#!/bin/bash # Derive lighter GGUFs from the F32 source-of-truth produced by convert.py. # omnivoice-base-F32.gguf -> BF16, Q8_0, Q4_K_M # omnivoice-tokenizer-F32.gguf -> BF16, Q8_0, Q4_K_M # # Three variants cover the useful precision range : BF16 for max precision # on CUDA, Q8_0 as the balanced default, Q4_K_M as the smallest variant # that still sounds correct. Q5_K_M / Q6_K were tested and dropped : their # size sits between Q4_K_M and Q8_0 with negligible perceptual gain. # # Quantization policy is centralized in tools/quantize.cpp should_quantize : # RVQ codebooks (quantizer.quantizers.*) and the fc / fc2 linear projections # wrapping them stay at F32 in every variant. Nearest-neighbor lookup is # sensitive to per-row quantization noise ; even BF16 mantissa truncation # drifts codes enough to break voice cloning. Conv weights stay at source # dtype and are cast to F16 at load time by gf_load_conv_f16 (ARM im2col # strict). Same policy as acestep.cpp keeping VAE-critical paths intact. set -eu Q="./build/quantize" quantize() { local src="$1" type="$2" local out="${src/-F32.gguf/-${type}.gguf}" if [ -f "$out" ]; then echo "[Skip] $out" else $Q "$src" "$out" "$type" fi } for src in models/omnivoice-base-F32.gguf models/omnivoice-tokenizer-F32.gguf; do quantize "$src" BF16 quantize "$src" Q8_0 quantize "$src" Q4_K_M done