ScrapeGoat-Music-Stage1 / train_hcf.py
Nathan9's picture
Upload 8 files
103c8f5 verified
import os
import json
import torch
import logging
from pathlib import Path
from dataclasses import dataclass
from typing import Optional, List, Dict, Tuple, Any
import transformers
from transformers import (
AutoModelForCausalLM,
AutoTokenizer,
TrainingArguments,
Trainer,
DataCollatorForLanguageModeling
)
from datasets import Dataset, load_dataset
import numpy as np
from accelerate import Accelerator
from safetensors import safe_open
from safetensors.torch import save_file, load_file
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
@dataclass
class TensorInfo:
"""Stores metadata about tensor indices and shape"""
shape: Tuple[int, ...]
dtype: str
indices: Optional[torch.Tensor] = None
hcf_patterns: Optional[Dict] = None
class SafeTensorHCFAnalyzer:
"""
Analyzes HCF patterns in model weights using SafeTensors format.
Handles efficient loading and analysis of large model weights.
"""
def __init__(self, tolerance: float = 1e-5):
self.tolerance = tolerance
self.tensor_info = {}
self.metadata = {}
def load_safetensor_file(self,
filepath: str,
device: str = 'cpu',
load_indices: bool = True) -> Dict[str, TensorInfo]:
"""
Load and parse a SafeTensor file with proper memory management.
Args:
filepath: Path to .safetensors file
device: Device to load tensors to
load_indices: Whether to load weight indices
Returns:
Dictionary mapping tensor names to their metadata
"""
try:
# First load metadata only to check structure
with safe_open(filepath, framework="pt") as f:
self.metadata = json.loads(f.metadata()) if f.metadata() else {}
# Load tensors efficiently
tensors = load_file(filepath, device=device)
for tensor_name, tensor in tensors.items():
self.tensor_info[tensor_name] = TensorInfo(
shape=tuple(tensor.shape),
dtype=str(tensor.dtype)
)
# Load indices if available in metadata
if load_indices and tensor_name in self.metadata:
if 'indices' in self.metadata[tensor_name]:
indices_data = self.metadata[tensor_name]['indices']
if isinstance(indices_data, list):
self.tensor_info[tensor_name].indices = torch.tensor(
indices_data, device=device
)
elif isinstance(indices_data, str) and os.path.exists(indices_data):
# Load indices from separate file if provided as path
self.tensor_info[tensor_name].indices = torch.load(indices_data)
return self.tensor_info
except Exception as e:
raise RuntimeError(f"Error loading SafeTensor file: {str(e)}")
def analyze_safetensor_weights(self,
filepath: str,
batch_size: int = 1000) -> Dict:
"""
Analyze weights from SafeTensor file in memory-efficient batches.
Args:
filepath: Path to .safetensors file
batch_size: Number of weights to process at once
Returns:
Analysis results including HCF patterns and optimization opportunities
"""
results = {
'tensor_hcfs': {},
'shared_patterns': [],
'optimization_suggestions': [],
'memory_impact': {}
}
# Process tensors in batches
with safe_open(filepath, framework="pt") as f:
for tensor_name in f.keys():
# Get tensor info
tensor_data = f.get_tensor(tensor_name)
tensor_size = np.prod(tensor_data.shape)
if tensor_name in self.tensor_info and self.tensor_info[tensor_name].indices is not None:
indices = self.tensor_info[tensor_name].indices
unique_indices = torch.unique(indices)
# Process each index group
tensor_hcfs = {}
for idx in unique_indices:
mask = (indices == idx)
indexed_weights = tensor_data[mask]
# Process in batches if needed
if len(indexed_weights) > batch_size:
hcf = self._process_large_weight_group(indexed_weights, batch_size)
else:
hcf = self._calculate_hcf(indexed_weights)
