NVIDIA · jantonguirao · Mar 2, 2026 · Mar 2, 2026 · Mar 2, 2026 · Mar 3, 2026
diff --git a/.github/workflows/lint.yml b/.github/workflows/lint.yml
@@ -19,7 +19,7 @@ jobs:
       - uses: actions/setup-python@v6
         with:
           python-version: '3.10'
-      - run: pip install flake8 bandit "black[jupyter]"==25.12.0
+      - run: pip install flake8 bandit "black[jupyter]"==26.1.0
       # keep black invocations separated so each invocation will pick appropriate configuration file
       # from the top dir used for it
       - run: black --check --verbose ${{ env.PYTHON_SECURITY_LINT_PATHS }} ${{ env.PYTHON_LINT_PATHS }} ${{ env.AUTOGRAPH_LINT_PATHS }}

diff --git a/DALI_DEPS_VERSION b/DALI_DEPS_VERSION
@@ -1 +1 @@
-f689931ff3fe42aa4ae420d977a725ad5e344b5b
+67cef5f4d1017999949e37f96a7359d8d05e8bbf
diff --git a/conda/dali_native_libs/recipe/meta.yaml b/conda/dali_native_libs/recipe/meta.yaml
@@ -73,8 +73,8 @@ requirements:
     - python-clang
     - git-lfs
   host:
-    - protobuf =6.31.1
-    - libprotobuf-static =6.31.1
+    - protobuf =6.33.5
+    - libprotobuf-static =6.33.5
     - libjpeg-turbo
     - dali-opencv
     - dali-ffmpeg
@@ -107,7 +107,7 @@ requirements:
     # dependency to install the right version on the libabseil (as protobuf depends on
     # libprotobuf-static and a newer version of libprotobuf-static may be available than
     # the protobuf was build with)
-    - libprotobuf =6.31.1
+    - libprotobuf =6.33.5
     - cfitsio
     - nvidia-nvimagecodec-cuda{{ environ.get('CUDA_VERSION', '') | replace(".","") }}
     - libnvcomp

diff --git a/conda/dali_python_bindings/recipe/meta.yaml b/conda/dali_python_bindings/recipe/meta.yaml
@@ -1,5 +1,5 @@
 # (C) Copyright IBM Corp. 2019. All Rights Reserved.
-# Copyright (c) 2019-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Copyright (c) 2019-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -72,11 +72,11 @@ requirements:
     - patchelf
     - python-clang
     - git-lfs
-    - black =25.12.0
-    - black-jupyter =25.12.0
+    - black =26.1.0
+    - black-jupyter =26.1.0
     - bandit
   host:
-    - libprotobuf-static =6.31.1
+    - libprotobuf-static =6.33.5
     - dali-opencv
     - dali-ffmpeg
     - python
@@ -97,7 +97,7 @@ requirements:
     # dependency to install the right version on the libabseil (as protobuf depends on
     # libprotobuf-static and a newer version of libprotobuf-static may be available than
     # the protobuf was build with)
-    - libprotobuf =6.31.1
+    - libprotobuf =6.33.5
     - future
     - astunparse >=1.6.0
     - gast >=0.3.3

diff --git a/dali/benchmark/caffe2_alexnet_bench.cc b/dali/benchmark/caffe2_alexnet_bench.cc
@@ -130,7 +130,7 @@ BENCHMARK_DEFINE_F(C2Alexnet, Caffe2Pipe)(benchmark::State& st) { // NOLINT
       benchmark::Counter::kIsRate);
 }
 
-static void PipeArgs(benchmark::internal::Benchmark *b) {
+static void PipeArgs(benchmark::Benchmark *b) {
   for (int executor = 2; executor < 3; ++executor) {
     for (int fast_resize = 0; fast_resize < 2; ++fast_resize) {
       for (int batch_size = 128; batch_size <= 128; batch_size += 32) {
@@ -258,7 +258,7 @@ BENCHMARK_DEFINE_F(C2Alexnet, HybridPipe)(benchmark::State& st) { // NOLINT
       benchmark::Counter::kIsRate);
 }
 
