[BUG] Get 'kernel not found' error when doing quantized_matmul with mxfp8

[soumya-batra]

Describe the bug
mixed precision kernel not found when matrix multiplying float and mxfp8/nvfp4 quantized matrices.

To Reproduce

Include code snippet

import mlx.core as mx
input = mx.random.normal(shape=(16,768))
wt = mx.random.normal(shape=(768,3*768))
wt_q , wt_s = mx.quantize(wt, mode='mxfp8')
qm = mx.quantized_matmul(input, wt_q, wt_s, mode='mxfp8', transpose=False)
mx.eval(qm)

get the error: RuntimeError: [metal::Device] Unable to load kernel mxfp8_qmm_n_float_gs_32_b_8_batch_0

happens with all of fp32, fp16, bf16 inputs.

Expected behavior
successful matmul operation

Desktop (please complete the following information):

• OS Version: [e.g. MacOS 15.6.1]
• Version [e.g. 0.7.0]

Additional context
Tried upgrading mlx, built from source, but error remains.

[awni]

Currently on Metal only mxfp4 is supported for quantized matmuls. We will likely add nvfp4 and mxfp8 in the future.