docs: update docs and release notes for Cosmos-Xenna 0.2.0 (PR #1571)

[lbliii]

Description

Updates documentation and release notes for the Cosmos-Xenna 0.2.0 bump introduced in #1571. Replaces the 26.02 fern release notes with a 26.04 skeleton containing the Xenna 0.2.0 entry, dependency updates, and breaking changes. Removes stale entire_gpu, nvdecs, and nvencs references from the Resources API reference, video abstractions, and custom model tutorial. Adds a 26.04 section to the non-fern release notes with matching content.

Usage

from nemo_curator.stages.resources import Resources

# Before (26.02): entire_gpu, nvdecs, nvencs were available
# resources = Resources(cpus=4.0, entire_gpu=True)

# After (26.04): use gpu_memory_gb or gpus only
resources = Resources(cpus=4.0, gpus=1.0)

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 updates documentation and release notes for the Cosmos-Xenna 0.2.0 bump introduced in #1571, replacing stale entire_gpu, nvdecs, and nvencs field references across the v26.04 Fern docs and the non-Fern release notes, and adding a FERN_TOKEN env var to the Fern autodocs CI workflow.

• Release notes (fern/versions/v26.04/pages/about/release-notes/index.mdx): The 26.02 content that was incorrectly living in the v26.04 versioned directory is replaced with a correct 26.04 skeleton covering the Cosmos-Xenna 0.2.0 changes.
• docs/about/release-notes/index.md: A new "What's New in 26.04" section is prepended, and version sub-headers are added to the Dependency Updates and Breaking Changes sections (addressing prior feedback).
• API reference / abstractions / tutorial: All three files consistently remove entire_gpu (and nvdecs/nvencs where present) and update descriptions to reflect the new two-option GPU model (gpu_memory_gb or gpus).
• CI workflow: A FERN_TOKEN environment variable is added to the fern docs md generate step, but it references secrets.DOCS_FERN_TOKEN — a different name than the secrets.FERN_TOKEN that the sibling publish-fern-docs.yml workflow documents and uses. If only FERN_TOKEN is configured as a repository secret, the CI step will receive an empty string and may not authenticate as intended.

Confidence Score: 4/5

• Safe to merge after verifying the FERN_TOKEN secret name used in the CI workflow.
• All documentation changes are accurate and well-structured. The one P1 finding is the inconsistent secret name (DOCS_FERN_TOKEN vs FERN_TOKEN) in the CI workflow, which could silently break the fern docs generation step in CI if DOCS_FERN_TOKEN is not separately configured as a repository secret.
• .github/workflows/fern-docs-ci.yml — verify that DOCS_FERN_TOKEN is a configured repository secret, or align it with the FERN_TOKEN name used in publish-fern-docs.yml.

Important Files Changed

Filename	Overview
.github/workflows/fern-docs-ci.yml	Added FERN_TOKEN env to the generate step, but uses secret name DOCS_FERN_TOKEN which is inconsistent with the FERN_TOKEN name documented and used in the publish workflow — could result in an empty token if only FERN_TOKEN is configured.
docs/about/release-notes/index.md	Added 26.04 "What's New" section and version sub-headers to Dependency Updates and Breaking Changes; content is accurate and well-structured.
fern/versions/v26.04/pages/about/release-notes/index.mdx	Correctly replaces the 26.02 content (which was misplaced in the v26.04 directory) with a focused 26.04 skeleton covering Cosmos-Xenna 0.2.0, dependency updates, and breaking changes.
fern/versions/v26.04/pages/api-reference/resources.mdx	Cleanly removes the entire_gpu field and its usage example; updated docstring, requires_gpu description, and warning block all consistently reflect the new two-option GPU model.
fern/versions/v26.04/pages/about/concepts/video/abstractions.mdx	Removed entire_gpu bullet and updated gpus description to reflect the new "one or more full GPUs" semantics.
fern/versions/v26.04/pages/curate-video/tutorials/pipeline-customization/add-cust-model.mdx	Removed the stale entire_gpu example reference from the resource management paragraph; no other issues.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[PR opens / fern changes] --> B[fern-docs-ci.yml triggered]
    B --> C[Checkout + Install Fern CLI]
    C --> D[fern docs md generate\nenv FERN_TOKEN from DOCS_FERN_TOKEN secret]
    D -->|Secret configured| E[Generate MDX with auth]
    D -->|Secret missing| F[Empty token - silent no-auth or failure]

