Ollama Copilot Fixer

A small Python CLI for one narrow job: make a short list of strong local coding models work cleanly with GitHub Copilot on 24 GB-class consumer hardware.

The main focus is Unsloth GGUF releases that already run in Ollama but still need the right template, Modelfile, and runtime defaults to behave correctly in Copilot, especially for tool calling.

See the huggingface repo that hs the most recent coding models for 24GB hardware: https://huggingface.co/johnml1135/models

Current Focus

• Qwen3.5 and Qwen3.6 Unsloth GGUFs in Ollama
• Nemotron 3 Nano 30B A3B class local setups
• clean Copilot tool calling
• practical local runs on 24 GB-class machines

This repo is not trying to support every GGUF on Hugging Face. It is intentionally focused on a few local coding models that are worth getting right.

What The CLI Does

• download a GGUF from Hugging Face or use a local GGUF
• merge sharded GGUFs when needed
• generate a Copilot-safe Ollama Modelfile
• create the model in Ollama
• publish the fixed package back to Hugging Face

Most Tested Path

The most validated setup in this repo is:

• unsloth/Qwen3.6-35B-A3B-GGUF:UD-IQ4_XS
• context length 131072
• published example: johnml1135/qwen36-35b-a3b-ud-iq4xs-128k-github-copilot

For Qwen3.5 and Qwen3.6, the key fix is using an Ollama-style Qwen template and safe no-think defaults so Copilot sees clean responses and structured tool calls instead of raw <think> output.

Requirements

• Python 3.10+
• Ollama installed and running
• python -m pip install -r requirements.txt
• hf auth login if you need Hugging Face downloads or publishing
• llama.cpp only if you need to merge sharded GGUFs

Quick Start

Install dependencies:

python -m pip install -r requirements.txt

If you need Hugging Face access:

hf auth login

Create the validated Qwen3.6 model locally:

python -m ollama_copilot_fixer \
	--model-source "hf.co/unsloth/Qwen3.6-35B-A3B-GGUF:UD-IQ4_XS" \
	--model-name "qwen36-35b-a3b-ud-iq4xs-128k" \
	--context-length 131072 \
	--probe-template

Create a Nemotron local model:

python -m ollama_copilot_fixer \
	--model-source "hf.co/unsloth/Nemotron-3-Nano-30B-A3B-GGUF:Q4_0" \
	--model-name "nemotron-copilot"

Use a GGUF that is already on disk:

python -m ollama_copilot_fixer --model-source "C:\\Models\\model.gguf"

Publish A Fixed Model

The repo also includes a local publish flow that uploads:

• the GGUF
• a Copilot-compatible Modelfile
• a generated README.md
• the release card from releases/

Example:

python -m ollama_copilot_fixer publish \
	--gguf-path "C:\\path\\to\\Qwen3.6-35B-A3B-UD-IQ4_XS.gguf" \
	--release-card "releases\\qwen36-35b-a3b-ud-iq4xs-128k.md" \
	--context-length 131072

Use releases/_template.md as the starting point for new release cards.

If a publish is interrupted, rerun the same command. The uploader now stages files in the local cache and resumes chunk uploads automatically.

Useful Commands

python -m ollama_copilot_fixer --help
python -m ollama_copilot_fixer cache info
python -m ollama_copilot_fixer cache clear

Notes

• If a model comes as split GGUF shards, install llama.cpp and point --llama-cpp-path at llama-gguf-split.
• For Qwen3.5 and Qwen3.6, --probe-template is the safest path when trying a new quant or variant.
• Some Nemotron GGUFs still emit plain-text tool markup through Ollama. When that happens, NVIDIA NIM is often the better runtime.

Non-Goals

This repo does not retune weights, fix broken GGUF exports, or guarantee every upstream model will become Copilot-safe. It focuses on a small set of practical local coding models and making the Unsloth plus Ollama path work reliably with GitHub Copilot.