-static void HybridPipeArgs(benchmark::internal::Benchmark *b) {
+static void HybridPipeArgs(benchmark::Benchmark *b) {
   for (int executor = 0; executor < 3; ++executor) {
     for (int batch_size = 128; batch_size <= 128; batch_size += 32) {
       for (int num_thread = 1; num_thread <= 4; ++num_thread) {

diff --git a/dali/benchmark/caffe_alexnet_bench.cc b/dali/benchmark/caffe_alexnet_bench.cc
@@ -132,7 +132,7 @@ BENCHMARK_DEFINE_F(Alexnet, CaffePipe)(benchmark::State& st) { // NOLINT
       benchmark::Counter::kIsRate);
 }
 
-static void PipeArgs(benchmark::internal::Benchmark *b) {
+static void PipeArgs(benchmark::Benchmark *b) {
   for (int executor = 2; executor < 3; ++executor) {
     for (int fast_resize = 0; fast_resize < 2; ++fast_resize) {
       for (int batch_size = 128; batch_size <= 128; batch_size += 32) {
@@ -260,7 +260,7 @@ BENCHMARK_DEFINE_F(Alexnet, HybridPipe)(benchmark::State& st) { // NOLINT
       benchmark::Counter::kIsRate);
 }
 
-static void HybridPipeArgs(benchmark::internal::Benchmark *b) {
+static void HybridPipeArgs(benchmark::Benchmark *b) {
   for (int executor = 0; executor < 3; ++executor) {
     for (int batch_size = 128; batch_size <= 128; batch_size += 32) {
       for (int num_thread = 1; num_thread <= 4; ++num_thread) {

diff --git a/dali/benchmark/checkpointing_bench.cc b/dali/benchmark/checkpointing_bench.cc
@@ -79,7 +79,7 @@ class CheckpointingOverhead : public DALIBenchmark {
   }
 };
 
-static void Args(benchmark::internal::Benchmark *b) {
+static void Args(benchmark::Benchmark *b) {
   const std::vector<CheckpointingPolicy> policies = {
     CheckpointingPolicy::Disabled,
     CheckpointingPolicy::Enabled,

diff --git a/dali/benchmark/copy_bench.cc b/dali/benchmark/copy_bench.cc
@@ -18,7 +18,7 @@
 
 namespace dali {
 
-static void CopyGPUArgs(benchmark::internal::Benchmark *b) {
+static void CopyGPUArgs(benchmark::Benchmark *b) {
   for (int batch_size = 256; batch_size >= 1; batch_size /= 2) {
     for (int H = 2000; H >= 500; H /= 2) {
       int W = H;

diff --git a/dali/benchmark/crop_bench.cc b/dali/benchmark/crop_bench.cc
@@ -18,7 +18,7 @@
 
 namespace dali {
 
-static void CropCPUArgs(benchmark::internal::Benchmark *b) {
+static void CropCPUArgs(benchmark::Benchmark *b) {
   int batch_size = 1;
   for (int H = 1000; H >= 500; H /= 2) {
     int W = H, C = 3;
@@ -54,7 +54,7 @@ BENCHMARK_REGISTER_F(OperatorBench, CropCPU)->Iterations(1000)
 ->UseRealTime()
 ->Apply(CropCPUArgs);
 
-static void CropGPUArgs(benchmark::internal::Benchmark *b) {
+static void CropGPUArgs(benchmark::Benchmark *b) {
   for (int batch_size = 128; batch_size >= 1; batch_size /= 4) {
     for (int H = 2000; H >= 500; H /= 2) {
       int W = H, C = 3;

diff --git a/dali/benchmark/crop_mirror_normalize_bench.cc b/dali/benchmark/crop_mirror_normalize_bench.cc
@@ -18,7 +18,7 @@
 
 namespace dali {
 
-static void CropMirrorNormalizeCPUArgs(benchmark::internal::Benchmark *b) {
+static void CropMirrorNormalizeCPUArgs(benchmark::Benchmark *b) {
   int batch_size = 8;
   int mean = 128, std = 1;
   for (auto &dtype : {DALI_FLOAT}) {
@@ -76,7 +76,7 @@ BENCHMARK_REGISTER_F(OperatorBench, CropMirrorNormalizeCPU)->Iterations(500)
 ->UseRealTime()
 ->Apply(CropMirrorNormalizeCPUArgs);
 
