[Sampling Replay 3/8] Replay sampling support in actor scoring

[nanjiangwill]

Stack position

Part 3/8 of the bounded sampling-support replay series. Parts 1/8 and 2/8 are merged. #3321 has merged SGLang v0.5.20 into main.

Layer-only diff: 12754e9...nanjiangwill:feat/sampling-replay-actor-scoring

Review umbrella: #3354.

Summary

• Reconstruct transported rollout sampling supports on the trainer.
• Apply support normalization only to actor scores in Megatron and FSDP.
• Keep reference, teacher, value, and full-vocabulary entropy semantics unchanged.
• Preserve generic Megatron forward callbacks that do not consume sampling support.

Why this layer is separate

This PR is the trainer-side correctness boundary. It lets reviewers evaluate TP/CP mask construction, actor-only attribution, and gradients without also reviewing HTTP and session-server behavior.

The default unbounded sampling path remains unchanged. SGLang capture, session propagation, user-facing validation, GPU CI, and documentation are separate layers in the stack.

Validation

Focused tests cover Megatron and FSDP actor selection, transport reconstruction, tensor-parallel support normalization, gradients, entropy, and generic forward callbacks.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[nanjiangwill]

Closing this layer-specific review PR in favor of the single integrated sampling-replay PR #3354. The implementation and validation remain available there.