tensor_hcfs[idx.item()] = hcf
results['tensor_hcfs'][tensor_name] = tensor_hcfs
# Find optimization opportunities
patterns = self._analyze_weight_patterns(tensor_data, indices)
self.tensor_info[tensor_name].hcf_patterns = patterns
# Calculate potential memory savings
savings = self._estimate_memory_savings(patterns, tensor_data.dtype)
results['memory_impact'][tensor_name] = {
'original_size': tensor_size * tensor_data.element_size(),
'potential_savings': savings
}
# Find shared patterns across tensors
results['shared_patterns'] = self._find_shared_patterns()
results['optimization_suggestions'] = self._generate_optimization_suggestions(results)
return results
def _calculate_hcf(self, weights: torch.Tensor) -> float:
"""Calculate HCF for a tensor of weights, with tolerance for floating point"""
# Implementation placeholder - actual implementation would depend on specific needs
if len(weights) == 0:
return 0.0
return 1.0 # Simplified for example
def _gcd_float(self, a: float, b: float) -> float:
"""Calculate greatest common divisor for floating point numbers"""
# Implementation placeholder
return min(a, b) # Simplified for example
def _process_large_weight_group(self,
weights: torch.Tensor,
batch_size: int) -> float:
"""Process large weight groups in batches to manage memory."""
current_hcf = None
for i in range(0, len(weights), batch_size):
batch = weights[i:i + batch_size]
batch_hcf = self._calculate_hcf(batch)
if current_hcf is None:
current_hcf = batch_hcf
elif batch_hcf > self.tolerance:
current_hcf = self._gcd_float(current_hcf, batch_hcf)
return current_hcf if current_hcf is not None else 0.0
def _analyze_weight_patterns(self,
weights: torch.Tensor,
indices: torch.Tensor) -> Dict:
"""Analyze weight patterns within indexed groups."""
patterns = {}
unique_indices = torch.unique(indices)
for idx in unique_indices:
mask = (indices == idx)
pattern_weights = weights[mask]
patterns[idx.item()] = {
'mean': float(pattern_weights.mean()),
'std': float(pattern_weights.std()),
'size': len(pattern_weights),
'hcf': self._calculate_hcf(pattern_weights)
}
return patterns
def _estimate_memory_savings(self, patterns: Dict, dtype: torch.dtype) -> int:
"""Estimate potential memory savings from patterns"""
# Implementation placeholder
return sum(p['size'] for p in patterns.values()) // 2 # Simplified estimate
def _find_shared_patterns(self) -> List[Dict]:
"""Find patterns that could be shared across tensors."""
shared_patterns = []
pattern_groups = {}
for tensor_name, info in self.tensor_info.items():
if info.hcf_patterns:
for idx, pattern in info.hcf_patterns.items():
# Create pattern signature
signature = f"{pattern['mean']:.4f}_{pattern['std']:.4f}"
if signature not in pattern_groups:
pattern_groups[signature] = []
pattern_groups[signature].append({
'tensor': tensor_name,
'index': idx,
'pattern': pattern
})
# Find groups with similar patterns
for signature, group in pattern_groups.items():
if len(group) > 1:
shared_patterns.append({
'signature': signature,
'occurrences': group,
'potential_savings': sum(p['pattern']['size'] for p in group[1:])
})
return shared_patterns
def _generate_optimization_suggestions(self, results: Dict) -> List[Dict]:
"""Generate optimization suggestions based on analysis"""
# Implementation placeholder
suggestions = []
for tensor_name, impact in results['memory_impact'].items():
if impact['potential_savings'] > 1000000: # If savings > 1MB
suggestions.append({
'tensor': tensor_name,
'suggestion': 'Consider weight quantization',
'impact': f"Save {impact['potential_savings'] / 1024 / 1024:.2f}MB"
})
