Fix: Remove unsupported attn_implementation parameter
Browse files- run_cloud_training.py +503 -0
run_cloud_training.py
ADDED
@@ -0,0 +1,503 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
# -*- coding: utf-8 -*-
|
3 |
+
|
4 |
+
"""
|
5 |
+
Fine-tuning script for DeepSeek-R1-Distill-Qwen-14B-bnb-4bit using unsloth
|
6 |
+
RESEARCH TRAINING PHASE ONLY - No output generation
|
7 |
+
WORKS WITH PRE-TOKENIZED DATASET - No re-tokenization
|
8 |
+
"""
|
9 |
+
|
10 |
+
import os
|
11 |
+
import json
|
12 |
+
import logging
|
13 |
+
import argparse
|
14 |
+
import numpy as np
|
15 |
+
from dotenv import load_dotenv
|
16 |
+
import torch
|
17 |
+
from datasets import load_dataset
|
18 |
+
import transformers
|
19 |
+
from transformers import AutoTokenizer, TrainingArguments, Trainer, AutoModelForCausalLM, AutoConfig
|
20 |
+
from transformers.data.data_collator import DataCollatorMixin
|
21 |
+
from peft import LoraConfig
|
22 |
+
from unsloth import FastLanguageModel
|
23 |
+
|
24 |
+
# Disable flash attention globally
|
25 |
+
os.environ["TRANSFORMERS_NO_FLASH_ATTENTION"] = "1"
|
26 |
+
|
27 |
+
# Check if tensorboard is available
|
28 |
+
try:
|
29 |
+
import tensorboard
|
30 |
+
TENSORBOARD_AVAILABLE = True
|
31 |
+
except ImportError:
|
32 |
+
TENSORBOARD_AVAILABLE = False
|
33 |
+
print("Tensorboard not available. Will skip tensorboard logging.")
|
34 |
+
|
35 |
+
# Configure logging
|
36 |
+
logging.basicConfig(
|
37 |
+
level=logging.INFO,
|
38 |
+
format='%(asctime)s - %(name)s - %(levelname)s - %(message)s',
|
39 |
+
handlers=[
|
40 |
+
logging.StreamHandler(),
|
41 |
+
logging.FileHandler("training.log")
|
42 |
+
]
|
43 |
+
)
|
44 |
+
logger = logging.getLogger(__name__)
|
45 |
+
|
46 |
+
# Default dataset path - use the correct path with username
|
47 |
+
DEFAULT_DATASET = "George-API/phi4-cognitive-dataset"
|
48 |
+
|
49 |
+
def load_config(config_path):
|
50 |
+
"""Load the transformers config from JSON file"""
|
51 |
+
logger.info(f"Loading config from {config_path}")
|
52 |
+
with open(config_path, 'r') as f:
|
53 |
+
config = json.load(f)
|
54 |
+
return config
|
55 |
+
|
56 |
+
def load_and_prepare_dataset(dataset_name, config):
|
57 |
+
"""
|
58 |
+
Load and prepare the dataset for fine-tuning.
|
59 |
+
Sort entries by prompt_number as required.