    G[Push to docs-website] --> H[publish-fern-docs.yml triggered]
    H --> I[Checkout + Install Fern CLI]
    I --> J[fern docs md generate - no token]
    J --> K[fern generate --docs\nenv FERN_TOKEN from FERN_TOKEN secret]
    K --> L[Docs published]

Loading

_{Reviews (4): Last reviewed commit: "docs: address PR feedback on release not..." | Re-trigger Greptile}

[greptile-apps]

Ray version format may be a typo

The version 2.54 appears inconsistently with standard Ray semantic versioning (e.g., 2.5.4 or 2.10.0). If this is intended to be 2.5.4, it should be corrected here and in the matching entry in docs/about/release-notes/index.md (lines 23 and 120).

Suggested change

	- **Ray 2.54**: Updated Ray dependency to version 2.54 for compatibility with Cosmos-Xenna 0.2.0.
	## Dependency Updates
	- **Transformers**: Pinned to 4.55.2 for stability and compatibility
	- **vLLM**: Updated to 0.14.1 with video pipeline compatibility fixes
	- **FFmpeg**: Upgraded to 8.0.1 for enhanced multimedia processing
	- **Security Patches**:
	- Addressed CVEs in aiohttp, urllib3, python-multipart, setuptools
	- Removed vulnerable thirdparty aiohttp file from Ray
	- Updated to secure dependency versions
	## Bug Fixes
	- Fixed fasttext predict call compatibility with numpy>2
	- Fixed broken NeMo Framework documentation links
	- Fixed MegatronTokenizerWriter to download only necessary tokenizer files
	- Fixed ID generator blocking issues for large-scale processing
	- Fixed vLLM API compatibility with video captioning pipeline
	- Fixed Gliner tutorial examples and SDG workflow bugs
	- Improved semantic deduplication unit test reliability
	## Infrastructure & Developer Experience
	- **Secrets Detection**: Automated secret scanning in CI/CD workflows
	- **Dependabot Integration**: Automatic dependency update pull requests
	- **Enhanced Install Tests**: Comprehensive installation validation across environments
	- **AWS Runner Support**: CI/CD execution on AWS infrastructure
	- **Docker Optimization**: Improved layer caching and build times with uv
	- **Code Linting**: Standardized code quality checks with markdownlint and pre-commit hooks
	- **Cursor Rules**: Development guidelines and patterns for IDE assistance
	- **Cosmos-Xenna**: Updated from 0.1.2 to 0.2.0 with simplified resource model
	- **Ray**: Updated to 2.54
	- **Xenna-managed CUDA devices**: Xenna now manages CUDA device visibility directly, replacing the previous Ray-managed approach.
	- **Ray 2.5.4**: Updated Ray dependency to version 2.5.4 for compatibility with Cosmos-Xenna 0.2.0.
	## Dependency Updates
	- **Cosmos-Xenna**: Updated from 0.1.2 to 0.2.0 with simplified resource model
	- **Ray**: Updated to 2.5.4

[greptile-apps]

Mixed-version Dependency Updates section lacks version attribution

The "## Dependency Updates" section now contains both 26.04 entries (Cosmos-Xenna, Ray) and 26.02 entries (Transformers, vLLM, FFmpeg, etc.) without any version headers to distinguish them. Readers cannot tell which dependencies changed in which release. The same applies to the "## Breaking Changes" section at line 148, which mixes 26.04 and 26.02 breaking changes.

Consider adding version sub-headers to make attribution clear:

## Dependency Updates

### 26.04
- **Cosmos-Xenna**: Updated from 0.1.2 to 0.2.0 with simplified resource model
- **Ray**: Updated to 2.54

### 26.02
- **Transformers**: Pinned to 4.55.2 ...