return suggestions
@dataclass
class TrainingStatistics:
"""Statistics collected during HCF-aware training"""
memory_savings: int = 0
quantization_error: float = 0.0
convergence_rate: float = 0.0
epoch: int = 0
batch_count: int = 0
def update(self, batch_stats: Dict[str, Any]):
"""Update statistics with batch results"""
self.memory_savings += batch_stats.get('memory_savings', 0)
self.quantization_error = batch_stats.get('quantization_error', self.quantization_error)
self.convergence_rate = batch_stats.get('convergence_rate', self.convergence_rate)
self.batch_count += 1
class HCFTrainingOptimizer(torch.optim.Adam):
"""
Optimizer with HCF-awareness for more efficient training
"""
def __init__(self,
params,
lr=0.001,
betas=(0.9, 0.999),
eps=1e-8,
weight_decay=0,
weight_quantization=True,
maintain_patterns=True):
super().__init__(params, lr, betas, eps, weight_decay)
self.weight_quantization = weight_quantization
self.maintain_patterns = maintain_patterns
self.analyzer = SafeTensorHCFAnalyzer()
self.stats = {'memory_savings': 0, 'quantization_error': 0.0}
def step(self, closure=None):
"""Perform optimization step with HCF awareness"""
# Run standard optimization step
loss = super().step(closure)
# Apply HCF optimizations if enabled
if self.weight_quantization:
self._apply_weight_quantization()
if self.maintain_patterns:
self._maintain_weight_patterns()
return loss
def _apply_weight_quantization(self):
"""Apply dynamic weight quantization using HCF patterns"""
savings = 0
total_error = 0.0
for group in self.param_groups:
for p in group['params']:
if p.grad is None or not p.requires_grad:
continue
# Apply weight quantization logic based on HCF analysis
# This is a simplified placeholder - real implementation would be more complex
if p.dim() > 1: # Only apply to matrices/tensors
# Find suitable quantization factor
factor = torch.max(torch.abs(p.data)) / 127 # 8-bit quantization example
# Quantize weights
quantized = torch.round(p.data / factor) * factor
# Calculate error and savings
error = torch.mean((p.data - quantized)**2).item()
savings += p.numel() * (p.element_size() - 1) # Assuming 8-bit savings
# Apply quantized weights
p.data.copy_(quantized)
total_error += error
# Update statistics
self.stats['memory_savings'] = savings
self.stats['quantization_error'] = total_error
def _maintain_weight_patterns(self):
"""Maintain efficient weight patterns identified by HCF analysis"""
# Placeholder for pattern maintenance logic
# Real implementation would analyze weight matrices and enforce patterns
pass
def get_stats(self):
"""Get current optimization statistics"""
return self.stats
class HCFAwareTrainer:
"""
Trainer that incorporates HCF analysis for better training efficiency
"""
def __init__(self, model, optimizer):
self.model = model
self.optimizer = optimizer
self.analyzer = SafeTensorHCFAnalyzer()
def train_epoch(self, train_loader, criterion, epoch):
"""Train one epoch with HCF awareness"""
self.model.train()
stats = TrainingStatistics(epoch=epoch)
for batch_idx, batch in enumerate(train_loader):
# Get data
inputs, targets = self._prepare_batch(batch)
# Forward pass
self.optimizer.zero_grad()
outputs = self.model(inputs)
loss = criterion(outputs, targets)
# Backward pass
loss.backward()
# Optimize with HCF awareness
self.optimizer.step()
# Get batch statistics
batch_stats = self.optimizer.get_stats()
stats.update(batch_stats)
# Log progress
if batch_idx % 50 == 0:
logger.info(f"Epoch {epoch} | Batch {batch_idx}/{len(train_loader)} | "
f"Memory Savings: {stats.memory_savings/1024/1024:.2f}MB | "
f"Quantization Error: {stats.quantization_error:.6f}")