|
60 |
+
NO TOKENIZATION - DATASET IS ALREADY TOKENIZED
|
61 |
+
"""
|
62 |
+
# Use the default dataset path if no specific path is provided
|
63 |
+
if dataset_name == "phi4-cognitive-dataset":
|
64 |
+
dataset_name = DEFAULT_DATASET
|
65 |
+
|
66 |
+
logger.info(f"Loading dataset: {dataset_name}")
|
67 |
+
|
68 |
+
try:
|
69 |
+
# Load dataset
|
70 |
+
dataset = load_dataset(dataset_name)
|
71 |
+
|
72 |
+
# Extract the split we want to use (usually 'train')
|
73 |
+
if 'train' in dataset:
|
74 |
+
dataset = dataset['train']
|
75 |
+
|
76 |
+
# Get the dataset config
|
77 |
+
dataset_config = config.get("dataset_config", {})
|
78 |
+
sort_field = dataset_config.get("sort_by_field", "prompt_number")
|
79 |
+
sort_direction = dataset_config.get("sort_direction", "ascending")
|
80 |
+
|
81 |
+
# Sort the dataset by prompt_number
|
82 |
+
logger.info(f"Sorting dataset by {sort_field} in {sort_direction} order")
|
83 |
+
if sort_direction == "ascending":
|
84 |
+
dataset = dataset.sort(sort_field)
|
85 |
+
else:
|
86 |
+
dataset = dataset.sort(sort_field, reverse=True)
|
87 |
+
|
88 |
+
# Add shuffle with fixed seed if specified
|
89 |
+
if "shuffle_seed" in dataset_config:
|
90 |
+
shuffle_seed = dataset_config.get("shuffle_seed")
|
91 |
+
logger.info(f"Shuffling dataset with seed {shuffle_seed}")
|
92 |
+
dataset = dataset.shuffle(seed=shuffle_seed)
|
93 |
+
|
94 |
+
# Print dataset structure for debugging
|
95 |
+
logger.info(f"Dataset loaded with {len(dataset)} entries")
|
96 |
+
logger.info(f"Dataset columns: {dataset.column_names}")
|
97 |
+
|
98 |
+
# Print a sample entry to understand structure
|
99 |
+
if len(dataset) > 0:
|
100 |
+
sample = dataset[0]
|
101 |
+
logger.info(f"Sample entry structure: {list(sample.keys())}")
|
102 |
+
if 'conversations' in sample:
|
103 |
+
logger.info(f"Sample conversations structure: {sample['conversations'][:1]}")
|
104 |
+
|
105 |
+
return dataset
|
106 |
+
|
107 |
+
except Exception as e:
|
108 |
+
logger.error(f"Error loading dataset: {str(e)}")
|
109 |
+
logger.info("Available datasets in the Hub:")
|
110 |
+
# Print a more helpful error message
|
111 |
+
print(f"Failed to load dataset: {dataset_name}")
|
112 |
+
print(f"Make sure the dataset exists and is accessible.")
|
113 |
+
print(f"If it's a private dataset, ensure your HF_TOKEN has access to it.")
|
114 |
+
raise
|
115 |
+
|
116 |
+
def tokenize_string(text, tokenizer):
|
117 |
+
"""Tokenize a string using the provided tokenizer"""
|
118 |
+
if not text:
|
119 |
+
return []
|
120 |
+
|
121 |
+
# Tokenize the text
|
122 |
+
tokens = tokenizer.encode(text, add_special_tokens=False)
|
123 |
+
return tokens
|
124 |
+
|
125 |
+
# Data collator for pre-tokenized dataset
|
126 |
+
class PreTokenizedCollator(DataCollatorMixin):
|
127 |
+
"""
|
128 |
+
Data collator for pre-tokenized datasets.
|
129 |
+
Expects input_ids and labels already tokenized.
|
130 |
+
"""
|
131 |
+
def __init__(self, pad_token_id=0, tokenizer=None):
|
132 |
+
self.pad_token_id = pad_token_id
|
133 |
+
self.tokenizer = tokenizer # Keep a reference to the tokenizer for string conversion
|
134 |
+
|
135 |
+
def __call__(self, features):
|
136 |
+
# Print a sample feature to understand structure
|
137 |
+
if len(features) > 0:
|
138 |
+
logger.info(f"Sample feature keys: {list(features[0].keys())}")
|
139 |
+
|
140 |
+
# Extract input_ids from conversations if needed
|
141 |
+
processed_features = []
|
142 |
+
for feature in features:
|
143 |
+
# If input_ids is not directly available, try to extract from conversations
|
144 |
+
if 'input_ids' not in feature and 'conversations' in feature:
|
145 |
+
# Extract from conversations based on your dataset structure
|
146 |
+
conversations = feature['conversations']
|
147 |
+
|
148 |
+
# Debug the conversations structure
|
149 |
+
logger.info(f"Conversations type: {type(conversations)}")
|
150 |
+
if isinstance(conversations, list) and len(conversations) > 0:
|
151 |
+
logger.info(f"First conversation type: {type(conversations[0])}")
|
152 |
+
logger.info(f"First conversation: {conversations[0]}")
|
153 |
+
|
154 |
+
# Try different approaches to extract input_ids
|
155 |
+
if isinstance(conversations, list) and len(conversations) > 0:
|
156 |
+
# Case 1: If conversations is a list of dicts with 'content' field
|
157 |
+
if isinstance(conversations[0], dict) and 'content' in conversations[0]:
|
158 |
+
content = conversations[0]['content']
|
159 |
+
logger.info(f"Found content field: {type(content)}")
|
160 |
+
|
161 |
+
# If content is a string, tokenize it
|
162 |
+
if isinstance(content, str) and self.tokenizer:
|
163 |
+
logger.info(f"Tokenizing string content: {content[:50]}...")
