chore: import upstream snapshot with attribution

samples/python/non_zero_plugin/README.md

@@ -0,0 +1,110 @@
+# Python-based NonZero Plugin for TensorRT using IPluginV3
+
+## Description
+
+This sample, `non_zero_plugin`, implements a Python-based plugin for the NonZero operation, configurable to use a `CUDA Python` or `PyTorch` backend.
+
+NonZero is an operation where the non-zero indices of the input tensor is found.
+
+## How does this sample work?
+
+This sample creates and runs a TensorRT engine built from a network containing a single NonZeroPlugin node. It demonstrates how
+custom layers with data-dependent output shapes can be implemented and added to a TensorRT network using Python.
+
+### Implementing a NonZero plugin using IPluginV3 interface
+
+Until `IPluginV3` (and associated interfaces), TensorRT plugins could not have outputs whose shapes depended on the input values (they could only depend
+on input shapes). `IPluginV3OneBuild` which exposes a build capability for `IPluginV3`, provides support for such data-dependent output shapes.
+
+`NonZeroPlugin` in this sample is written to handle 2-D input tensors of shape $R \times C$. Assume that the tensor contains $K$ non-zero elements and that the
+non-zero indices are required in a row ordering (each set of indices in its own row). Then the output shape would be $K \times 2$.
+
+The output shapes are expressed to the TensorRT builder through the `IPluginV3OneBuild.get_output_shapes()` API. Expressing the second dimension of the output is
+straightforward:
+```
+# output_dims[0] = trt.DimsExprs(2)
+output_dims[0][1] = exprBuilder.constant(2)
+```
+
+The extent of each data-dependent dimension in the plugin must be expressed in terms of a *_size tensor_*. A size tensor is a scalar output of type
+`trt.int32` or `trt.int64` that must be added as one of the plugin outputs. In this case, it is sufficient to declare one size tensor to denote the extent of the
+first dimension of the non-zero indices output. To declare a size tensor, one must provide an upper-bound and optimum value for its extent as `IDimensionExpr`s. These can be formed through the `IExprBuilder` argument passed to the `IPluginV3OneBuild.get_output_shapes()` method.
+ - For unknown inputs, the upper-bound is the total number of elements in the input
+	```
+	upper_bound = exprBuilder.operation(trt.DimensionOperation.PROD, inputs[0][0], inputs[0][1])
+	```
+ - A good estimate for the optimum is that half of the elements are non-zero
+	```
+	opt_value = exprBuilder.operation(trt.DimensionOperation.FLOOR_DIV, upper_bound, exprBuilder.constant(2))
+	```
+
+Now we can declare the size tensor using the `IExprBuilder.declare_size_tensor()` method, which also requires the specification of the output index at which the size tensor would reside. Let us place it after the non-zero indices output:
+```
+num_non_zero_size_tensor = exprBuilder.declare_size_tensor(1, opt_value, upper_bound)
+```
+
+Now we are ready to specify the extent of the first dimension of the non-zero indices output:
+```
+# output_dims[0] = trt.DimsExprs(0)
+output_dims[0][0] = num_non_zero_size_tensor
+```
+Note that the size tensor is declared to be a scalar (0-D):
+
+### Creating network and building the engine
+
+To add the plugin to the network, the `INetworkDefinition::add_plugin_v3()` method must be used.
+
+Similar to `IPluginCreator` used for V2 plugins, V3 plugins must be accompanied by the registration of a plugin creator implementing the `IPluginCreatorV3One` interface.
+
+## Running the sample
+
+1.  Run the sample to create a TensorRT inference engine and run inference:
+    `python3 non_zero_plugin.py [-h] [--precision {fp32,fp16}] [--backend {cuda_python,torch}] [--net_type {onnx,inetdef}]`
+
+2.  Verify that the sample ran successfully. If the sample runs successfully you should see the following message:
+     ```
+    Inference result correct!
