NVFP4QuantizeMethod.quantize_input() and apply() asserted bf16/fp16 inputs
and crashed on the fp32 pre-attention norm output in eager mode (without the
torch.compile fusion that keeps it bf16):
AssertionError: only allow bf16/fp16 inputs to fp4 linear, got torch.float32
The FP4 linear emits bf16 regardless (mm_fp4 out dtype is bf16), and the
sibling fastvideo/layers/fp4linear.py already casts non-bf16/fp16 inputs, so
cast fp32 -> bf16 here too instead of failing. Lets NVFP4Config inference run
in eager mode (repro: LTX2-distilled on DGX Spark / sm_121, no compile).
Raghav
2e8eb75c11