release: prepare 0.38.0 - #904
Open
michalharakal wants to merge 1 commit into
Open
Conversation
Bump version 0.37.0 -> 0.38.0 (gradle.properties, docs/antora.yml, README quickstart). Promote CHANGELOG [Unreleased] to [0.38.0]. The [Unreleased] section only carried the dynamic-dimension work (#891), so this also writes up the narrow-float line that merged without changelog entries: the shared BF16/FP16 codec and KEEP_NATIVE loading (#886), the zero-copy transpose for input-major narrow weights (#888/#895), the native FFM FP16 kernel and the missing matmulFp16 accessor behind it (#887/#896), and read-once weight tiling in both native narrow kernels (#897). Records the measured result -- both narrow formats now beat the FP32 SGEMM, where FP16 used to be 2-18x slower -- and the finding that these kernels are compute-bound on the FMA chain at batch 16 rather than bandwidth-bound, so the next win is not more layout work. Two behaviour notes worth a reader's attention are called out rather than buried: Fp16Codec.decode now quiets NaN, which changes exactly the 1022 signaling patterns and nothing else, and a row-major narrow buffer still takes the generic transpose path on purpose. Refresh README "What's New" around the two headline features and add a Contributors (0.38.0) section. No tag yet.
|
📖 Documentation Preview The documentation has been built successfully for this PR. Generated Files:
Artifacts:
This comment will be updated automatically when the PR is updated. |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Bump version 0.37.0 -> 0.38.0 (gradle.properties, docs/antora.yml, README quickstart). Promote CHANGELOG [Unreleased] to [0.38.0].
The [Unreleased] section only carried the dynamic-dimension work (#891), so this also writes up the narrow-float line that merged without changelog entries: the shared BF16/FP16 codec and KEEP_NATIVE loading (#886), the zero-copy transpose for input-major narrow weights (#888/#895), the native FFM FP16 kernel and the missing matmulFp16 accessor behind it (#887/#896), and read-once weight tiling in both native narrow kernels (#897). Records the measured result -- both narrow formats now beat the FP32 SGEMM, where FP16 used to be 2-18x slower -- and the finding that these kernels are compute-bound on the FMA chain at batch 16 rather than bandwidth-bound, so the next win is not more layout work.
Two behaviour notes worth a reader's attention are called out rather than buried: Fp16Codec.decode now quiets NaN, which changes exactly the 1022 signaling patterns and nothing else, and a row-major narrow buffer still takes the generic transpose path on purpose.
Refresh README "What's New" around the two headline features and add a Contributors (0.38.0) section.
No tag yet.