---

name: prompt-tuning description: "Operate Petals prompt-tuning, deep prompt-tuning, and adapter-aware training workflows without relying on the original repository checkout." disable-model-invocation: true metadata: disco-role: operating license: MIT

Petals Prompt Tuning

Use this sub-skill for distributed prompt tuning or deep prompt tuning on a Petals model, including causal-LM chat tuning, sequence-classification tuning, and prompt tuning on top of a preloaded PEFT/LoRA adapter.

Route generation-only work to client-inference, server adapter flags to server-swarms, and low-level LoRA/block conversion to distributed-blocks.

Procedure

1. Pick a recipe in references/workflows.md.
2. Choose tuning_mode="ptune" or "deep_ptune"; always set pre_seq_len > 0.
3. Load a distributed model constructor with pre_seq_len and tuning_mode.
4. Preprocess data into fixed-length input_ids and labels; omit zero-valued attention masks before Petals forwards.
5. Print trainable parameters. Expected trainables are prompt embeddings and, for classification, the classifier head.
6. Use a normal PyTorch optimizer over requires_grad parameters only.
7. Keep Hugging Face, dataset, W&B, CUDA, and swarm requirements explicit.

Read references/api-reference.md, references/troubleshooting.md, and use python scripts/prompt_tuning_skeleton.py --help for no-download planning.