[arnavkomaragiri]

PR Dependencies

Depends on: NVIDIA-NeMo/RL#4144.

Required by: NVIDIA-NeMo/RL#4146.

Summary

Carry exact-call and tree metadata through the transfer queue without broadcasting a logical episode padded to its concatenated length.

• Preserve non-tensor layouts through KV metadata slicing and DP sharding.
• Plan physical call/tree-fragment rows before sharding.
• Fetch and broadcast bounded microbatch payloads, reconstructing routing data against the same execution layout.
• Keep token-aligned and tree-edge-aligned field lengths separate.
• Return packed logprob rows for logical reassembly by the dependent policy layer.
• Add optional payload/padding diagnostics for transport memory review.

Ordinary unpacked rows retain the existing transport path. This layer does not enable packed rollout production on its own.

Stack

Depends on #4144. The model execution and rollout integration layers follow.

Validation

Matching preshard and leader-broadcast regressions are included. Prior integrated validation covered async TQ/R3/CP2 and six actual two-rank Gloo checks in the container; the clean 1x16 gate retained all generated tokens.

Login-side RL test collection is blocked by missing upstream nemo-lens. Scoped repository formatting hooks, Python parsing, privacy scanning, and whitespace checks are run here; native container tests/typechecking remain required before readiness.

Draft only. No benchmark recipes, datasets, private deployment details, or live-job changes are included. Tested RL base: 81115337c62ec8e2c34ff069d0659e406da6c601; latest-main reconciliation is pending.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.