/claim #82 /fixes #82

This PR implements an offline DeepSeek model loader for inference as requested in the feature request.

Features

• Loads DeepSeek models directly from HuggingFace
• Supports both full and quantized versions
• Implements memory optimization techniques
• Dynamically detects model file structure
• Supports multiple model formats (.safetensors, .bin, .pt, .ckpt)

Implementation Details

1. Created a modular architecture with separate components:

   • DeepSeekLoader: Core loading functionality with memory optimization
   • DeepSeekTokenizer: Text encoding/decoding
   • DeepSeekWrapper: High-level interface following project patterns

2. Implemented memory optimization techniques:

   • Chunk-based loading to reduce memory footprint
   • Int8 quantization for reduced memory usage
   • Efficient tensor management with device control

3. Added dynamic model discovery:

   • Automatically detects model file structure
   • Supports different weight file formats
   • Handles various tokenizer configurations

4. Created comprehensive tests:

   • Model initialization and loading tests
   • Chat interface tests
   • Code generation tests
   • Quantization tests
   • Memory efficiency tests

5. Added example usage for easy integration

Components

• DeepSeekLoader: Handles model loading with memory optimization
• DeepSeekTokenizer: Handles tokenization for input/output
• DeepSeekWrapper: Provides a unified interface
• Tests: Verify model loading, quantization, and memory efficiency
• Example: Demonstrates usage

Dependencies

• torch
• safetensors
• huggingface_hub
• numpy

This implementation avoids high-level libraries like transformers as requested.