HuggingFace Model Transfer Workflows

This repository provides GitHub Actions workflows for downloading, converting, and archiving machine learning models.

Workflows

1. Transfer HuggingFace Models (transfer.yml)

Downloads a complete HuggingFace model repository and uploads it as a GitHub release.

Use case: Archive or mirror a model from HuggingFace to GitHub.

Inputs:

• model_name (required): HuggingFace model repository name (e.g., meta-llama/Llama-2-7b-hf)
• split_size (optional): Archive split size in MB (default: 800, max ~2GB per GitHub release asset)
• hf_token (optional): HuggingFace token for private models

Example:

model_name: unsloth/Qwen3.5-8B-GGUF
split_size: 800

---

2. Transfer GGUF Models (gguf.yml)

Downloads specific GGUF files from HuggingFace using file patterns, with optional merge of split files.

Use case: Download specific quantization variants (e.g., Q4_K_M) from GGUF model repos.

Inputs:

• model_name (required): HuggingFace model repo (e.g., unsloth/Qwen3.5-27B-GGUF)
• file_filter (required): Glob pattern(s) for files (e.g., *Q4_K_M.*, or *Q4_K_M.*;*Q8_0.* for multiple)
• split_size (optional): Archive split size in MB (default: 800)
• merge_gguf (optional): Set to true to merge split GGUF files (e.g., model-00001-of-00002.gguf + model-00002-of-00002.gguf → single file)

Example:

model_name: unsloth/Qwen3.5-27B-GGUF
file_filter: *Q4_K_M.*
merge_gguf: true
split_size: 800

---

3. Ollama to GGUF (ollama2gguf.yml)

Pulls a model from Ollama registry, converts it to GGUF format, and uploads to GitHub.

Use case: Convert Ollama models (e.g., phi3:mini, llama3.2:latest) to GGUF format for use with llama.cpp.

Inputs:

• model_name (required): Ollama model name (e.g., qwen2.5:7b, phi3:mini, llama3.2:latest)
• split_size (optional): Archive split size in MB (default: 800)

Example:

model_name: qwen2.5:7b
split_size: 800

---

How to Use

1. Navigate to the Actions tab in your GitHub repository
2. Select the desired workflow from the left sidebar
3. Click Run workflow
4. Fill in the required inputs
5. Click Run workflow again to start

Notes

• All workflows run on windows-latest runners
• Large models are split into 7z volumes to stay under GitHub's 2GB asset limit
• Releases include SHA256 checksums where applicable
• For private HuggingFace models, ensure HF_TOKEN secret is configured in repository settings