Bump up nvComp version to 5.3.0.16

cmake/Dependencies.common.cmake

@@ -167,7 +167,7 @@ if(BUILD_NVCOMP)
     list(APPEND DALI_LIBS ${nvcomp_LIBS})
   else()
     message(STATUS "Found nvCOMP: ${nvcomp_INCLUDE_DIR}.")
-    set(DALI_INSTALL_REQUIRES_NVCOMP "\'nvidia-libnvcomp-cu${CUDA_VERSION_MAJOR} == 5.2.0.13\',")
+    set(DALI_INSTALL_REQUIRES_NVCOMP "\'nvidia-libnvcomp-cu${CUDA_VERSION_MAJOR} == 5.3.0.16\',")
     message(STATUS "Adding nvComp requirement as: ${DALI_INSTALL_REQUIRES_NVCOMP}")
   endif()
 endif()

dali/operators/decoder/inflate/inflate.cc

@@ -54,6 +54,12 @@ concatenating compressed frames from the corresponding sequences.::
     .NumOutput(1)
     .AddArg(inflate::shapeArgName, "The shape of the output (inflated) chunk.", DALI_INT_VEC, true)
     .AddOptionalTypeArg(inflate::dTypeArgName, "The output (inflated) data type.", DALI_UINT8)
+    .AddOptionalArg<bool>(inflate::checkOutputSizeArgName,
+                          R"code(If True, validates before decompression that the requested output
+buffers are large enough for the compressed data.
+
+This validation synchronizes the GPU stream and is disabled by default.)code",
+                          false)
     .AddOptionalArg<std::vector<int>>(inflate::offsetArgName,
                                       R"code(A list of offsets within the input sample
 describing where the consecutive chunks begin.

dali/operators/decoder/inflate/inflate_gpu.cc

@@ -35,7 +35,9 @@ namespace inflate {
 
 class InflateOpGpuLZ4Impl : public InflateOpImplBase<GPUBackend> {
  public:
-  explicit InflateOpGpuLZ4Impl(const OpSpec &spec) : InflateOpImplBase<GPUBackend>{spec} {}
+  explicit InflateOpGpuLZ4Impl(const OpSpec &spec)
+      : InflateOpImplBase<GPUBackend>{spec},
+        check_output_size_{spec.GetArgument<bool>(inflate::checkOutputSizeArgName)} {}
 
   void RunImpl(Workspace &ws) override {
     const auto &input = ws.template Input<GPUBackend>(0);
@@ -57,6 +59,32 @@ class InflateOpGpuLZ4Impl : public InflateOpImplBase<GPUBackend> {
     auto [in_sizes, in, out_sizes, out] = scratchpad.ToContiguousGPU(
         stream, params_.GetInChunkSizes(), input_ptrs_, inflated_sizes_, inflated_ptrs_);
 
+    if (check_output_size_) {
+      // Query the decoded sizes before decompression. nvCOMP's decompression API documents an
+      // insufficient output buffer as undefined behaviour for some backends, so relying on its
+      // per-chunk status is not safe here.
+      CUDA_CALL(nvcompBatchedLZ4GetDecompressSizeAsync(in, in_sizes, actual_out_sizes,
+                                                        total_chunks_num, stream));
+      std::vector<size_t> decoded_sizes(total_chunks_num);
+      CUDA_CALL(cudaMemcpyAsync(decoded_sizes.data(), actual_out_sizes,
+                                total_chunks_num * sizeof(size_t), cudaMemcpyDeviceToHost, stream));
+      CUDA_CALL(cudaStreamSynchronize(stream));
+      size_t flat_chunk_idx = 0;
+      const auto &chunks_per_sample = params_.GetChunksNumPerSample();
+      for (int sample_idx = 0; sample_idx < chunks_per_sample.num_samples(); sample_idx++) {
+        auto num_chunks = chunks_per_sample[sample_idx].num_elements();
+        for (int chunk_idx = 0; chunk_idx < num_chunks; chunk_idx++, flat_chunk_idx++) {
+          DALI_ENFORCE(
+              decoded_sizes[flat_chunk_idx] <= inflated_sizes_[flat_chunk_idx],
+              make_string("Output buffer for inflated chunk ", chunk_idx, " in sample ", sample_idx,
+                          " is too small: it has ", inflated_sizes_[flat_chunk_idx],
+                          " bytes, but the compressed input expands to ",
+                          decoded_sizes[flat_chunk_idx],
+                          " bytes. Check the `shape` and `dtype` arguments."));
+        }
+      }
+    }
+
     size_t tempSize;
     CUDA_CALL(nvcompBatchedLZ4DecompressGetTempSizeAsync(
         total_chunks_num,
@@ -131,6 +159,7 @@ class InflateOpGpuLZ4Impl : public InflateOpImplBase<GPUBackend> {
   std::vector<const void *> input_ptrs_;
   std::vector<void *> inflated_ptrs_;
   std::vector<size_t> inflated_sizes_;
+  bool check_output_size_;
 };
 
