Disabling xnnpack except for dynamic quantlized linear (pytorch#391)

kimishpatel · malfet · commit a25e23356b18 · 2024-07-17T09:55:42.000-07:00
Summary:
In executorch this path was slower compared to using optimized ops. We
did not debug that further, but for now just disabled it.

Test Plan:
CI

Reviewers:

Subscribers:

Tasks:

Tags:
diff --git a/export_et.py b/export_et.py
@@ -97,7 +97,6 @@ def export_model(model, device, output_path, args=None) -> str:  # noqa: C901
             edge_compile_config=edge_config,
         )
     edge_manager = edge_manager.to_backend(XnnpackDynamicallyQuantizedPartitioner())
-    edge_manager = edge_manager.to_backend(XnnpackPartitioner())
     export_program = edge_manager.to_executorch(
         ExecutorchBackendConfig(
             extract_constant_segment=True,

Original file line number	Diff line number	Diff line change
`@@ -97,7 +97,6 @@ def export_model(model, device, output_path, args=None) -> str: # noqa: C901`
`97`	`97`	`edge_compile_config=edge_config,`
`98`	`98`	`)`
`99`	`99`	`edge_manager = edge_manager.to_backend(XnnpackDynamicallyQuantizedPartitioner())`
`100`		`- edge_manager = edge_manager.to_backend(XnnpackPartitioner())`
`101`	`100`	`export_program = edge_manager.to_executorch(`
`102`	`101`	`ExecutorchBackendConfig(`
`103`	`102`	`extract_constant_segment=True,`