[QNN] Concatenate operator #3730
Conversation
|
Thank you @anijain2305 , please remind me when this PR rebased to a merged requantize head :) |
anijain2305
force-pushed
the
qnn_concatenate
branch
2 times, most recently
from
43be829
to
b232102
Compare
anijain2305
force-pushed
the
qnn_concatenate
branch
from
b232102
to
20e375a
Compare
|
@u99127 @ZihengJiang @vinx13 Please review. This (hopefully) should be quick. |
|
|
||
| # Find the dtype of the input expr. This is required for the requantize op. Since, this is | ||
| # concatenate op, the dtype of the input is same as dtype of the output. | ||
| data0 = relay.transform.infer_type(data[0]) |
There was a problem hiding this comment.
@anijain2305 This API has been removed in favor of InferType, please update and send another PR. Also I'm considering that requantize can by default use same output dtype if null is provided, so that we can avoid type inference here
There was a problem hiding this comment.
Just added a PR for fixing InferType. I think it missed CI because CI passed 3 days ago and InferType changes happened after that.
Regarding requantize, it makes sense. Let me think little more tomorrow and finalize this.
|
It seems the new added tests will raise error. http://ci.tvm.ai:8080/blue/organizations/jenkins/tvm/detail/PR-3762/5/pipeline Since this breaks CI for other PRs, could you raise a PR to fix it soon? @anijain2305 |
This reverts commit f06ef4f.
|
@ZihengJiang @vinx13 Added the following quick fix for InferType |
@FrozenGene @u99127 @jackwish @tqchen
Pinging for review. The code changes will be very small after we have requantize merged in.