:
|
995 |
-
log_info(f"WARNING: Sample {i} missing 'conversations' field")
|
996 |
-
elif sample["conversations"] is None:
|
997 |
-
log_info(f"WARNING: Sample {i} has None 'conversations' field")
|
998 |
-
elif not isinstance(sample["conversations"], list):
|
999 |
-
log_info(f"WARNING: Sample {i} has non-list 'conversations' field: {type(sample['conversations'])}")
|
1000 |
-
|
1001 |
-
# Calculate batch size based on device availability
|
1002 |
-
if getattr(training_args, "no_cuda", False):
|
1003 |
-
batch_size = training_args.per_device_train_batch_size
|
1004 |
-
else:
|
1005 |
-
batch_size = max(training_args.per_device_train_batch_size * max(1, NUM_GPUS), 1)
|
1006 |
-
|
1007 |
-
log_info(f"Using sequential sampler with batch size {batch_size}")
|
1008 |
-
|
1009 |
-
# Return DataLoader with sequential sampler and extra error handling
|
1010 |
-
try:
|
1011 |
-
return torch.utils.data.DataLoader(
|
1012 |
-
dataset,
|
1013 |
-
batch_size=batch_size,
|
1014 |
-
sampler=sequential_sampler, # Always use sequential sampler
|
1015 |
-
collate_fn=data_collator,
|
1016 |
-
drop_last=training_args.dataloader_drop_last,
|
1017 |
-
num_workers=training_args.dataloader_num_workers,
|
1018 |
-
pin_memory=training_args.dataloader_pin_memory,
|
1019 |
-
)
|
1020 |
-
except Exception as e:
|
1021 |
-
log_info(f"Error creating DataLoader: {str(e)}")
|
1022 |
-
# Try again with minimal settings
|
1023 |
-
log_info("Attempting to create DataLoader with minimal settings")
|
1024 |
-
return torch.utils.data.DataLoader(
|
1025 |
-
dataset,
|
1026 |
-
batch_size=1, # Minimal batch size
|
1027 |
-
sampler=sequential_sampler,
|
1028 |
-
collate_fn=data_collator,
|
1029 |
-
num_workers=0, # No parallel workers
|
1030 |
-
pin_memory=False,
|
1031 |
-
)
|
1032 |
-
|
1033 |
-
# Override the get_train_dataloader method
|
1034 |
-
trainer.get_train_dataloader = custom_get_train_dataloader
|
1035 |
-
|
1036 |
-
# Start training
|
1037 |
-
log_info("=== Starting Training ===")
|
1038 |
-
try:
|
1039 |
-
# Empty cache again right before training
|
1040 |
-
if CUDA_AVAILABLE:
|
1041 |
-
torch.cuda.empty_cache()
|
1042 |
-
log_info("Cleared CUDA cache before training")
|
1043 |
-
|
1044 |
-
# Display compact training info
|
1045 |
-
total_steps = int((len(dataset) / (per_device_batch_size * NUM_GPUS * gradient_accumulation_steps)) * training_args.num_train_epochs)
|
1046 |
-
log_info(f"Training plan: {len(dataset)} examples over {training_args.num_train_epochs} epochs ≈ {total_steps} steps")
|
1047 |
-
|
1048 |
-
trainer.train()
|
1049 |
-
log_info("Training completed successfully!")
|
1050 |
-
|
1051 |
-
# Save the final model
|
1052 |
-
log_info("Saving final model...")
|
1053 |
-
trainer.save_model()
|
1054 |
-
log_info(f"Model saved to {training_args.output_dir}")
|
1055 |
-
|
1056 |
-
# Push to hub if enabled
|
1057 |
-
if transformers_config.get("huggingface_hub", {}).get("push_to_hub", False):
|
1058 |
-
hub_id = transformers_config.get("huggingface_hub", {}).get("hub_model_id", "model")
|
1059 |
-
log_info(f"Pushing model to Hugging Face Hub as {hub_id}...")
|
1060 |
-
trainer.push_to_hub()
|
1061 |
-
log_info("Model successfully pushed to Hub")
|
1062 |
-
|
1063 |
-
# Update the Hugging Face Space with current code
|
1064 |
-
if os.environ.get("HF_TOKEN") and os.environ.get("HF_USERNAME") and os.environ.get("HF_SPACE_NAME"):
|
1065 |
-
update_huggingface_space()
|
1066 |
-
|
1067 |
-
return 0
|
1068 |
-
except Exception as e:
|
1069 |
-
logger.error(f"Training failed with error: {str(e)}")
|
1070 |
-
# Log CUDA memory info if available in compact format
|
1071 |
-
if CUDA_AVAILABLE:
|
1072 |
-
memory_info = []
|
1073 |
-
for i in range(NUM_GPUS):
|
1074 |
-
allocated = torch.cuda.memory_allocated(i) / 1024**2
|
1075 |
-
reserved = torch.cuda.memory_reserved(i) / 1024**2
|
1076 |
-
max_mem = torch.cuda.max_memory_allocated(i) / 1024**2
|
1077 |
-
memory_info.append(f"GPU {i}: {allocated:.1f}MB/{reserved:.1f}MB (max: {max_mem:.1f}MB)")
|
1078 |
-
logger.error(f"GPU memory at failure: {', '.join(memory_info)}")
|
1079 |
-
raise
|
1080 |
|
|
|
|
|
|
|
1081 |
except Exception as e:
|
1082 |
-
logger.error(f"
|
|
|
|
|
|
|
|
|
1083 |
return 1
|
1084 |
|
1085 |
except Exception as e:
|
1086 |
logger.error(f"Error in main function: {str(e)}")
|
|
|
1087 |
return 1
|
1088 |
|
1089 |
if __name__ == "__main__":
|
|
|
184 |
raise
|
185 |
|
186 |
def parse_args():
|
187 |
+
"""
|
188 |
+
Parse command line arguments for the training script.
|
189 |
+
|
190 |
+
Returns:
|
191 |
+
argparse.Namespace: The parsed command line arguments
|
192 |
+
"""
|
193 |
+
parser = argparse.ArgumentParser(description="Run training for language models")
|
194 |
+
parser.add_argument(
|
195 |
+
"--config_file",
|
196 |
+
type=str,
|
197 |
+
default=None,
|
198 |
+
help="Path to the configuration file (default: transformers_config.json in script directory)"
|
199 |
+
)
|
200 |
+
parser.add_argument(
|
201 |
+
"--seed",
|
202 |
+
type=int,
|
203 |
+
default=None,
|
204 |
+
help="Random seed for reproducibility (default: based on current time)"
|
205 |
+
)
|
206 |
+
parser.add_argument(
|
207 |
+
"--log_level",
|
208 |
+
type=str,
|
209 |
+
choices=["debug", "info", "warning", "error", "critical"],
|
210 |
+
default="info",
|
211 |
+
help="Logging level (default: info)"
|
212 |
+
)
|
213 |
return parser.parse_args()
|
214 |
|
215 |
def load_model_and_tokenizer(config):
|
216 |
+
"""
|
217 |
+
Load the model and tokenizer according to the configuration.