-static void CropMirrorNormalizeGPUArgs(benchmark::internal::Benchmark *b) {
+static void CropMirrorNormalizeGPUArgs(benchmark::Benchmark *b) {
   int mean = 128, std = 1;
   for (int batch_size : {1, 8, 128, 256}) {
     for (int H = 1000; H >= 250; H /= 2) {

diff --git a/dali/benchmark/decoder_bench.cc b/dali/benchmark/decoder_bench.cc
@@ -85,7 +85,7 @@ class DecoderBench : public DALIBenchmark {
   }
 };
 
-static void PipeArgs(benchmark::internal::Benchmark *b) {
+static void PipeArgs(benchmark::Benchmark *b) {
   for (int batch_size = 1; batch_size <= 128; batch_size <<= 1) {
     for (int num_thread = 1; num_thread <= 4; ++num_thread) {
       b->Args({batch_size, num_thread});

diff --git a/dali/benchmark/file_reader_alexnet_bench.cc b/dali/benchmark/file_reader_alexnet_bench.cc
@@ -131,7 +131,7 @@ BENCHMARK_DEFINE_F(FileReaderAlexnet, CaffePipe)(benchmark::State& st) { // NOLI
       benchmark::Counter::kIsRate);
 }
 
-static void PipeArgs(benchmark::internal::Benchmark *b) {
+static void PipeArgs(benchmark::Benchmark *b) {
   for (int executor = 2; executor < 3; ++executor) {
     for (int fast_resize = 0; fast_resize < 2; ++fast_resize) {
       for (int batch_size = 128; batch_size <= 128; batch_size += 32) {

diff --git a/dali/benchmark/file_reader_fast_forward_bench.cc b/dali/benchmark/file_reader_fast_forward_bench.cc
@@ -90,7 +90,7 @@ BENCHMARK_DEFINE_F(FileReaderFastForward, FastForward)(benchmark::State& st) { /
   }
 }
 
-static void Args(benchmark::internal::Benchmark *b) {
+static void Args(benchmark::Benchmark *b) {
   for (int dataset_size = 100; dataset_size <= 1000000; dataset_size *= 100) {
     for (int frac = 0; frac <= 10; frac++) {
       // Without random shuffle

diff --git a/dali/benchmark/flip_bench.cc b/dali/benchmark/flip_bench.cc
@@ -18,7 +18,7 @@
 
 namespace dali {
 
-static void FlipGPUArgs(benchmark::internal::Benchmark *b) {
+static void FlipGPUArgs(benchmark::Benchmark *b) {
   for (int batch_size = 256; batch_size >= 1; batch_size /= 2) {
     for (int H = 2000; H >= 500; H /= 2) {
       int W = H, C = 3;

diff --git a/dali/benchmark/gaussian_blur_bench.cc b/dali/benchmark/gaussian_blur_bench.cc
@@ -18,7 +18,7 @@
 
 namespace dali {
 
-static void GaussianBlurGPUArgs(benchmark::internal::Benchmark *b) {
+static void GaussianBlurGPUArgs(benchmark::Benchmark *b) {
   for (int batch_size = 256; batch_size >= 1; batch_size /= 2) {
     for (int H = 1000; H >= 500; H /= 2) {
       int W = H, C = 3;

diff --git a/dali/benchmark/one_hot_bench.cc b/dali/benchmark/one_hot_bench.cc
@@ -18,15 +18,15 @@
 
 namespace dali {
 
-static void OneHotGPUAlternateAxis(benchmark::internal::Benchmark *b, int batch_size, int h, int w,
+static void OneHotGPUAlternateAxis(benchmark::Benchmark *b, int batch_size, int h, int w,
                                    int num_classes) {
   int axis = -1;
   b->Args({batch_size, h, w, axis, num_classes});
   axis = 1;
   b->Args({batch_size, h, w, axis, num_classes});
 }
 
-static void OneHotGPUArgs(benchmark::internal::Benchmark *b) {
+static void OneHotGPUArgs(benchmark::Benchmark *b) {
   for (int batch_size = 256; batch_size >= 1; batch_size /= 2) {
     int h = 64, w = h;
     int num_classes = 500;

diff --git a/dali/benchmark/resnet50_bench.cc b/dali/benchmark/resnet50_bench.cc
@@ -134,7 +134,7 @@ BENCHMARK_DEFINE_F(RN50, C2Pipe)(benchmark::State& st) { // NOLINT
       benchmark::Counter::kIsRate);
 }
 
