§ 02How it works
Four commands. A running GPU.
Describe the workload, set a budget. slipa picks the cheapest live spot GPU, provisions it, runs your container, handles eviction recovery, and hands back the result.
# install + connect
$ pip install slipa
$ slipa configure
# boot a GPU, prove CUDA, mark complete. ~$0.05.
$ slipa run hello-gpu --duration 60
# fine-tune an open LLM with a per-run budget ceiling.
$ slipa run finetune \
--model meta-llama/Llama-3.1-8B \
--dataset tatsu-lab/alpaca \
--method qlora \
--max-budget 5.00
# stream a prompt dataset through a model; predictions to R2.
$ slipa run batch-inference \
--model meta-llama/Llama-3.1-8B \
--dataset-url r2://bucket/prompts.jsonl \
--max-budget 2.00
# RLHF rollouts: N completions per prompt, optional reward scoring.
$ slipa run rl-rollouts \
--policy-model meta-llama/Llama-3.1-8B \
--prompts-url r2://bucket/prompts.jsonl \
--completions-per-prompt 8 \
--max-budget 4.00
# follow, download.
$ slipa logs <id>
$ slipa download <id>