|
218 |
+
|
219 |
+
Args:
|
220 |
+
config (dict): Complete configuration dictionary
|
221 |
+
|
222 |
+
Returns:
|
223 |
+
tuple: (model, tokenizer) - The loaded model and tokenizer
|
224 |
+
"""
|
225 |
+
# Extract model configuration
|
226 |
+
model_config = get_config_value(config, "model", {})
|
227 |
+
model_name = get_config_value(model_config, "name", "unsloth/phi-4-unsloth-bnb-4bit")
|
228 |
+
use_fast_tokenizer = get_config_value(model_config, "use_fast_tokenizer", True)
|
229 |
+
trust_remote_code = get_config_value(model_config, "trust_remote_code", True)
|
230 |
+
model_revision = get_config_value(config, "model_revision", "main")
|
231 |
+
|
232 |
+
# Unsloth configuration
|
233 |
+
unsloth_config = get_config_value(config, "unsloth", {})
|
234 |
+
unsloth_enabled = get_config_value(unsloth_config, "enabled", True)
|
235 |
+
|
236 |
+
# Tokenizer configuration
|
237 |
+
tokenizer_config = get_config_value(config, "tokenizer", {})
|
238 |
+
max_seq_length = min(
|
239 |
+
get_config_value(tokenizer_config, "max_seq_length", 2048),
|
240 |
+
4096 # Maximum supported by most models
|
241 |
+
)
|
242 |
+
add_eos_token = get_config_value(tokenizer_config, "add_eos_token", True)
|
243 |
+
chat_template = get_config_value(tokenizer_config, "chat_template", None)
|
244 |
+
padding_side = get_config_value(tokenizer_config, "padding_side", "right")
|
245 |
+
|
246 |
+
log_info(f"Loading model: {model_name} (revision: {model_revision})")
|
247 |
+
log_info(f"Max sequence length: {max_seq_length}")
|
248 |
+
|
249 |
try:
|
250 |
+
if unsloth_enabled and unsloth_available:
|
251 |
+
log_info("Using Unsloth for 4-bit quantized model and LoRA")
|
252 |
+
# Load using Unsloth
|
253 |
+
from unsloth import FastLanguageModel
|
254 |
+
model, tokenizer = FastLanguageModel.from_pretrained(
|
255 |
+
model_name=model_name,
|
256 |
+
max_seq_length=max_seq_length,
|
257 |
+
dtype=get_config_value(config, "torch_dtype", "bfloat16"),
|
258 |
+
revision=model_revision,
|
259 |
+
trust_remote_code=trust_remote_code,
|
260 |
+
use_flash_attention_2=get_config_value(config, "use_flash_attention", True)
|
261 |
+
)
|
262 |
|
263 |
+
# Configure tokenizer settings
|
264 |
+
tokenizer.padding_side = padding_side
|
265 |
+
if add_eos_token and tokenizer.eos_token is None:
|
266 |
+
log_info("Setting EOS token")
|
267 |
+
tokenizer.add_special_tokens({"eos_token": "</s>"})
|
268 |
+
|
269 |
+
# Set chat template if specified
|
270 |
+
if chat_template:
|
271 |
+
log_info(f"Setting chat template: {chat_template}")
|
272 |
+
if hasattr(tokenizer, "chat_template"):
|
273 |
+
tokenizer.chat_template = chat_template
|
274 |
+
else:
|
275 |
+
log_info("Tokenizer does not support chat templates, using default formatting")
|
276 |
+
|
277 |
+
# Apply LoRA
|
278 |
+
lora_r = get_config_value(unsloth_config, "r", 16)
|
279 |
+
lora_alpha = get_config_value(unsloth_config, "alpha", 32)
|
280 |
+
lora_dropout = get_config_value(unsloth_config, "dropout", 0)
|
281 |
+
target_modules = get_config_value(unsloth_config, "target_modules",
|
282 |
+
["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"])
|
283 |
+
|
284 |
+
log_info(f"Applying LoRA with r={lora_r}, alpha={lora_alpha}, dropout={lora_dropout}")
|
285 |
+
model = FastLanguageModel.get_peft_model(
|
286 |
+
model,
|
287 |
+
r=lora_r,
|
288 |
+
target_modules=target_modules,
|
289 |
+
lora_alpha=lora_alpha,
|
290 |
+
lora_dropout=lora_dropout,
|
291 |
+
bias="none",
|
292 |
+
use_gradient_checkpointing=get_config_value(config, "training.gradient_checkpointing", True),
|
293 |
+
random_state=0,
|
294 |
+
max_seq_length=max_seq_length,
|
295 |
+
modules_to_save=None
|
296 |
+
)
|
297 |
else:
|
298 |
+
# Standard HuggingFace loading
|
299 |
+
log_info("Using standard HuggingFace model loading (Unsloth not available or disabled)")
|
300 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
301 |
|
302 |
+
# Load tokenizer first
|
303 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
304 |
+
model_name,
|
305 |
+
trust_remote_code=trust_remote_code,
|
306 |
+
use_fast=use_fast_tokenizer,
|
307 |
+
revision=model_revision,
|
308 |
+
padding_side=padding_side
|
309 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
310 |
|
311 |
+
# Configure tokenizer settings
|
312 |
+
if add_eos_token and tokenizer.eos_token is None:
|
313 |
+
log_info("Setting EOS token")
|
314 |
+
tokenizer.add_special_tokens({"eos_token": "</s>"})
|
315 |
+
|
316 |
+
# Set chat template if specified
|
317 |
+
if chat_template:
|
318 |
+
log_info(f"Setting chat template: {chat_template}")
|
319 |
+
if hasattr(tokenizer, "chat_template"):
|
320 |
+
tokenizer.chat_template = chat_template
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
321 |
else:
|
322 |
+
log_info("Tokenizer does not support chat templates, using default formatting")
|
323 |
+
|
324 |
+
# Now load model with updated tokenizer
|
325 |
+
model = AutoModelForCausalLM.from_pretrained(
|
326 |
+
model_name,
|
327 |
+
trust_remote_code=trust_remote_code,
|
328 |
+
revision=model_revision,
|
329 |
+
torch_dtype=torch.bfloat16 if get_config_value(config, "torch_dtype", "bfloat16") == "bfloat16" else torch.float16,
|
330 |
+
device_map="auto" if CUDA_AVAILABLE else None
|
331 |
+
)
|
332 |
+
|
333 |
+
# Apply PEFT/LoRA if enabled but using standard loading
|
334 |
+
if peft_available and get_config_value(unsloth_config, "enabled", True):
|
335 |
+
log_info("Applying standard PEFT/LoRA configuration")
|
336 |
+
from peft import LoraConfig, get_peft_model
|
337 |
+
|
338 |
+
lora_r = get_config_value(unsloth_config, "r", 16)
|
339 |
+
lora_alpha = get_config_value(unsloth_config, "alpha", 32)
|
340 |
+
lora_dropout = get_config_value(unsloth_config, "dropout", 0)
|
341 |
+
target_modules = get_config_value(unsloth_config, "target_modules",
|
342 |
+
["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"])
|
343 |
+
|
344 |
+
log_info(f"Applying LoRA with r={lora_r}, alpha={lora_alpha}, dropout={lora_dropout}")
|
345 |
+
lora_config = LoraConfig(
|
346 |
+
r=lora_r,
|
347 |
+
lora_alpha=lora_alpha,
|
348 |
+
target_modules=target_modules,
|
349 |
+
lora_dropout=lora_dropout,
|
350 |
+
bias="none",
|
351 |
+
task_type="CAUSAL_LM"
|
352 |
+
)
|
353 |
+
model = get_peft_model(model, lora_config)
|
354 |
|
355 |
+
# Print model summary
|
356 |
+
log_info(f"Model loaded successfully: {model.__class__.__name__}")
|
357 |
+
if hasattr(model, "print_trainable_parameters"):
|
358 |
+
model.print_trainable_parameters()
|
359 |
+
else:
|
360 |
+
total_params = sum(p.numel() for p in model.parameters())
|
361 |
+
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
|
362 |
+
log_info(f"Model has {total_params:,} parameters, {trainable_params:,} trainable ({trainable_params/total_params:.2%})")
|
363 |
|
364 |
return model, tokenizer
|
365 |
+
|
366 |
except Exception as e:
|
367 |
+
log_info(f"Error loading model: {str(e)}")
|
368 |
+
traceback.print_exc()
|
369 |
+
return None, None
|
370 |
|
371 |
+
def load_dataset_with_mapping(config):
|
372 |
+
"""
|
373 |
+
Load dataset from Hugging Face or local files and apply necessary transformations.