|
164 |
+
feature['input_ids'] = self.tokenizer.encode(content, add_special_tokens=False)
|
165 |
+
# If content is already a list of integers, use it directly
|
166 |
+
elif isinstance(content, list) and all(isinstance(x, int) for x in content):
|
167 |
+
feature['input_ids'] = content
|
168 |
+
# If content is already tokenized in some other format
|
169 |
+
else:
|
170 |
+
logger.warning(f"Unexpected content format: {type(content)}")
|
171 |
+
|
172 |
+
# Case 2: If conversations is a list of dicts with 'input_ids' field
|
173 |
+
elif isinstance(conversations[0], dict) and 'input_ids' in conversations[0]:
|
174 |
+
feature['input_ids'] = conversations[0]['input_ids']
|
175 |
+
|
176 |
+
# Case 3: If conversations itself contains the input_ids
|
177 |
+
elif all(isinstance(x, int) for x in conversations):
|
178 |
+
feature['input_ids'] = conversations
|
179 |
+
|
180 |
+
# Case 4: If conversations is a list of strings
|
181 |
+
elif all(isinstance(x, str) for x in conversations) and self.tokenizer:
|
182 |
+
# Join all strings and tokenize
|
183 |
+
full_text = " ".join(conversations)
|
184 |
+
feature['input_ids'] = self.tokenizer.encode(full_text, add_special_tokens=False)
|
185 |
+
|
186 |
+
# Ensure input_ids is a list of integers
|
187 |
+
if 'input_ids' in feature:
|
188 |
+
# If input_ids is a string, tokenize it
|
189 |
+
if isinstance(feature['input_ids'], str) and self.tokenizer:
|
190 |
+
logger.info(f"Converting string input_ids to tokens: {feature['input_ids'][:50]}...")
|
191 |
+
feature['input_ids'] = self.tokenizer.encode(feature['input_ids'], add_special_tokens=False)
|
192 |
+
# If input_ids is not a list, convert it
|
193 |
+
elif not isinstance(feature['input_ids'], list):
|
194 |
+
try:
|
195 |
+
feature['input_ids'] = list(feature['input_ids'])
|
196 |
+
except:
|
197 |
+
logger.error(f"Could not convert input_ids to list: {type(feature['input_ids'])}")
|
198 |
+
|
199 |
+
processed_features.append(feature)
|
200 |
+
|
201 |
+
# If we still don't have input_ids, log an error
|
202 |
+
if len(processed_features) > 0 and 'input_ids' not in processed_features[0]:
|
203 |
+
logger.error(f"Could not find input_ids in features. Available keys: {list(processed_features[0].keys())}")
|
204 |
+
if 'conversations' in processed_features[0]:
|
205 |
+
logger.error(f"Conversations structure: {processed_features[0]['conversations'][:1]}")
|
206 |
+
raise ValueError("Could not find input_ids in dataset. Please check dataset structure.")
