Property EnableCpuWeightRepacking
EnableCpuWeightRepacking
Whether weights placed in host memory may be repacked into
CPU-optimized layouts (a large matrix-multiplication speedup
for CPU inference). Repacked layouts serve forward inference
only: fine-tuning cannot read them, so a model intended for
TRAINING on the CPU must be loaded with this set to
false. Models fully offloaded to a GPU are unaffected
either way. Defaults to true.
public bool EnableCpuWeightRepacking { get; set; }