+    ```
+
+### Sample `--help` options
+
+To see the full list of available options and their descriptions, use the `-h` or `--help` command line option.
+
+
+# Additional resources
+
+The following resources provide a deeper understanding about the V3 TensorRT plugins and the NonZero operation:
+
+**NonZero**
+- [ONNX: NonZero](https://onnx.ai/onnx/operators/onnx__NonZero.html)
+
+**C++-based NonZero Plugin sample**
+- [NonZero C++ Plugin](../../sampleNonZeroPlugin/)
+
+**TensorRT plugins**
+- [Extending TensorRT with Custom Layers](https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#extending)
+- [TensorRT Python-based Plugins](https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/#add_custom_layer_python)
+
+**Other documentation**
+- [Introduction To NVIDIA’s TensorRT Samples](https://docs.nvidia.com/deeplearning/sdk/tensorrt-sample-support-guide/index.html#samples)
+- [Working With TensorRT Using The Python API](https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/#python_topics)
+- [NVIDIA’s TensorRT Documentation Library](https://docs.nvidia.com/deeplearning/sdk/tensorrt-archived/index.html)
+
+# License
+
+For terms and conditions for use, reproduction, and distribution, see the [TensorRT Software License Agreement](https://docs.nvidia.com/deeplearning/sdk/tensorrt-sla/index.html) documentation.
+
+# Changelog
+
+October 2025
+Migrate to strongly typed APIs.
+
+August 2025
+Removed support for Python versions < 3.10.
+
+April 2024
+This is the first version of this `README.md` file.
+
+# Known issues
+
+There are no known issues in this sample.

samples/python/non_zero_plugin/non_zero_plugin.py

@@ -0,0 +1,352 @@
+#
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import onnx_graphsurgeon as gs
+import numpy as np
+import onnx
+import os
+import sys
+
+import tensorrt as trt
+from polygraphy.backend.trt import (
+    CreateConfig,
+    EngineFromNetwork,
+    NetworkFromOnnxPath,
+    TrtRunner,
+    create_network,
+    engine_from_network,
+)
+
+import argparse
+
+from polygraphy import mod
+
+sys.path.insert(1, os.path.join(os.path.dirname(os.path.realpath(__file__)), os.pardir))
+from plugin_utils import cuda_call, KernelHelper, UnownedMemory, volume
+
+
+cuda = mod.lazy_import("cuda.bindings.driver")
+cudart = mod.lazy_import("cuda.bindings.runtime")
+nvrtc = mod.lazy_import("cuda.bindings.nvrtc")
+
+torch = mod.lazy_import("torch")
+cp = mod.lazy_import("cupy")
+
+non_zero_half_kernel = r'''
+#include <cuda_fp16.h>
+extern "C" __global__
+void find_non_zero_indices_half(
+    half const* X, int* indices, unsigned long long* count, int R, int C)
+{
+    static_assert(sizeof(unsigned long long) == 8U, "unsigned long long must be 8 bytes in NVCC");
+    int row = blockIdx.x * blockDim.x + threadIdx.x;
+
+    // Check if the row index is within bounds
+    if (row < R)
+    {
+        for (int col = 0; col < C; ++col)
+        {
+            half const z = static_cast<half>(0.F);
+            if (X[col + C * row] != z)
+            {
+                // Increment count atomically and get the previous value
+                unsigned long long index = atomicAdd(count, 1ULL);
+                indices[2 * index] = row;
+                indices[2 * index + 1] = col;
+            }
+        }
+    }
+}
+'''
+
+non_zero_float_kernel = r'''
+extern "C" __global__
+void find_non_zero_indices_float(
+    float const* X, int* indices, unsigned long long* count, int R, int C)
+{
+    static_assert(sizeof(unsigned long long) == 8U, "unsigned long long must be 8 bytes in NVCC");
