xnnpack io

Author: lucylq

backends/xnnpack/partition/config/gemm_configs.py

@@ -122,6 +122,7 @@ def _get_weight_deps(
         self, node: torch.fx.Node, ep: ExportedProgram, precision: ConfigPrecisionType
     ) -> Tuple[bool, List[torch.fx.Node]]:
         gemm_deps = []
+        breakpoint()
         if precision == ConfigPrecisionType.FP32:
             # First find the weight
             weight_node = get_input_node(node, self.weight_idx)
@@ -272,6 +273,17 @@ def _get_weight_deps(
 
         return super()._get_weight_deps(node, ep, precision)
 
+    def _get_bias_deps(
+        self, node: torch.fx.Node, ep: ExportedProgram, precision: ConfigPrecisionType
+    ) -> Tuple[bool, List[torch.fx.Node]]:
+        if precision == ConfigPrecisionType.FP32 and self.force_fp32_dynamic_linear:
+            # if force fp32_dynamic_linear is on and we detected this as fp32, then we
+            # do not partition the weight node
+            breakpoint()
+            return (True, [])
+
+        return super()._get_bias_deps(node, ep, precision)
+
     def supported_precision_types(self):
         return [
             ConfigPrecisionType.DYNAMIC_QUANT,

examples/xnnpack/aot_compiler.py

@@ -93,6 +93,7 @@
         # TODO(T165162973): This pass shall eventually be folded into quantizer
         model = quantize(model, example_inputs)
 
+    breakpoint()
     edge = to_edge_transform_and_lower(
         ep,
         partitioner=[XnnpackPartitioner()],
@@ -110,6 +111,8 @@
         config=ExecutorchBackendConfig(extract_delegate_segments=False)
     )
 
+    breakpoint()
+
     if args.etrecord is not None:
         generate_etrecord(args.etrecord, edge_copy, exec_prog)
         logging.info(f"Saved ETRecord to {args.etrecord}")