tenstorrent / tt-metal

:metal: TT-NN operator library, and TT-Metalium low level kernel programming model.
Apache License 2.0
473 stars 75 forks source link

test_untilize_with_unpadding fails on BH #14548

Open abhullar-tt opened 2 weeks ago

abhullar-tt commented 2 weeks ago

Test config:

params += [
    pytest.param(
        [[1, 1, 128, 32]],
        {
            "dtype": [ttnn.bfloat16],
            "layout": [ttnn.TILE_LAYOUT],
            "input_mem_config": [
                ttnn.MemoryConfig(
                    ttnn.TensorMemoryLayout.HEIGHT_SHARDED,
                    ttnn.BufferType.L1,
                    ttnn.ShardSpec(create_grid(1, 2), [64, 32], ttnn.ShardOrientation.ROW_MAJOR, False),
                )
            ],
            "output_mem_config": ttnn.MemoryConfig(
                ttnn.TensorMemoryLayout.HEIGHT_SHARDED,
                ttnn.BufferType.L1,
                ttnn.ShardSpec(create_grid(1, 2), [64, 16], ttnn.ShardOrientation.ROW_MAJOR, False),
            ),
            "output_tensor_start": [0, 0, 0, 0],
            "output_tensor_end": [0, 0, 127, 15],
        },
    )
]

results in bad PCC. if 32 -> 64 and 16->32 and 15->31 the test passes.

it looks like untilize with unpadding op is missing an assert for this case

abhullar-tt commented 2 weeks ago

Will merge the test update for BH to run this suite on post commit and then re-assign to TM team so the op can be generalized for this shape as well. FYI @ntarafdar