+    int row = blockIdx.x * blockDim.x + threadIdx.x;
+
+    // Check if the row index is within bounds
+    if (row < R)
+    {
+        for (int col = 0; col < C; ++col)
+        {
+            if (X[col + C * row] != 0.F)
+            {
+                // Increment count atomically and get the previous value
+                unsigned long long index = atomicAdd(count, 1ULL);
+                indices[2 * index] = row;
+                indices[2 * index + 1] = col;
+            }
+        }
+    }
+}
+'''
+
+class NonZeroPlugin(trt.IPluginV3, trt.IPluginV3OneCore, trt.IPluginV3OneBuild, trt.IPluginV3OneRuntime):
+    def __init__(self, backend = None):
+        trt.IPluginV3.__init__(self)
+        trt.IPluginV3OneCore.__init__(self)
+        trt.IPluginV3OneBuild.__init__(self)
+        trt.IPluginV3OneRuntime.__init__(self)
+
+        self.num_outputs = 2
+        self.plugin_namespace = ""
+        self.plugin_name = "NonZeroPlugin"
+        self.plugin_version = "1"
+
+        if backend is not None:
+            self.backend = backend.tobytes().decode("utf-8")
+        else:
+            self.backend = "cuda_python"
+
+        self.cuDevice = None
+
+    def get_capability_interface(self, type):
+        return self
+
+    def get_output_data_types(self, input_types):
+        return [trt.DataType.INT32, trt.DataType.INT64]
+
+    def get_output_shapes(self, inputs, shape_inputs, exprBuilder):
+        # First output is 2-D
+        # Second output is a size tensor, which must be declared a scalar (0-D)
+        output_dims = [trt.DimsExprs(2), trt.DimsExprs(0)]
+
+        upper_bound = exprBuilder.operation(trt.DimensionOperation.PROD, inputs[0][0], inputs[0][1])
+        opt_value = exprBuilder.operation(trt.DimensionOperation.FLOOR_DIV, upper_bound, exprBuilder.constant(2))
+        num_non_zero_size_tensor = exprBuilder.declare_size_tensor(1, opt_value, upper_bound)
+
+        output_dims[0][0] = num_non_zero_size_tensor
+        output_dims[0][1] = exprBuilder.constant(2)
+
+        return output_dims
+
+    def get_fields_to_serialize(self):
+        return trt.PluginFieldCollection(
+            [
+                trt.PluginField(
+                    "backend", self.backend.encode(), trt.PluginFieldType.CHAR
+                )
+            ]
+        )
+
+    def configure_plugin(self, inp, out):
+        if self.backend == "cuda_python":
+            self.cuDevice = cuda_call(cuda.cuDeviceGet(0))
+
+    def on_shape_change(self, inp, out):
+        if self.backend == "cuda_python":
+            self.cuDevice = cuda_call(cuda.cuDeviceGet(0))
+
+    def supports_format_combination(self, pos, in_out, num_inputs):
+        assert num_inputs == 1
+        assert pos < len(in_out)
+
+        type_ok = False
+
+        # first input should be float16 or float32
+        if pos == 0:
+            type_ok = in_out[0].desc.type == trt.DataType.FLOAT or in_out[0].desc.type == trt.DataType.HALF
+        elif pos == 1:
+            type_ok = in_out[1].desc.type == trt.DataType.INT32
+        else: # pos == 2
+            # size tensor outputs must be NCHW INT64
+            type_ok = in_out[2].desc.type == trt.DataType.INT64
+
+        return in_out[pos].desc.format == trt.TensorFormat.LINEAR and type_ok
+
+    def enqueue(self, input_desc, output_desc, inputs, outputs, workspace, stream):
+        inp_dtype = trt.nptype(input_desc[0].type)
+
+        if self.backend == "cuda_python":
+            R = input_desc[0].dims[0]
+            C = input_desc[0].dims[1]
+
+            blockSize = 256
+            numBlocks = int((C + blockSize - 1) // blockSize)
