| """Weight-only int8 (per-channel symmetric) of an encoder package: the 24-layer fp16 encoder does not fit the ANE | |
| compiler on M1, the int8 one does (as FluidAudio's Parakeet encoders). Usage: python quantize.py <in.mlpackage> <out.mlpackage>""" | |
| import sys | |
| import coremltools as ct | |
| import coremltools.optimize.coreml as cto | |
| m = ct.models.MLModel(sys.argv[1], skip_model_load=True) | |
| cfg = cto.OptimizationConfig(global_config=cto.OpLinearQuantizerConfig(mode="linear_symmetric", dtype="int8", granularity="per_channel")) | |
| cto.linear_quantize_weights(m, cfg).save(sys.argv[2]) | |