Isuru Fernando

icon indicating an email [email protected]

icon company @openteams-ai icon location Vancouver, Canada

Results 928 comments of Isuru Fernando

Invalidate StorageImpl instances when tensor is overwritten with cudagraphs

@pytorchbot merge

Consider using macholib instead of native otool and install_name_tool

You can get `install_name_tool` and `otool` from https://github.com/tpoechtrager/cctools-port. You can use `ldid` instead of `codesign` on Linux.

Delocate "@rpath" when search paths include "/usr/lib/swift"

Can you send the output with `output -l` (lowercase `L`)

Delocate "@rpath" when search paths include "/usr/lib/swift"

Using `@rpath` paths isn't a bug, but it complicates the logic. > I think the only way workaround for delocate would be to explicitly treat paths matching @rpath/libswift* the same...

add an option for install directory of mini{forge,conda}

cc @finnagin

triton generates unnecessary shared memory stores/loads

Based on a suggestion from @peterbell10 I removed `AtomicRMWOp` at https://github.com/openai/triton/blob/0ba87e2ff35f703f84040400554702ee55476cdb/lib/Dialect/TritonGPU/Transforms/RemoveLayoutConversions.cpp#L192 which resulted in the PTX not having any shared memory loads/stores. This resulted in the triton generated kernel to...

triton generates unnecessary shared memory stores/loads

Is there a case where removing `AtomicRMWOp` as a layout anchor can result in incorrect code?

triton generates unnecessary shared memory stores/loads

I'm using pytorch `v2.3.0-rc6`

ENH: Don't add system library directories to rpath

You are assuming that LIBDIR is in the dynamic linker's default search directories which is not true for some python distributions.

ENH: Don't add system library directories to rpath

> Is there a way to access the dynamic linker's default search directories from python? I don't think there is any. I guess there are couple of ways to allow...