-static void PipeArgs(benchmark::internal::Benchmark *b) {
+static void PipeArgs(benchmark::Benchmark *b) {
   for (int executor = 2; executor < 3; ++executor) {
     for (int fast_resize = 0; fast_resize < 2; ++fast_resize) {
       for (int batch_size = 128; batch_size <= 128; batch_size += 32) {
@@ -266,7 +266,7 @@ BENCHMARK_DEFINE_F(RN50, HybridPipe)(benchmark::State& st) { // NOLINT
       benchmark::Counter::kIsRate);
 }
 
-static void HybridPipeArgs(benchmark::internal::Benchmark *b) {
+static void HybridPipeArgs(benchmark::Benchmark *b) {
   for (int executor = 0; executor < 3; ++executor) {
     for (int batch_size = 128; batch_size <= 128; batch_size += 32) {
       for (int num_thread = 1; num_thread <= 4; ++num_thread) {
@@ -375,7 +375,7 @@ BENCHMARK_DEFINE_F(RN50, nvJPEGPipe)(benchmark::State& st) { // NOLINT
       benchmark::Counter::kIsRate);
 }
 
-static void nvjpegPipeArgs(benchmark::internal::Benchmark *b) {
+static void nvjpegPipeArgs(benchmark::Benchmark *b) {
   for (int executor = 0; executor < 3; ++executor) {
     for (int batch_size = 128; batch_size <= 128; batch_size += 32) {
       for (int num_thread = 1; num_thread <= 4; ++num_thread) {

diff --git a/dali/benchmark/resnet50_nvjpeg_bench.cc b/dali/benchmark/resnet50_nvjpeg_bench.cc
@@ -115,7 +115,7 @@ BENCHMARK_DEFINE_F(RealRN50, nvjpegPipe)(benchmark::State& st) { // NOLINT
       benchmark::Counter::kIsRate);
 }
 
-static void PipeArgs(benchmark::internal::Benchmark *b) {
+static void PipeArgs(benchmark::Benchmark *b) {
   for (int executor = 2; executor < 3; ++executor) {
     for (int batch_size = 128; batch_size <= 128; batch_size += 32) {
       for (int num_thread = 1; num_thread <= 4; ++num_thread) {

diff --git a/dali/benchmark/slice_kernel_bench.cc b/dali/benchmark/slice_kernel_bench.cc
@@ -76,7 +76,7 @@ class SliceBenchCPU : public DALIBenchmark {
   }
 };
 
-static void SliceKernelArgs_OnlySlice_CPU(benchmark::internal::Benchmark *b) {
+static void SliceKernelArgs_OnlySlice_CPU(benchmark::Benchmark *b) {
   for (int H = 1000; H >= 500; H /= 2) {
     int W = H, C = 3;
     int crop_h = 9 * H / 10;
@@ -94,7 +94,7 @@ BENCHMARK_REGISTER_F(SliceBenchCPU, Slice_CPU_OnlySlice)->Iterations(1000)
 ->UseRealTime()
 ->Apply(SliceKernelArgs_OnlySlice_CPU);
 
-static void SliceKernelArgs_OnlyPad_CPU(benchmark::internal::Benchmark *b) {
+static void SliceKernelArgs_OnlyPad_CPU(benchmark::Benchmark *b) {
   for (int H = 1000; H >= 500; H /= 2) {
     int W = H, C = 3;
     int crop_h = 9 * H / 10;
@@ -114,7 +114,7 @@ BENCHMARK_REGISTER_F(SliceBenchCPU, Slice_OnlyPad_CPU)->Iterations(1000)
 ->UseRealTime()
 ->Apply(SliceKernelArgs_OnlyPad_CPU);
 
-static void SliceKernelArgs_SliceAndPad_CPU(benchmark::internal::Benchmark *b) {
+static void SliceKernelArgs_SliceAndPad_CPU(benchmark::Benchmark *b) {
   for (int H = 1000; H >= 500; H /= 2) {
     int W = H, C = 3;
     int crop_h = 9 * H / 10;

diff --git a/dali/benchmark/slice_kernel_bench.cu b/dali/benchmark/slice_kernel_bench.cu
@@ -87,7 +87,7 @@ class SliceBenchGPU : public DALIBenchmark {
   }
 };
 
