Vedisasi's picture
Upload folder using huggingface_hub
54c5666 verified
# ULTRATHINK Environment Configuration
# Copy this file to .env and customize for your environment
# ===== Training Directories =====
ULTRATHINK_OUTPUT_DIR=./outputs
ULTRATHINK_CACHE_DIR=./cache
ULTRATHINK_DATA_DIR=./data
ULTRATHINK_CHECKPOINT_DIR=./checkpoints
# ===== Distributed Training =====
# Set these when using torchrun or accelerate
MASTER_ADDR=localhost
MASTER_PORT=29500
WORLD_SIZE=1
RANK=0
LOCAL_RANK=0
# TensorBoard
TENSORBOARD_LOG_DIR=./runs
# ===== GPU Settings =====
# Specify which GPUs to use (comma-separated)
CUDA_VISIBLE_DEVICES=0,1,2,3
# CUDA memory allocation
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
# ===== Performance Settings =====
# Number of CPU threads for data loading
OMP_NUM_THREADS=8
# Tokenizers parallelism (set to false if you see warnings)
TOKENIZERS_PARALLELISM=false
# Torch compile mode
TORCHDYNAMO_DISABLE=1
# ===== Hugging Face =====
# Hugging Face Hub token for private models/datasets
HF_TOKEN=your_huggingface_token_here
HF_HOME=./hf_cache
# ===== DeepSpeed =====
# DeepSpeed configuration
DEEPSPEED_CONFIG=./deepspeed_config_zero2.json
# ===== Security =====
# Maximum file upload size (MB)
MAX_FILE_SIZE_MB=1000
# Allowed checkpoint directories (colon-separated)
ALLOWED_CHECKPOINT_DIRS=./checkpoints:./outputs
# ===== Development =====
# Logging level
LOG_LEVEL=INFO # DEBUG, INFO, WARNING, ERROR
# Debug mode
DEBUG=false
# ===== Model Defaults =====
# Default model configuration
DEFAULT_MODEL_SIZE=small
DEFAULT_BATCH_SIZE=8
DEFAULT_SEQ_LENGTH=512
DEFAULT_LEARNING_RATE=3e-4
# ===== Dataset Settings =====
# Default dataset
DEFAULT_DATASET=wikitext
DATASET_CACHE_DIR=./dataset_cache
# Streaming settings
DATASET_STREAMING=false
DATASET_NUM_PROC=4