Cherry pick Revert "Remove nvenc/dec for xenna 0.1.6 (#1202)" (#1374) for r1.1.0

[ayushdg]

This reverts commit c4805ae.

Description

Usage

# Add snippet demonstrating usage

Checklist

• I am familiar with the Contributing Guide.
• New or Existing tests cover these changes.
• The documentation is up to date with these changes.

[greptile-apps]

Greptile Summary

This PR reverts commit c4805ae ("Remove nvenc/dec for xenna 0.1.6") to restore NVDEC/NVENC hardware encoder support for video processing stages. The revert downgrades cosmos-xenna from 0.1.8 to 0.1.2 and re-introduces the nvdecs and nvencs resource attributes.

Key Changes:

• Restored nvdecs and nvencs fields in the Resources dataclass for explicit hardware encoder/decoder allocation
• Updated ClipTranscodingStage to calculate NVENC resources from GPU info instead of using fractional GPU allocation
• Changed cosmos_xenna imports from pipelines.private.resources to ray_utils.resources (matches 0.1.2 API)
• Added validation in Ray Data adapter to reject nvdecs/nvencs usage (not supported by Ray Data)
• Removed the automatic gpus = 1.0 assignment when entire_gpu=True from the Resources.__post_init__ method
• Updated pipeline description to display NVDEC/NVENC resource information

Confidence Score: 4/5

• This is a clean revert of a previous commit to restore NVDEC/NVENC support with an older cosmos-xenna version; changes are straightforward and well-scoped.
• The PR is a revert that restores previously working functionality. The code changes are consistent with reverting to cosmos-xenna 0.1.2 API. One potential concern is the removal of the entire_gpu -> gpus = 1.0 assignment, which could affect stages using entire_gpu=True in the Ray Data backend.
• nemo_curator/stages/resources.py - verify entire_gpu=True behavior still works correctly without automatic gpus assignment

Important Files Changed

Filename	Overview
nemo_curator/stages/resources.py	Added nvdecs and nvencs fields for NVDEC/NVENC hardware encoder support; removed automatic gpus=1.0 assignment when entire_gpu is True.
nemo_curator/backends/xenna/adapter.py	Updated imports to use cosmos_xenna.ray_utils.resources instead of cosmos_xenna.pipelines.private.resources; added nvdecs, nvencs, entire_gpu fields to XennaResources.
nemo_curator/backends/experimental/ray_data/adapter.py	Added validation to raise error when nvdecs/nvencs are requested without GPUs in Ray Data backend (Ray Data doesn't support these hardware resources directly).
nemo_curator/stages/video/clipping/clip_extraction_stages.py	Updated to use cosmos_xenna GPU info functions for NVENC resource calculation; now allocates nvencs and gpu_memory_gb instead of fractional gpus.
pyproject.toml	Downgraded cosmos-xenna dependency from 0.1.8 to 0.1.2 as part of the revert.

Sequence Diagram

sequenceDiagram
    participant User
    participant Pipeline
    participant ClipTranscodingStage
    participant Resources
    participant XennaAdapter
    participant cosmos_xenna

    User->>Pipeline: Create pipeline with video stages
    Pipeline->>ClipTranscodingStage: __post_init__()
    ClipTranscodingStage->>cosmos_xenna: _get_local_gpu_info()
    cosmos_xenna-->>ClipTranscodingStage: GPU info (name, etc.)
    ClipTranscodingStage->>cosmos_xenna: _make_gpu_resources_from_gpu_name()
    cosmos_xenna-->>ClipTranscodingStage: num_nvencs
    ClipTranscodingStage->>Resources: Create(nvencs, gpu_memory_gb)
    Resources-->>ClipTranscodingStage: Resources instance
    
    User->>Pipeline: run(executor)
    Pipeline->>XennaAdapter: adapt stage
    XennaAdapter->>Resources: Get cpus, gpus, nvdecs, nvencs, entire_gpu
    XennaAdapter->>cosmos_xenna: XennaResources(cpus, gpus, nvdecs, nvencs, entire_gpu)
    cosmos_xenna-->>XennaAdapter: Resource allocation

Loading