gpt/configs

Configuration Files

This directory contains configuration files for different GPT versions and training scenarios.

Version-Specific Configurations

• gpt2_config.py: Configuration for GPT-2 models (124M-1.5B parameters)
• gpt3_config.py: Configuration for GPT-3 models (175B parameters)
• gpt3.5_config.py: Configuration for GPT-3.5 models (~350B parameters)
• gpt4_config.py: Configuration for GPT-4 models with MoE architecture
• gpt4.5_config.py: Configuration for GPT-4.5 models with enhanced MoE
• gpt5_config.py: Configuration for theoretical GPT-5 models with unlimited context

Configuration Parameters

Common Parameters

• block_size: Maximum context length
• vocab_size: Vocabulary size
• n_layer: Number of transformer layers
• n_head: Number of attention heads
• n_embd: Embedding dimension
• dropout: Dropout probability
• bias: Whether to use bias in Linear and LayerNorm layers

MoE-Specific Parameters

• use_moe: Whether to use Mixture of Experts
• num_experts: Total number of experts
• num_active_experts: Number of experts active per token
• expert_capacity_factor: Capacity factor for load balancing

Unlimited Context Parameters

• billion_token_context: Enable billion-token context support
• use_sparse_attention: Use sparse attention mechanisms
• preserve_begin_tokens: Number of beginning tokens to preserve
• preserve_end_tokens: Number of end tokens to preserve
• token_eviction_strategy: Strategy for evicting tokens

Usage

Configurations are loaded through the version manager:

from utils.version_manager import load_model_config

# Load configuration for a specific version
config = load_model_config('gpt4')

# Access and modify configuration parameters
print(f"Context length: {config.block_size}")
config.block_size = 16384  # Override context length

For training, configurations can be specified as command-line arguments:

python train.py --config configs/gpt4_config.py --block_size=32768