Host Ir Lowering: foundation #3535
Conversation
|
!test --diff |
samnordmann
changed the title
|
!test |
|
!test |
|
@wujingyue the CI failures seen on the previous commits seem to be caused by a too small tolerance rate on single device test. It doesn't seem related to the present PR |
| auto& mesh = tv->getDeviceMesh().vector(); | ||
| std::copy(mesh.begin(), mesh.end(), std::inserter(ret, ret.end())); | ||
| } else { | ||
| ret.insert(0); |
There was a problem hiding this comment.
Why is this change needed?
There was a problem hiding this comment.
involvedDevices is now called from HostIrEvaluator which can take in single device fusion where TVs don't have device mesh. In a single device Fusion, it is natural that this function returns device 0
Co-authored-by: Jingyue Wu <[email protected]>
| auto& mesh = tv->getDeviceMesh().vector(); | ||
| std::copy(mesh.begin(), mesh.end(), std::inserter(ret, ret.end())); | ||
| } else { | ||
| ret.insert(0); |
There was a problem hiding this comment.
involvedDevices is now called from HostIrEvaluator which can take in single device fusion where TVs don't have device mesh. In a single device Fusion, it is natural that this function returns device 0
…ann/Fuser into host_ir_lowering_first_refactor
|
!test |
What
Refactor Host Ir lowering. https://jirasw.nvidia.com/browse/NVFUSER-105
Dependent on:
Why
How
multidevice/lower_communicationtohost_ir/lowerMultiDeviceExecutor's constructor logic tohost_ir/lowerI divided in incremental commits to ease the review. The relevant commits start at move HostIr lowering to host ir folder and special class, the commits before that belong to #3524