# End of epoch analysis
self._analyze_model_weights()
return stats
def _prepare_batch(self, batch):
"""Prepare batch data for training"""
# Implementation depends on dataset structure
if isinstance(batch, dict):
inputs = batch.get('input_ids')
targets = batch.get('labels', inputs)
else:
# Assume batch is a tuple of (inputs, targets)
inputs, targets = batch
return inputs, targets
def _analyze_model_weights(self):
"""Analyze model weights for patterns and optimizations"""
# Save model to temporary safetensor file for analysis
model_path = "temp_model.safetensors"
tensors = {name: param for name, param in self.model.named_parameters()}
save_file(tensors, model_path)
# Analyze weights
results = self.analyzer.analyze_safetensor_weights(model_path)
# Log findings
logger.info(f"Weight Analysis: Found {len(results['shared_patterns'])} shared patterns")
logger.info(f"Potential memory savings: "
f"{sum(i['potential_savings'] for i in results['memory_impact'].values())/1024/1024:.2f}MB")
# Clean up
if os.path.exists(model_path):
os.remove(model_path)
@dataclass
class ModelConfig:
name: str
model_id: str
tokenizer_id: str
CONFIGS = {
"7b": ModelConfig(
name="7b",
model_id="scrapegoat/ScrapeGoat-Music-Stage1",
tokenizer_id="scrapegoat/ScrapeGoat-Music-Stage1"
),
"1b": ModelConfig(
name="1b",
model_id="scrapegoat/ScrapeGoat-Music-Stage2",
tokenizer_id="scrapegoat/ScrapeGoat-Music-Stage2"
)
}
class MusicFineTuner:
def __init__(
self,
model_size: str,
dataset_path: str,
output_dir: str,
device: str = "auto",
batch_size: int = 4,
gradient_accumulation_steps: int = 4,
learning_rate: float = 1e-5,
num_epochs: int = 3,
use_hcf: bool = True
):
self.config = CONFIGS[model_size]
self.dataset_path = Path(dataset_path)
self.output_dir = Path(output_dir)
self.device = self._setup_device(device)
self.use_hcf = use_hcf
self.training_args = TrainingArguments(
output_dir=str(self.output_dir),
per_device_train_batch_size=batch_size,
gradient_accumulation_steps=gradient_accumulation_steps,
learning_rate=learning_rate,
num_train_epochs=num_epochs,
logging_steps=100,
save_steps=1000,
evaluation_strategy="steps",
eval_steps=500,
save_total_limit=3,
load_best_model_at_end=True,
gradient_checkpointing=True,
fp16=torch.cuda.is_available(),
optim="adamw_torch"
)
def _setup_device(self, device: str) -> str:
if device == "auto":
if torch.cuda.is_available():
return "cuda"
elif torch.backends.mps.is_available():
return "mps"
else:
return "cpu"
return device
def _load_model_and_tokenizer(self):
logger.info(f"Loading model {self.config.model_id}")
# Determine dtype based on device
dtype = torch.bfloat16 if self.device == "cuda" else torch.float32
model = AutoModelForCausalLM.from_pretrained(
self.config.model_id,
torch_dtype=dtype,
device_map="auto" if self.device == "cuda" else None,
attn_implementation="flash_attention_2" if self.device == "cuda" else "eager"
)
tokenizer = AutoTokenizer.from_pretrained(self.config.tokenizer_id)
return model, tokenizer
def _prepare_dataset(self, tokenizer):
logger.info("Preparing dataset")
with open(self.dataset_path / "metadata" / "dataset_info.json") as f:
metadata = json.load(f)
def generate_text(item):
return f"Genre: {item['genre']}\nDuration: {item['duration']:.2f}s\nTitle: {item['title']}\nArtist: {item['artist']}\n"
texts = [generate_text(item) for item in metadata["files"]]
dataset = Dataset.from_dict({"text": texts})
def tokenize(examples):
return tokenizer(
examples["text"],
truncation=True,
padding="max_length",
max_length=512,
return_tensors="pt"
)
tokenized_dataset = dataset.map(
tokenize,
