Feat/nccl #846
Feat/nccl #846
Conversation
| ) -> ComputeClient<Server, Channel> { | ||
| // To get the supported WMMA features, and memory properties, we have to initialize the server immediately. | ||
| #[cfg(not(feature = "nccl"))] | ||
| cudarc::driver::result::init().unwrap(); |
There was a problem hiding this comment.
cudarc::driver::result::init().unwrap(); is not longer handled with nccl feat
nathanielsimard
left a comment
nathanielsimard
left a comment
There was a problem hiding this comment.
I've got a few comments, but it looks quite clean.
I think a bit of docs on the output: Option<&Handle> would help.
Also, LazyNccl could become NcclGroup maybe? Since that's how you name the variables.
| @@ -0,0 +1,586 @@ | |||
| //* WITH THIS MODULE `cudarc::driver::result::init().unwrap();` IS NOT LONGER HANDLES BY `CudaRuntime::create_clien()` | |||
| pub fn all_reduce<N: Float + CubeElement>( | ||
| &self, | ||
| input: &Handle, | ||
| output: Option<&Handle>, |
There was a problem hiding this comment.
When the output should be provided?
There was a problem hiding this comment.
Deleted it, was inplace or not inplace ...
| pub fn broadcast<N: Float + CubeElement>( | ||
| &self, | ||
| input: &Handle, | ||
| output: Option<&Handle>, |
| pub fn reduce<N: Float + CubeElement>( | ||
| &self, | ||
| input: &Handle, | ||
| output: Option<&Handle>, |
| cudarc::nccl::result::group_end().unwrap(); | ||
| } | ||
|
|
||
| /// Assumes the `handles` vecotor got a `Handle` for each device in order of the rank. |
There was a problem hiding this comment.
typo: vecotor it's a nice one 😄
There was a problem hiding this comment.
i am not only dyslexic but also bad typer ... at least i understand code xD
Was thinking to make a Runtime Management on your side, its on the Use-AI.rs site for now. So fixed it like asked :) |
2 participants
So now its working.