Add CUDA 13.4 build configuration

[JanuszL]

Category:

Other (Configuration)

Description:

Add CUDA 13.4 dependency-image configurations for x86_64 and aarch64, make
CUDA 13.4 the default Docker build version, and update the compilation
documentation.

Additional information:

Affected modules and functionalities:

• Docker CUDA dependency images for x86_64 and aarch64
• Docker build CUDA-version selection and validation
• Compilation documentation

Key points relevant for the review:

• Both architecture-specific Dockerfiles pin the CUDA 13.4 and cuFile package
  versions.
• The Dockerfiles have no cuda-repo.nvidia.com sources; they use only the
  public NVIDIA developer-download repository.

Tests:

• Existing tests apply
  • build process
• New tests added
  • Python tests
  • GTests
  • Benchmark
  • Other
• N/A

Checklist

Documentation

• Existing documentation applies
• Documentation updated
  • Docstring
  • Doxygen
  • RST
  • Jupyter
  • Other
• N/A

DALI team only

Requirements

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

REQ IDs: N/A

JIRA TASK: DALI-4810

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge; no outstanding correctness, security, or repository-rule issue was identified.

Summary

• Pins the CUDA, cuFile, and nvCOMP versions for both architectures.
• Extends CUDA-version validation and corrects the build-script help text.
• Updates compilation documentation to identify CUDA 13.4 as the supported default.
• Removes the temporary nvImageCodec reproducer introduced after the prior review.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  Selection["CUDA_VERSION=13.4"] --> BuildScript["docker/build.sh"]
  BuildScript --> X86["CUDA 13.4 x86_64 dependency image"]
  BuildScript --> ARM["CUDA 13.4 aarch64 dependency image"]
  X86 --> Builder["DALI builder"]
  ARM --> Builder
  Builder --> Wheels["Architecture-specific DALI wheels"]

Loading

_{Reviews (5) · Last reviewed commit: "Address CUDA 13.4 configuration review f..."}

[JanuszL]

@greptile review

[dali-automaton]

CI MESSAGE: [67060476]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [67060476]: BUILD FAILED

[dali-automaton]

CI MESSAGE: [67108784]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [67108784]: BUILD FAILED

[dali-automaton]

CI MESSAGE: [67108784]: BUILD PASSED

[jantonguirao]

Reviewed the diff (2 new Dockerfiles + build.sh + docs/compilation.rst).

Verified the version bump is consistent with the existing 13.0–13.3 pattern: Dockerfile.cuda134.{x86_64,aarch64}.deps follow the same structure as Dockerfile.cuda133.*.deps (same package list, same NVCOMP archive URL pattern, same fatbinary wrapper — just rewritten as individual echo >> lines instead of one embedded \n-escaped string, which is a robustness improvement, not a behavior change). docker/build.sh's Dockerfile selection is by filename convention (Dockerfile.cuda${CUDA_VER}.${ARCH}.deps), which resolves correctly for CUDA_VER=134. Docs (compilation.rst) and the CUDA_VERSION validation list were both updated consistently.

All three Greptile findings from the first review round (wrong default in help text, nvCOMP version mismatch 5.2.0.10 vs 5.3.0.16, missing terminal punctuation in the error message) are fixed in the follow-up commit ed15c41 — confirmed against the current diff, not just the bot's say-so. The later Greptile comments about docker/repro-nvimgcodec-issue55/* are stale/misattributed — those files aren't part of this PR's diff at all.

Two pre-existing, out-of-scope observations (not blocking, not introduced by this PR, so no inline comments):

• Dockerfile.cuda133.*.deps is missing libnvjpeg/libnvjpeg-dev (present in 13.0–13.2 and correctly restored in the new 13.4 files) — looks like a regression that crept into 13.3, worth a follow-up fix there.
• The CREATE_RUNNER=YES path in build.sh (~line 325) only sets CUDA_IMAGE_NAME for CUDA_VER 120/130 and falls through to "Not supported CUDA version" for everything else, including 131/132/133 and now 134 — predates this PR (introduced 2026-07-13), so out of scope here.

No blocking issues. Approving.