Repository navigation
Conversation
…CUDA graph Admit --enable-encoder-swa-bounded-replay with the breakable prefill backend (and attention DP). The fold's row floor, replay-row compressor skip and the request-window layout become bucket-sized metadata refreshed in place per replay; padding rows take a spare window group and never commit; the DP prefill-graph vote counts the folded replay rows.
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Motivation
--enable-encoder-swa-bounded-replayrejected every prefill CUDA graph backend, so DSv4.1 with encoder replay (admitted with DP attention since #43813) ran eager prefill. This admits the breakable backend, the same way #43736 did for the decoder replay.Modifications
disabled/breakable.encoder_swa_*fold fields.refresh_for_breakable_cuda_graph_replay_copiesrequest_window_layoutin place; under the graph the layout uses a fixed group count (req_to_token_pool.size) and the workspace is reserved at the first capture.window_layout(pad=...): padding rows (DP fill, graph bucket fill;out_loc == 0) take a spare group and never commit. Grouped with the last request they shortened its commit range and committed garbage at its ring slot 0; this was latent in the eager DP path too.DSV4Metadata.low_ratio_replay_rowsmask.Accuracy / Performance
Not yet run on GPU; CPU unit tests only (the new cases fail on main). Validation on DSv4.1 Flash with
--enable-encoder-swa-bounded-replay --cuda-graph-backend-prefill breakable --attn-dp-size Nis pending.Checklist
CI States
Latest PR Test (Base): ❌ Run #38985250998
Latest PR Test (Extra): ❌ Run #38985250955
Latest PR Test (AMD ROCm 10): ❌ Run #38985251059