Let quantizers add custom meta data to GraphModule #2711
Conversation
Make the 'custom' meta data field in the GraphModule be propageted with any data that is added by the quantizer. Signed-off-by: Per Åstrand <[email protected]>
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/2711
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 3b487e3 with merge base 5d99ce4 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
|
@msaroufim, @cpuhrsch: You are marked as code owners, so reaching out to you. Who do I ping to get the workflow running and reviewers added? |
| @@ -116,13 +116,17 @@ def calibrate(model, data_loader): | |||
| model = quantizer.transform_for_annotation(model) | |||
| quantizer.annotate(model) | |||
| quantizer.validate(model) | |||
| # Store the 'custom' meta data if any added by quantizer | |||
| annotated_meta = model.meta.get("custom", {}) | |||
There was a problem hiding this comment.
what's the intended use for this?
There was a problem hiding this comment.
I want to pass on a list of the annotated nodes to the partitioner and then on to the exported programs graph_module in order to validate the quantization folding decisions in the backend when we start to mix INT and FP. The ET part of it (passing on the 'custom' meta to the edge program) is also needed, but otherwise the 'custom' field is handled in the re-exports to survive all the way to that point.
Make the 'custom' meta data field in the GraphModule be propageted with any data that is added by the quantizer.