AMDMIGraphX

AMDMIGraphX copied to clipboard

pfultz2

Propagate layout in reshape operator and broadcasting in binary operators

1

pfultz2

Remove rocmlir unsupported reduce types

9

We are going to enable f16 for reduce in this rocmlir PR: https://github.com/ROCm/rocMLIR/pull/1722 However, the code here assumes bf16 and fp8 are supported and that's not the case.

dhernandez0

Issue with int8 for MaxPool

1

Fixes #2991

taylding-amd

Warnings upon tuning information mismatch for Convolutions

6

This PR adds `MIGRAPHX_ENABLE_TUNING_WARNINGS` environment flag. Based on this flag it would warn/error out if saved Tuning information has a different version, and architecture setup compared to the machine on...

umangyadav

Related PR: https://github.com/ROCm/AMDMIGraphX/pull/3595

ahsan-ca

There's a push to have dockers come/go to a private repo for CI. Today github actions can build a docker image to be used for builds but that is published...

causten

Efficiently support mixed precision w8a16 models

1

Generalize the work to efficiently support w4a16, to support w8a16 (weights in an 8 bit format, e.g. int8, eventually fp8); activations in fp16 (eventually also bf16). "Efficiently" here means that...

hgaspar

hipblaslt_ext::getAllAlgos returns solutions, none of which are supported

5

`hipblaslt_ext::getAllAlgos` returns algos which can then be validated using `hipblaslt_ext::matmulIsAlgoSupported`. For some cases, all the algos returned are invalid. This seems to be an issue that hipBLASLt team needs to...

ahsan-ca

Questions about installing MIGraphX on WSL2

3

So I followed steps from [https://rocm.docs.amd.com/projects/radeon/en/latest/docs/install/native_linux/install-migraphx.html](url), I installed Radeon software for WSL with ROCm and Pytouch strictly following those steps and everything works fine, and I start to try to...

Mhyho