feat: Add BF16 tensor support via dlpack

README.md

@@ -1557,6 +1557,10 @@ input0 = pb_utils.Tensor.from_dlpack("INPUT0", pytorch_tensor)
 This method only supports contiguous Tensors that are in C-order. If the tensor
 is not C-order contiguous an exception will be raised.
 
+For python models with input or output tensors of type BFloat16 (BF16), the
+`as_numpy()` method is not supported, and the `from_dlpack` and `to_dlpack`
+methods must be used instead.
+
 ## `pb_utils.Tensor.is_cpu() -> bool`
 
 This function can be used to check whether a tensor is placed in CPU or not.

src/pb_stub_utils.cc

@@ -1,4 +1,4 @@
-// Copyright 2021-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// Copyright 2021-2024, 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
@@ -168,6 +168,8 @@ triton_to_pybind_dtype(TRITONSERVER_DataType data_type)
       dtype_numpy = py::dtype(py::format_descriptor<uint8_t>::format());
       break;
     case TRITONSERVER_TYPE_BF16:
+      // NOTE: Currently skipping this call via `if (BF16)` check, but may
+      // want to better handle this or set some default/invalid dtype.
       throw PythonBackendException("TYPE_BF16 not currently supported.");
     case TRITONSERVER_TYPE_INVALID:
       throw PythonBackendException("Dtype is invalid.");
@@ -240,6 +242,10 @@ triton_to_dlpack_type(TRITONSERVER_DataType triton_dtype)
     case TRITONSERVER_TYPE_BYTES:
       throw PythonBackendException(
           "TYPE_BYTES tensors cannot be converted to DLPack.");
+    case TRITONSERVER_TYPE_BF16:
+      dl_code = DLDataTypeCode::kDLBfloat;
+      dt_size = 16;
+      break;
 
     default:
       throw PythonBackendException(
@@ -301,6 +307,15 @@ dlpack_to_triton_type(const DLDataType& data_type)
     }
   }
 
+  if (data_type.code == DLDataTypeCode::kDLBfloat) {
+    if (data_type.bits != 16) {
+      throw PythonBackendException(
+          "Expected BF16 tensor to have 16 bits, but had: " +
+          std::to_string(data_type.bits));
+    }
+    return TRITONSERVER_TYPE_BF16;
+  }
+
   return TRITONSERVER_TYPE_INVALID;
 }
 }}}  // namespace triton::backend::python

src/pb_tensor.cc

@@ -1,4 +1,4 @@
-// Copyright 2021-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// Copyright 2021-2024, 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
@@ -152,7 +152,10 @@ PbTensor::PbTensor(
 #ifdef TRITON_PB_STUB
   if (memory_type_ == TRITONSERVER_MEMORY_CPU ||
       memory_type_ == TRITONSERVER_MEMORY_CPU_PINNED) {
-    if (dtype != TRITONSERVER_TYPE_BYTES) {
+    if (dtype == TRITONSERVER_TYPE_BF16) {
+      // No native numpy representation for BF16. DLPack should be used instead.
+      numpy_array_ = py::none();
+    } else if (dtype != TRITONSERVER_TYPE_BYTES) {
       py::object numpy_array =
           py::array(triton_to_pybind_dtype(dtype_), dims_, (void*)memory_ptr_);
       numpy_array_ = numpy_array.attr("view")(triton_to_numpy_type(dtype_));
@@ -512,12 +515,18 @@ PbTensor::Name() const
 const py::array*
 PbTensor::AsNumpy() const
 {
-  if (IsCPU()) {
-    return &numpy_array_;
-  } else {
+  if (!IsCPU()) {
     throw PythonBackendException(
         "Tensor is stored in GPU and cannot be converted to NumPy.");
   }
+
+  if (dtype_ == TRITONSERVER_TYPE_BF16) {
+    throw PythonBackendException(
+        "Tensor dtype is BF16 and cannot be converted to NumPy. Use "
+        "to_dlpack() and from_dlpack() instead.");
+  }
+
+  return &numpy_array_;
 }
 #endif  // TRITON_PB_STUB
 
@@ -643,7 +652,10 @@ PbTensor::PbTensor(
 #ifdef TRITON_PB_STUB
   if (memory_type_ == TRITONSERVER_MEMORY_CPU ||
       memory_type_ == TRITONSERVER_MEMORY_CPU_PINNED) {
-    if (dtype_ != TRITONSERVER_TYPE_BYTES) {
+    if (dtype_ == TRITONSERVER_TYPE_BF16) {
+      // No native numpy representation for BF16. DLPack should be used instead.
+      numpy_array_ = py::none();
+    } else if (dtype_ != TRITONSERVER_TYPE_BYTES) {
       py::object numpy_array =
           py::array(triton_to_pybind_dtype(dtype_), dims_, (void*)memory_ptr_);
       numpy_array_ = numpy_array.attr("view")(triton_to_numpy_type(dtype_));