+
+            d_in = np.array([inputs[0]], dtype=np.uint64)
+            d_out_0 = np.array([outputs[0]], dtype=np.uint64)
+            d_out_1 = np.array([outputs[1]], dtype=np.uint64)
+
+            args = [d_in, d_out_0, d_out_1, np.array(R, dtype=np.uint32), np.array(C, dtype=np.uint32)]
+            kernelArgs = np.array([arg.ctypes.data for arg in args], dtype=np.uint64)
+
+            stream_ptr = np.array([stream], dtype=np.uint64)
+
+            if inp_dtype == np.float32:
+                kernelHelper = KernelHelper(non_zero_float_kernel, int(self.cuDevice))
+                _non_zero_float_kernel = kernelHelper.getFunction(b'find_non_zero_indices_float')
+                cuda_call(cuda.cuLaunchKernel(_non_zero_float_kernel,
+                                            numBlocks, 1, 1,
+                                            blockSize, 1, 1,
+                                            0,
+                                            stream_ptr,
+                                            kernelArgs, 0))
+            elif inp_dtype == np.float16:
+                kernelHelper = KernelHelper(non_zero_half_kernel, int(self.cuDevice))
+                _non_zero_half_kernel = kernelHelper.getFunction(b'find_non_zero_indices_half')
+                cuda_call(cuda.cuLaunchKernel(_non_zero_half_kernel,
+                                            numBlocks, 1, 1,
+                                            blockSize, 1, 1,
+                                            0,
+                                            stream_ptr,
+                                            kernelArgs, 0))
+            else:
+                raise ValueError("inp_dtype not valid")
+
+        elif self.backend == "torch":
+            inp_mem = UnownedMemory(inputs[0], input_desc[0].dims, inp_dtype)
+
+            out_mem = UnownedMemory(
+                outputs[0], 2 * volume(input_desc[0].dims), np.int32
+            )
+
+            out_1_mem = UnownedMemory(outputs[1], 1, np.int64)
+
+            a_t = torch.as_tensor(inp_mem.d, device="cuda")
+            out = torch.nonzero(a_t)
+
+            out_mem.d[: volume(out.shape)] = cp.reshape(cp.asarray(out), (-1,))
+            cp.copyto(out_1_mem.d, cp.reshape(cp.asarray([out.shape[0]]), (-1,)))
+
+        else:
+            raise ValueError(f"backend not valid: {self.backend}")
+
+    def attach_to_context(self, context):
+        return self.clone()
+
+    def set_tactic(self, tactic):
+        pass
+
+    def clone(self):
+        cloned_plugin = NonZeroPlugin()
+        cloned_plugin.__dict__.update(self.__dict__)
+        return cloned_plugin
+
+    #
+    # The following defaults take effect since the respective methods are not overriden
+    #
+
+    # def get_valid_tactics(self):
+    #     return []
+
+    # def get_workspace_size(self, input_desc, output_desc):
+    #     return 0
+
+    # def destroy(self):
+    #     pass
+
+
+class NonZeroPluginCreator(trt.IPluginCreatorV3One):
+    def __init__(self):
+        trt.IPluginCreatorV3One.__init__(self)
+        self.name = "NonZeroPlugin"
+        self.plugin_namespace = ""
+        self.plugin_version = "1"
+        self.field_names = trt.PluginFieldCollection(
+            [trt.PluginField("backend", np.array([]), trt.PluginFieldType.CHAR)]
+        )
+
+    def create_plugin(self, name, fc, phase):
+        backend = None
+        for f in fc:
+            if f.name == "backend":
+                backend = f.data[:-1] if f.data[-1] == 0 else f.data
