| --- |
| license: apache-2.0 |
| base_model: unsloth/Qwen2.5-Coder-32B-Instruct-bnb-4bit |
| language: |
| - en |
| library_name: unsloth |
| tags: |
| - amd |
| - rocm |
| - hip |
| - cuda |
| - code-generation |
| - lablab-ai |
| - ghost-coder |
| - mi300x |
| - unsloth |
| --- |
| |
| # Ghost-Coder: Autonomous CUDA-to-HIP Translator 👻 |
|
|
| **Ghost-Coder** is a specialized, agent-ready LLM designed to bridge the gap between NVIDIA's CUDA and AMD's open ROCm ecosystem. Developed for the **Lablab.ai AMD Developer Hackathon (2026)**, this model serves as the "brain" of a self-healing agentic workflow that translates, compiles, and iterates on GPU kernels. |
|
|
| ## 🚀 Overview |
| Ghost-Coder isn't just a translator; it’s an engineer. By fine-tuning **Qwen2.5-Coder-32B** specifically on the **CASS (CUDA-to-HIP)** mapping dataset, we've enabled a model that understands the deep structural nuances of GPU programming, from shared memory primitives to warp-level synchronization. |
|
|
| ### 💎 Hardware & Framework |
| - **Training Hardware:** AMD Instinct™ MI300X VF (192GB HBM3) |
| - **Framework:** [Unsloth](https://github.com/unslothai/unsloth) (Optimized for 2x faster ROCm fine-tuning) |
| - **Optimization:** 4-bit QLoRA with a 4096 context window. |
|
|
| ## 🧠 Model Highlights |
| - **High-Fidelity Mapping:** Precise translation of `cuda*` APIs to their corresponding `hip*` counterparts. |
| - **Agentic Ready:** Optimized to parse `hipcc` compiler error logs and self-correct syntax or logic errors in real-time. |
| - **Massive Scale:** Leveraging the 32B parameter Qwen2.5-Coder foundation for superior C++ reasoning compared to smaller 7B models. |
|
|
| ## 🛠️ Training Specifications |
| To ensure maximum generalization and prevent overfitting, the model underwent a high-throughput training sprint: |
|
|
| | Parameter | Configuration | |
| | :--- | :--- | |
| | **Total Steps** | 200 (Optimized Sprint) | |
| | **Global Batch Size** | 64 | |
| | **Learning Rate** | 2e-4 | |
| | **VRAM Utilization** | ~158GB / 192GB | |
| | **Dataset** | 12,800+ Curated CUDA-to-HIP Pairs | |
|
|
| ## 🏁 Intended Use (The Ghost-Harness) |
| This model is designed to work within the **Ghost-Harness** agentic loop: |
| 1. **Input:** User provides a raw `.cu` (CUDA) file. |
| 2. **Action:** Ghost-Coder generates a `.cpp` (HIP) translation. |
| 3. **Validation:** The harness runs `hipcc` on the output. |
| 4. **Self-Healing:** If compilation fails, the error logs are fed back to Ghost-Coder for an iterative fix. |
|
|
| ## 📝 Acknowledgments |
| Special thanks to **AMD** for the world-class compute and **Lablab.ai** for hosting the "Build Across the AI Stack" challenge. |
|
|
| --- |
| *Developed by Talha* |
