TorchInductor _scaled_mm_v2 lowering had no template or extern path for non-trivial swizzle patterns and asserted during compilation, breaking blockwise MXFP8/NVFP4 compile tests on Blackwell.
For swizzled scale layouts, defer to the eager _scaled_mm_v2 op before mm_args. The trivial-swizzle path is unchanged.
Authored with Claude Opus 4.8.
rel: 9eb8bcb
Pull Request resolved: #186384
Approved by: https://github.com/drisspg
SOCIAL SHARE CARD GENERATOR