Resolve the KV flags (--kv-quant, its TurboQuant spelling, --quantized-kv-start, --fused-sdpa) against the artifact's kv_config.json. fusedSdpa picks the affine read kernels (on: tiled-causal, off: unfused); unset, they follow the scheme's kind.
--kv-quant
--quantized-kv-start
--fused-sdpa
kv_config.json
fusedSdpa
Resolve the KV flags (
--kv-quant, its TurboQuant spelling,--quantized-kv-start,--fused-sdpa) against the artifact'skv_config.json.fusedSdpapicks the affine read kernels (on: tiled-causal, off: unfused); unset, they follow the scheme's kind.