trunk/312ef3ee1cd45fbee99f2e4e8f6035f05ceeebed: Add BF16x9 precision mode for CUDA FP32 matmul (#195301)
- PyTorch: 413 events in the last 90 days
- PyTorch: 410th Release in the last 90 days
- Previous: earlier the same day · trunk/129e6f1d40be5e428a664436d97803f70aa747c0: [inductor] Enable MXFP8 nested-reduction test on XPU (#194864)
What happened
Human Note Cublas has a cool bf16x9 mode for fp32 gemms; in theory it should be more accurate than tf32; ill keep prodding to get some intersting data showing this SOL went crazy (although in a good way) and made sure this doesnt break things hence the largish amount of changes .. indeed these flags are insidiuous Most of the changes were in the pt2 stack where we need to properly guard/ensure caches are working cor…
Summary assembled by rule from the sources below