Ted Themistokleous

Marking as revisit as we've pushed off build/maintaining ROCm 6.1 + Onnxruntime wheels to devops team.

Can't get UT zip from a stable place. Closing this out but ROCm 6.2 CI update allows us to just build MIGraphX EP which is a smaller overhead (~5 min)...

Will approve when all green. Lets see what springs a leak

This still needed or do we want to close this and aim to merge in 6.2 once it's ready?

> > This still needed or do we want to close this and aim to merge in 6.2 once it's ready? > > We should try to get this merged...

Waiting for DLM piece to run and finish

Kicking off another run. Fixed an issue upstream with something that was occurring with our calibration table. Hoping to get an answer by this evening

> > Without this we attempt to get mlir to try to create int8 x uint8 which fail at compile. > > Can you explain this in more detail? When...

> But why are we getting mixed types like this in the first place? Because the implimentation of dynamicquantizelinear breaks it up into separate instructions which hardcode uint8 as part...

Moving to draft. Will craft better matcher after tomorrow morning's coffee. @lakhinderwalia I see your point from Monday's meeting. Looks like we can just target and replace q_min/q_max as well...