perf(r3): expand replay routes to local layers only#1910
Open
dyurk-lila wants to merge 3 commits into
Open
Conversation
Fix routed-expert replay (R3) correctness for RL training and introduce a shared token-metadata layout that later routed-expert and sampler-support work builds on. - Scope global RouterReplay state to one Megatron pipeline schedule so a forward-only logprob pass can no longer leak backward replay state into the next training schedule (clear before the schedule and in `finally`). - Keep `rollout_expert_indices` ragged and treat its length as the captured-prefix length. Derive a `router_padding_mask` after left padding that marks alignment padding and the uncaptured trajectory suffix, and carry it through the training data, replay experiences, microbatch padding, and the Megatron model call. - Build one `TokenMetadataLayout` per microbatch and apply it to both routes and the padding mask. Generic construction, alignment, next-token shifting, and packed-output restoration live in `skyrl/utils/token_metadata.py`. - Pass Megatron's `padding_mask` through the model and apply a narrow compatibility shim so `[tokens]` masks broadcast over experts in expert-bias accounting. - Slice every per-trajectory generator field generically during dynamic-sampling replacement and filtering so route metadata stays attached to its trajectory. Synthetic padding rows use distinct dummy experts `[0, ..., topk - 1]`; the mask excludes them from expert-bias accounting while preserving Megatron's dropless `tokens * topk` dispatcher invariant.
Store routed-expert (R3) generation data as compact NumPy arrays instead of large nested Python lists, and send it over the network base64-encoded alongside its shape and dtype. Expert IDs are compacted to the smallest safe uint8/int16/int32 dtype, vLLM responses and client responses use orjson, and preprocessing accepts the decoded NumPy route arrays directly. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Under pipeline parallelism each rank replays only its local router layers, but the Megatron worker eagerly expanded the full global-layer routed-expert tensor to int32 before replay setup, allocating a large device temporary for unused layers. Keep routed-expert IDs in their compact dtype through whole-batch device movement, index_select the current PP stage's router layers before metadata alignment, and perform the single int32 conversion inside _split_replay_indices so only the bounded PP-local slice is materialized as int32. Also validate the 4D replay-indices shape up front. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This was referenced Jul 16, 2026
Contributor
|
Warning Gemini encountered an error creating the review. You can try again by commenting |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
Routed-expert IDs are stored in a compact integer dtype, but the Megatron worker
previously expanded the entire global-layer tensor to
int32before replay setup.Under pipeline parallelism, each rank consumes only its local router layers, so that
eager expansion allocates a large
int32temporary on device for layers the ranknever uses.
Implementation
device movement (drop the eager
.to(torch.int32)in the Megatron worker forwardpaths).
index_selectonly those layers before metadata alignment, via a new
_get_local_router_layer_indiceshelper (patchedlayer_numberwhen available,offset-based fallback otherwise).
int32conversion inside_split_replay_indices,immediately before installing local replay data, so only the bounded PP-local
slice is ever materialized as
int32.setup_per_microbatch_replay_forward.This leaves
Experience.to_device(), batch generation, and normal batch ownershipunchanged. Part of a small series of routed-expert replay improvements.
Testing
Replay tests cover packed multi-row alignment, PP-local layer selection before
layout, compact
uint8/int16/int32sources, requiredint32replay data,padding-mask alignment, and sequence-parallel mask layout.
py_compile, Ruff(0.11.9), and Black (24.10.0) all pass on the changed files.