feat: Establish initial kLLM project structure with core configurations, dependencies, and Docker setup.

Dockerfile

@@ -0,0 +1,38 @@
+FROM python:3.10-slim
+
+# Set working directory
+WORKDIR /app
+
+# Install system dependencies
+RUN apt-get update && apt-get install -y \
+    git \
+    git-lfs \
+    && rm -rf /var/lib/apt/lists/* \
+    && git lfs install
+
+# Copy requirements first for caching
+COPY requirements.txt .
+
+# Install Python dependencies
+RUN pip install --no-cache-dir -r requirements.txt
+
+# Copy application code
+COPY . .
+
+# Create necessary directories
+RUN mkdir -p /app/checkpoints /app/data
+
+# Set environment variables
+ENV PYTHONUNBUFFERED=1
+ENV GRADIO_SERVER_NAME=0.0.0.0
+ENV GRADIO_SERVER_PORT=7860
+
+# Expose port for Gradio
+EXPOSE 7860
+
+# Health check
+HEALTHCHECK --interval=30s --timeout=30s --start-period=5s --retries=3 \
+    CMD curl -f http://localhost:7860/ || exit 1
+
+# Run the Gradio app
+CMD ["python", "app.py"]

requirements.txt

@@ -0,0 +1,34 @@
+# kLLM - Custom LLM Training & Inference Dependencies
+# Hugging Face Ecosystem
+transformers>=4.36.0
+datasets>=2.16.0
+accelerate>=0.25.0
+huggingface_hub>=0.20.0
+
+# Training with RLHF
+trl>=0.7.0
+peft>=0.7.0
+
+# Quantization & Memory Optimization
+bitsandbytes>=0.41.0
+
+# Deep Learning
+torch>=2.1.0
+safetensors>=0.4.0
+
+# Web Interface & API
+gradio>=4.0.0
+fastapi>=0.104.0
+uvicorn>=0.24.0
+
+# Configuration
+pyyaml>=6.0
+python-dotenv>=1.0.0
+
+# Monitoring (optional)
+wandb>=0.16.0
+tensorboard>=2.15.0
+
+# Utilities
+numpy>=1.24.0
+tqdm>=4.66.0

src/__init__.py

@@ -0,0 +1,2 @@
+"""kLLM - Custom LLM Training Package"""
+__version__ = "0.1.0"

src/config.py

@@ -0,0 +1,145 @@
+"""
+kLLM Configuration
+Central configuration for model training and inference.
+"""
+import os
+from dataclasses import dataclass, field
+from typing import Optional, List
+
+# Hugging Face Configuration
+HF_TOKEN = os.getenv("HF_TOKEN", "")
+HF_USERNAME = "imkrish"
+MODEL_REPO = f"{HF_USERNAME}/kLLM"
+CHECKPOINT_REPO = f"{HF_USERNAME}/kLLM-checkpoints"
+
+# Base Model Options
+BASE_MODELS = {
+    "tinyllama": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
+    "phi2": "microsoft/phi-2",
+    "mistral": "mistralai/Mistral-7B-Instruct-v0.2",
+    "llama2": "meta-llama/Llama-2-7b-chat-hf",
+}
+
+# Default base model
+DEFAULT_BASE_MODEL = "tinyllama"
+
+
+@dataclass
+class ModelConfig:
+    """Model architecture configuration."""
+    base_model: str = BASE_MODELS[DEFAULT_BASE_MODEL]
+    model_name: str = "kLLM"
+    max_length: int = 2048
+    load_in_4bit: bool = True
+    load_in_8bit: bool = False
+    trust_remote_code: bool = True
+    use_flash_attention: bool = False
+
+
+@dataclass
+class LoRAConfig:
+    """LoRA (Low-Rank Adaptation) configuration for efficient fine-tuning."""
+    r: int = 16  # LoRA rank
+    lora_alpha: int = 32  # LoRA alpha
+    lora_dropout: float = 0.05
+    bias: str = "none"
+    task_type: str = "CAUSAL_LM"
+    target_modules: List[str] = field(default_factory=lambda: [
+        "q_proj", "k_proj", "v_proj", "o_proj",
+        "gate_proj", "up_proj", "down_proj"
+    ])
+
+
+@dataclass
+class SFTConfig:
+    """Supervised Fine-Tuning configuration."""
+    output_dir: str = "./checkpoints/sft"
+    num_train_epochs: int = 3
+    per_device_train_batch_size: int = 4
+    per_device_eval_batch_size: int = 4
+    gradient_accumulation_steps: int = 4
+    learning_rate: float = 2e-4
+    weight_decay: float = 0.01
+    warmup_ratio: float = 0.03
+    lr_scheduler_type: str = "cosine"
+    max_seq_length: int = 2048
+    
+    # Logging & Saving
+    logging_steps: int = 10
+    save_strategy: str = "epoch"  # Save every epoch
+    save_total_limit: int = 5
+    evaluation_strategy: str = "epoch"
+    
+    # Push to Hub
+    push_to_hub: bool = True
+    hub_model_id: str = f"{HF_USERNAME}/kLLM-sft"
+    hub_token: str = HF_TOKEN
+    
+    # Mixed precision
+    fp16: bool = False
+    bf16: bool = True  # Better for modern GPUs
+
+
+@dataclass
+class RewardConfig:
+    """Reward Model configuration for RLHF."""
+    output_dir: str = "./checkpoints/reward"
+    num_train_epochs: int = 1
+    per_device_train_batch_size: int = 4
+    gradient_accumulation_steps: int = 4
+    learning_rate: float = 1e-5
+    
+    # Save settings
+    save_strategy: str = "epoch"
+    push_to_hub: bool = True
+    hub_model_id: str = f"{HF_USERNAME}/kLLM-reward"
+
+
+@dataclass 
+class RLHFConfig:
+    """RLHF/PPO Training configuration."""
+    output_dir: str = "./checkpoints/rlhf"
+    
+    # PPO specific
+    ppo_epochs: int = 4
+    mini_batch_size: int = 4
+    batch_size: int = 16
+    learning_rate: float = 1e-5
+    
+    # KL penalty
+    init_kl_coef: float = 0.2
+    target_kl: float = 6.0
+    
+    # Reward scaling
+    reward_baseline: float = 0.0
+    
+    # Training
+    num_train_epochs: int = 1
+    max_steps: int = 1000
+    
+    # Checkpointing - save every epoch
+    save_strategy: str = "epoch"
+    save_steps: int = 100
+    push_to_hub: bool = True
+    hub_model_id: str = f"{HF_USERNAME}/kLLM"
+
+
+@dataclass
+class InferenceConfig:
+    """Inference configuration."""
+    model_path: str = MODEL_REPO
+    max_new_tokens: int = 512
+    temperature: float = 0.7
+    top_p: float = 0.9
+    top_k: int = 50
+    repetition_penalty: float = 1.1
+    do_sample: bool = True
+
+
+# Global configs
+model_config = ModelConfig()
+lora_config = LoRAConfig()
+sft_config = SFTConfig()
+reward_config = RewardConfig()
+rlhf_config = RLHFConfig()
+inference_config = InferenceConfig()

src/data/__init__.py

@@ -0,0 +1 @@
+"""Data processing modules for kLLM."""

src/inference/__init__.py

@@ -0,0 +1 @@
+"""Inference modules for kLLM."""

src/training/__init__.py

@@ -0,0 +1 @@
+"""Training modules for kLLM."""