Add On-Policy Distillation (OPSD) Trainer backend in DeepSpeed#8027

Open

PKUWZP wants to merge 8 commits into

deepspeedai:masterfrom

PKUWZP:zhipwang_opd_pr

Commits on Jul 2, 2026

Add OPSD example: config, divergence losses, utils + tests
PKUWZP
committed
Add OPSD frozen teacher with CPU logit cache + tests
PKUWZP
committed
Add OPSD trainer, hybrid-engine rollout, and end-to-end entry point
PKUWZP
committed
Add OPSD vLLM rollout scaffold, Qwen2/Qwen3 weight bridges, and README
PKUWZP
committed
feat(rollout): OPSD rollout engine with graph capture, vLLM backend

delock
authored and
PKUWZP
committed
Use ROLLOUT_VISIBLE_DEVICE env var for vLLM GPU placement; rename vllm_dtype to engine_dtype

delock
authored and
PKUWZP
committed
Fix formatting and CPU unit-test checks for OPSD rollout
PKUWZP
committed
Remove Microsoft Corporation copyright line from OPSD file headers
PKUWZP
committed