Training Machine Provisioning

The training-machine-provisioning skill stands up an AWS EC2 instance and environment to LoRA fine-tune JoyAI-VL-Interaction-Preview into a custom proactive speak/silent specialist — watch a video stream and speak up only for a specific target event you describe, staying silent otherwise. It parallels Getting Started's inference-machine-provisioning skill, but provisions a machine sized for training rather than serving.

Create a new training machine

This stands up a g6e.2xlarge (1× L40S, 48 GB VRAM, 64 GB RAM) by default — the smallest single-GPU shape that satisfies both the GPU and system-RAM guidance derived for this task — then installs LLaMA-Factory (the same SFT framework upstream's own recipe is built with) plus MLflow for experiment tracking, and clones the base checkpoint. It stops there.

This provisions the machine and environment, not the pipeline or the dataset

This skill only gets the box ready — no fine-tuning code lands on it yet, and no training data exists. Both come from the follow-up lora-finetune skill.

note

g6e.2xlarge is a derived, not benchmarked default — see the skill's references/hardware-and-cost.md for the VRAM/RAM math it's derived from.

What's Next

Once the machine and environment are ready, the lora-finetune skill picks up from here: given a local video and a one-line target-event description, it prepares the labeled dataset, smoke-tests the pipeline, and runs the actual LoRA fine-tune — with a live MLflow dashboard to watch it progress and compare against past runs.