Ipanfilo/wheel build action#529
Conversation
|
Can we test the action on a fork? |
| git submodule update --init --recursive --depth 1 \ | ||
| 3rdparty/aotriton \ | ||
| 3rdparty/aiter \ | ||
| 3rdparty/hipify_torch |
There was a problem hiding this comment.
Won't the command already update all submodules without specifying them individually? It might be a pain point if we later modify our submodules e.g. adding qola or a separate CK for GEMM
There was a problem hiding this comment.
It is indeed maintenance burden, but it allows to skip quite heavy cublass and especially cutlass modules
| force_build_image: | ||
| type: boolean | ||
| default: true | ||
| upload_image: | ||
| type: boolean | ||
| default: false |
There was a problem hiding this comment.
Considering these defaults, when would we actually create/use the cache mechanism here?
There was a problem hiding this comment.
After adding those parameters I reconsidered benefits of image caching. It eliminates few minutes of rocm downloading/installing at the cost of bigger docker image downloading that depending on network can eat all or most of this time. So it is not in priority now
|
/claude summary |
3 participants
Description
Workflow to build TE wheels - useful to test this functionality and can also be used for nightly distributions
WIP: https://github.com/ROCm/frameworks-internal/issues/14193
Type of change
Changes
Checklist: