← ClaudeAtlas

autodsparklisted

Train a DSpark speculative-decoding drafter for a user's own LLM with the AutoDSpark CLI, so the model generates faster in vLLM. Use when the user wants to speed up inference of a local or fine-tuned model with speculative decoding, train a DSpark / draft / speculator model, or asks about AutoDSpark runs, plans, logs or reports.
Zihao-Wu06/AutoDSpark · ★ 1 · AI & Automation · score 74
Install: claude install-skill Zihao-Wu06/AutoDSpark
# Train a DSpark drafter with AutoDSpark AutoDSpark plans and runs the whole pipeline (regenerate data with the target model, train the drafter with vLLM Speculators, benchmark, report). Your job is to drive it safely, explain what it is doing in plain words, and fix problems when a stage fails. ## Ground rules - Training uses GPUs for hours. **Show the plan and get the user's go-ahead before `autodspark run`** (anything other than `--dry-run`). Mention the preset, sample count and GPU layout. - Start with `--preset quick` on a new machine or model unless the user asks otherwise. It checks the pipeline in minutes; then offer `standard`. - Never delete a run folder, checkpoints or the `~/.cache/autodspark` environments without asking. - Run long commands in the background and follow the logs; don't block on them. - Only pass `--trust-remote-code` (or accept a plan with `trust_remote_code: true`) when the user trusts that model's code. ## Workflow 1. **Check the machine.** `autodspark doctor --model <model>` - No NVIDIA GPU: stop and tell the user training needs a Linux machine with NVIDIA GPUs. - Environments missing: ask, then run `autodspark setup` (needs `uv`; downloads vLLM and torch, several GB, and the Speculators scripts from GitHub). Safe to re-run after a network failure; existing environments are reused. - Work folder on a network filesystem: suggest `--workdir` on a local disk. 2. **Make the plan.** `autodspark plan --model <mode