|
374 |
+
|
375 |
+
Args:
|
376 |
+
config (dict): Dataset configuration dictionary
|
377 |
+
|
378 |
+
Returns:
|
379 |
+
Dataset: The loaded and processed dataset
|
380 |
+
"""
|
381 |
+
# Extract dataset configuration
|
382 |
+
dataset_info = get_config_value(config, "dataset", {})
|
383 |
+
dataset_name = get_config_value(dataset_info, "name", None)
|
384 |
+
dataset_split = get_config_value(dataset_info, "split", "train")
|
385 |
+
|
386 |
+
# Data formatting configuration
|
387 |
+
formatting_config = get_config_value(config, "data_formatting", {})
|
388 |
+
|
389 |
+
if not dataset_name:
|
390 |
+
raise ValueError("Dataset name not specified in config")
|
391 |
+
|
392 |
+
log_info(f"Loading dataset: {dataset_name} (split: {dataset_split})")
|
393 |
+
|
394 |
try:
|
395 |
+
# Load dataset from Hugging Face or local path
|
396 |
+
from datasets import load_dataset
|
|
|
397 |
|
398 |
+
# Check if it's a local path or Hugging Face dataset
|
399 |
+
if os.path.exists(dataset_name) or os.path.exists(os.path.join(os.getcwd(), dataset_name)):
|
400 |
+
log_info(f"Loading dataset from local path: {dataset_name}")
|
401 |
+
# Local dataset - check if it's a directory or file
|
402 |
+
if os.path.isdir(dataset_name):
|
403 |
+
# Directory - look for data files
|
404 |
+
dataset = load_dataset(
|
405 |
+
"json",
|
406 |
+
data_files={"train": os.path.join(dataset_name, "*.json")},
|
407 |
+
split=dataset_split
|
408 |
+
)
|
409 |
+
else:
|
410 |
+
# Single file
|
411 |
+
dataset = load_dataset(
|
412 |
+
"json",
|
413 |
+
data_files={"train": dataset_name},
|
414 |
+
split=dataset_split
|
415 |
+
)
|
416 |
+
else:
|
417 |
+
# Hugging Face dataset
|
418 |
+
log_info(f"Loading dataset from Hugging Face: {dataset_name}")
|
419 |
+
dataset = load_dataset(dataset_name, split=dataset_split)
|
420 |
|
421 |
+
log_info(f"Dataset loaded with {len(dataset)} examples")
|
422 |
|
423 |
+
# Check if dataset contains required fields
|
424 |
+
required_fields = ["conversations"]
|
425 |
+
missing_fields = [field for field in required_fields if field not in dataset.column_names]
|
426 |
+
|
427 |
+
if missing_fields:
|
428 |
+
log_info(f"WARNING: Dataset missing required fields: {missing_fields}")
|
429 |
+
log_info("Attempting to map dataset structure to required format")
|
430 |
|
431 |
+
# Implement conversion logic based on dataset structure
|
432 |
+
if "messages" in dataset.column_names:
|
433 |
+
log_info("Converting 'messages' field to 'conversations' format")
|
434 |
+
dataset = dataset.map(
|
435 |
+
lambda x: {"conversations": x["messages"]},
|
436 |
+
remove_columns=["messages"]
|
437 |
+
)
|
438 |
+
elif "text" in dataset.column_names:
|
439 |
+
log_info("Converting plain text to conversations format")
|
440 |
+
dataset = dataset.map(
|
441 |
+
lambda x: {"conversations": [{"role": "user", "content": x["text"]}]},
|
442 |
+
remove_columns=["text"]
|
443 |
+
)
|
444 |
+
else:
|
445 |
+
raise ValueError(f"Cannot convert dataset format - missing required fields and no conversion path available")
|
446 |
+
|
447 |
+
# Log dataset info
|
448 |
+
log_info(f"Dataset has {len(dataset)} examples and columns: {dataset.column_names}")
|
449 |
+
|
450 |
+
# Show a few examples for verification
|
451 |
+
for i in range(min(3, len(dataset))):
|
452 |
+
example = dataset[i]
|
453 |
+
log_info(f"Example {i}:")
|
454 |
+
for key, value in example.items():
|
455 |
+
if key == "conversations":
|
456 |
+
log_info(f" conversations: {len(value)} messages")
|
457 |
+
# Show first message only to avoid cluttering logs
|
458 |
+
if value and len(value) > 0:
|
459 |
+
first_msg = value[0]
|
460 |
+
if isinstance(first_msg, dict) and "content" in first_msg:
|
461 |
+
content = first_msg["content"]
|
462 |
+
log_info(f" First message: {content[:50]}..." if len(content) > 50 else f" First message: {content}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
463 |
else:
|
464 |
+
log_info(f" {key}: {value}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
465 |
|
466 |
return dataset
|
467 |
|
468 |
except Exception as e:
|
469 |
+
log_info(f"Error loading dataset: {str(e)}")
|
470 |
+
traceback.print_exc()
|
471 |
+
return None
|
472 |
|
473 |
def format_phi_chat(messages, dataset_config):
|
474 |
"""Format messages according to phi-4's chat template and dataset config.
