Add Encoder_v2.mlmodelc: int8-linear re-quantization fixing token corruption (FluidAudio#760)
#4
by alexwengg - opened
int8 per-channel linear quantization of the fp16 encoder export (568M, vs 425M 6-bit LUT original). The original Encoder.mlmodelc's 6-bit LUT palettization noise corrupts tokens under specific right-context (e.g. Ukrainian 'фоновий процес' -> 'фоновиц' on the issue #760 repro); the int8-linear rebuild decodes all repro cuts identically to fp32 and parakeet-mlx. Original file left untouched for older library versions; FluidAudio PR #872 adds Swift-side preference for Encoder_v2 with fallback. config.json now records actual weight formats per encoder file.
alexwengg changed pull request status to open
alexwengg changed pull request status to merged