-static void SliceKernelArgs_GPU_OnlySlice(benchmark::internal::Benchmark *b) {
+static void SliceKernelArgs_GPU_OnlySlice(benchmark::Benchmark *b) {
   for (int H = 1000; H >= 500; H /= 2) {
     int W = H, C = 3;
     int crop_h = 9 * H / 10;
@@ -106,7 +106,7 @@ BENCHMARK_REGISTER_F(SliceBenchGPU, Slice_GPU_OnlySlice)->Iterations(1000)
 ->UseRealTime()
 ->Apply(SliceKernelArgs_GPU_OnlySlice);
 
-static void SliceKernelArgs_GPU_OnlyPad(benchmark::internal::Benchmark *b) {
+static void SliceKernelArgs_GPU_OnlyPad(benchmark::Benchmark *b) {
   for (int H = 1000; H >= 500; H /= 2) {
     int W = H, C = 3;
     int crop_h = 9 * H / 10;
@@ -127,7 +127,7 @@ BENCHMARK_REGISTER_F(SliceBenchGPU, Slice_GPU_OnlyPad)->Iterations(1000)
 ->UseRealTime()
 ->Apply(SliceKernelArgs_GPU_OnlyPad);
 
-static void SliceKernelArgs_GPU_SliceAndPad(benchmark::internal::Benchmark *b) {
+static void SliceKernelArgs_GPU_SliceAndPad(benchmark::Benchmark *b) {
   for (int H = 1000; H >= 500; H /= 2) {
     int W = H, C = 3;
     int crop_h = 9 * H / 10;

diff --git a/dali/benchmark/thread_pool_bench.cc b/dali/benchmark/thread_pool_bench.cc
@@ -19,7 +19,7 @@ namespace dali {
 
 class ThreadPoolBench : public DALIBenchmark {};
 
-static void ThreadPoolArgs(benchmark::internal::Benchmark *b) {
+static void ThreadPoolArgs(benchmark::Benchmark *b) {
   int batch_size = 64;
   int work_size_min = 400;
   int work_size_max = 10000;

diff --git a/dali/benchmark/transpose_bench.cc b/dali/benchmark/transpose_bench.cc
@@ -26,7 +26,7 @@ static std::vector<int> permutations3[] = {
   {2, 1, 0},
 };
 
-static void TransposeGPUArgs(benchmark::internal::Benchmark *b) {
+static void TransposeGPUArgs(benchmark::Benchmark *b) {
   for (int batch_size = 256; batch_size >= 1; batch_size /= 2) {
     for (int H : dimensions) {
       for (int W : dimensions) {

diff --git a/dali/benchmark/transpose_cpu_bench.cc b/dali/benchmark/transpose_cpu_bench.cc
@@ -52,7 +52,7 @@ std::tuple<TensorShape<>, std::vector<int>> GetCase(int id) {
   return cases[id];
 }
 
-static void CustomArguments(benchmark::internal::Benchmark* b) {
+static void CustomArguments(benchmark::Benchmark* b) {
   for (unsigned int i = 0; i < sizeof(cases) / sizeof(*cases); i++) {
     for (int scale = 1; scale <= 8; scale *= 2) {
       b->Args({i, scale});

diff --git a/dali/benchmark/warp_affine_bench.cc b/dali/benchmark/warp_affine_bench.cc
@@ -18,7 +18,7 @@
 
 namespace dali {
 
-static void WarpAffineCPUArgs(benchmark::internal::Benchmark *b) {
+static void WarpAffineCPUArgs(benchmark::Benchmark *b) {
   for (int batch_size = 16; batch_size >= 1; batch_size /= 2) {
     for (int H = 2048; H >= 256; H /= 2) {
       int W = H, C = 3;
@@ -55,7 +55,7 @@ BENCHMARK_REGISTER_F(OperatorBench, WarpAffineCPU)->Iterations(50)
 ->Apply(WarpAffineCPUArgs);
 
