Repository navigation
Conversation
bsbodden
changed the title
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Superseded by PR #100. The BF16-only control fails Bonsai batch parity, and its performance precision/allocation limitations remain recorded. Runtime work is incorporated in the combined correction; no release was published from this PR. Immutable source and protocol:
audit/2026-10-12/bf16-hotfix-control(405639cb). Unique test reports are preserved in Models' scratch archive.BF16 single-vector and batched projections can change their summation order after JIT compilation. This draft isolates that independently reproduced arithmetic defect on the published v0.1.29 base. It includes two-query weight reuse, existing shared row dispatch and the executor safeguards required by that dispatch; activation implementations and unrelated kernels remain unchanged.
This is not a fix for the current Bonsai native fixture failure. The unchanged real Bonsai tests failed on both architectures with the exact cb3dab3 candidate dependency in Models run 38319438508, proving this BF16-only patch is insufficient. The subsequent full activation-only candidate passed both fixtures on both architectures in Models run 38321333816, establishing that this BF16 change is not required to fix those Bonsai failures. The earlier local Bonsai pass used an unhashed mutable classes directory and cannot establish an isolated dependency fix.
Fresh local validation passed 45 BF16/dispatch/executor tests, including twelve cold/hot SIMD-width cases, batch allocation, shared/confined memory and cross-pool nested dispatch. The minimal fixed-tree alternative is retained as a rejected checkpoint because earlier x86/ARM measurements found slowdowns.
The frozen actual-artifact twelve-pair kernel workflow 38319356761 completed on AMD EPYC 9V74 and ARM Neoverse-N2 (SVE active). All twelve latency confidence intervals favor the candidate, but five cases fail the predeclared precision gate. ARM small-batch allocation also increases from 5.03 to 52.78 bytes per call. The workflow correctly fails; this candidate is not performance-qualified. Raw results retain every pair and do not establish whole-model performance. No publication or merge is justified by the present evidence. Useful BF16 work will be consolidated with the continuing development branch after review.