Add div decomposition in ArmQuantizer #5267
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/5267
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 2 Unrelated FailuresAs of commit 104d3a8 with merge base 20a157f ( NEW FAILURE - The following job has failed:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
Erik-Lundell
added
partner: arm
| @@ -45,3 +50,8 @@ def transform_to_backend_pipeline( | |||
| self.add_pass(AnnotateChannelsLastDimOrder()) | |||
|
|
|||
| return self._transform(graph_module) | |||
|
|
|||
| def transform_for_annotation_pipeline(self, graph_module: torch.fx.GraphModule): | |||
| self.add_pass(DecomposeDivPass()) | |||
There was a problem hiding this comment.
I guess the motivation for this API is clear and this is allowed, however I do want to make sure modifying graph in the Quantizer is something we are OK with from general ET lowering point of view. I also see SDPA as a valid use case.
cc @kimishpatel, @mergennachin - any comments?
There was a problem hiding this comment.
Ok - I'll resolve the merge conflict and mark this as ready for review tomorrow unless I get new comments.
Erik-Lundell
force-pushed
the
op_div
branch
from
2edce3d to
af8f22b
Compare
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
still failing with |
Erik-Lundell
force-pushed
the
op_div
branch
from
124fc23 to
cd6dea4
Compare
I was not able to find this error in the logs of the failed jobs here on GH and could not reproduce this locally but I rebased and forced-push and now only one job fails from what I can see |
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
|
||
| if is_quant_node: | ||
| input = inputs[0] | ||
| input_qargs = get_quant_node_args(node.args[0]) |
There was a problem hiding this comment.
New internal error
executorch/backends/arm/operators/op_reciprocal.py:44:46 Incompatible parameter type [6]: In call get_quant_node_args, for 1st positional argument, expected Node but got Union[None, List[typing.Any], Dict[str, typing.Any], bool, complex, float, int, range, slice, str, SymBool, SymFloat, SymInt, device, dtype, layout, memory_format, OpOverload, Tensor, Node, typing.Tuple[typing.Any, ...]].
There was a problem hiding this comment.
Ok, I assume this is due to some internal extra type checker? I realized that get_quant_node_args was called differently than in other node visitors, hopefully fixing that solves the error.
|
Also rebase when you fix this 😓 |
Implements node visitor and tests Quantized by one_to_one annotator Signed-off-by: Erik Lundell <[email protected]> Change-Id: I3f25096a8b908d7c25b6cd83bd7edb6871b145ab
Implements pass that decomposes aten.div to aten.reciprocal and aten.mul. This is done in the Quantizer get quantization annotation on the decomposed operators. Add infra for passes in ArmQuantizer Signed-off-by: Erik Lundell <[email protected]> Change-Id: Idd1698dc5fc82ab42b68094b405fb3a08804a45e
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@digantdesai merged this pull request in c5fdebd. |
Suggestion or what a pass to decompose a div in ArmQuantizer might look like.