CoolFace
Apppublic

alst10/ROCmFPX-my-model

sourceHugging Facecc-by-4.0updated 23d agoView on Hugging Face
0likes
App README

🚀 ROCmFPX Model Converter (alst10/ROCmFPX-my-model)

This Hugging Face Space provides an automated pipeline and GUI interface to convert standard machine learning weights (Safetensors, Hugging Face repositories, standard GGUF binaries, and ONNX models) into experimental ROCmFPX GGUF formats.

The resulting GGUF binaries are optimized for inference on AMD architectures (such as Strix Halo, RDNA3/4, and CDNA).

✨ Features

  • —Hugging Face / Safetensors Support: Direct conversion from HF Repo IDs or local Safetensors weight directories.
  • —GGUF Re-quantization: Downscale or convert standard GGUF models into specialized ROCmFPX formats.
  • —ONNX Unpacking: Unpack ONNX weight initializers into standard format for ROCmFPX pipeline processing.
  • —Pure CPU Execution: Model quantization runs entirely on system CPU RAM. No GPU hardware is required for conversion.

📊 Supported ROCmFPX Formats

Quantization PresetTarget Bitwidth / BPWDescription
Q4_0_ROCMFP44.00 bpwRecommended: Native 4-bit ROCm FP format for AMD APUs/GPUs.
Q2_0_ROCMFPX2.50 bpwUltra-compact 2-bit format with dual UE4M3 scaling vectors.
Q6_0_ROCMFPX6.00 bpwHigh-precision 6-bit floating-point format.
Q8_0_ROCMFPX8.00 bpw8-bit precision with minimal quality degradation.

🛠️ Usage

  1. 1.Open the Gradio Web UI in your Space browser (https://huggingface.co/spaces/alst10/ROCmFPX-my-model).
  2. 2.Select your input tab:
  3. 3.Hugging Face / Safetensors: Paste the target model Repo ID (e.g., Qwen/Qwen2.5-1.5B-Instruct).
  4. 4.Existing GGUF: Upload your .gguf file directly.
  5. 5.ONNX: Upload your .onnx graph definition.
  6. 6.Choose your desired ROCmFPX Quant Preset (Q4_0_ROCMFP4, Q2_0_ROCMFPX, etc.).
  7. 7.Click Convert and download your compiled ROCmFPX .gguf file once complete.

📄 License & Component Attribution