Add lift scalar to constant tensor pass#8313
Conversation
Differential Revision: D69318149
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/8313
Note: Links to docs will display an error until the docs builds have been completed. ❌ 3 New FailuresAs of commit 71eb924 with merge base 77f18b2 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
This pull request was exported from Phabricator. Differential Revision: D69318149 |
This PR needs a
|
haowhsu-quic
left a comment
There was a problem hiding this comment.
This is super helpful for us instead of tackling constant inside op builder. Will try make it more general and do some clean up jobs in codebase.
Thank you for this PR.
|
#8642 is merged instead |
Differential Revision: D69318149
Sometime the op in the graph will receive a scalar input, and quantization can only deal with tensor. In this case, the recommended workflow is to lift the scalar constant to be a tensor, and annotate them, then these constants will have corresponding q/dq nodes.
This PR is mostly a demonstration regarding where to insert the pass and how the pass should look like