|
207 |
+
|
208 |
+
# Determine max length in this batch
|
209 |
+
batch_max_len = max(len(x["input_ids"]) for x in processed_features)
|
210 |
+
|
211 |
+
# Initialize batch tensors
|
212 |
+
batch = {
|
213 |
+
"input_ids": torch.ones((len(processed_features), batch_max_len), dtype=torch.long) * self.pad_token_id,
|
214 |
+
"attention_mask": torch.zeros((len(processed_features), batch_max_len), dtype=torch.long),
|
215 |
+
"labels": torch.ones((len(processed_features), batch_max_len), dtype=torch.long) * -100 # -100 is ignored in loss
|
216 |
+
}
|
217 |
+
|
218 |
+
# Fill batch tensors
|
219 |
+
for i, feature in enumerate(processed_features):
|
220 |
+
input_ids = feature["input_ids"]
|
221 |
+
seq_len = len(input_ids)
|
222 |
+
|
223 |
+
# Convert to tensor if it's a list
|
224 |
+
if isinstance(input_ids, list):
|
225 |
+
input_ids = torch.tensor(input_ids, dtype=torch.long)
|
226 |
+
|
227 |
+
# Copy data to batch tensors
|
228 |
+
batch["input_ids"][i, :seq_len] = input_ids
|
229 |
+
batch["attention_mask"][i, :seq_len] = 1
|
230 |
+
|
231 |
+
# If there are labels, use them, otherwise use input_ids
|
232 |
+
if "labels" in feature:
|
233 |
+
labels = feature["labels"]
|
234 |
+
if isinstance(labels, list):
|
235 |
+
labels = torch.tensor(labels, dtype=torch.long)
|
236 |
+
batch["labels"][i, :len(labels)] = labels
|
237 |
+
else:
|
238 |
+
batch["labels"][i, :seq_len] = input_ids
|
239 |
+
|
240 |
+
return batch
|
241 |
+
|
242 |
+
def create_training_marker(output_dir):
|
243 |
+
"""Create a marker file to indicate training is active"""
|
244 |
+
# Create in current directory for app.py to find
|
245 |
+
with open("TRAINING_ACTIVE", "w") as f:
|
246 |
+
f.write(f"Training active in {output_dir}")
|
247 |
+
|
248 |
+
# Also create in output directory
|
249 |
+
os.makedirs(output_dir, exist_ok=True)
|
250 |
+
with open(os.path.join(output_dir, "RESEARCH_TRAINING_ONLY"), "w") as f:
|
251 |
+
f.write("This model is for research training only. No interactive outputs.")
|
252 |
+
|
253 |
+
def remove_training_marker():
|
254 |
+
"""Remove the training marker file"""
|
255 |
+
if os.path.exists("TRAINING_ACTIVE"):
|
256 |
+
os.remove("TRAINING_ACTIVE")
|
257 |
+
logger.info("Removed training active marker")
|
258 |
+
|
259 |
+
def load_model_safely(model_name, max_seq_length, dtype=None):
|
260 |
+
"""
|
261 |
+
Load the model in a safe way that works with Qwen models
|
262 |
+
by trying different loading strategies.
|
263 |
+
"""
|
264 |
+
try:
|
265 |
+
logger.info(f"Attempting to load model with unsloth optimizations: {model_name}")
|
266 |
+
# First try the standard unsloth loading
|
267 |
+
try:
|
268 |
+
# Try loading with unsloth but without the problematic parameter
|
269 |
+
logger.info("Loading model with flash attention DISABLED")
|
270 |
+
model, tokenizer = FastLanguageModel.from_pretrained(
|
271 |
+
model_name=model_name,
|
272 |
+
max_seq_length=max_seq_length,
|
273 |
+
dtype=dtype,
|
274 |
+
load_in_4bit=True, # This should work for already quantized models
|
275 |
+
use_flash_attention=False # Explicitly disable flash attention
|
276 |
+
)
|
277 |
+
logger.info("Model loaded successfully with unsloth with 4-bit quantization and flash attention disabled")
|
278 |
+
return model, tokenizer
|
279 |
+
|
280 |
+
except TypeError as e:
|
281 |
+
# If we get a TypeError about unexpected keyword arguments
|
282 |
+
if "unexpected keyword argument" in str(e):
|
283 |
+
logger.warning(f"Unsloth loading error with 4-bit: {e}")
|
284 |
+
logger.info("Trying alternative loading method for Qwen model...")
