Set USE_XCCL default as OFF #1318
Conversation
| @@ -38,7 +38,7 @@ list(APPEND CMAKE_MODULE_PATH ${TORCH_XPU_OPS_ROOT}/cmake/Modules) | |||
| include(${TORCH_XPU_OPS_ROOT}/cmake/SYCL.cmake) | |||
| include(${TORCH_XPU_OPS_ROOT}/cmake/BuildFlags.cmake) | |||
|
|
|||
| option(USE_XCCL "Build with XCCL support" ON) | |||
| option(USE_XCCL "Build with XCCL support" OFF) | |||
There was a problem hiding this comment.
Currently, XCCL backend does not get ready in PyTorch release. So we disable it by default (you still can build XCCL with USE_XCCL=1) and will enable again in PyTorch 2.8 release by our expectation.
There was a problem hiding this comment.
That's quite unfortunate since a number of projects awaits distributed support fromXPU. What are outstanding issues preventing to enable XCCL by default?
|
@Chao1Han Please add |
|
@gujinghui Please be aware that we plan to disable XCCL by default (also a suggestion from Validation) and enable when we are ready for XCCL binary release. |
No description provided.