|
|
|
592 |
# Return empty batch if no valid examples
|
593 |
return {k: [] for k in batch}
|
594 |
|
595 |
+
def log_gpu_memory_usage(step=None, frequency=50, clear_cache_threshold=0.9, label=None):
|
596 |
+
"""
|
597 |
+
Log GPU memory usage statistics with optional cache clearing
|
598 |
+
|
599 |
+
Args:
|
600 |
+
step: Current training step (if None, logs regardless of frequency)
|
601 |
+
frequency: How often to log when step is provided
|
602 |
+
clear_cache_threshold: Fraction of memory used that triggers cache clearing (0-1)
|
603 |
+
label: Optional label for the log message (e.g., "Initial", "Error", "Step")
|
604 |
+
"""
|
605 |
+
if not CUDA_AVAILABLE:
|
606 |
+
return
|
607 |
+
|
608 |
+
# Only log every 'frequency' steps if step is provided
|
609 |
+
if step is not None and frequency > 0 and step % frequency != 0:
|
610 |
+
return
|
611 |
+
|
612 |
+
# Get memory usage for each GPU
|
613 |
+
memory_info = []
|
614 |
+
for i in range(NUM_GPUS):
|
615 |
+
allocated = torch.cuda.memory_allocated(i) / (1024 ** 2) # MB
|
616 |
+
reserved = torch.cuda.memory_reserved(i) / (1024 ** 2) # MB
|
617 |
+
max_mem = torch.cuda.max_memory_allocated(i) / (1024 ** 2) # MB
|
618 |
+
|
619 |
+
# Calculate percentage of reserved memory that's allocated
|
620 |
+
usage_percent = (allocated / reserved) * 100 if reserved > 0 else 0
|
621 |
+
memory_info.append(f"GPU {i}: {allocated:.1f}MB/{reserved:.1f}MB ({usage_percent:.1f}%, max: {max_mem:.1f}MB)")
|
622 |
+
|
623 |
+
# Automatically clear cache if over threshold
|
624 |
+
if clear_cache_threshold > 0 and reserved > 0 and (allocated / reserved) > clear_cache_threshold:
|
625 |
+
log_info(f"Clearing CUDA cache for GPU {i} - high utilization ({allocated:.1f}/{reserved:.1f}MB)")
|
626 |
+
with torch.cuda.device(i):
|
627 |
+
torch.cuda.empty_cache()
|
628 |
+
|
629 |
+
prefix = f"{label} " if label else ""
|
630 |
+
log_info(f"{prefix}GPU Memory: {', '.join(memory_info)}")
|
631 |
+
|
632 |
class LoggingCallback(TrainerCallback):
|
633 |
def __init__(self, model=None, dataset=None):
|
634 |
super().__init__()
|
635 |
self.training_started = time.time()
|
636 |
self.last_log_time = time.time()
|
637 |
+
self.last_step_time = None
|
638 |
+
self.step_durations = []
|
639 |
+
self.best_loss = float('inf')
|
640 |
self.model = model
|
641 |
self.dataset = dataset
|
642 |
|
643 |
def on_train_begin(self, args, state, control, **kwargs):
|
644 |
+
"""Called at the beginning of training"""
|
645 |
+
try:
|
646 |
+
log_info(f"=== Training started at {time.strftime('%Y-%m-%d %H:%M:%S')} ===")
|
647 |
+
|
648 |
+
# Log model info if available
|
649 |
+
if self.model is not None:
|
650 |
+
total_params = sum(p.numel() for p in self.model.parameters())
|
651 |
+
trainable_params = sum(p.numel() for p in self.model.parameters() if p.requires_grad)
|
652 |
+
log_info(f"Model parameters: {total_params/1e6:.2f}M total, {trainable_params/1e6:.2f}M trainable")
|
653 |
+
|
654 |
+
# Log dataset info if available
|
655 |
+
if self.dataset is not None:
|
656 |
+
log_info(f"Dataset size: {len(self.dataset)} examples")
|
657 |
+
|
658 |
+
# Log important training parameters for visibility
|
659 |
+
total_batch_size = args.per_device_train_batch_size * args.gradient_accumulation_steps * NUM_GPUS
|
660 |
+
total_steps = int(len(self.dataset or []) / (args.per_device_train_batch_size * NUM_GPUS * args.gradient_accumulation_steps) * args.num_train_epochs)
|
661 |
+
log_info(f"Training plan: {len(self.dataset or [])} examples over {args.num_train_epochs} epochs ≈ {total_steps} steps")
|
662 |
+
log_info(f"Batch size: {args.per_device_train_batch_size} × {args.gradient_accumulation_steps} steps × {NUM_GPUS} GPUs = {total_batch_size} total")
|
663 |
+
|
664 |
+
# Log initial GPU memory usage with label
|
665 |
+
log_gpu_memory_usage(label="Initial")
|
666 |
+
except Exception as e:
|
667 |
+
logger.warning(f"Error logging training begin statistics: {str(e)}")
|
668 |
+
|
669 |
+
def on_step_end(self, args, state, control, **kwargs):
|
670 |
+
"""Called at the end of each step"""
|
671 |
+
try:
|
672 |
+
if state.global_step == 1 or state.global_step % args.logging_steps == 0:
|
673 |
+
# Track step timing
|
674 |
+
current_time = time.time()
|
675 |
+
if self.last_step_time:
|
676 |
+
step_duration = current_time - self.last_step_time
|
677 |
+
self.step_durations.append(step_duration)
|
678 |
+
# Keep only last 100 steps for averaging
|
679 |
+
if len(self.step_durations) > 100:
|
680 |
+
self.step_durations.pop(0)
|
681 |
+
avg_step_time = sum(self.step_durations) / len(self.step_durations)
|
682 |
+
log_info(f"Step {state.global_step}: {step_duration:.2f}s (avg: {avg_step_time:.2f}s)")
|
683 |
+
|
684 |
+
self.last_step_time = current_time
|
685 |
+
|
686 |
+
# Log GPU memory usage with step number
|
687 |
+
log_gpu_memory_usage(state.global_step, args.logging_steps)
|
688 |
+
|
689 |
+
# Log loss
|
690 |
+
if state.log_history:
|
691 |
+
latest_logs = state.log_history[-1] if state.log_history else {}
|
692 |
+
if "loss" in latest_logs:
|
693 |
+
loss = latest_logs["loss"]
|
694 |
+
log_info(f"Step {state.global_step} loss: {loss:.4f}")
|
695 |
+
|
696 |
+
# Track best loss
|
697 |
+
if loss < self.best_loss:
|
698 |
+
self.best_loss = loss
|
699 |
+
log_info(f"New best loss: {loss:.4f}")
|
700 |
+
except Exception as e:
|
701 |
+
logger.warning(f"Error logging step end statistics: {str(e)}")
|
702 |
+
|
703 |
+
def on_train_end(self, args, state, control, **kwargs):
|
704 |
+
"""Called at the end of training"""
|
705 |
+
try:
|
706 |
+
# Calculate training duration
|
707 |
+
training_time = time.time() - self.training_started
|
708 |
+
hours, remainder = divmod(training_time, 3600)
|
709 |
+
minutes, seconds = divmod(remainder, 60)
|
710 |
+
|
711 |
+
log_info(f"=== Training completed at {time.strftime('%Y-%m-%d %H:%M:%S')} ===")
|
712 |
+
log_info(f"Training duration: {int(hours)}h {int(minutes)}m {int(seconds)}s")
|
713 |