|
285 |
+
|
286 |
+
# Try loading with different parameters for Qwen model
|
287 |
+
model, tokenizer = FastLanguageModel.from_pretrained(
|
288 |
+
model_name=model_name,
|
289 |
+
max_seq_length=max_seq_length,
|
290 |
+
dtype=dtype,
|
291 |
+
use_flash_attention=False, # Explicitly disable flash attention
|
292 |
+
)
|
293 |
+
logger.info("Model loaded successfully with unsloth using alternative method")
|
294 |
+
return model, tokenizer
|
295 |
+
else:
|
296 |
+
# Re-raise if it's a different type error
|
297 |
+
raise
|
298 |
+
|
299 |
+
except Exception as e:
|
300 |
+
# Fallback to standard loading if unsloth methods fail
|
301 |
+
logger.warning(f"Unsloth loading failed: {e}")
|
302 |
+
logger.info("Falling back to standard Hugging Face loading...")
|
303 |
+
|
304 |
+
# Disable flash attention in transformers config
|
305 |
+
config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
|
306 |
+
if hasattr(config, "use_flash_attention"):
|
307 |
+
config.use_flash_attention = False
|
308 |
+
logger.info("Disabled flash attention in model config")
|
309 |
+
|
310 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
|
311 |
+
model = AutoModelForCausalLM.from_pretrained(
|
312 |
+
model_name,
|
313 |
+
config=config,
|
314 |
+
device_map="auto",
|
315 |
+
torch_dtype=dtype or torch.float16,
|
316 |
+
load_in_4bit=True
|
317 |
+
)
|
318 |
+
logger.info("Model loaded successfully with standard HF loading and flash attention disabled")
|
319 |
+
return model, tokenizer
|
320 |
+
|
321 |
+
def train(config_path, dataset_name, output_dir):
|
322 |
+
"""Main training function - RESEARCH TRAINING PHASE ONLY"""
|
323 |
+
# Load environment variables
|
324 |
+
load_dotenv()
|
325 |
+
config = load_config(config_path)
|
326 |
+
|
327 |
+
# Extract configs
|
328 |
+
model_config = config.get("model_config", {})
|
329 |
+
training_config = config.get("training_config", {})
|
330 |
+
hardware_config = config.get("hardware_config", {})
|
331 |
+
lora_config = config.get("lora_config", {})
|
332 |
+
dataset_config = config.get("dataset_config", {})
|
333 |
+
|
334 |
+
# Override flash attention setting to disable it
|
335 |
+
hardware_config["use_flash_attention"] = False
|
336 |
+
logger.info("Flash attention has been DISABLED due to GPU compatibility issues")
|
337 |
+
|
338 |
+
# Verify this is training phase only
|
339 |
+
training_phase_only = dataset_config.get("training_phase_only", True)
|
340 |
+
if not training_phase_only:
|
341 |
+
logger.warning("This script is meant for research training phase only")
|
342 |
+
logger.warning("Setting training_phase_only=True")
|
343 |
+
|
344 |
+
# Verify dataset is pre-tokenized
|
345 |
+
logger.info("IMPORTANT: Using pre-tokenized dataset - No tokenization will be performed")
|
346 |
+
|
347 |
+
# Set the output directory
|
348 |
+
output_dir = output_dir or training_config.get("output_dir", "fine_tuned_model")
|
349 |
+
os.makedirs(output_dir, exist_ok=True)
|
350 |
+
|
351 |
+
# Create training marker
|
352 |
+
create_training_marker(output_dir)
|
353 |
+
|
354 |
+
try:
|
355 |
+
# Print configuration summary
|
356 |
+
logger.info("RESEARCH TRAINING PHASE ACTIVE - No output generation")
|
357 |
+
logger.info("Configuration Summary:")
|
358 |
+
model_name = model_config.get("model_name_or_path")
|
359 |
+
logger.info(f"Model: {model_name}")
|
360 |
+
logger.info(f"Dataset: {dataset_name if dataset_name != 'phi4-cognitive-dataset' else DEFAULT_DATASET}")
