File size: 9,061 Bytes
7275aef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
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
226
227
228
229
230
231
232
233
234
235
236
237
238
239
# pipelines/distributed_trainer.py

import torch
import torch.distributed as dist
from torch.utils.data import DataLoader, DistributedSampler
from transformers import (
    AutoTokenizer, AutoModelForCausalLM, 
    TrainingArguments, Trainer, DataCollatorForLanguageModeling,
    BitsAndBytesConfig
)
from peft import prepare_model_for_kbit_training, LoraConfig, get_peft_model
from distributed_utils import RankZeroOnly
from typing import Dict, List, Tuple, Optional
import json
from pathlib import Path

class DistributedProductionTrainer:
    """Production trainer with proper distributed training support"""
    
    def __init__(self, config: Dict):
        self.config = config
        self.output_dir = Path("runs/humigence")
        self.output_dir.mkdir(parents=True, exist_ok=True)
        
        # Distributed training setup
        self.ddp = config.get("ddp", False)
        self.rank = config.get("rank", 0)
        self.world_size = config.get("world_size", 1)
        self.is_main = config.get("is_main", True)
        self.device = torch.device(config.get("device", "cuda:0"))
        
        # Training configuration
        self.base_model = config["base_model"]
        self.training_recipe = config["training_recipe"]
        self.learning_rate = float(config.get("learning_rate", "2e-4"))
        self.num_epochs = int(config.get("num_train_epochs", "1"))
        self.batch_size = int(config.get("per_device_train_batch_size", "2"))
        self.gradient_accumulation = int(config.get("gradient_accumulation_steps", "4"))
        
        # Initialize components
        self.tokenizer = None
        self.model = None
        self.trainer = None
        
    def load_model_and_tokenizer(self):
        """Load model and tokenizer with proper device placement"""
        with RankZeroOnly(self.is_main) as rank_zero:
            rank_zero.print(f"[blue]🤖 Loading model: {self.base_model}[/blue]")
        
        # Load tokenizer
        self.tokenizer = AutoTokenizer.from_pretrained(self.base_model, trust_remote_code=True)
        self.tokenizer.pad_token = self.tokenizer.eos_token
        
        # Load model with proper device placement
        if self.ddp:
            # For DDP, load model to CPU first, then move to device
            self.model = AutoModelForCausalLM.from_pretrained(
                self.base_model,
                device_map=None,  # Load to CPU
                trust_remote_code=True,
                torch_dtype=torch.bfloat16 if "BF16" in self.training_recipe else torch.float16
            )
            # Move to device
            self.model = self.model.to(self.device)
        else:
            # Single GPU training
            self.model = AutoModelForCausalLM.from_pretrained(
                self.base_model,
                device_map="auto",
                trust_remote_code=True,
                torch_dtype=torch.bfloat16 if "BF16" in self.training_recipe else torch.float16
            )
        
        # Apply LoRA if needed
        if "LoRA" in self.training_recipe:
            self._apply_lora()
        
        # Setup DDP if needed
        if self.ddp:
            self.model = torch.nn.parallel.DistributedDataParallel(
                self.model,
                device_ids=[self.device.index],
                output_device=self.device.index
            )
            with RankZeroOnly(self.is_main) as rank_zero:
                rank_zero.print(f"[blue]✅ Model wrapped with DDP (rank {self.rank})[/blue]")
    
    def _apply_lora(self):
        """Apply LoRA configuration to the model"""
        lora_config = LoraConfig(
            r=16,
            lora_alpha=32,
            target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
            lora_dropout=0.1,
            bias="none",
            task_type="CAUSAL_LM"
        )
        
        self.model = get_peft_model(self.model, lora_config)
        
        with RankZeroOnly(self.is_main) as rank_zero:
            rank_zero.print("[blue]✅ LoRA configuration applied[/blue]")
    
    def prepare_datasets(self, train_data: List[Dict], val_data: List[Dict], test_data: List[Dict]):
        """Prepare datasets with distributed sampling"""
        # Convert to datasets
        from datasets import Dataset
        
        train_dataset = Dataset.from_list(train_data)
        val_dataset = Dataset.from_list(val_data)
        test_dataset = Dataset.from_list(test_data)
        
