Tests whether uninitialized padding rows in NVFP4 swizzled block
scales (from scaled_fp4_quant's torch.empty) can contaminate real
rows' GEMM output via flashinfer mm_fp4.
Key scenarios:
- Poison padding scale rows with NaN
- Poison padding scale rows with max finite values (overflow)
- Poison padding activation rows
- Simulate CUDA graph replay with persistent scale tensor
These target the suspected root cause of NaN in NVFP4 models with
CUDA graphs, where capture sizes (1,2,4,8,...) are not multiples
of 128, leaving uninitialized padding in the swizzled scale layout.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
