cp: [benchmarking] Adds --use-host-curator-benchmarking and --skip-curator-image-build* options (1910) into r1.2.0

[svcnvidia-nemo-ci]

beep boop [🤖]: Hi @rlratzel 👋,

we've cherry picked #1910 into  for you! 🚀

Please review and approve this cherry pick by your convenience!

[svcnvidia-nemo-ci]

/ok to test fee405b

[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]

Greptile Summary

This cherry-pick from #1910 adds two new options to the benchmarking tooling: --skip-curator-image-build / --skip-curator-image-build-and-pull for build_docker.sh (enabling CI workflows where the curator image is already available), and --use-host-curator-benchmarking for run.sh (mounting only the host's benchmarking/ directory rather than the full repo). It also enables GPU video-codec capabilities inside the benchmarking container.

• build_docker.sh: Replaces the original glob-based flag detection with a proper case loop; adds two mutually exclusive skip flags that allow pulling or reusing an existing curator image instead of rebuilding it.
• gen_runscript_vars.py: Adds --use-host-curator-benchmarking, which mounts only $HOST_CURATOR_DIR/benchmarking into the container, useful for iterating on benchmarking scripts against an already-installed Curator package.
• run.sh: Adds NVIDIA_DRIVER_CAPABILITIES=compute,utility,video so NVENC/NVDEC video codec paths are available to benchmarks that need them.

Confidence Score: 4/5

The changes are self-contained to benchmarking tooling and carry no risk to the core Curator library or production data paths.

All three files are benchmarking-only scripts. The new skip flags in build_docker.sh include a proper mutual-exclusion guard, the new --use-host-curator-benchmarking path construction mirrors the existing pattern, and the NVIDIA_DRIVER_CAPABILITIES addition is a straightforward env-var extension. The only finding is a duplicated comment that misleadingly labels the validation block — a cosmetic issue with no runtime impact.

No files require special attention; all changes are isolated to the benchmarking tooling directory.

Important Files Changed

Filename	Overview
benchmarking/tools/build_docker.sh	Refactors argument parsing to a proper case loop and adds --skip-curator-image-build / --skip-curator-image-build-and-pull flags with mutual-exclusion guard; minor duplicate comment on the validation block.
benchmarking/tools/gen_runscript_vars.py	Adds --use-host-curator-benchmarking flag with mutual-exclusion check against --use-host-curator; correctly constructs the benchmarking-only volume mount using the same .absolute() pattern as the existing flag.
benchmarking/tools/run.sh	Adds --env=NVIDIA_DRIVER_CAPABILITIES=compute,utility,video to expose GPU video-codec capabilities (NVENC/NVDEC) inside the benchmarking container.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A([build_docker.sh invoked]) --> B{Parse CLI flags}
    B --> C{Both skip flags set?}
    C -- Yes --> D[Error & exit 1]
    C -- No --> E{--skip-curator-image-build-and-pull?}
    E -- Yes --> F[Use existing local CURATOR_IMAGE]
    E -- No --> G{--skip-curator-image-build?}
    G -- Yes --> H[docker pull CURATOR_IMAGE]
    G -- No --> I[docker build CURATOR_IMAGE]
    I --> J{--tag-as-latest?}
    J -- Yes --> K[docker tag curator :latest]
    J -- No --> L
    K --> L[docker build CURATOR_BENCHMARKING_IMAGE]
    F --> L
    H --> L
    L --> M{--tag-as-latest?}
    M -- Yes --> N[docker tag benchmarking :latest]
    M -- No --> O([Done])
    N --> O

    subgraph run.sh
        P([run.sh invoked]) --> Q[gen_runscript_vars.py]
        Q --> R{--use-host-curator?}
        R -- Yes --> S[Mount full HOST_CURATOR_DIR]
        R -- No --> T{--use-host-curator-benchmarking?}
        T -- Yes --> U[Mount HOST_CURATOR_DIR/benchmarking only]
        T -- No --> V[No curator mount]
        S --> W[docker run with NVIDIA_DRIVER_CAPABILITIES=compute,utility,video]
        U --> W
        V --> W
    end

Loading

_{Reviews (1): Last reviewed commit: "[benchmarking] Adds `--use-host-curator-..." | Re-trigger Greptile}

[greptile-apps]

The comment on line 55 is a copy-paste duplicate of the comment on line 61. The first block's comment should describe its actual purpose — validating that the two skip flags aren't combined — rather than describing the action that follows later.

Suggested change

	# Either pull, build, or skip the standard NeMo Curator image
	if ${SKIP_CURATOR_BUILD} && ${SKIP_CURATOR_BUILD_AND_PULL}; then
	echo "Error: --skip-curator-image-build and --skip-curator-image-build-and-pull cannot be combined."
	exit 1
	fi
	if ${TAG_AS_LATEST}; then
	# Tag image as <name>:latest, where <name> is the part of CURATOR_IMAGE before the colon
	docker tag "${CURATOR_IMAGE}" "${CURATOR_IMAGE%%:*}:latest"
	# Either pull, build, or skip the standard NeMo Curator image
	# Validate that conflicting skip flags are not combined
	if ${SKIP_CURATOR_BUILD} && ${SKIP_CURATOR_BUILD_AND_PULL}; then
	echo "Error: --skip-curator-image-build and --skip-curator-image-build-and-pull cannot be combined."
	exit 1
	fi
	# Either pull, build, or skip the standard NeMo Curator image