 }  // namespace inflate

dali/operators/decoder/inflate/inflate_params.h

@@ -36,6 +36,7 @@ constexpr static const char *offsetArgName = "chunk_offsets";
 constexpr static const char *sizeArgName = "chunk_sizes";
 constexpr static const char *layoutArgName = "layout";
 constexpr static const char *sequenceLayoutArgName = "sequence_axis_name";
+constexpr static const char *checkOutputSizeArgName = "check_output_size";
 
 enum class InflateAlg {
   LZ4

dali/test/python/operator_1/test_inflate.py

@@ -1,4 +1,4 @@
-# Copyright (c) 2022-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Copyright (c) 2022-2026, 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.
@@ -129,6 +129,44 @@ def test_sample_inflate():
             seed += 1
 
 
+@has_operator("decoders.inflate")
+@restrict_platform(min_compute_cap=6.0)
+def test_rejects_insufficient_output_buffer():
+    sample = np.full((13, 7), 42, dtype=np.int64)
+    deflated = sample_to_lz4(sample)
+
+    @pipeline_def(batch_size=1, num_threads=4, device_id=0)
+    def pipeline():
+        compressed = fn.external_source(source=lambda: deflated, batch=False)
+        # dtype defaults to uint8, which is too small for the int64 LZ4 payload.
+        return fn.decoders.inflate(compressed.gpu(), shape=sample.shape, check_output_size=True)
+
+    with assert_raises(
+        RuntimeError, glob="Output buffer for inflated chunk 0 in sample 0 is too small"
+    ):
+        pipeline().run()
+
+
+@has_operator("decoders.inflate")
+@restrict_platform(min_compute_cap=6.0)
+def test_checks_sufficient_output_buffer():
+    sample = np.full((13, 7), 42, dtype=np.int64)
+    deflated = sample_to_lz4(sample)
+
+    @pipeline_def(batch_size=1, num_threads=4, device_id=0)
+    def pipeline():
+        compressed = fn.external_source(source=lambda: deflated, batch=False)
+        return fn.decoders.inflate(
+            compressed.gpu(),
+            shape=sample.shape,
+            dtype=types.INT64,
+            check_output_size=True,
+        )
+
+    (inflated,) = pipeline().run()
+    check_batch(inflated, [sample], batch_size=1)
+
+
 def _test_scalar_shape(dtype, shape, layout):
     def sample_source(sample_info):
         sample_size = np.prod(shape)

dali/test/python/test_dali_variable_batch_size.py

@@ -1460,7 +1460,7 @@ def inflate_pipline(max_batch_size, inputs, device):
         def piepline():
             defalted = fn.external_source(source=input_data)
             shape = fn.external_source(source=input_shape)
-            return fn.decoders.inflate(defalted.gpu(), shape=shape)
+            return fn.decoders.inflate(defalted.gpu(), shape=shape, dtype=types.INT64)
 
         return piepline(batch_size=max_batch_size, num_threads=4, device_id=0)
 

docker/Dockerfile.cuda129.aarch64.deps

@@ -12,7 +12,7 @@ RUN curl -LO https://developer.download.nvidia.com/compute/cuda/12.9.1/local_ins
     rm -f cuda_*.run;
 
