Fix the precision issue when slice/add are connected.

espressif · Jan 2, 2025 · f2ffc61 · f2ffc61
1 parent 34b6d8a
commit f2ffc61
Show file tree

Hide file tree

Showing 3 changed files with 9 additions and 9 deletions.
diff --git a/ppq/api/espdl_interface.py b/ppq/api/espdl_interface.py
@@ -177,6 +177,11 @@ def espdl_quantize_onnx(
         BaseGraph:      The Quantized Graph, containing all information needed for backend execution
     """    
 
+    model = onnx.load(onnx_import_file)
+    model_sim, check = simplify(model)
+    if check:
+        onnx.save(model_sim, onnx_import_file)
+
     export_path = os.path.dirname(os.path.abspath(espdl_export_file))
     os.makedirs(export_path, exist_ok=True)
 
@@ -357,11 +362,6 @@ def espdl_quantize_torch(
         opset_version=13,
         do_constant_folding=True,
     )
-
-    model = onnx.load(onnx_file_path)
-    model_sim, check = simplify(model)
-    if check:
-        onnx.save(model_sim, onnx_file_path)
 
     # step2: quantize onnx model and export espdl model
     return espdl_quantize_onnx(

diff --git a/ppq/executor/torch.py b/ppq/executor/torch.py
@@ -527,9 +527,9 @@ def __forward(
                                         for input, config in zip(inputs, input_configs)])
 
                 # PATCH 20220208
-                for idx, var in enumerate(operation.inputs):
-                    if var.name in output_names:
-                        result_collector[output_names.index(var.name)] = inputs[idx]
+                # for idx, var in enumerate(operation.inputs):
+                #     if var.name in output_names:
+                #         result_collector[output_names.index(var.name)] = inputs[idx]
 
                 # invoking pre-forward hook
                 if operation_runtime_hook is not None:

diff --git a/ppq/quantization/optim/refine.py b/ppq/quantization/optim/refine.py
@@ -545,7 +545,7 @@ def optimize(self, graph: BaseGraph, **kwargs) -> None:
                 for up_op in graph.get_upstream_operations(operation):
                     if not isinstance(up_op, QuantableOperation): continue
 
-                    if len(graph.get_downstream_operations(up_op)) != 1 and not self.force_overlap: continue
+                    if (up_op.type in PASSIVE_OPERATIONS or len(graph.get_downstream_operations(up_op)) != 1) and not self.force_overlap: continue
                     # for cfg, var in up_op.config_with_variable:
                     for cfg, var in zip(up_op.config.output_quantization_config, up_op.outputs):
                         if operation in var.dest_ops: