Add OPSD (On-Policy Distillation) training example#1002

Merged

delock merged 25 commits into

Jul 15, 2026

Commits on Jul 2, 2026

Add OPSD (On-Policy Distillation) training example
delock
committed
Use ROLLOUT_VISIBLE_DEVICE env var for vLLM GPU placement; rename vllm_dtype to engine_dtype
delock
committed

Commits on Jul 3, 2026

Remove vLLM path, absorb trainer/config/losses/utils/benchmarks from DeepSpeed
delock
committed
Move teacher.py and data.py from DeepSpeed to DeepSpeedExamples
delock
committed
Extend RolloutConfig with app-level generation knobs; clean vLLM remnants from JSON
delock
committed
Keep only bench_decode_1p1r; add --graph-capture flag and engine wrapper fix
delock
committed
Fix distillation temperature from 0 to 1.0 in smoke and production configs
delock
committed
Fix trailing commas in JSON configs; remove unused smoke_ds_zero3.json
delock
committed

Commits on Jul 12, 2026

Commits on Jul 13, 2026