Repository navigation
ggml-cuda: fattn do not check oob if k_VKQ_0 = (k_VKQ_max - nbatch_fa) - #30572
Merged
JohannesGaessler merged 1 commit intoOct 18, 2026
Merged
Conversation
thelittlefireman
changed the title
github-actions
Bot
added
ggml
am17an
approved these changes
Oct 16, 2026
JohannesGaessler
approved these changes
Oct 18, 2026
wispoffates
added a commit
to wispoffates/llama.cpp
that referenced
this pull request
Oct 24, 2026
…e no-oob loop) -- p100-series keeps the v5.25 path Gate R26: with ggml-org#30572 the gemma4 prefill KLD vs the F32 reference went 0.00187 -> 0.00367 (pf16 0.00266 -> 0.00287) and the gemma4 triage reply changed; decode (vec kernel) identical. Mathematically the same tile, but gemma4's fp16 tile-FA prefill on Pascal is rounding-sensitive, and the KV views of a prefill (512/1024/1536/2048) are exact multiples of nbatch_fa = exactly the case it moves. We declined it as a micro-opt at the 10-16 PR check; it arrived with the rebase. This reverts commit 5aeac37. Assisted-by: Claude Opus 5.5 (Claude Code)
edwardyoon
pushed a commit
to edwardyoon/focus-llama
that referenced
this pull request
Oct 24, 2026
wispoffates
added a commit
to wispoffates/llama.cpp
that referenced
this pull request
Oct 25, 2026
…e no-oob loop) -- p100-series keeps the v5.25 path Gate R26: with ggml-org#30572 the gemma4 prefill KLD vs the F32 reference went 0.00187 -> 0.00367 (pf16 0.00266 -> 0.00287) and the gemma4 triage reply changed; decode (vec kernel) identical. Mathematically the same tile, but gemma4's fp16 tile-FA prefill on Pascal is rounding-sensitive, and the KV views of a prefill (512/1024/1536/2048) are exact multiples of nbatch_fa = exactly the case it moves. We declined it as a micro-opt at the 10-16 PR check; it arrived with the rebase. This reverts commit 5aeac37. Assisted-by: Claude Opus 5.5 (Claude Code)
wispoffates
added a commit
to wispoffates/llama.cpp
that referenced
this pull request
Oct 26, 2026
…e no-oob loop) -- p100-series keeps the v5.25 path Gate R26: with ggml-org#30572 the gemma4 prefill KLD vs the F32 reference went 0.00187 -> 0.00367 (pf16 0.00266 -> 0.00287) and the gemma4 triage reply changed; decode (vec kernel) identical. Mathematically the same tile, but gemma4's fp16 tile-FA prefill on Pascal is rounding-sensitive, and the KV views of a prefill (512/1024/1536/2048) are exact multiples of nbatch_fa = exactly the case it moves. We declined it as a micro-opt at the 10-16 PR check; it arrived with the rebase. This reverts commit 5aeac37. Assisted-by: Claude Opus 5.5 (Claude Code)
auRose94
pushed a commit
to auRose94/llama.cpp
that referenced
this pull request
Oct 28, 2026
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.
Overview
Do not check oob if k_VKQ_0 = (k_VKQ_max - nbatch_fa)
Additional information
For example if k_VKQ_max = 256 and nbatch_fa = 128 :
Before:
After:
Both tile are full and does not need to be oob check.
Requirements