+
log_info(f"Final step: {state.global_step}")
|
714 |
+
log_info(f"Best loss: {self.best_loss:.4f}")
|
715 |
+
|
716 |
+
# Log final GPU memory usage
|
717 |
+
log_gpu_memory_usage(label="Final")
|
718 |
+
except Exception as e:
|
719 |
+
logger.warning(f"Error logging training end statistics: {str(e)}")
|
720 |
|
721 |
+
# Other callback methods with proper error handling
|
722 |
+
def on_save(self, args, state, control, **kwargs):
|
723 |
+
"""Called when a checkpoint is saved"""
|
724 |
+
try:
|
725 |
+
log_info(f"Saving checkpoint at step {state.global_step}")
|
726 |
+
except Exception as e:
|
727 |
+
logger.warning(f"Error in on_save: {str(e)}")
|
728 |
+
|
729 |
+
def on_log(self, args, state, control, **kwargs):
|
730 |
+
"""Called when a log is created"""
|
731 |
+
pass
|
732 |
|
733 |
+
def on_evaluate(self, args, state, control, **kwargs):
|
734 |
+
"""Called when evaluation is performed"""
|
735 |
+
pass
|
|
|
|
|
736 |
|
737 |
+
# Only implement the methods we actually need, remove the others
|
738 |
+
def on_prediction_step(self, args, state, control, **kwargs):
|
739 |
+
"""Called when prediction is performed"""
|
740 |
+
pass
|
741 |
+
|
742 |
+
def on_save_model(self, args, state, control, **kwargs):
|
743 |
+
"""Called when model is saved"""
|
744 |
+
try:
|
745 |
+
# Log memory usage after saving
|
746 |
+
log_gpu_memory_usage(label=f"Save at step {state.global_step}")
|
747 |
+
except Exception as e:
|
748 |
+
logger.warning(f"Error in on_save_model: {str(e)}")
|
749 |
+
|
750 |
+
def on_epoch_end(self, args, state, control, **kwargs):
|
751 |
+
"""Called at the end of an epoch"""
|
752 |
+
try:
|
753 |
+
epoch = state.epoch
|
754 |
+
log_info(f"Completed epoch {epoch:.2f}")
|
755 |
+
log_gpu_memory_usage(label=f"Epoch {epoch:.2f}")
|
756 |
+
except Exception as e:
|
757 |
+
logger.warning(f"Error in on_epoch_end: {str(e)}")
|
758 |
+
|
759 |
+
def on_step_begin(self, args, state, control, **kwargs):
|
760 |
+
"""Called at the beginning of a step"""
|
761 |
+
pass
|
762 |
|
763 |
def check_dependencies():
|
764 |
+
"""
|
765 |
+
Check for required and optional dependencies, ensuring proper versions and import order.
|
766 |
+
Returns True if all required dependencies are present, False otherwise.
|
767 |
+
"""
|
768 |
+
# Define required packages with versions and descriptions
|
769 |
required_packages = {
|
770 |
+
"unsloth": {"version": ">=2024.3", "feature": "fast 4-bit quantization and LoRA"},
|
771 |
+
"transformers": {"version": ">=4.38.0", "feature": "core model functionality"},
|
772 |
+
"peft": {"version": ">=0.9.0", "feature": "parameter-efficient fine-tuning"},
|
773 |
+
"accelerate": {"version": ">=0.27.0", "feature": "multi-GPU training"}
|
774 |
}
|
775 |
|
776 |
+
# Optional packages that enhance functionality
|
777 |
+
optional_packages = {
|
778 |
+
"flash_attn": {"feature": "faster attention computation"},
|
779 |
+
"bitsandbytes": {"feature": "quantization support"},
|
780 |
+
"optimum": {"feature": "model optimization"},
|
781 |
+
"wandb": {"feature": "experiment tracking"}
|
782 |
+
}
|
783 |
+
|
784 |
+
# Store results
|
785 |
+
missing_packages = []
|
786 |
+
package_versions = {}
|
787 |
+
order_issues = []
|
788 |
+
|
789 |
+
# Check required packages
|
790 |
+
log_info("Checking required dependencies...")
|
791 |
+
for package, info in required_packages.items():
|
792 |
+
version_req = info["version"]
|
793 |
+
feature = info["feature"]
|
794 |
+
|
795 |
try:
|
796 |
+
# Special handling for packages we've already checked
|
797 |
if package == "unsloth" and not unsloth_available:
|
798 |
+
missing_packages.append(f"{package}{version_req}")
|
799 |
+
log_info(f"❌ {package} - {feature} MISSING")
|
800 |
+
continue
|
801 |
elif package == "peft" and not peft_available:
|
802 |
+
missing_packages.append(f"{package}{version_req}")
|
803 |
+
log_info(f"❌ {package} - {feature} MISSING")
|
804 |
+
continue
|
805 |
+
|
806 |
+
# Try to import and get version
|
807 |
+
module = __import__(package)
|
808 |
+
version = getattr(module, "__version__", "unknown")
|
809 |
+
package_versions[package] = version
|
810 |
+
log_info(f"✅ {package} v{version} - {feature}")
|
811 |
+
|
812 |
except ImportError:
|
813 |
+
missing_packages.append(f"{package}{version_req}")
|
814 |
+
log_info(f"❌ {package} - {feature} MISSING")
|
815 |
|
816 |
+
# Check optional packages
|
817 |
+
log_info("\nChecking optional dependencies...")
|
818 |
+
for package, info in optional_packages.items():
|
819 |
+
feature = info["feature"]
|
820 |
+
try:
|
821 |
+
__import__(package)
|
822 |
+
log_info(f"✅ {package} - {feature} available")
|
823 |
+
except ImportError:
|
824 |
+
log_info(f"⚠️ {package} - {feature} not available")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
825 |
|
826 |
+
# Check import order for optimal performance
|
827 |
+
if "transformers" in package_versions and "unsloth" in package_versions:
|
828 |
+
try:
|
829 |
+
import sys
|
830 |
+
modules = list(sys.modules.keys())
|
831 |
+
transformers_idx = modules.index("transformers")
|
832 |
+
unsloth_idx = modules.index("unsloth")
|
833 |
+
|
834 |
+
if transformers_idx < unsloth_idx:
|
835 |
+
order_issue = "⚠️ For optimal performance, import unsloth before transformers"
|
836 |
+
order_issues.append(order_issue)
|
837 |
+
log_info(order_issue)
|
838 |
+
else:
|
839 |
+
log_info("✅ Import order: unsloth before transformers (optimal)")
|
840 |
+
except (ValueError, IndexError) as e:
|
841 |
+
log_info(f"⚠️ Could not verify import order: {str(e)}")
|
842 |
|
843 |
# Report missing required packages
|
844 |
if missing_packages:
|
845 |
+
log_info("\n❌ Critical dependencies missing:")
|
846 |
for pkg in missing_packages:
|
847 |
+
log_info(f" - {pkg}")
|
848 |
+
log_info("Please install missing dependencies with:")
|
849 |
+
log_info(f" pip install {' '.join(missing_packages)}")
|
850 |
return False
|
851 |
|
852 |
+
log_info("\n✅ All required dependencies satisfied!")
