yelite opened a new pull request, #14350:
URL: https://github.com/apache/tvm/pull/14350
This PR adds new check logic to the BYOC patterns for cutlass, to ensure
intermediate variables in the fused region aren't used outside.
For example,
```python
@R.function
def main(x: R.Tensor((128, 128), "float16"), w: R.Tensor((128, 128),
"float16")):
with R.dataflow():
lv = R.matmul(x, w)
lv1 = R.power(lv, R.const(2.0, "float16"))
lv2 = R.add(lv, lv1)
R.output(lv2)
return lv2
```
Without the new check logic, this function would be fused as biased matmul,
which isn't correct because the bias (`lv1`) depends on the result of matmul.
cc @vinx13 @masahi
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
