saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-29 20:14:01 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

[Quant tool] Extend support for QDQ type conversion at graph output (#20841)

Browse source 
### Description
Allows mixed-precision overrides that adds a QDQ quantization type
conversion sequence at a graph output that **is not** consumed by other
nodes. This is not a common use-case but should handle it instead of
raising an error.

#### Example
Original model

![image](https://github.com/microsoft/onnxruntime/assets/19691973/4c9c3bb0-4ca1-4213-9259-9d0506ed22f2)

mixed-precision overrides:
```python
        mixed_prec_overrides = {
            "input_0": [{"quant_type": QuantType.QUInt16}],
            "op_0_out": [
                {
                    "quant_type": QuantType.QUInt16,
                    "convert": {"quant_type": QuantType.QUInt8},
                }
            ],
        }
        quantize_static(
            float_model_path,
            qdq_model_path,
            data_reader,
            quant_format=QuantFormat.QDQ,
            activation_type=QuantType.QUInt8,
            op_types_to_quantize=[node.op_type for node in float_model.graph.node],
            extra_options={
                "TensorQuantOverrides": mixed_prec_overrides,
            },
        )
```

QDQ model:

![image](https://github.com/microsoft/onnxruntime/assets/19691973/804fc89b-4a00-43bc-a4ff-21edd6f27e98)

### Motivation and Context
This scenario is arising for certain quantization configurations. Should
handle it gracefully.
This commit is contained in:
Adrian Lizarraga 2024-05-28 21:27:54 -07:00 committed by GitHub
parent d44be41e1c
commit 3044aa8743
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 57 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
onnxruntime/python/tools/quantization/qdq_quantizer.py
View file

@ -636,8 +636,12 @@ class QDQQuantizer(BaseQuantizer):
| +-> <Graph output>
|
+-> DQ1' ---> Q2 ---> DQ2 ---> <Consumers of converted type>
5) Tensor T is a graph output that is not consumed by any other nodes.
<Producer> ---> Q1 ---> DQ1 ---> Q2 ---> DQ2 ---> <Graph output>
"""
tensor_recv_nodes = set([node.name for node in self.tensor_to_its_receiving_nodes[tensor_name]])
tensor_recv_nodes = set([node.name for node in self.tensor_to_its_receiving_nodes.get(tensor_name, [])])
if (
self.dedicated_qdq_pair

52
onnxruntime/test/python/quantization/test_qdq.py
View file

@ -1335,6 +1335,58 @@ class TestQDQMixedPrecision(TestQDQFormat):
self.assertIn("op_0_out", graph_outputs)
def test_add_tensor_qdq_ops_case_5(self):
"""
Tensor T is a graph output without any consumers.
<Producer> ---> Q1 ---> DQ1 ---> Q2 ---> DQ2 ---> <Graph output>
"""
float_model_path = os.path.join(self._tmp_dir_path, "case_5.onnx")
qdq_model_path = os.path.join(self._tmp_dir_path, "case_5.qdq.onnx")
# Build model with input_0 -> op_0 -> op_0_out
# The graph output has no consumers.
float_model = self.build_test_model_for_add_qdq_ops(0, True)
onnx.save_model(float_model, float_model_path)
data_reader = self.input_feeds(3, {"input_0": (1, 2, 3)}, np.float32)
mixed_prec_overrides = {
"input_0": [{"quant_type": QuantType.QUInt16}],
"op_0_out": [
{
"quant_type": QuantType.QUInt16,
"convert": {"quant_type": QuantType.QUInt8},
}
],
}
quantize_static(
float_model_path,
qdq_model_path,
data_reader,
quant_format=QuantFormat.QDQ,
activation_type=QuantType.QUInt8,
op_types_to_quantize=[node.op_type for node in float_model.graph.node],
extra_options={
"TensorQuantOverrides": mixed_prec_overrides,
},
)
# Expect the following QDQ model:
# input_0 --> Q_16 --> DQ_16 --> op_0 --> Q_16 --> DQ_16 --> Q_8 --> DQ_8 --> output_0
qdq_node_counts = {"QuantizeLinear": 3, "DequantizeLinear": 3}
check_op_type_count(self, qdq_model_path, **qdq_node_counts)
qdq_model = onnx.load_model(qdq_model_path)
onnx.checker.check_model(qdq_model, True)
initializers = {init.name: init for init in qdq_model.graph.initializer}
# Check zero-point data types
orig_zp_init = initializers["op_0_out_zero_point"]
self.assertEqual(orig_zp_init.data_type, onnx.TensorProto.UINT16)
convert_zp_init = initializers["op_0_out_zero_point_convert"]
self.assertEqual(convert_zp_init.data_type, onnx.TensorProto.UINT8)
def build_test_model_1(self, shape):
"""
Returns the following float32 model.