batched=True,
remove_columns=dataset.column_names
)
return tokenized_dataset
def train(self):
# Create output directory
self.output_dir.mkdir(parents=True, exist_ok=True)
# Load model and tokenizer
model, tokenizer = self._load_model_and_tokenizer()
# Prepare dataset
dataset = self._prepare_dataset(tokenizer)
# Split dataset
dataset = dataset.train_test_split(test_size=0.1)
if self.use_hcf:
logger.info("Using HCF-aware training")
# Create custom HCF optimizer
optimizer = HCFTrainingOptimizer(
model.parameters(),
lr=self.training_args.learning_rate,
weight_quantization=True,
maintain_patterns=True
)
# Create HCF trainer
hcf_trainer = HCFAwareTrainer(model, optimizer)
# Create custom training loop
train_loader = torch.utils.data.DataLoader(
dataset["train"],
batch_size=self.training_args.per_device_train_batch_size,
shuffle=True
)
# Training loop with HCF awareness
criterion = torch.nn.CrossEntropyLoss()
for epoch in range(int(self.training_args.num_train_epochs)):
stats = hcf_trainer.train_epoch(train_loader, criterion, epoch)
# Log training metrics
logger.info(f"Epoch {epoch} completed")
logger.info(f"Memory Savings: {stats.memory_savings/1024/1024:.2f}MB")
logger.info(f"Quantization Error: {stats.quantization_error:.6f}")
logger.info(f"Convergence Rate: {stats.convergence_rate:.4f}")
# Save checkpoint
self._save_hcf_checkpoint(model, tokenizer, epoch)
else:
# Use standard HuggingFace Trainer
logger.info("Using standard training")
trainer = Trainer(
model=model,
args=self.training_args,
train_dataset=dataset["train"],
eval_dataset=dataset["test"],
data_collator=DataCollatorForLanguageModeling(tokenizer, mlm=False),
)
# Train
logger.info("Starting training")
trainer.train()
# Save final model
logger.info("Saving model")
model.save_pretrained(str(self.output_dir / "final_model"))
tokenizer.save_pretrained(str(self.output_dir / "final_model"))
def _save_hcf_checkpoint(self, model, tokenizer, epoch):
"""Save checkpoint with HCF metadata"""
checkpoint_dir = self.output_dir / f"checkpoint-{epoch}"
checkpoint_dir.mkdir(exist_ok=True)
# Save model and tokenizer
model.save_pretrained(str(checkpoint_dir))
tokenizer.save_pretrained(str(checkpoint_dir))
# Analyze and save HCF metadata
analyzer = SafeTensorHCFAnalyzer()
# Save tensors to analyze
model_path = str(checkpoint_dir / "model.safetensors")
if os.path.exists(model_path):
results = analyzer.analyze_safetensor_weights(model_path)
# Save analysis results
with open(checkpoint_dir / "hcf_analysis.json", "w") as f:
json.dump(results, f, indent=2)
logger.info(f"Saved checkpoint at {checkpoint_dir}")
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--model_size", type=str, choices=["1b", "7b"], required=True)
parser.add_argument("--dataset_path", type=str, required=True)
parser.add_argument("--output_dir", type=str, required=True)
parser.add_argument("--device", type=str, default="auto")
parser.add_argument("--batch_size", type=int, default=4)
parser.add_argument("--gradient_accumulation_steps", type=int, default=4)
parser.add_argument("--learning_rate", type=float, default=1e-5)
parser.add_argument("--num_epochs", type=int, default=3)
parser.add_argument("--use_hcf", action="store_true", help="Enable HCF-aware training")
args = parser.parse_args()
fine_tuner = MusicFineTuner(
model_size=args.model_size,
dataset_path=args.dataset_path,
output_dir=args.output_dir,
device=args.device,
batch_size=args.batch_size,
gradient_accumulation_steps=args.gradient_accumulation_steps,
learning_rate=args.learning_rate,
num_epochs=args.num_epochs,
use_hcf=args.use_hcf
)
fine_tuner.train()