ExoticsLabs/deepseek-r1-distill-qwen-1.5b-Fast

DeepSeek-R1 CPU Optimization & GGUF Conversion Guide

This project focuses on the high-performance optimization of the DeepSeek-R1-Distill-Qwen-1.5B model for local CPU inference. By bypassing standard library overhead and implementing a custom encoding layer, we achieve significant throughput improvements on consumer-grade hardware.

🚀 Project Overview

Standard LLM deployments often suffer from excessive memory overhead and slow tokenization when running on CPUs. This implementation solves these bottlenecks through:

1. Custom Raw Encoding: A library-free Byte-Level BPE processor optimized for DeepSeek/Qwen token artifacts.
2. Hardware-Level Threading: Explicit control over the PyTorch execution engine to match physical CPU core counts.
3. GGUF Quantization: Conversion to the GGUF format for industry-standard CPU inference via llama.cpp.

🛠 Technical Architecture

1. The Custom RawProcessor

The core of the speed optimization is the RawProcessor. Unlike standard tokenizers that rely on heavy Rust bindings or complex dependency chains, this processor handles the specific whitespace markers (Ġ for spaces and Ċ for newlines) natively.

• Mechanism: Greedy longest-match BPE traversal.
• Artifact Handling: Maps byte-level representations back to standard UTF-8 whitespace to ensure reasoning chains (like those in DeepSeek-R1) render correctly without specialized rendering logic.

2. CPU Inference Tuning

To prevent thread contention and maximize throughput:

• torch.set_num_threads() is dynamically set to the environment's physical core count.
• low_cpu_mem_usage=True is utilized during model instantiation to prevent memory spikes.
• The model is forced into float32 or bfloat16 depending on the instruction set availability (AVX-512/AMX).

📦 Deployment Formats

Native PyTorch (.pt)

The optimized state dictionary including the custom processing logic. Ideal for Python-based research environments.

GGUF Format (.gguf)

The model has been converted using a specialized HuggingFace-to-GGUF pipeline.

• Quantization: F16 (Floating Point 16) for maximum precision.
• Compatibility: Fully compatible with llama.cpp, Ollama, and LM Studio.
• Integrated Vocab: The custom BPE merge rules and special tokens (<｜User｜>, <｜Assistant｜>, and <think>) are baked into the GGUF metadata.

📈 Performance Metrics

• Latency: ~0.5s - 1.0s per token (hardware dependent).
• Thread Efficiency: Linear scaling observed up to 4 physical cores.
• Stability: Repetition penalties (1.2x) applied to prevent reasoning loop degradation.

📝 How to Use

1. Load the GGUF: Use any GGUF-compatible runner.
2. Prompt Template:

   <｜User｜>{prompt}<｜Assistant｜><think>
   

3. Inference: Ensure you set the thread count to match your CPU's physical (not logical) cores for optimal speed.

---

This documentation outlines the methodology for efficient local LLM deployment without relying on high-end GPU infrastructure.