|
361 |
+
logger.info(f"Output directory: {output_dir}")
|
362 |
+
logger.info("IMPORTANT: Using already 4-bit quantized model - not re-quantizing")
|
363 |
+
|
364 |
+
# Load and prepare the dataset
|
365 |
+
dataset = load_and_prepare_dataset(dataset_name, config)
|
366 |
+
|
367 |
+
# Initialize tokenizer (just for model initialization, not for tokenizing data)
|
368 |
+
logger.info("Loading tokenizer (for model initialization only, not for tokenizing data)")
|
369 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
370 |
+
model_name,
|
371 |
+
trust_remote_code=True
|
372 |
+
)
|
373 |
+
tokenizer.pad_token = tokenizer.eos_token
|
374 |
+
|
375 |
+
# Initialize model with unsloth
|
376 |
+
logger.info("Initializing model with unsloth (preserving 4-bit quantization)")
|
377 |
+
max_seq_length = training_config.get("max_seq_length", 2048)
|
378 |
+
|
379 |
+
# Create LoRA config directly
|
380 |
+
logger.info("Creating LoRA configuration")
|
381 |
+
lora_config_obj = LoraConfig(
|
382 |
+
r=lora_config.get("r", 16),
|
383 |
+
lora_alpha=lora_config.get("lora_alpha", 32),
|
384 |
+
lora_dropout=lora_config.get("lora_dropout", 0.05),
|
385 |
+
bias=lora_config.get("bias", "none"),
|
386 |
+
target_modules=lora_config.get("target_modules", ["q_proj", "k_proj", "v_proj", "o_proj"])
|
387 |
+
)
|
388 |
+
|
389 |
+
# Initialize model with our safe loading function
|
390 |
+
logger.info("Loading pre-quantized model safely")
|
391 |
+
dtype = torch.float16 if hardware_config.get("fp16", True) else None
|
392 |
+
model, tokenizer = load_model_safely(model_name, max_seq_length, dtype)
|
393 |
+
|
394 |
+
# Try different approaches to apply LoRA
|
395 |
+
logger.info("Applying LoRA to model")
|
396 |
+
|
397 |
+
# Skip unsloth's method and go directly to PEFT
|
398 |
+
logger.info("Using standard PEFT method to apply LoRA")
|
399 |
+
from peft import get_peft_model
|
400 |
+
model = get_peft_model(model, lora_config_obj)
|
401 |
+
logger.info("Successfully applied LoRA with standard PEFT")
|
402 |
+
|
403 |
+
# No need to format the dataset - it's already pre-tokenized
|
404 |
+
logger.info("Using pre-tokenized dataset - skipping tokenization step")
|
405 |
+
training_dataset = dataset
|
406 |
+
|
407 |
+
# Configure reporting backends with fallbacks
|
408 |
+
reports = []
|
409 |
+
if TENSORBOARD_AVAILABLE:
|
410 |
+
reports.append("tensorboard")
|
411 |
+
logger.info("Tensorboard available and enabled for reporting")
|
412 |
+
else:
|
413 |
+
logger.warning("Tensorboard not available - metrics won't be logged to tensorboard")
|
414 |
+
|
415 |
+
if os.getenv("WANDB_API_KEY"):
|
416 |
+
reports.append("wandb")
|
417 |
+
logger.info("Wandb API key found, enabling wandb reporting")
|
418 |
+
|
419 |
+
# Default to "none" if no reporting backends are available
|
420 |
+
if not reports:
|
421 |
+
reports = ["none"]
|
422 |
+
logger.warning("No reporting backends available - training metrics won't be logged")
|
423 |
+
|
424 |
+
# Set up training arguments with flash attention disabled
|
425 |
+
training_args = TrainingArguments(
|
426 |
+
output_dir=output_dir,
|
427 |
+
num_train_epochs=training_config.get("num_train_epochs", 3),
|
428 |
+
per_device_train_batch_size=training_config.get("per_device_train_batch_size", 2),
|
429 |
+
gradient_accumulation_steps=training_config.get("gradient_accumulation_steps", 4),
|
430 |
+
learning_rate=training_config.get("learning_rate", 2e-5),
|
