quic-sanirudh commented on issue #16489:
URL: https://github.com/apache/tvm/issues/16489#issuecomment-1923998891
I tried this out and it looks like the issue is that dtype has to be
specified when creating the `decl_buffer` for specialization. The default dtype
would become float32 and the const version directly specified float16.
Something like this
```python
gemm_sp = gemm.specialize(
{
data: tvm.tir.decl_buffer((4096, 4096), dtype="float16"),
weight: tvm.tir.decl_buffer((4096, 4096), dtype="float16"),
}
)
```
However the printed IR for the specialized version is misleading as it shows
the dtype as "float16", which I understand is why this issue was created, so
that probably has to be fixed.
--
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]
