[PROTON] Simplify proton viewer APIs for bench_mlp analysis #6452
Conversation
| "HIP": { | ||
| "gfx90a": lambda width, **kwargs: 383e12 / (width / 8), | ||
| "gfx941": lambda width, **kwargs: 2614.9e12 / (width / 8), | ||
| "gfx942": lambda width, **kwargs: 2614.9e12 / (width / 8), |
There was a problem hiding this comment.
@antiagainst shall we drop a 0 or None value here for gfx950?
There was a problem hiding this comment.
Yeah. The spec is unavailable right now. For mi300x and mi325x you can see the spec in https://github.com/triton-lang/triton/pull/6513/files#diff-5e6a8d3fc5ad9de85fc09ead926355cc19497d3056c368d463bf4c626ce68540. Can drop gfx941 here given that's deprecated.
There was a problem hiding this comment.
So only "gfx90a" and "gfx942" at this moment?
There was a problem hiding this comment.
Also cc @ptillet for any additional comments
There was a problem hiding this comment.
Yup--gfx90a for mi210/mi250, gfx942 for mi300/mi325. gfx940 and gfx941 was deprecated, see llvm/llvm-project#126763.
Jokeren
force-pushed
the
keren/bench-refactor
branch
from
d0893dd to
f55c57f
Compare
|
@ptillet Any other concerns with the changes? |
Jokeren
force-pushed
the
keren/bench-refactor
branch
from
338cf51 to
8a73a20
Compare
| name = "triton_kernels" | ||
| version = "1.0.0" | ||
| dependencies = ["torch", "numpy", "pytest"] | ||
| dependencies = ["torch", "numpy", "pytest", "llnl-hatchet"] |
There was a problem hiding this comment.
This dependency is causing problems internally because it depends on numpy 1. It appears that hatchet is not actually used inside triton_kernels here, so I think it should be possible to remove this dependency.
No description provided.