Skip to content

Quantize softmax probabilities S into [0, 255] uint8_t and utilize unsigned-by-signed (uint8 * int8) matrix multiplication for SV accumulation, improving quantization precision. Add SIMD emulation fallback using SumOfMulQuadAccumulate and TableLookupBytes. - #1013

Merged
copybara-service[bot] merged 1 commit into
devfrom
test_976294440
Sep 8, 2026

Commits