Bump up nvComp version to 5.3.0.16

[JanuszL]

Category:

Bug fix

Description:

• Bumps nvCOMP to 5.3.0.16.
• Fixes test_dali_variable_batch_size.test_inflate, which compressed int64
  samples while relying on the Inflate default output type of uint8.
• Adds the optional check_output_size argument to decoders.inflate.
  When enabled, it validates the decoded LZ4 size before decompression and
  raises an error if the requested output buffer is too small.

Additional information:

Affected modules and functionalities:

• nvCOMP dependency definitions and CUDA Dockerfiles
• GPU LZ4 Inflate operator
• Inflate operator tests

Key points relevant for the review:

• check_output_size is disabled by default to preserve existing behavior and
  avoid an additional GPU stream synchronization.
• When enabled, it prevents calling nvCOMP with an output buffer that is
  smaller than the compressed payload expands to.

Tests:

• Existing tests apply
  • operator_1.test_inflate (141 tests)
  • test_dali_variable_batch_size.test_inflate
• New tests added
  • Python tests
    • test_rejects_insufficient_output_buffer
• GTests
• Benchmark
• Other

Checklist

Documentation

• Existing documentation applies
• Documentation updated
  • Docstring
• RST
• Jupyter
• Other

DALI team only

Requirements

• Implements new requirements
• Affects existing requirements
• N/A

REQ IDs: N/A

JIRA TASK: N/A

[JanuszL]

@stiepan did we have a discussion about this unsafe behavior and potential switch for that or I mixing things?

[greptile-apps]

Greptile Summary

This PR bumps nvCOMP from 5.2.0.x to 5.3.0.16 across CMake, Dockerfiles, and the stub generator, and adds an optional check_output_size argument to decoders.inflate that pre-validates output buffer sizes before GPU LZ4 decompression. It also fixes a latent bug in the variable-batch-size test where int64 samples were compressed but inflated with the default uint8 output type.

• check_output_size flag calls nvcompBatchedLZ4GetDecompressSizeAsync, copies results to host, synchronises the stream, and raises a descriptive error if any chunk's output buffer is too small; the feature is opt-in to avoid the extra synchronisation in normal paths.
• Variable-batch-size test fix adds dtype=types.INT64 to match the int64 test samples, resolving the dtype mismatch that caused test_inflate to fail.
• Two new Python tests cover both the rejection path and the positive path for check_output_size=True.

Confidence Score: 4/5

Safe to merge after the docstring is corrected; the GPU decompression logic and stream ordering are sound.

The new check_output_size path has correct stream ordering and actionable error messages. One user-facing docstring says compressed data where it should say decompressed data, which would mislead users about what the flag actually checks. All other changes are straightforward.

dali/operators/decoder/inflate/inflate.cc - the check_output_size docstring needs compressed corrected to decompressed.

Important Files Changed

Filename	Overview
dali/operators/decoder/inflate/inflate.cc	Adds check_output_size optional bool arg to operator spec; docstring incorrectly says compressed data instead of decompressed data.
dali/operators/decoder/inflate/inflate_gpu.cc	Implements check_output_size_ path: queries decoded sizes via nvcompBatchedLZ4GetDecompressSizeAsync, D2H copies, stream-syncs, then validates each chunk; logic and stream ordering are correct.
dali/operators/decoder/inflate/inflate_params.h	Adds checkOutputSizeArgName constant; straightforward and consistent with existing arg-name declarations.
dali/test/python/operator_1/test_inflate.py	Adds two new tests covering rejection and positive paths for check_output_size=True; both use real value comparisons via check_batch. Copyright year updated to 2026.
dali/test/python/test_dali_variable_batch_size.py	Fixes test_inflate by adding dtype=types.INT64 to match the int64 test samples, resolving the silent type mismatch that caused the test to fail.
cmake/Dependencies.common.cmake	Version pin updated from 5.2.0.13 to 5.3.0.16; straightforward one-line change.
internal_tools/stub_generator/nvcomp.json	Adds nvcompBatchedLZ4GetDecompressSizeAsync stub entry; matches the new call site in inflate_gpu.cc.

Sequence Diagram

sequenceDiagram
    participant U as User Pipeline
    participant I as InflateOpGpuLZ4Impl
    participant nv as nvCOMP (GPU)
    participant H as Host (CPU)

    U->>I: RunImpl(ws)
    I->>I: SetupInChunks / SetupOutChunks
    I->>nv: ToContiguousGPU (in_sizes, in, out_sizes, out)

    alt "check_output_size == true"
        I->>nv: nvcompBatchedLZ4GetDecompressSizeAsync(stream)
        I->>H: cudaMemcpyAsync D2H decoded_sizes (stream)
        I->>I: cudaStreamSynchronize
        I->>I: "Validate each chunk decoded_size <= inflated_size"
        note over I,H: Raises DALI_ENFORCE if too small
    end

    I->>nv: nvcompBatchedLZ4DecompressGetTempSizeAsync
    I->>nv: nvcompBatchedLZ4DecompressAsync(stream)
    I->>nv: FillTheTails zero-pad oversized output
    nv-->>U: decompressed output tensor

Loading

_{Reviews (4): Last reviewed commit: "Add missing nvCOMP LZ4 dynamic stub" | Re-trigger Greptile}

[dali-automaton]

CI MESSAGE: [59212184]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [59212200]: BUILD STARTED

[JanuszL]

Addressed all four Greptile findings in 901733d:

• updated the modified Python test copyright year;
• moved check_output_size_ into the GPU-only LZ4 implementation;
• made the validation error identify both sample and chunk-within-sample;
• added a successful check_output_size=True regression test.

Validated with operator_1.test_inflate and test_dali_variable_batch_size.test_inflate (143 tests total).

[dali-automaton]

CI MESSAGE: [59214762]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [59214775]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [59214762]: BUILD FAILED

[dali-automaton]

CI MESSAGE: [59214775]: BUILD FAILED

[dali-automaton]

CI MESSAGE: [59239054]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [59239066]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [59239066]: BUILD PASSED

[dali-automaton]

CI MESSAGE: [59239054]: BUILD PASSED