feat(backend): commonMain scalar Q5_1/Q5_0/Q4_K/Q6_K kernels + SPI (N…#710
Merged
Merged
Conversation
…ative parity) Part of #708. Brings quantized matmul to Kotlin/Native (and JS/WASM), which previously only had FP32/BF16/Q8_0/Q4_0 scalar kernels — Q4_K/Q6_K/Q5_x were JVM-only (Panama/FFM), so on non-JVM targets packed-quant matmul had no kernel. SPI (skainet-backend-api, commonMain): - New Q5_1MatmulKernel / Q5_0MatmulKernel / Q6KMatmulKernel interfaces (block-major `(blockIdx*outputDim+o)*BYTES_PER_BLOCK`, exact dequant in kdoc). - KernelProvider: matmulQ5_1()/matmulQ5_0()/matmulQ6K() accessors (default null) + supports() keys for "Q5_1"/"Q5_0"/"Q6_K". Scalar kernels (skainet-backend-cpu, commonMain — available on every target): - ScalarQ5_1/Q5_0/Q4_K/Q6_KMatmulKernel, math ported from JvmQuantizedVectorKernels / DequantOps (Q4_K get_scale_min_k4 + sub-block codeSum*scale - inputSum*offset; Q6_K ql/qh 6-bit reassembly). Shared decodeHalf() FP16 helper. - ScalarKernelProvider now overrides matmulQ4K/Q6K/Q5_1/Q5_0 → the scalar floor carries every packed format. Test: ScalarPackedKernelParityTest (commonTest) validates each kernel's matmul against an independent inline dequant; passes on jvmTest AND linuxX64Test, proving Native packed-matmul correctness (relative tol for the FP reassociation of the per-sub-block accumulation). Note: dispatch wiring (so ops.matmul routes packed tensors to these kernels on non-JVM) + non-JVM provider registration land in follow-up commits; this commit is the kernels + SPI surface. Co-Authored-By: Claude Opus 4.8 <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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
…ative parity)
Part of #708. Brings quantized matmul to Kotlin/Native (and JS/WASM), which previously only had FP32/BF16/Q8_0/Q4_0 scalar kernels — Q4_K/Q6_K/Q5_x were JVM-only (Panama/FFM), so on non-JVM targets packed-quant matmul had no kernel.
SPI (skainet-backend-api, commonMain):
(blockIdx*outputDim+o)*BYTES_PER_BLOCK, exact dequant in kdoc).Scalar kernels (skainet-backend-cpu, commonMain — available on every target):
Test: ScalarPackedKernelParityTest (commonTest) validates each kernel's matmul against an independent inline dequant; passes on jvmTest AND linuxX64Test, proving Native packed-matmul correctness (relative tol for the FP reassociation of the per-sub-block accumulation).
Note: dispatch wiring (so ops.matmul routes packed tensors to these kernels on non-JVM) + non-JVM provider registration land in follow-up commits; this commit is the kernels + SPI surface.