+        return NonZeroPlugin(backend)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--precision', type=str, default="fp32", choices=["fp32", "fp16"])
+    parser.add_argument("--backend", type=str, default="torch", choices=["cuda_python", "torch"])
+    parser.add_argument('--net_type', type=str, default="onnx", choices=["onnx", "inetdef"])
+
+    args = parser.parse_args()
+
+    if args.backend == "cuda_python":
+        # Initialize CUDA and create default context
+        cuda_call(cudart.cudaFree(0))
+
+    elif args.backend == "torch":
+        # Initialize CUDA and create default context
+        torch.cuda.init()
+
+    precision = np.float32 if args.precision == "fp32" else np.float16
+
+    inp_shape = (128, 128)
+    X = np.random.normal(size=inp_shape).astype(precision)
+    # Zero out a random set of indices
+    indices = np.random.choice(np.prod(inp_shape), replace=False, size=np.random.randint(0, np.prod(inp_shape) + 1))
+    X[np.unravel_index(indices, inp_shape)] = 0
+
+    # Register plugin creator
+    plg_registry = trt.get_plugin_registry()
+    my_plugin_creator = NonZeroPluginCreator()
+    plg_registry.register_creator(my_plugin_creator, "")
+
+    if args.net_type == "onnx":
+        # create ONNX model
+        onnx_path = "test_NonZeroPlugin.onnx"
+        inputX = gs.Variable(name="X", shape=inp_shape, dtype=precision)
+        Y = gs.Variable(name="Y", dtype=np.int32)
+        Y_num = gs.Variable(name="Y_num", dtype=np.int64)
+        nonZeroPluginNode = gs.Node(
+            name="NonZeroPlugin",
+            op="NonZeroPlugin",
+            inputs=[inputX],
+            outputs=[Y, Y_num],
+            attrs={"backend": args.backend.encode()},
+        )
+        graph = gs.Graph(nodes=[nonZeroPluginNode], inputs=[inputX], outputs=[Y], opset=16)
+        onnx.save(gs.export_onnx(graph), onnx_path)
+
+        # build engine
+        build_engine = EngineFromNetwork(
+            NetworkFromOnnxPath(onnx_path, strongly_typed=True), CreateConfig()
+        )
+    else:
+        # Create plugin object
+        builder, network = create_network(strongly_typed=True)
+        plg_creator = plg_registry.get_creator("NonZeroPlugin", "1", "")
+        plugin_fields_list = [
+            trt.PluginField("backend", args.backend.encode(), trt.PluginFieldType.CHAR)
+        ]
+        pfc = trt.PluginFieldCollection(plugin_fields_list)
+        plugin = plg_creator.create_plugin("NonZeroPlugin", pfc, trt.TensorRTPhase.BUILD)
+
+        # Populate network
+        inputX = network.add_input(name="X", dtype=trt.float32 if precision==np.float32 else trt.float16, shape=inp_shape)
+        out = network.add_plugin_v3([inputX], [], plugin)
+        out.get_output(0).name = "Y"
+        network.mark_output(tensor=out.get_output(0))
+        build_engine = engine_from_network((builder, network), CreateConfig())
+
+    # Compare against Numpy's nonzero
+    Y_ref = np.transpose(np.nonzero(X))
+
+    # Run
+    with TrtRunner(build_engine, "trt_runner") as runner:
+        outputs = runner.infer({"X": X})
+        Y = outputs["Y"]
+        Y = Y[np.lexsort(np.fliplr(Y).T)]
+
+        if np.allclose(Y, Y_ref):
+            print("Inference result correct!")
+        else:
+            print("Inference result incorrect!")

samples/python/non_zero_plugin/requirements.txt

@@ -0,0 +1,13 @@
+cuda-python==12.9.0
+cupy-cuda12x
+torch
+--extra-index-url https://pypi.ngc.nvidia.com
+polygraphy>=0.50.1
+colored
+numpy==1.26.4
+--extra-index-url https://pypi.ngc.nvidia.com
+onnx-graphsurgeon
+pywin32; platform_system == "Windows"
+pyyaml==6.0.3
+requests==2.32.4
+tqdm==4.66.4