|
|
|
|
|
|
|
853 |
return True
|
854 |
|
855 |
+
def get_config_value(config, path, default=None):
|
856 |
+
"""
|
857 |
+
Safely get a nested value from a config dictionary using a dot-separated path.
|
858 |
+
|
859 |
+
Args:
|
860 |
+
config: The configuration dictionary
|
861 |
+
path: Dot-separated path to the value (e.g., "training.optimizer.lr")
|
862 |
+
default: Default value to return if path doesn't exist
|
863 |
+
|
864 |
+
Returns:
|
865 |
+
The value at the specified path or the default value
|
866 |
+
"""
|
867 |
+
if not config:
|
868 |
+
return default
|
869 |
+
|
870 |
+
parts = path.split('.')
|
871 |
+
current = config
|
872 |
+
|
873 |
+
for part in parts:
|
874 |
+
if isinstance(current, dict) and part in current:
|
875 |
+
current = current[part]
|
876 |
+
else:
|
877 |
+
return default
|
878 |
+
|
879 |
+
return current
|
880 |
+
|
881 |
def update_huggingface_space():
|
882 |
"""Update the Hugging Face Space with the current code."""
|
883 |
log_info("Updating Hugging Face Space...")
|
|
|
955 |
logger.warning(f"Error validating Hugging Face credentials: {str(e)}")
|
956 |
return False
|
957 |
|
958 |
+
def setup_environment(args):
|
959 |
+
"""
|
960 |
+
Set up the training environment including logging, seed, and configurations.
|
961 |
+
|
962 |
+
Args:
|
963 |
+
args: Command line arguments
|
964 |
+
|
965 |
+
Returns:
|
966 |
+
tuple: (transformers_config, seed) - The loaded configuration and random seed
|
967 |
+
"""
|
968 |
+
# Load environment variables first
|
969 |
+
load_env_variables()
|
970 |
+
|
971 |
+
# Set random seed for reproducibility
|
972 |
+
seed = args.seed if args.seed is not None else int(time.time()) % 10000
|
973 |
+
set_seed(seed)
|
974 |
+
log_info(f"Using random seed: {seed}")
|
975 |
+
|
976 |
+
# Load configuration
|
977 |
+
base_path = os.path.dirname(os.path.abspath(__file__))
|
978 |
+
config_file = args.config_file or os.path.join(base_path, "transformers_config.json")
|
979 |
+
|
980 |
+
if not os.path.exists(config_file):
|
981 |
+
raise FileNotFoundError(f"Config file not found: {config_file}")
|
982 |
+
|
983 |
+
log_info(f"Loading configuration from {config_file}")
|
984 |
+
transformers_config = load_configs(config_file)
|
985 |
+
|
986 |
+
# Set up hardware environment variables if CUDA is available
|
987 |
+
if CUDA_AVAILABLE:
|
988 |
+
memory_fraction = get_config_value(transformers_config, "hardware.system_settings.cuda_memory_fraction", 0.75)
|
989 |
+
if memory_fraction < 1.0:
|
990 |
+
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = f"max_split_size_mb:128,expandable_segments:True"
|
991 |
+
log_info(f"Set CUDA memory allocation limit to expandable with max_split_size_mb:128")
|
992 |
+
|
993 |
+
# Check dependencies before proceeding
|
994 |
+
if not check_dependencies():
|
995 |
+
raise RuntimeError("Critical dependencies missing")
|
996 |
+
|
997 |
+
return transformers_config, seed
|
998 |
+
|
999 |
+
def setup_model_and_tokenizer(config):
|
1000 |
+
"""
|
1001 |
+
Load and configure the model and tokenizer.
|
1002 |
+
|
1003 |
+
Args:
|
1004 |
+
config: Complete configuration dictionary
|
1005 |
+
|
1006 |
+
Returns:
|
1007 |
+
tuple: (model, tokenizer) - The loaded model and tokenizer
|
1008 |
+
"""
|
1009 |
+
log_info("Loading model and tokenizer...")
|
1010 |
+
model, tokenizer = load_model_and_tokenizer(config)
|
1011 |
+
|
1012 |
+
if model is None or tokenizer is None:
|
1013 |
+
raise ValueError("Failed to load model or tokenizer")
|
1014 |
+
|
1015 |
+
log_info(f"Model loaded successfully: {model.__class__.__name__}")
|
1016 |
+
log_info(f"Tokenizer loaded: {tokenizer.__class__.__name__} (vocab size: {tokenizer.vocab_size})")
|
1017 |
+
|
1018 |
+
return model, tokenizer
|
1019 |
+
|
1020 |
+
def setup_dataset_and_collator(config, tokenizer):
|
1021 |
+
"""
|
1022 |
+
Load and configure the dataset and data collator.
|
1023 |
+
|
1024 |
+
Args:
|
1025 |
+
config: Complete configuration dictionary
|
1026 |
+
tokenizer: The tokenizer for the data collator
|
1027 |
+
|
1028 |
+
Returns:
|
1029 |
+
tuple: (dataset, data_collator) - The loaded dataset and configured data collator
|
1030 |
+
"""
|
1031 |
+
dataset_config = get_config_value(config, "dataset", {})
|
1032 |
+
|
1033 |
+
log_info("Loading dataset...")
|
1034 |
+
dataset = load_dataset_with_mapping(dataset_config)
|
1035 |
+
|
1036 |
+
# Validate dataset
|
1037 |
+
if dataset is None:
|
1038 |
+
raise ValueError("Dataset is None! Cannot proceed with training.")
|
1039 |
+
|
1040 |
+
if not hasattr(dataset, '__len__') or len(dataset) == 0:
|
1041 |
+
raise ValueError("Dataset is empty! Cannot proceed with training.")
|
1042 |
+
|
1043 |
+
log_info(f"Dataset loaded with {len(dataset)} examples")
|
1044 |
+
|
1045 |
+
# Create data collator
|
1046 |
+
data_collator = SimpleDataCollator(tokenizer, dataset_config)
|
1047 |
+
|
1048 |
+
return dataset, data_collator
|
1049 |
+
|
1050 |
+
def create_training_arguments(config, dataset):
|
1051 |
+
"""
|
1052 |
+
Create and configure training arguments for the Trainer.
