[yifeif-nv]

Background

The family isolation refactor removed the model detail page's configuration and command reference, and the Supported Models table now renders checkpoint and family names as plain text. Users cannot inspect the settings behind a listed model.

Exit Criteria

Every listed family and checkpoint recipe opens a detail page with its current parameters, defaults, restrictions, source references, and explicitly scoped verification results.

Implementation

• Restore family links and add a detail route for each checkpoint recipe, with searchable support inventory and complete task/family navigation.
• Generate configuration reference from current family Python/C++ declarations and CLI contracts. Preserve exact manifest settings, alternate defaults, complex declarations, and links pinned to the source revision. Load configuration data separately for each family.
• Check every recipe through the current manifest catalog, CLI serializer/parser, typed build request, and conservative request-only source guards. Check testcase request mapping separately; neither result claims native runtime or GPU qualification.
• Run rendered-page checks as part of the website build so missing links, parameters, declarations, or verification results fail validation.

Change categories

• Documentation only
• CI or developer tooling

Validation

Commands and Results

• npm --prefix website run test:model-support under Node 20.20.2: 18 Node tests and 21 Python tests passed, including complete current-family/recipe coverage and adversarial extraction cases.
• npm --prefix website run build under Node 20.20.2: passed, including 34 diagrams and rendered links/configuration checks for 135 families, 303 recipes, and 31 tasks.
• PYTHONPATH=core/builder:apps/benchmark:. python3 -m tools.model_ci validate: passed for all 135 families.
• PYTHONPATH=core/builder:apps/benchmark:. python3 -m pytest core/builder/tests/test_family_cli.py core/builder/tests/test_build_cli.py -q: 63 passed.
• Ruff on the four new Python scripts and git diff --check: passed.
• Local Chromium smoke: checkpoint search, family/checkpoint navigation, expanded CLI parameters, issue visibility, task index, and mobile layout passed with no browser runtime exceptions.

Hardware, Environment, and Revisions

The validated website changes are committed as 9ad9d51ffbaac6ab4cd24352090e26ad2586fcd4, based on audited family source 383271563a10dfacaf1360ded8ec8a59f7ce7dbe. Local validation used Linux/aarch64, Python 3.12.3 and Node 20.20.2 with the existing website lockfile. GPU, CUDA and TensorRT were not used. Checkpoint revisions are preserved from each manifest; checkpoint weights were not downloaded.

Not Run / Remaining Gaps

TensorRT builds, native runtime configuration acceptance, GPU execution, parity, and performance were not run. The two build-input discrepancies and fourteen incomplete benchmark request mappings are described below and shown on the affected pages. Required remote premerge has not been established; this is a draft PR.

Contributor Self-Review

• I have completed a self-review of this change.

Reviewed the complete website-only diff, historical behavior, parser/control-flow edge cases, generated HTML coverage, and browser behavior. No family implementation or test threshold changes are included.

Notes For Future Readers

Start with the inventory plugin, then the extraction scripts and parameter renderer. The new reference contains 783 configuration declarations. Three families have no extracted typed configuration declaration; their actual build request, manifest, restrictions, and source links remain available.

Of 303 recipes, 301 pass the limited build-input checks. Two existing boundaries are shown on their pages: sam-vit-base-tp4 declares a max_sequence_length rejected by its family, and sam2-public-core-l0 uses mixed_bf16_fp32, accepted by its Python request but not expressible through the standard CLI precision choices. No family implementation or test passing criteria were changed.

Of 386 testcase requests, 372 map through the current CPU benchmark adapters. Seven need an adapter and seven Parakeet probes need family fixture preparation. These incomplete mappings are not model-support failures.

Risk level

• Low

The change affects documentation generation and navigation. Configuration extraction remains scoped source/CPU evidence, preserves complete source declarations for complex cases, and does not change model runtime behavior.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}