Open athul-bos-semi opened 4 months ago
@jliangTT for the TM team to triage , this should be part of interleave2sharded , sharded2interleave work
fyi @tarafdarTT @yan-zaretskiy
@davorchap , please advice on priority
@davorchap , should @athul-bos-semi use the partial sharding code like in https://github.com/tenstorrent/tt-metal/blob/main/tests/tt_eager/python_api_testing/unit_testing/misc/test_sharded.py#L764
?
When I tried to perform ttnn.MaxPool2d operation on High-resolution feature map, the L1 memory ran out. So I instead split the feature map into four quarters and performed ttnn.MxPool2d on each of them, then combined their feature maps. But it is not possible to do the splitting and merging operations on device, hence I had to do them on host.
Alternative solution
Alternatives considered Copying from and into a tensor slice would solve this issue, as described in Feature Request #7045. Copying from a tensor can be accomplished using slicing.