search

tenstorrent / tt-forge-fe

The TT-Forge FE is a graph compiler designed to optimize and transform computational graphs for deep learning models, enhancing their performance and efficiency.
https://docs.tenstorrent.com/tt-forge-fe/
Apache License 2.0
20 stars 3 forks source link

Deprecate sparse matmul in BTS model #749

Open kamalrajkannan78 opened 12 hours ago

kamalrajkannan78 commented 12 hours ago

Description

BTS model faces AssertionError: Inner dimensions don't match for sparse matmul . sparse matmul should be deprecated and as it arises from decompose_resize2d, so removal will throw unsupported op - resize2d error at MLIR stage.

Trace

forge/forge/op/eval/forge/resize.py:359: in decompose
    decompose_resize2d(attr, dc, inputs, resize_method)
forge/forge/op/eval/forge/resize.py:295: in decompose_resize2d
    decompose_downsample_2d(attr, dc, inputs, resize_method)
forge/forge/op/eval/forge/resize.py:246: in decompose_downsample_2d
    result = dc.op("sparse_matmul", [dident_tensor, activations])