feat(fsdp): add bounded sampler-support replay - #1914
Open
dyurk-lila wants to merge 9 commits into
Open
Conversation
Contributor
|
Warning Gemini encountered an error creating the review. You can try again by commenting |
dyurk-lila
force-pushed
the
upstream/sample-support-fsdp
branch
from
July 17, 2026 23:14
20d9a75 to
fa43c6e
Compare
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>
Extract the single-request HTTP generation path out of RemoteInferenceClient into a standalone RemoteInferenceGenerator and a RemoteGenerateResult dataclass. RemoteInferenceClient now owns an internal generator and delegates session management, _post, and _generate_single to it. This is a pure refactor with no functionality change: endpoint routing, retry/backoff, cache_salt handling, serialization, and lifecycle behavior are all preserved. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
For multi-turn rollouts, build routed-expert (R3) trace data incrementally as the conversation grows instead of re-gathering the whole conversation's routes on every turn. - Add `routed_experts_prompt_starts` to `InferenceEngineInput` and thread a per-request `routed_experts_prompt_start` through `RemoteInferenceClient` and `RemoteInferenceGenerator.generate()`, forwarded as a sampling param so the engine only returns routes for the newly generated suffix. - Introduce `TokenMetadataTrace` (token-aligned array accumulator) and `RoutedExpertTrace`, which records each generation's routes and finalizes a full per-token routed-expert array with loss-mask-aware terminal padding. - Track a `RoutedExpertTrace` on `AgentLoopState` and record routes per turn, replacing the previous whole-conversation re-gather in `SkyRLGymGenerator.agent_loop`. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add the core of bounded sampler-support replay across data plumbing, inference capture, and the Megatron dense-path trainer. - Inference: /skyrl/v1/generate gains a return_sample_support flag that requests flat_logprobs from vLLM; the fixed-width rows are decoded into per-token support IDs and surfaced through RemoteInferenceGenerator, RemoteGenerateResult, and InferenceEngineOutput.rollout_sample_support. - Data plumbing: support rows ride the shared TokenMetadataTrace through multi-turn accumulation, dynamic filtering, truncation, replay buffering, and microbatch padding; preprocessing validates support width, trailing -1 padding, int32 vocab IDs, and presence of every loss-bearing token. - Config: reject sampler settings whose rollout distribution cannot be replayed exactly (temperature > 0, top_k > 1, repetition penalty 1.0, no arbitrary additional_kwargs). - Megatron: renormalize sampled-token scores over recorded support with fixed-shape gathers; fused LM-head path projects only selected candidate pairs and chunks by logprobs_chunk_size; controller-packed rows reuse TokenMetadataLayout. New utils/sample_support_replay.py holds the core. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…om its top-k support (NovaSky-AI#66) * fix(sample-support): repair rows where the sampled token is absent from its top-k support vLLM's approximate top-k/top-p pivot can let a sampled survivor rank just beyond top_k, so it is missing from the captured support row and the downstream sample-support invariant hard-crashes the run. When the sampled id is absent, overwrite the weakest support member with it, preserving width==top_k, single-occurrence, and trailing padding. Emit one aggregated warning per call. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * style(sample-support): apply black formatting to regression test --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Extend bounded sampler-support replay to the FSDP backend so policy and reference forwards apply the same support-conditioned normalization as Megatron, keeping recomputed logprobs on the same distribution. - Enable replay for trainer.strategy=fsdp with the existing sampler-support constraints; thread support IDs and the response loss mask through policy training, policy recomputation, and reference forwards only when enabled. - Right-align the response loss mask into the full token layout and keep sampled IDs, support IDs, and the mask aligned through padding removal, next-token shifting, and Ulysses SP slicing; reuse padding-removal indices. - Generalize Ulysses input partitioning to [batch, seq, ...] metadata with a configurable padding value, avoiding a second validity-mask partition. - Share the backend-neutral aligned replay helper for support normalization and synthetic-EOS handling; preserve feature-disabled behavior. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
dyurk-lila
force-pushed
the
upstream/sample-support-fsdp
branch
from
July 20, 2026 20:25
fa43c6e to
eefcccc
Compare
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
2 participants
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
The bounded sampler-support replay feature records vLLM's sampler support and replays it during training. The Megatron backend already applies this support-conditioned normalization; the FSDP policy and reference forwards must do the same, or their recomputed logprobs will represent a different distribution.
Implementation
trainer.strategy=fsdpwhile retaining the sampler-support constraints.[batch, seq, ...]metadata and allow support padding to use-1, avoiding a second full validity-mask partition.A loss-bearing EOS may be appended after vLLM returns, so that token intentionally has an empty support row. FSDP computes its ordinary full-vocabulary logprob explicitly; other loss-bearing rows must have recorded support. Loss-masked empty rows remain neutral.
The trainer representation stays the dense
[batch, seq, support_width]contract.Testing
pytest tests/backends/skyrl_train/distributed/test_ulysses_token_metadata.py tests/backends/skyrl_train/utils/test_torch_utils.py tests/backends/skyrl_train/workers/test_fsdp_dense_sample_support.py tests/train/test_config.pygit diff --checkpass.