Revert "Remove nvenc/dec for xenna 0.1.6"

[ayushdg]

Reverts #1202 due to some issues we see with Xenna's GPU usage vs the allocated resources via Ray

[greptile-apps]

Greptile Summary

This PR reverts #1202 to restore NVENC/NVDEC hardware encoder/decoder support for video transcoding stages.

• Dependency Downgrade: Reverts cosmos-xenna from 0.1.8 to 0.1.2 to use the older resource allocation API
• Resource Model: Re-adds nvdecs and nvencs fields to the Resources dataclass for fine-grained GPU hardware unit allocation
• GPU Partitioning: Restores dynamic GPU resource calculation in ClipTranscodingStage that queries actual GPU capabilities via _get_local_gpu_info() and _make_gpu_resources_from_gpu_name()
• Backend Compatibility: Updates Xenna adapter to pass nvdec/nvenc/entire_gpu fields to XennaResources; adds validation in Ray Data adapter to reject unsupported nvdec/nvenc requests
• Import Path Changes: Updates imports from cosmos_xenna.pipelines.private.resources to cosmos_xenna.ray_utils.resources

The revert addresses GPU resource allocation issues observed with Xenna's GPU usage vs allocated resources via Ray.

Confidence Score: 4/5

• This is a clean revert of a previous PR with no custom modifications. The dependency downgrade is the primary risk factor.
• Score of 4 because: (1) This is a straightforward revert with no manual conflict resolution, (2) The changes are well-scoped to GPU resource handling, (3) Tests were appropriately updated. Minor concerns: dependency downgrade may affect other features added in cosmos-xenna 0.1.8, and the MockGpuResources class appears unused.
• nemo_curator/stages/video/clipping/clip_extraction_stages.py - uses cosmos_xenna internal APIs for GPU detection which may fail without GPUs present

Important Files Changed

Filename	Overview
nemo_curator/stages/resources.py	Re-adds nvdecs and nvencs fields to Resources dataclass for NVIDIA hardware encoder/decoder support. Removes the entire_gpu auto-conversion to gpus=1.0 logic. Updates docstrings accordingly.
nemo_curator/stages/video/clipping/clip_extraction_stages.py	Restores GPU resource partitioning logic using _get_local_gpu_info() and _make_gpu_resources_from_gpu_name() from cosmos_xenna. When hardware encoding is enabled, dynamically calculates nvencs and gpu_memory_gb based on available GPU resources.
nemo_curator/backends/xenna/adapter.py	Changes import paths from cosmos_xenna.pipelines.private.resources to cosmos_xenna.ray_utils.resources for compatibility with cosmos-xenna 0.1.2. Re-adds nvdecs, nvencs, and entire_gpu fields to XennaResources mapping.
pyproject.toml	Downgrades cosmos-xenna dependency from 0.1.8 to 0.1.2 to revert to older API that supports nvdecs/nvencs resource allocation.
tests/stages/video/clipping/test_clip_transcoding_stage.py	Adds MockGpuResources class (currently unused). Removes two hwaccel tests that validated the old gpus-based resource allocation.

Sequence Diagram

sequenceDiagram
    participant User as User
    participant Pipeline as Pipeline
    participant ClipTranscodingStage as ClipTranscodingStage
    participant Resources as Resources
    participant XennaAdapter as XennaStageAdapter
    participant Xenna as XennaResources

    User->>Pipeline: run()
    Pipeline->>ClipTranscodingStage: __post_init__()
    
    alt h264_nvenc or use_hwaccel
        ClipTranscodingStage->>ClipTranscodingStage: _get_local_gpu_info()
        ClipTranscodingStage->>ClipTranscodingStage: _make_gpu_resources_from_gpu_name()
        ClipTranscodingStage->>Resources: Resources(nvencs=..., gpu_memory_gb=...)
    else CPU encoding
        ClipTranscodingStage->>Resources: Resources(cpus=...)
    end
    
    Pipeline->>XennaAdapter: required_resources
    XennaAdapter->>Resources: Get cpus, gpus, nvdecs, nvencs, entire_gpu
    XennaAdapter->>Xenna: XennaResources(cpus, gpus, nvdecs, nvencs, entire_gpu)
    Xenna-->>XennaAdapter: Resource allocation
    XennaAdapter-->>Pipeline: Execute stage with resources

Loading

[greptile-apps]

Additional Comments (1)

1. tests/stages/video/clipping/test_clip_transcoding_stage.py, line 31-42 (link)

   style: Both MockGpuInfo and MockGpuResources classes are defined but never used in any tests. Consider removing them or adding tests that utilize these mocks.

_{7 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}