Lowered transpose op; working training of matmul. #158
Conversation
pybuda/test/mlir/test_ops.py
Outdated
|
|
||
| def forward(self, a, b): | ||
| c = a + b | ||
| return torch.transpose(c, 1, 2) |
There was a problem hiding this comment.
Can we remove the add operation, and just transpose the single input? Just to be inline with the theme in this test file...
pybuda/test/mlir/test_training.py
Outdated
| @@ -7,51 +7,52 @@ | |||
|
|
|||
| import pybuda | |||
| import pybuda.config | |||
| from pybuda.op.eval.common import compare_with_golden_pcc | |||
|
|
|||
| def test_torch_training(): | |||
| class MultParam(nn.Module): | |||
There was a problem hiding this comment.
Rename to MatMulParam, for example...
pybuda/test/mlir/test_training.py
Outdated
| golden = model(inputs) | ||
| output = tt_model(inputs) | ||
|
|
||
| if not torch.allclose(output[0], golden, rtol=1e-1): | ||
| raise ValueError("Output does not match the golden output") | ||
| #print(f"golden = {golden}, output = {output[0]}") |
There was a problem hiding this comment.
Please remove the prints which were commented out.
pybuda/test/mlir/test_training.py
Outdated
| if not torch.allclose(output[0], golden, rtol=1e-1): | ||
| raise ValueError("Output does not match the golden output") | ||
| #print(f"golden = {golden}, output = {output[0]}") | ||
| oputput = [co.to("cpu") for co in output] |
There was a problem hiding this comment.
oops, tnx
pybuda/test/mlir/test_ops.py
Outdated
|
|
||
| def forward(self, a, b): | ||
| c = a + b | ||
| return torch.transpose(c, 1, 2) |
There was a problem hiding this comment.
Can we add more test cases? E.g. to test out different transpose dims, different shapes (e.g. 7, 41, etc.) and also potentially different ranks (e.g. 2, 3 and 4).
There was a problem hiding this comment.
Definitely makes sense. Also should apply to other ops. test_ops.py will blow up in size soon; we should consider separating it into multiple files.
vladimirjovanovicTT
force-pushed
the
vladimirjovanovic/lower_transpose_op
branch
3 times, most recently
from
2d6b5f0 to
e3e033f
Compare
Improved tests, fixed bug in tests. Cleaned up training test.
vladimirjovanovicTT
force-pushed
the
vladimirjovanovic/lower_transpose_op
branch
from
e3e033f to
40baa5c
Compare
Lowered transpose op, added test.
Modified training to support matmul.
#23 #21 #15