SmolFactory / config /train_smollm3_h100_lightweight.py
Tonic's picture
attempt to fix bfloat16 issue
6c63876 verified
"""
SmolLM3 H100 Lightweight Training Configuration
Optimized for rapid training on H100 with 80K Hermes-FR samples
"""
import os
from dataclasses import dataclass
from typing import Optional
from config.train_smollm3 import SmolLM3Config
@dataclass
class SmolLM3ConfigH100Lightweight(SmolLM3Config):
"""Configuration for SmolLM3 fine-tuning on OpenHermes-FR dataset - H100 Lightweight"""
# Model configuration - optimized for H100
model_name: str = "HuggingFaceTB/SmolLM3-3B"
max_seq_length: int = 8192 # Increased for better context understanding
use_flash_attention: bool = True
use_gradient_checkpointing: bool = True # Enabled for memory efficiency
# Training configuration - H100 optimized for rapid training
batch_size: int = 16 # Larger batch size for H100
gradient_accumulation_steps: int = 4 # Reduced for faster updates
learning_rate: float = 8e-6 # Slightly higher for rapid convergence
weight_decay: float = 0.01
warmup_steps: int = 50 # Reduced warmup for rapid training
max_iters: int = None # Will be calculated based on epochs
eval_interval: int = 50 # More frequent evaluation
log_interval: int = 5 # More frequent logging
save_interval: int = 200 # More frequent saving
# Optimizer configuration - optimized for rapid training
optimizer: str = "adamw_torch"
beta1: float = 0.9
beta2: float = 0.95
eps: float = 1e-8
# Scheduler configuration - faster learning
scheduler: str = "cosine"
min_lr: float = 2e-6 # Higher minimum LR
# Mixed precision - Using fp16 for better compatibility
# Note: bf16 can cause issues on some GPU setups, fp16 is more universally supported
fp16: bool = False
bf16: bool = True
# Logging and saving - more frequent for rapid training
save_steps: int = 200
eval_steps: int = 50
logging_steps: int = 5
save_total_limit: Optional[int] = 2 # Keep fewer checkpoints
# Evaluation
eval_strategy: str = "steps"
metric_for_best_model: str = "eval_loss"
greater_is_better: bool = False
load_best_model_at_end: bool = True
# OpenHermes-FR Dataset configuration with sampling
dataset_name: str = "legmlai/openhermes-fr"
dataset_split: str = "train"
input_field: str = "prompt"
target_field: str = "completion"
filter_bad_entries: bool = False
bad_entry_field: str = "bad_entry"
sample_size: int = 80000 # 80K samples for lightweight training
sample_seed: int = 42 # For reproducibility
# Data configuration (not used for HF datasets but kept for compatibility)
data_dir: str = "my_dataset"
train_file: str = "train.json"
validation_file: Optional[str] = "validation.json"
test_file: Optional[str] = None
# Chat template configuration
use_chat_template: bool = True
chat_template_kwargs: dict = None
# Trackio monitoring configuration
enable_tracking: bool = True
trackio_url: Optional[str] = None
trackio_token: Optional[str] = None
log_artifacts: bool = True
log_metrics: bool = True
log_config: bool = True
experiment_name: Optional[str] = None
# HF Datasets configuration
hf_token: Optional[str] = None
dataset_repo: Optional[str] = None
# H100-specific optimizations
dataloader_num_workers: int = 4 # Optimized for H100
dataloader_pin_memory: bool = True
dataloader_prefetch_factor: int = 2
# Memory optimizations for rapid training
max_grad_norm: float = 1.0
group_by_length: bool = True # Group similar length sequences
# Training duration calculations
# With 80k datapoints and effective batch size of 64:
# Steps per epoch = 80,000 / 64 = 1,250 steps
# For 1 epoch: 1,250 steps
# For 2 epochs: 2,500 steps
def __post_init__(self):
if self.chat_template_kwargs is None:
self.chat_template_kwargs = {
"enable_thinking": False,
"add_generation_prompt": True,
"no_think_system_message": True
}
# Validate configuration
if self.fp16 and self.bf16:
raise ValueError("Cannot use both fp16 and bf16")
if self.max_seq_length > 131072: # 128k limit
raise ValueError("max_seq_length cannot exceed 131072")
# Calculate training statistics
effective_batch_size = self.batch_size * self.gradient_accumulation_steps
steps_per_epoch = self.sample_size // effective_batch_size # For 80k dataset
epochs_for_max_iters = self.max_iters / steps_per_epoch if self.max_iters else 1
print(f"=== H100 Lightweight Training Configuration ===")
print(f"Effective batch size: {effective_batch_size}")
print(f"Steps per epoch: ~{steps_per_epoch}")
print(f"Training for ~{epochs_for_max_iters:.1f} epochs")
print(f"Total training steps: {self.max_iters or 'auto'}")
print(f"Learning rate: {self.learning_rate}")
print(f"Mixed precision: {'fp16' if self.fp16 else 'bf16'}")
print(f"Max sequence length: {self.max_seq_length}")
print(f"Gradient checkpointing: {self.use_gradient_checkpointing}")
print(f"Dataset sample size: {self.sample_size}")
print("=" * 50)
# Set default experiment name if not provided
if self.experiment_name is None:
self.experiment_name = "smollm3_h100_lightweight"
def get_config(config_path: str) -> SmolLM3ConfigH100Lightweight:
"""Load configuration from file or return default"""
if os.path.exists(config_path):
# Load from file if it exists
import importlib.util
spec = importlib.util.spec_from_file_location("config_module", config_path)
config_module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(config_module)
if hasattr(config_module, 'config'):
return config_module.config
else:
# Try to find a config class
for attr_name in dir(config_module):
attr = getattr(config_module, attr_name)
if isinstance(attr, SmolLM3ConfigH100Lightweight):
return attr
# Return default configuration
return SmolLM3ConfigH100Lightweight()
# Default configuration instance
config = SmolLM3ConfigH100Lightweight()