Repository navigation
hrx: pin llama.cpp to the shared-memory FA tile visibility (engine#314/315) - #347
Merged
Merged
Conversation
bong-water-water-bong
marked this pull request as ready for review
October 7, 2026 11:08
bong-water-water-bong
enabled auto-merge (squash)
October 7, 2026 11:09
…4/315) Moves third_party/llama.cpp cf4dfd801e -> 26330cc20, a merge that keeps the engine#314 emission fix and drops the cost it introduced. Context: cf4dfd801e removed the two unordered subgroup collectives in the decode-split FA kernel by forcing the staging path on every key tile (the engine#314 always-stage fallback). Its own commit message noted it "pays the staging loop and two subgroup barriers on every key tile instead of only the tiles that need masking". The merged commit takes f9ff06675 (open as 1bit-MONSTER/llama.cpp#95) instead, which keeps the conditional staging and derives the visibility predicates from shared memory, so no collective remains: real subgroup.vote in the kernel: 0 merged kernel file is byte-identical to f9ff06675 Measured on Strix Halo with the pressure harness (3 neighbour servers so each arm runs at gtt_start ~102 GiB, 66 requests per arm, control and guard-off arms paired inside one box state): on cf4dfd801e the guard-off arm faulted with all-NaN logits (count=154880 of 154880, contiguous=yes, the whole output vector unwritten) 1 time in 7 arms; on the pre-#314 revision f2099e9b the same harness produced 0 NaN in 11 arms, including 5 of 5 clean in the guard-off cell. That is consistent with the unconditional staging being implicated, and is NOT significant (about 59% by chance at the ~10% base rate), so this change is justified by removing a known unconditional cost while keeping the correctness fix, not by a measured cure. The engine#315 residual fault is therefore not claimed fixed. Its acceptance criterion remains 12 consecutive arms with 0 NaN and 66/66 requests at gtt_start >= 95 GiB, judged on the guard's count/contiguous=yes case rather than on frequency. Also note the automatic route has been dead: bump-hrx has failed on 12 consecutive scheduled runs (last success 2026-09-24), so this is a hand pin move, which .gitmodules documents as the practice ("hrx-system is rebased by hand") and which every recent pin move has used. hrx-system is unchanged at 98d05d94a9f9.
bong-water-water-bong
force-pushed
the
bump-hrx-315
branch
from
October 7, 2026 11:20
dbee9ab to
1ca8847
Compare
This was referenced Oct 7, 2026
bong-water-water-bong
added a commit
that referenced
this pull request
Oct 9, 2026
…a76c18 long-prompt result (#364) * docs: note the engine#315 pin movement in KNOWN-ISSUES.md The entry cites 26330cc20490 as the pin engine@main ships; main now pins e44c9d01a4d5 with hrx-system 98d05d94 (#361 reverted in #363). The study's measurements still describe what ships: 98d05d94 is the same runtime, and 26330cc20490 is an ancestor of e44c9d01a4d5. * docs: the engine#315 pin note covers #373 and the hrx-system 4ba76c18 long-prompt result Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> --------- Co-authored-by: bong-water-water-bong <bong-water-water-bong@users.noreply.github.com> Co-authored-by: bong-water-water-bong <bong-water-water-bong@1bit.gg> Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
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.
hrx: pin llama.cpp to the shared-memory FA tile visibility (engine#314/315)
Moves third_party/llama.cpp cf4dfd801e -> 26330cc20, a merge that keeps the engine#314
emission fix and drops the cost it introduced.
Context: cf4dfd801e removed the two unordered subgroup collectives in the decode-split FA
kernel by forcing the staging path on every key tile (the engine#314 always-stage fallback).
Its own commit message noted it "pays the staging loop and two subgroup barriers on every key
tile instead of only the tiles that need masking". The merged commit takes f9ff06675 (open as
1bit-MONSTER/llama.cpp#95) instead, which keeps the conditional staging and derives the
visibility predicates from shared memory, so no collective remains:
real subgroup.vote in the kernel: 0
merged kernel file is byte-identical to f9ff06675
Measured on Strix Halo with the pressure harness (3 neighbour servers so each arm runs at
gtt_start ~102 GiB, 66 requests per arm, control and guard-off arms paired inside one box
state): on cf4dfd801e the guard-off arm faulted with all-NaN logits (count=154880 of 154880,
contiguous=yes, the whole output vector unwritten) 1 time in 7 arms; on the pre-#314 revision
f2099e9b the same harness produced 0 NaN in 11 arms, including 5 of 5 clean in the guard-off
cell. That is consistent with the unconditional staging being implicated, and is NOT
significant (about 59% by chance at the ~10% base rate), so this change is justified by
removing a known unconditional cost while keeping the correctness fix, not by a measured
cure.
The engine#315 residual fault is therefore not claimed fixed. Its acceptance criterion
remains 12 consecutive arms with 0 NaN and 66/66 requests at gtt_start >= 95 GiB, judged on
the guard's count/contiguous=yes case rather than on frequency.
Also note the automatic route has been dead: bump-hrx has failed on 12 consecutive scheduled
runs (last success 2026-09-24), so this is a hand pin move, which .gitmodules documents as
the practice ("hrx-system is rebased by hand") and which every recent pin move has used.
hrx-system is unchanged at 98d05d94a9f9.