test: Add BF16 test for python backend (#7483)

qa/L0_backend_python/python_test.py

@@ -365,6 +365,36 @@ def test_bool(self):
                 self.assertIsNotNone(output0)
                 self.assertTrue(np.all(output0 == input_data))
 
+    def test_bf16(self):
+        model_name = "identity_bf16"
+        shape = [2, 2]
+        with self._shm_leak_detector.Probe() as shm_probe:
+            with httpclient.InferenceServerClient(
+                f"{_tritonserver_ipaddr}:8000"
+            ) as client:
+                # NOTE: Client will truncate FP32 to BF16 internally
+                # since numpy has no built-in BF16 representation.
+                np_input = np.ones(shape, dtype=np.float32)
+                inputs = [
+                    httpclient.InferInput(
+                        "INPUT0", np_input.shape, "BF16"
+                    ).set_data_from_numpy(np_input)
+                ]
+                result = client.infer(model_name, inputs)
+
+                # Assert that Triton correctly returned a BF16 tensor.
+                response = result.get_response()
+                triton_output = response["outputs"][0]
+                triton_dtype = triton_output["datatype"]
+                self.assertEqual(triton_dtype, "BF16")
+
+                np_output = result.as_numpy("OUTPUT0")
+                self.assertIsNotNone(np_output)
+                # BF16 tensors are held in FP32 when converted to numpy due to
+                # lack of native BF16 support in numpy, so verify that.
+                self.assertEqual(np_output.dtype, np.float32)
+                self.assertTrue(np.allclose(np_output, np_input))
+
     def test_infer_pytorch(self):
         # FIXME: This model requires torch. Because windows tests are not run in a docker
         # environment with torch installed, we need to think about how we want to install

qa/L0_backend_python/test.sh

@@ -95,6 +95,9 @@ fi
 mkdir -p models/identity_fp32/1/
 cp ../python_models/identity_fp32/model.py ./models/identity_fp32/1/model.py
 cp ../python_models/identity_fp32/config.pbtxt ./models/identity_fp32/config.pbtxt
+mkdir -p models/identity_bf16/1/
+cp ../python_models/identity_bf16/model.py ./models/identity_bf16/1/model.py
+cp ../python_models/identity_bf16/config.pbtxt ./models/identity_bf16/config.pbtxt
 RET=0
 
 cp -r ./models/identity_fp32 ./models/identity_uint8

qa/python_models/identity_bf16/config.pbtxt

@@ -0,0 +1,51 @@
+# Copyright 2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#  * Redistributions of source code must retain the above copyright
+#    notice, this list of conditions and the following disclaimer.
+#  * Redistributions in binary form must reproduce the above copyright
+#    notice, this list of conditions and the following disclaimer in the
+#    documentation and/or other materials provided with the distribution.
+#  * Neither the name of NVIDIA CORPORATION nor the names of its
+#    contributors may be used to endorse or promote products derived
+#    from this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+# PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+backend: "python"
+max_batch_size: 64
+
+input [
+  {
+    name: "INPUT0"
+    data_type: TYPE_BF16
+    dims: [ -1 ]
+  }
+]
+
+output [
+  {
+    name: "OUTPUT0"
+    data_type: TYPE_BF16
+    dims: [ -1 ]
+  }
+]
+
+instance_group [
+  {
+    count: 1
+    kind : KIND_CPU
+  }
+]

qa/python_models/identity_bf16/model.py

@@ -0,0 +1,88 @@
+# Copyright 2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#  * Redistributions of source code must retain the above copyright
+#    notice, this list of conditions and the following disclaimer.
+#  * Redistributions in binary form must reproduce the above copyright
+#    notice, this list of conditions and the following disclaimer in the
+#    documentation and/or other materials provided with the distribution.
+#  * Neither the name of NVIDIA CORPORATION nor the names of its
+#    contributors may be used to endorse or promote products derived
+#    from this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+# PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+import json
+
+import torch
+import triton_python_backend_utils as pb_utils
+
+
+class TritonPythonModel:
+    def initialize(self, args):
+        # You must parse model_config. JSON string is not parsed here
+        self.model_config = json.loads(args["model_config"])
+
+        # Get tensor configurations for testing/validation
+        self.input0_config = pb_utils.get_input_config_by_name(
+            self.model_config, "INPUT0"
+        )
+        self.output0_config = pb_utils.get_output_config_by_name(
+            self.model_config, "OUTPUT0"
+        )
+
+    def validate_bf16_tensor(self, tensor, tensor_config):
+        # I/O datatypes can be queried from the model config if needed
+        dtype = tensor_config["data_type"]
+        if dtype != "TYPE_BF16":
+            raise Exception(f"Expected a BF16 tensor, but got {dtype} instead.")
+
+        # Converting BF16 tensors to numpy is not supported, and DLPack
+        # should be used instead via to_dlpack and from_dlpack.
+        try:
+            _ = tensor.as_numpy()
+        except pb_utils.TritonModelException as e:
+            expected_error = "tensor dtype is bf16 and cannot be converted to numpy"
+            assert expected_error in str(e).lower()
+        else:
+            raise Exception("Expected BF16 conversion to numpy to fail")
+
+    def execute(self, requests):
+        """
+        Identity model in Python backend with example BF16 and PyTorch usage.
+        """
+        responses = []
+        for request in requests:
+            input_tensor = pb_utils.get_input_tensor_by_name(request, "INPUT0")
+
+            # Numpy does not support BF16, so use DLPack instead.
+            bf16_dlpack = input_tensor.to_dlpack()
+
+            # OPTIONAL: The tensor can be converted to other dlpack-compatible
+            # frameworks like PyTorch and TensorFlow with their dlpack utilities.
+            torch_tensor = torch.utils.dlpack.from_dlpack(bf16_dlpack)
+
+            # When complete, convert back to a pb_utils.Tensor via DLPack.
+            output_tensor = pb_utils.Tensor.from_dlpack(
+                "OUTPUT0", torch.utils.dlpack.to_dlpack(torch_tensor)
+            )
+            responses.append(pb_utils.InferenceResponse([output_tensor]))
+
+            # NOTE: The following helper function is for testing and example
+            # purposes only, you should remove this in practice.
+            self.validate_bf16_tensor(input_tensor, self.input0_config)
+            self.validate_bf16_tensor(output_tensor, self.output0_config)
+
+        return responses