local llm training

The problem

Fine-tuning LLMs on consumer hardware fails because of silent version mismatches and disk constraints. Standard scripts often assume standard environments, leading to VRAM fragmentation or broken imports on homelab nodes.

What it does

• Probes remote GPU nodes for live Python environment facts, disk availability, and resident VRAM usage before proposing a training plan.
• Orchestrates QLoRA and PEFT training stacks on high-capacity storage paths to bypass full system partitions.
• Enforces strict dependency pinning for Unsloth and Transformers to avoid common import and tokenizer errors.
• Manages training windows around hardware render blocks to prevent VRAM collisions with other services.

Frameworks & tools

PyTorch, PEFT, QLoRA, Unsloth, Transformers 4.56.1, Ollama, and SSH-based Linux environments.

Why this beats prompting it yourself

Standard LLMs will recommend "all-linear" target modules or latest package versions that break the Unsloth/PEFT dependency chain. This skill uses hard-won version pins and environment-specific disk pathing that general-purpose prompts miss.

Use cases

• Deploying a fine-tuning stack on a multi-disk GPU node with restricted system partitions.
• Fine-tuning Qwen3.5-9B models on generic corpora using QLoRA.
• Scheduling automated training jobs that respect hardware-locked rendering windows.

Known limitations

Does not support training on served quants like GGUF files. Requires manual user approval for package installations and model pulls on the target node.