|
1053 |
+
|
1054 |
+
Args:
|
1055 |
+
config: Complete configuration dictionary
|
1056 |
+
dataset: The dataset to determine total steps
|
1057 |
+
|
1058 |
+
Returns:
|
1059 |
+
TrainingArguments: Configured training arguments
|
1060 |
+
"""
|
1061 |
+
# Extract configuration sections
|
1062 |
+
training_config = get_config_value(config, "training", {})
|
1063 |
+
hardware_config = get_config_value(config, "hardware", {})
|
1064 |
+
huggingface_config = get_config_value(config, "huggingface_hub", {})
|
1065 |
+
distributed_config = get_config_value(config, "distributed_training", {})
|
1066 |
+
|
1067 |
+
# Extract key training parameters
|
1068 |
+
per_device_batch_size = get_config_value(training_config, "per_device_train_batch_size", 4)
|
1069 |
+
gradient_accumulation_steps = get_config_value(training_config, "gradient_accumulation_steps", 8)
|
1070 |
+
learning_rate = get_config_value(training_config, "learning_rate", 2e-5)
|
1071 |
+
num_train_epochs = get_config_value(training_config, "num_train_epochs", 3)
|
1072 |
+
|
1073 |
+
# Extract hardware settings
|
1074 |
+
dataloader_workers = get_config_value(hardware_config, "system_settings.dataloader_num_workers",
|
1075 |
+
get_config_value(distributed_config, "dataloader_num_workers", 2))
|
1076 |
+
pin_memory = get_config_value(hardware_config, "system_settings.dataloader_pin_memory", True)
|
1077 |
+
|
1078 |
+
# BF16/FP16 settings - ensure only one is enabled
|
1079 |
+
use_bf16 = get_config_value(training_config, "bf16", False)
|
1080 |
+
use_fp16 = get_config_value(training_config, "fp16", False) if not use_bf16 else False
|
1081 |
+
|
1082 |
+
# Configure distributed training
|
1083 |
+
fsdp_config = get_config_value(distributed_config, "fsdp_config", {})
|
1084 |
+
fsdp_enabled = get_config_value(fsdp_config, "enabled", False)
|
1085 |
+
|
1086 |
+
ddp_config = get_config_value(distributed_config, "ddp_config", {})
|
1087 |
+
ddp_find_unused_parameters = get_config_value(ddp_config, "find_unused_parameters", False)
|
1088 |
+
|
1089 |
+
# Set up FSDP args if enabled
|
1090 |
+
fsdp_args = None
|
1091 |
+
if fsdp_enabled and NUM_GPUS > 1:
|
1092 |
+
from accelerate import FullyShardedDataParallelPlugin
|
1093 |
+
from torch.distributed.fsdp.fully_sharded_data_parallel import (
|
1094 |
+
FullOptimStateDictConfig, FullStateDictConfig
|
1095 |
+
)
|
1096 |
+
|
1097 |
+
fsdp_plugin = FullyShardedDataParallelPlugin(
|
1098 |
+
sharding_strategy=get_config_value(fsdp_config, "sharding_strategy", "FULL_SHARD"),
|
1099 |
+
mixed_precision_policy=get_config_value(fsdp_config, "mixed_precision", "BF16"),
|
1100 |
+
state_dict_config=FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
|
1101 |
+
optim_state_dict_config=FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True),
|
1102 |
+
)
|
1103 |
+
|
1104 |
+
fsdp_args = {
|
1105 |
+
"fsdp": fsdp_plugin,
|
1106 |
+
"fsdp_transformer_layer_cls_to_wrap": ["LlamaDecoderLayer", "PhiDecoderLayer"]
|
1107 |
+
}
|
1108 |
+
|
1109 |
+
# Create and return training arguments
|
1110 |
+
training_args = TrainingArguments(
|
1111 |
+
output_dir=get_config_value(config, "checkpointing.output_dir", "./results"),
|
1112 |
+
overwrite_output_dir=True,
|
1113 |
+
num_train_epochs=num_train_epochs,
|
1114 |
+
per_device_train_batch_size=per_device_batch_size,
|
1115 |
+
gradient_accumulation_steps=gradient_accumulation_steps,
|
1116 |
+
learning_rate=learning_rate,
|
1117 |
+
weight_decay=get_config_value(training_config, "weight_decay", 0.01),
|
1118 |
+
max_grad_norm=get_config_value(training_config, "max_grad_norm", 1.0),
|
1119 |
+
warmup_ratio=get_config_value(training_config, "warmup_ratio", 0.03),
|
1120 |
+
lr_scheduler_type=get_config_value(training_config, "lr_scheduler_type", "cosine"),
|
1121 |
+
logging_steps=get_config_value(training_config, "logging_steps", 10),
|
1122 |
+
save_strategy=get_config_value(config, "checkpointing.save_strategy", "steps"),
|
1123 |
+
save_steps=get_config_value(config, "checkpointing.save_steps", 500),
|
1124 |
+
save_total_limit=get_config_value(config, "checkpointing.save_total_limit", 3),
|
1125 |
+
bf16=use_bf16,
|
1126 |
+
fp16=use_fp16,
|
1127 |
+
push_to_hub=get_config_value(huggingface_config, "push_to_hub", False),
|
1128 |
+
hub_model_id=get_config_value(huggingface_config, "hub_model_id", None),
|
1129 |
+
hub_strategy=get_config_value(huggingface_config, "hub_strategy", "every_save"),
|
1130 |
+
hub_private_repo=get_config_value(huggingface_config, "hub_private_repo", True),
|
1131 |
+
gradient_checkpointing=get_config_value(training_config, "gradient_checkpointing", True),
|
1132 |
+
dataloader_pin_memory=pin_memory,
|
1133 |
+
optim=get_config_value(training_config, "optim", "adamw_torch"),
|
1134 |
+
ddp_find_unused_parameters=ddp_find_unused_parameters,
|
1135 |
+
dataloader_drop_last=False,
|
1136 |
+
dataloader_num_workers=dataloader_workers,
|
1137 |
+
no_cuda=False if CUDA_AVAILABLE else True,
|
1138 |
+
**({} if fsdp_args is None else fsdp_args)
|
1139 |
+
)
|
1140 |
+
|
1141 |
+
log_info("Training arguments created successfully")
|
1142 |
+
return training_args
|
1143 |
+
|
1144 |
+
def configure_custom_dataloader(trainer, dataset, config, training_args):
|
1145 |
+
"""
|
1146 |
+
Configure a custom dataloader for the trainer if needed.
