Janani Sriram

Add plots variable along 1 input parameter to jagged_sum operator in TritonBench

3

Summary: Add plotting functionality to the `jagged_sum` operator in TritonBench, enabling the creation of line plots for any set of benchmarks variable along one of the following input parameters: `B`,...

I'm attempting to modify the PT2 Inductor Triton template for FP8 GEMMs running [`(BlockWise1x128, BlockWise1x128)` scaling](https://github.com/pytorch/pytorch/blob/main/torch/_inductor/kernel/mm.py#L648-L790) to use `tl.dot_scaled` (instead of `tl.dot`) (fyi, I run into the same problem with...

[Inductor][FP8] Simplify load_scales in main loop scaled_mm

2

Summary: Remove `load_scales` function from jinja template for `scaled_mm` main loop scaling, which handles block-wise scaling variants. All main loop scaling variants need to load the pointer to the scale...