On current main, Pi0.5 FP8 crashes on SM89 (4090 / RTX 6000 Ada / L40):
AttributeError: 'GemmRunner' object has no attribute 'fp8_nt_dev'. Did you mean: 'fp8_nn_dev'?
Looks like the SM89 FP8 path from #26 got dropped in #30 (d8d5de7) when the shared GEMM files were rewritten — but the frontend still selects the nk layout and calls fp8_nt_dev. (fp8_layout=kn isn't a workaround — it just hits the original code=15 from #21.)
Restoring the fp8_nt_dev bits from #26 fixes it; tested on RTX 6000 Ada (sanity PASS, P50 20.8 ms). Just flagging it — happy to send a PR.
On current
main, Pi0.5 FP8 crashes on SM89 (4090 / RTX 6000 Ada / L40):Looks like the SM89 FP8 path from #26 got dropped in #30 (
d8d5de7) when the shared GEMM files were rewritten — but the frontend still selects thenklayout and callsfp8_nt_dev. (fp8_layout=knisn't a workaround — it just hits the original code=15 from #21.)Restoring the
fp8_nt_devbits from #26 fixes it; tested on RTX 6000 Ada (sanity PASS, P50 20.8 ms). Just flagging it — happy to send a PR.