|
1147 |
+
|
1148 |
+
Args:
|
1149 |
+
trainer: The Trainer instance to configure
|
1150 |
+
dataset: The dataset to use
|
1151 |
+
config: Complete configuration dictionary
|
1152 |
+
training_args: The training arguments
|
1153 |
+
|
1154 |
+
Returns:
|
1155 |
+
None (modifies trainer in-place)
|
1156 |
+
"""
|
1157 |
+
dataset_config = get_config_value(config, "dataset", {})
|
1158 |
+
|
1159 |
+
# Check if we need a custom dataloader
|
1160 |
+
if get_config_value(dataset_config, "data_loading.sequential_processing", True):
|
1161 |
+
log_info("Using custom sequential dataloader")
|
1162 |
+
|
1163 |
+
# Create sequential sampler to maintain dataset order
|
1164 |
+
sequential_sampler = torch.utils.data.SequentialSampler(dataset)
|
1165 |
+
log_info("Sequential sampler created")
|
1166 |
+
|
1167 |
+
# Define custom dataloader getter
|
1168 |
+
def custom_get_train_dataloader():
|
1169 |
+
"""Create a custom dataloader that maintains dataset order"""
|
1170 |
+
# Get configuration values
|
1171 |
+
batch_size = training_args.per_device_train_batch_size
|
1172 |
+
drop_last = get_config_value(dataset_config, "data_loading.drop_last", False)
|
1173 |
+
num_workers = training_args.dataloader_num_workers
|
1174 |
+
pin_memory = training_args.dataloader_pin_memory
|
1175 |
+
prefetch_factor = get_config_value(dataset_config, "data_loading.prefetch_factor", 2)
|
1176 |
+
persistent_workers = get_config_value(dataset_config, "data_loading.persistent_workers", False)
|
1177 |
+
|
1178 |
+
# Create DataLoader with sequential sampler
|
1179 |
+
return DataLoader(
|
1180 |
+
dataset,
|
1181 |
+
batch_size=batch_size,
|
1182 |
+
sampler=sequential_sampler,
|
1183 |
+
collate_fn=trainer.data_collator,
|
1184 |
+
drop_last=drop_last,
|
1185 |
+
num_workers=num_workers,
|
1186 |
+
pin_memory=pin_memory,
|
1187 |
+
prefetch_factor=prefetch_factor if num_workers > 0 else None,
|
1188 |
+
persistent_workers=persistent_workers if num_workers > 0 else False,
|
1189 |
+
)
|
1190 |
+
|
1191 |
+
# Override the default dataloader
|
1192 |
+
trainer.get_train_dataloader = custom_get_train_dataloader
|
1193 |
+
|
1194 |
+
def run_training(trainer, tokenizer, training_args):
|
1195 |
+
"""
|
1196 |
+
Run the training process and handle model saving.
|
1197 |
+
|
1198 |
+
Args:
|
1199 |
+
trainer: Configured Trainer instance
|
1200 |
+
tokenizer: The tokenizer to save with the model
|
1201 |
+
training_args: Training arguments
|
1202 |
+
|
1203 |
+
Returns:
|
1204 |
+
int: 0 for success, 1 for failure
|
1205 |
+
"""
|
1206 |
+
log_info("Starting training...")
|
1207 |
+
trainer.train()
|
1208 |
+
|
1209 |
+
log_info("Training complete! Saving final model...")
|
1210 |
+
trainer.save_model()
|
1211 |
+
tokenizer.save_pretrained(training_args.output_dir)
|
1212 |
+
|
1213 |
+
# Push to Hub if configured
|
1214 |
+
if training_args.push_to_hub:
|
1215 |
+
log_info(f"Pushing model to Hugging Face Hub: {training_args.hub_model_id}")
|
1216 |
+
trainer.push_to_hub()
|
1217 |
+
|
1218 |
+
log_info("Training completed successfully!")
|
1219 |
+
return 0
|
1220 |
+
|
1221 |
def main():
|
1222 |
+
"""
|
1223 |
+
Main entry point for the training script.
|
1224 |
+
|
1225 |
+
Returns:
|
1226 |
+
int: 0 for success, non-zero for failure
|
1227 |
+
"""
|
1228 |
# Set up logging
|
1229 |
logger.info("Starting training process")
|
1230 |
|
1231 |
try:
|
1232 |
+
# Parse command line arguments
|
|
|
|
|
|
|
|
|
|
|
1233 |
args = parse_args()
|
1234 |
|
1235 |
+
# Set up environment and load configuration
|
1236 |
+
transformers_config, seed = setup_environment(args)
|
1237 |
|
1238 |
+
# Load model and tokenizer
|
|
|
|
|
|
|
1239 |
try:
|
1240 |
+
model, tokenizer = setup_model_and_tokenizer(transformers_config)
|
|
|
|
|
|
|
1241 |
except Exception as e:
|
1242 |
+
logger.error(f"Error setting up model: {str(e)}")
|
1243 |
return 1
|
1244 |
|
1245 |
+
# Load dataset and create data collator
|
1246 |
+
try:
|
1247 |
+
dataset, data_collator = setup_dataset_and_collator(transformers_config, tokenizer)
|
1248 |
+
except Exception as e:
|
1249 |
+
logger.error(f"Error setting up dataset: {str(e)}")
|
1250 |
+
return 1
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1251 |
|
1252 |
+
# Configure training arguments
|
1253 |
+
try:
|
1254 |
+
training_args = create_training_arguments(transformers_config, dataset)
|
1255 |
+
except Exception as e:
|
1256 |
+
logger.error(f"Error configuring training arguments: {str(e)}")
|
1257 |
+
return 1
|
1258 |
|
1259 |
+
# Initialize trainer with callbacks
|
1260 |
+
log_info("Initializing Trainer")
|
1261 |
+
trainer = Trainer(
|
1262 |
+
model=model,
|
1263 |
+
args=training_args,
|
1264 |
+
train_dataset=dataset,
|
1265 |
+
data_collator=data_collator,
|
1266 |
+
callbacks=[LoggingCallback(model=model, dataset=dataset)],
|
1267 |
+
)
|
1268 |
|
1269 |
+
# Configure custom dataloader if needed
|
1270 |
try:
|
1271 |
+
configure_custom_dataloader(trainer, dataset, transformers_config, training_args)
|
1272 |
+
except Exception as e:
|
1273 |
+
logger.error(f"Error configuring custom dataloader: {str(e)}")
|
1274 |
+
return 1
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1275 |
|
1276 |
+
# Run training process
|
1277 |
+
try:
|
1278 |
+
return run_training(trainer, tokenizer, training_args)
|
1279 |
except Exception as e:
|
1280 |
+
logger.error(f"Training failed with error: {str(e)}")
|
1281 |
+
# Log GPU memory for debugging
|
1282 |
+
log_gpu_memory_usage(label="Error")
|
1283 |
+
# Print full stack trace
|
1284 |
+
traceback.print_exc()
|
1285 |
return 1
|
1286 |
|
1287 |
except Exception as e:
|
1288 |
logger.error(f"Error in main function: {str(e)}")
|
1289 |
+
traceback.print_exc()
|
1290 |
return 1
|
1291 |
|
1292 |
if __name__ == "__main__":
|