Han-Chung Wang
Han-Chung Wang
> [@hanhanW](https://github.com/hanhanW) I am wondering what the plan here is to move forward? Do we have a breakdown of all pieces that are needed somewhere? (Sorry if missed it somewhere...
Good news, I made some progress with https://github.com/iree-org/iree/pull/22814. I made few small examples with encodings, and we are able to do `compiler->tool->runtime` trick for encodings. I'll look at llama models...
Update: there are few issues in CPU codegen that need to support GPU relayout behavior better, but we can bypass it with flags. However, I ran into https://github.com/iree-org/iree/issues/22941 issue, which...
I'd like to get a status update for mxfp4 MLIR ukernel. What is the current status and remaining work? (I'm just check in the status. We can discuss more tomorrow,...
This is awesome, @Yu-Zhewen ! Does the numbers for `dt` include relayout overheads?
cc @krzysz00 @Muzammiluddin-Syed-ECE who used to work on scaled matmul things.
I think you now want to pin this issue in `Issues`, but not https://github.com/iree-org/iree/issues/21868
> I think you now want to pin this issue in `Issues`, but not [#21868](https://github.com/iree-org/iree/issues/21868) > > (I already pinned the issue.)
The PR depends on https://github.com/iree-org/iree/pull/22886
> I think the problem is that you may "decompose" it too early. It is not common to create linalg ops in such methods. I can't comment much unless I...