 
-static void WarpAffineGPUArgs(benchmark::internal::Benchmark *b) {
+static void WarpAffineGPUArgs(benchmark::Benchmark *b) {
   for (int batch_size = 256; batch_size >= 1; batch_size /= 2) {
     for (int H = 2048; H >= 256; H /= 2) {
       int W = H, C = 3;

diff --git a/dali/pipeline/operator/checkpointing/checkpoint.cc b/dali/pipeline/operator/checkpointing/checkpoint.cc
@@ -93,7 +93,8 @@ std::string Checkpoint::SerializeToProtobuf(ExecutorBase &exec) const {
 void Checkpoint::DeserializeFromProtobuf(ExecutorBase &exec, std::string_view serialized_data) {
   Clear();
   dali_proto::Checkpoint checkpoint;
-  checkpoint.ParseFromArray(serialized_data.data(), serialized_data.size());
+  DALI_ENFORCE(checkpoint.ParseFromArray(serialized_data.data(), serialized_data.size()),
+               "Failed to deserialize checkpoint from protobuf.");
 
   for (int i = 0; i < checkpoint.cpts_size(); i++) {
     const auto &name = checkpoint.cpts(i).operator_name();

diff --git a/dali/pipeline/operator/checkpointing/snapshot_serializer.cc b/dali/pipeline/operator/checkpointing/snapshot_serializer.cc
@@ -59,7 +59,8 @@ std::string SerializeRNG(const std::vector<T> &snapshot) {
 template<class T>
 std::vector<T> DeserializeRNG(const std::string &data) {
   dali_proto::RNGSnapshotCPU proto_snapshot;
-  proto_snapshot.ParseFromString(data);
+  DALI_ENFORCE(proto_snapshot.ParseFromString(data),
+               "Failed to deserialize RNG snapshot from protobuf.");
   std::vector<T> snapshot;
   for (int i = 0; i < proto_snapshot.rng_size(); i++)
     snapshot.push_back(DeserializeFromString<T>(proto_snapshot.rng(i)));
@@ -96,7 +97,8 @@ std::string SnapshotSerializer::Serialize(const std::vector<curandState> &snapsh
 template<> DLL_PUBLIC
 std::vector<curandState> SnapshotSerializer::Deserialize(const std::string &data) {
   dali_proto::RNGSnapshotGPU proto_snapshot;
-  proto_snapshot.ParseFromString(data);
+  DALI_ENFORCE(proto_snapshot.ParseFromString(data),
+               "Failed to deserialize GPU RNG snapshot from protobuf.");
   auto str = proto_snapshot.rng();
   std::vector<curandState> snapshot(str.size() / sizeof(curandState));
   memcpy(snapshot.data(), str.data(), str.size());
@@ -115,7 +117,8 @@ std::string SnapshotSerializer::Serialize(const LoaderStateSnapshot &snapshot) {
 template<> DLL_PUBLIC
 LoaderStateSnapshot SnapshotSerializer::Deserialize(const std::string &data) {
   dali_proto::ReaderStateSnapshot proto_snapshot;
-  proto_snapshot.ParseFromString(data);
+  DALI_ENFORCE(proto_snapshot.ParseFromString(data),
+               "Failed to deserialize loader state snapshot from protobuf.");
   return LoaderStateSnapshot {
     DeserializeFromString<std::default_random_engine>(proto_snapshot.loader_state().rng()),
     proto_snapshot.loader_state().current_epoch(),

diff --git a/dali/python/nvidia/dali/_autograph/converters/call_trees.py b/dali/python/nvidia/dali/_autograph/converters/call_trees.py
@@ -29,7 +29,6 @@
 from nvidia.dali._autograph.pyct import templates
 from nvidia.dali._autograph.utils import ag_logging
 
-
 # TODO(mdan): Rename to FunctionCallsTransformer.
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		f689931ff3fe42aa4ae420d977a725ad5e344b5b
		67cef5f4d1017999949e37f96a7359d8d05e8bbf
Original file line number	Diff line number	Diff line change
Expand Up		@@ -29,7 +29,6 @@
		from nvidia.dali._autograph.pyct import templates
		from nvidia.dali._autograph.utils import ag_logging


		# TODO(mdan): Rename to FunctionCallsTransformer.


Expand Down