        # Tokenize datasets
        def tokenize_function(examples):
            return self.tokenizer(
                examples["text"],
                truncation=True,
                padding=False,
                max_length=512
            )
        
        train_dataset = train_dataset.map(tokenize_function, batched=True)
        val_dataset = val_dataset.map(tokenize_function, batched=True)
        test_dataset = test_dataset.map(tokenize_function, batched=True)
        
        # Create data collator
        self.data_collator = DataCollatorForLanguageModeling(
            tokenizer=self.tokenizer,
            mlm=False
        )
        
        # Create distributed samplers if needed
        if self.ddp:
            self.train_sampler = DistributedSampler(
                train_dataset,
                num_replicas=self.world_size,
                rank=self.rank,
                shuffle=True
            )
            self.val_sampler = DistributedSampler(
                val_dataset,
                num_replicas=self.world_size,
                rank=self.rank,
                shuffle=False
            )
        else:
            self.train_sampler = None
            self.val_sampler = None
        
        self.train_dataset = train_dataset
        self.val_dataset = val_dataset
        self.test_dataset = test_dataset
    
    def setup_training(self):
        """Setup training arguments and trainer"""
        # Training arguments
        training_args = TrainingArguments(
            output_dir=str(self.output_dir),
            per_device_train_batch_size=self.batch_size,
            per_device_eval_batch_size=self.batch_size,
            gradient_accumulation_steps=self.gradient_accumulation,
            num_train_epochs=self.num_epochs,
            learning_rate=self.learning_rate,
            logging_steps=10,
            save_steps=100,
            eval_steps=100,
            evaluation_strategy="steps",
            save_strategy="steps",
            load_best_model_at_end=True,
            metric_for_best_model="eval_loss",
            greater_is_better=False,
            ddp_find_unused_parameters=False,  # Important for DDP
            remove_unused_columns=False,
            dataloader_pin_memory=True,
            dataloader_num_workers=4,
        )
        
        # Create trainer
        self.trainer = Trainer(
            model=self.model,
            args=training_args,
            train_dataset=self.train_dataset,
            eval_dataset=self.val_dataset,
            data_collator=self.data_collator,
            tokenizer=self.tokenizer,
        )
        
        # Set samplers for distributed training
        if self.ddp:
            self.trainer.train_dataloader.sampler = self.train_sampler
            self.trainer.eval_dataloader.sampler = self.val_sampler
    
    def train(self):
        """Run training with proper distributed handling"""
        with RankZeroOnly(self.is_main) as rank_zero:
            rank_zero.print("[blue]🚀 Starting training...[/blue]")
        
        # Train the model
        self.trainer.train()
        
        # Save model (only on main process)
        if self.is_main:
            self.trainer.save_model()
            with RankZeroOnly(self.is_main) as rank_zero:
                rank_zero.print("[blue]💾 Model saved[/blue]")
        
        # Synchronize all processes
        if self.ddp:
            dist.barrier()
        
        return {"status": "success", "message": "Training completed"}
    
    def evaluate(self):
        """Run evaluation with proper distributed handling"""
        with RankZeroOnly(self.is_main) as rank_zero:
            rank_zero.print("[blue]🧪 Running evaluation...[/blue]")
        
        # Run evaluation
        eval_results = self.trainer.evaluate()
        
        # Gather results from all ranks if DDP
        if self.ddp:
            # Gather evaluation results
            gathered_results = [None] * self.world_size
            dist.all_gather_object(gathered_results, eval_results)
            
            # Average results across ranks
            if self.is_main:
                avg_results = {}
                for key in eval_results.keys():
                    if isinstance(eval_results[key], (int, float)):
                        values = [r[key] for r in gathered_results if r is not None]
                        avg_results[key] = sum(values) / len(values)
                    else:
                        avg_results[key] = eval_results[key]
                eval_results = avg_results
        
        return eval_results