Repository navigation
Conversation
The tail skip used to turn off whenever a helper was present. A file-tier miss is still skipped. A row a helper holds, or one whose bytes are already in the host copy, is still computed. With no helper the old skip stays. A peer GPU still disables the skip. The settings page matches that rule.
sakuumomo
marked this pull request as ready for review
October 21, 2026 07:16
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 and result: The file-tier tail skip turned off whenever a helper GPU was attached. A file-tier miss is skipped again while a helper is attached. A row the helper holds, or a row whose bytes are already in the host copy, is still computed. With no helper, a missed tail is skipped even when those bytes are already in RAM. A peer GPU still turns the skip off. The startup line and
docs/DETAILS.mdsay the same thing.Related discussion: none for this helper case. The skip itself is the existing
STRATA_ROUTE_TAIL_SKIPdefault.Approach: The dispatch no longer treats
remote_count > 0as a reason to disable the skip. A helper still refuses the skip for a row it holds and for a rowhost_copy()reports as already in a host copy (a resident complement, or an arena that is not a demand-paged file mapping). Peer GPUs keep the old disable. The default rank is unchanged. The SYCL port compiles its owngenerate.cppandexpert_source.cpp, so this pull request does not change that backend.Validation:
Passed: HIP and CUDA builds of
4c1f0ddaon Linux x86-64, through Nix, withSTRATA_BUILD_TESTS=OFFandSTRATA_PORTABLE=ON.HIP:
STRATA_ENABLE_HIP=ON,STRATA_ENABLE_CUDA=OFF,CMAKE_HIP_ARCHITECTURES=gfx1030,STRATA_PREFILL_MMQ=ON, HIP 7.2.3. Output/nix/store/c6zp3p826fryl6x7cgp3lsw16z56him8-strata-hip-0.2.1(strataandstrata-device). Thestratabinary contains thegfx1030name.CUDA:
STRATA_ENABLE_CUDA=ON,STRATA_ENABLE_HIP=OFF,STRATA_HYBRIDleft off,CMAKE_CUDA_ARCHITECTURES=89-real,STRATA_PREFILL_MMQ=OFF, CUDA 12.9.86, gcc 14.4.0. Output/nix/store/yr652vys91l94sj7dvxigsz6a2ynjgpy-strata-cuda-0.2.1(strataandstrata-device).The author served on two Linux machines, each with a helper GPU, and reports that both function as intended. The desktop is an RX 6800 and an RX 6900 XT on the HIP build. The laptop is an RTX 4060 and a Radeon 890M on the CUDA and HIP hybrid build. No prompt log or timing was attached to that report.
Failed: none observed.
Not run: a SYCL build of this commit, because the port does not compile these files. A new measurement of the skip.
Limits: The startup line no longer prints the +10..20% figure while a helper is attached, because that figure was measured without one.