431 |
+
lr_scheduler_type=training_config.get("lr_scheduler_type", "cosine"),
|
432 |
+
warmup_ratio=training_config.get("warmup_ratio", 0.03),
|
433 |
+
weight_decay=training_config.get("weight_decay", 0.01),
|
434 |
+
optim=training_config.get("optim", "adamw_torch"),
|
435 |
+
logging_steps=training_config.get("logging_steps", 10),
|
436 |
+
save_steps=training_config.get("save_steps", 200),
|
437 |
+
save_total_limit=training_config.get("save_total_limit", 3),
|
438 |
+
fp16=hardware_config.get("fp16", True),
|
439 |
+
bf16=hardware_config.get("bf16", False),
|
440 |
+
max_grad_norm=training_config.get("max_grad_norm", 0.3),
|
441 |
+
report_to=reports,
|
442 |
+
logging_first_step=training_config.get("logging_first_step", True),
|
443 |
+
disable_tqdm=training_config.get("disable_tqdm", False),
|
444 |
+
# Important: Don't remove columns that don't match model's forward method
|
445 |
+
remove_unused_columns=False
|
446 |
+
)
|
447 |
+
|
448 |
+
# Create trainer with pre-tokenized collator
|
449 |
+
trainer = Trainer(
|
450 |
+
model=model,
|
451 |
+
args=training_args,
|
452 |
+
train_dataset=training_dataset,
|
453 |
+
data_collator=PreTokenizedCollator(pad_token_id=tokenizer.pad_token_id, tokenizer=tokenizer),
|
454 |
+
)
|
455 |
+
|
456 |
+
# Start training
|
457 |
+
logger.info("Starting training - RESEARCH PHASE ONLY")
|
458 |
+
trainer.train()
|
459 |
+
|
460 |
+
# Save the model
|
461 |
+
logger.info(f"Saving model to {output_dir}")
|
462 |
+
trainer.save_model(output_dir)
|
463 |
+
|
464 |
+
# Save LoRA adapter separately for easier deployment
|
465 |
+
lora_output_dir = os.path.join(output_dir, "lora_adapter")
|
466 |
+
model.save_pretrained(lora_output_dir)
|
467 |
+
logger.info(f"Saved LoRA adapter to {lora_output_dir}")
|
468 |
+
|
469 |
+
# Save tokenizer for completeness
|
470 |
+
tokenizer_output_dir = os.path.join(output_dir, "tokenizer")
|
471 |
+
tokenizer.save_pretrained(tokenizer_output_dir)
|
472 |
+
logger.info(f"Saved tokenizer to {tokenizer_output_dir}")
|
473 |
+
|
474 |
+
# Copy config file for reference
|
475 |
+
with open(os.path.join(output_dir, "training_config.json"), "w") as f:
|
476 |
+
json.dump(config, f, indent=2)
|
477 |
+
|
478 |
+
logger.info("Training complete - RESEARCH PHASE ONLY")
|
479 |
+
return output_dir
|
480 |
+
|
481 |
+
finally:
|
482 |
+
# Always remove the training marker when done
|
483 |
+
remove_training_marker()
|
484 |
+
|
485 |
+
if __name__ == "__main__":
|
486 |
+
parser = argparse.ArgumentParser(description="Fine-tune Unsloth/DeepSeek-R1-Distill-Qwen-14B-4bit model (RESEARCH ONLY)")
|
487 |
+
parser.add_argument("--config", type=str, default="transformers_config.json",
|
488 |
+
help="Path to the transformers config JSON file")
|
489 |
+
parser.add_argument("--dataset", type=str, default="phi4-cognitive-dataset",
|
490 |
+
help="Dataset name or path")
|
491 |
+
parser.add_argument("--output_dir", type=str, default=None,
|
492 |
+
help="Output directory for the fine-tuned model")
|
493 |
+
|
494 |
+
args = parser.parse_args()
|
495 |
+
|
496 |
+
# Run training - Research phase only
|
497 |
+
try:
|
498 |
+
output_path = train(args.config, args.dataset, args.output_dir)
|
499 |
+
print(f"Research training completed. Model saved to: {output_path}")
|
500 |
+
except Exception as e:
|
501 |
+
logger.error(f"Training failed: {str(e)}")
|
502 |
+
remove_training_marker() # Clean up marker if training fails
|
503 |
+
raise
|