Speculative Decoding Draft-Model Training

Everything needed to take a target model from "no draft head" to "validated acceptance rate" lives in this directory. Work through the stages below in order for a new model; jump straight to a stage when you already know which one you need.

Two axes: stage and algorithm

The pipeline is the same shape for every draft-model algorithm — synthesize data, dump base-model hidden states, train the draft, benchmark acceptance rate. What changes between algorithms is which training script and recipe run, which knobs matter, and which failures are typical.

So this skill is split along those two axes, and you almost always read one file from each:

Read the stage file for what to do, and the algorithm sheet for the values to plug in. When a stage file says "see the algorithm sheet", it means the section of references/algorithms/<algorithm>.md with the matching heading.

Where the code lives

The launcher scripts in tools/launcher/common/{eagle3,specdec}/ are thin wrappers around examples/speculative_decoding/ — main.py (training), collect_hidden_states/ (dumps), scripts/{export_hf_checkpoint,ar_validate,quantize_drafter}.py. Read a wrapper for the flags it passes, that directory for what they do. Its README.md is also the non-launcher path: training invoked directly, without Slurm.

Stages

Review-logs and triage overlap by design: review-logs is the fast sweep across all tasks, triage is the deep dive into one failing task. Start with review-logs unless the user already knows which task broke.

Algorithms

All recipes live in modelopt_recipes/general/speculative_decoding/<algorithm>.yaml.

Everything below EAGLE3 is a DFlash variant, not a separate pipeline: same recipe_type: speculative_dflash, same training script, same dflash.* config namespace, selected by dflash_architecture_config.projector_type. Read references/algorithms/dflash.md first, then the variant's sheet for the delta.

LiLiCorr ships two recipes — lilicorr.yaml and lilicorr_conv.yaml, the latter adding DFlash2's sublayer convolutions — but one sheet covers both.

If the user's algorithm has no sheet yet, the stage procedures still apply — derive the missing values from an existing launcher example for that algorithm (tools/launcher/examples/*/*/hf_*_<algorithm>.yaml) and its recipe, then write the sheet as you go. references/algorithms/README.md defines what a sheet must contain.

End-to-end: a new model

1. Confirm the algorithm and find the closest existing launcher example.
2. Configure — write the pipeline YAML (references/stages/configure.md).

3. Preview with --dryrun, then submit:

   cd tools/launcher
   uv run launch.py --yaml examples/<Org>/<Model>/<config>.yaml --yes

4. Register the job and set up monitoring per the modelopt-monitor skill.
5. Review logs when it finishes (references/stages/review-logs.md).
6. Triage anything that failed (references/stages/triage.md), fix, re-run only the failed tasks onward via pipeline.task_N.skip=true.
7. Validate once all tasks pass (references/stages/validate.md).

Model-support gaps that need code changes land in modelopt/torch/speculative/ and require a separate ModelOpt PR — the pipeline YAML alone cannot fix an unrecognized architecture.