 RUN CUFILE_VERSION=1.14.1.1-1 && \
-    NVCOMP_VERSION=4.2.0.11-1 && \
+    NVCOMP_VERSION=5.3.0.16 && \
     CUDA_VERSION_MAJOR=12 && \
     CUDA_VERSION_MINOR=9 && \
     apt-get update && \
@@ -22,9 +22,9 @@ RUN CUFILE_VERSION=1.14.1.1-1 && \
     apt-get update && \
     apt-get install libcufile-dev-${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR}=${CUFILE_VERSION} -y  && \
     mkdir /tmp/nvcomp && cd /tmp/nvcomp && \
-    wget https://developer.download.nvidia.com/compute/nvcomp/redist/nvcomp/linux-sbsa/nvcomp-linux-sbsa-5.2.0.10_cuda12-archive.tar.xz && \
+    wget https://developer.download.nvidia.com/compute/nvcomp/redist/nvcomp/linux-sbsa/nvcomp-linux-sbsa-${NVCOMP_VERSION}_cuda12-archive.tar.xz && \
     tar -xf * && \
-    cp -r nvcomp-linux-sbsa-5.2.0.10_cuda12-archive/include/* /usr/local/cuda/include/ && \
-    cp -r nvcomp-linux-sbsa-5.2.0.10_cuda12-archive/lib/* /usr/local/cuda/lib64/ && \
+    cp -r nvcomp-linux-sbsa-${NVCOMP_VERSION}_cuda12-archive/include/* /usr/local/cuda/include/ && \
+    cp -r nvcomp-linux-sbsa-${NVCOMP_VERSION}_cuda12-archive/lib/* /usr/local/cuda/lib64/ && \
     cd / && rm -rf /tmp/nvcomp && \
     rm -rf /var/lib/apt/lists/*

docker/Dockerfile.cuda129.x86_64.deps

@@ -12,7 +12,7 @@ RUN curl -LO https://developer.download.nvidia.com/compute/cuda/12.9.1/local_ins
     rm -f cuda_*.run;
 
 RUN CUFILE_VERSION=1.14.1.1-1 && \
-    NVCOMP_VERSION=4.2.0.11-1 && \
+    NVCOMP_VERSION=5.3.0.16 && \
     CUDA_VERSION_MAJOR=12 && \
     CUDA_VERSION_MINOR=9 && \
     apt-get update && \
@@ -22,9 +22,9 @@ RUN CUFILE_VERSION=1.14.1.1-1 && \
     apt-get update && \
     apt-get install libcufile-dev-${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR}=${CUFILE_VERSION} -y  && \
     mkdir /tmp/nvcomp && cd /tmp/nvcomp && \
-    wget https://developer.download.nvidia.com/compute/nvcomp/redist/nvcomp/linux-x86_64/nvcomp-linux-x86_64-5.2.0.10_cuda12-archive.tar.xz && \
+    wget https://developer.download.nvidia.com/compute/nvcomp/redist/nvcomp/linux-x86_64/nvcomp-linux-x86_64-${NVCOMP_VERSION}_cuda12-archive.tar.xz && \
     tar -xf * && \
-    cp -r nvcomp-linux-x86_64-5.2.0.10_cuda12-archive/include/* /usr/local/cuda/include/ && \
-    cp -r nvcomp-linux-x86_64-5.2.0.10_cuda12-archive/lib/* /usr/local/cuda/lib64/ && \
+    cp -r nvcomp-linux-x86_64-${NVCOMP_VERSION}_cuda12-archive/include/* /usr/local/cuda/include/ && \
+    cp -r nvcomp-linux-x86_64-${NVCOMP_VERSION}_cuda12-archive/lib/* /usr/local/cuda/lib64/ && \
     cd / && rm -rf /tmp/nvcomp && \
     rm -rf /var/lib/apt/lists/*

