Every read unfused: mlx-lm's quantized SDPA port. The composition of a
uniform --kv-quant N (KV scheme affine, kernels: "unfused"), which
keeps mlx-lm parity, and of --fused-sdpa off. The unfused port takes any
bits, group size and activation dtype, so the dtype only names the
configuration.
Every read unfused: mlx-lm's quantized SDPA port. The composition of a uniform
--kv-quant N(KV schemeaffine,kernels: "unfused"), which keeps mlx-lm parity, and of--fused-sdpa off. The unfused port takes any bits, group size and activation dtype, so the dtype only names the configuration.