docker/Dockerfile.cuda133.aarch64.deps

@@ -8,6 +8,7 @@ RUN CUDA_SUBVERSION=13.3.1-1 && \
     CUFILE_VERSION=1.18.1.6-1 && \
     CUDA_VERSION_MAJOR=13 && \
     CUDA_VERSION_MINOR=3 && \
+    NVCOMP_VERSION=5.3.0.16 && \
     apt-get update && \
     apt-get install wget software-properties-common -y && \
     apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/sbsa/3bf863cc.pub && \
@@ -29,10 +30,10 @@ RUN CUDA_SUBVERSION=13.3.1-1 && \
         cuda-compat-${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR} \
         libcufile-dev-${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR}=${CUFILE_VERSION} && \
     mkdir /tmp/nvcomp && cd /tmp/nvcomp && \
-    wget https://developer.download.nvidia.com/compute/nvcomp/redist/nvcomp/linux-sbsa/nvcomp-linux-sbsa-5.2.0.10_cuda13-archive.tar.xz && \
+    wget https://developer.download.nvidia.com/compute/nvcomp/redist/nvcomp/linux-sbsa/nvcomp-linux-sbsa-${NVCOMP_VERSION}_cuda13-archive.tar.xz && \
     tar -xf * && \
-    cp -r nvcomp-linux-sbsa-5.2.0.10_cuda13-archive/include/* /usr/local/cuda/include/ && \
-    cp -r nvcomp-linux-sbsa-5.2.0.10_cuda13-archive/lib/* /usr/local/cuda/lib64/ && \
+    cp -r nvcomp-linux-sbsa-${NVCOMP_VERSION}_cuda13-archive/include/* /usr/local/cuda/include/ && \
+    cp -r nvcomp-linux-sbsa-${NVCOMP_VERSION}_cuda13-archive/lib/* /usr/local/cuda/lib64/ && \
     cd / && rm -rf /tmp/nvcomp && \
     mv /usr/local/cuda/bin/fatbinary /usr/local/cuda/bin/fatbinary_org && \
     # fatbinary removed one of the options while clang still uses it

docker/Dockerfile.cuda133.x86_64.deps

@@ -8,6 +8,7 @@ RUN CUDA_SUBVERSION=13.3.1-1 && \
     CUFILE_VERSION=1.18.1.6-1 && \
     CUDA_VERSION_MAJOR=13 && \
     CUDA_VERSION_MINOR=3 && \
+    NVCOMP_VERSION=5.3.0.16 && \
     apt-get update && \
     apt-get install wget software-properties-common -y && \
     apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/3bf863cc.pub && \
@@ -29,10 +30,10 @@ RUN CUDA_SUBVERSION=13.3.1-1 && \
         cuda-compat-${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR} \
         libcufile-dev-${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR}=${CUFILE_VERSION} && \
     mkdir /tmp/nvcomp && cd /tmp/nvcomp && \
-    wget https://developer.download.nvidia.com/compute/nvcomp/redist/nvcomp/linux-x86_64/nvcomp-linux-x86_64-5.2.0.10_cuda13-archive.tar.xz && \
+    wget https://developer.download.nvidia.com/compute/nvcomp/redist/nvcomp/linux-x86_64/nvcomp-linux-x86_64-${NVCOMP_VERSION}_cuda13-archive.tar.xz && \
     tar -xf * && \
-    cp -r nvcomp-linux-x86_64-5.2.0.10_cuda13-archive/include/* /usr/local/cuda/include/ && \
-    cp -r nvcomp-linux-x86_64-5.2.0.10_cuda13-archive/lib/* /usr/local/cuda/lib64/ && \
+    cp -r nvcomp-linux-x86_64-${NVCOMP_VERSION}_cuda13-archive/include/* /usr/local/cuda/include/ && \
+    cp -r nvcomp-linux-x86_64-${NVCOMP_VERSION}_cuda13-archive/lib/* /usr/local/cuda/lib64/ && \
     cd / && rm -rf /tmp/nvcomp && \
     mv /usr/local/cuda/bin/fatbinary /usr/local/cuda/bin/fatbinary_org && \
     # fatbinary removed one of the options while clang still uses it

internal_tools/stub_generator/nvcomp.json

@@ -8,6 +8,7 @@
    "functions": {
       "nvcompBatchedLZ4DecompressGetTempSizeAsync": {},
       "nvcompBatchedLZ4DecompressAsync": {},
+      "nvcompBatchedLZ4GetDecompressSizeAsync": {},
       "nvcompGetStatusString": {
          "return_type":"const char*",
          "not_found_error":"\"(nvcompGetStatusString not available)\""