docs: replace the compatibility CUDA checker with a generated support matrix

[dagil-nvidia]

Overview

The Compatibility page opened with Find a Compatible Release, an interactive picker (RunsWhereWizard) that asked the reader to select a backend and a CUDA driver generation before showing anything. This replaces it with the matrix itself, collapsed in a native Fern <Accordion>.

Two problems motivated the change. The picker gated a plain lookup behind two choices, one of which — your CUDA driver generation — is frequently the thing the reader came to the page to find out. And the complete CUDA-and-driver-by-release table had become unreachable from the sidebar: it survived only in reference/releases-data.mdx (marked hidden: true in index.yml) and in the <llms-only> agent twin, neither of which a browsing reader will find.

What Changed

• RunsWhereWizard is unmounted from compatibility.mdx, along with its import and the heading and prose that introduced it. The component file is left in place — compatibility.mdx was its only consumer, but deleting it is the author's call, not this PR's.
• A collapsed <Accordion> takes its slot, under a new ## Release Support Matrix heading, carrying CUDA toolkit and minimum driver per backend for all 60 rows in CUDA_HISTORY.

The Table Cannot Drift

The numbers are generated from docs/fern/components/releases.data.ts, the single source of truth the other reference pages already use. No version data is hand-copied into the MDX, so the matrix regenerates at every release bump like the rest of the reference content.

gen_llms_tables.py previously supported one generated span per page. It now keys spans by marker name, so compatibility.mdx carries a human-visible support-matrix span alongside its existing llms-tables twin. Both call the same cuda_table() renderer. The new span fails closed when its markers are absent rather than appending itself to the end of the file, keeping placement under human control.

Scope

The matrix covers every released line — stable releases and their patches. Platform previews (-dev.N) and model-specific builds are excluded; the CUDA_NOTES bullets beneath the table cover their toolkit support in prose. CUDA_HISTORY carries no preview or model-build rows today, so that filter is currently a no-op, but it is expressed on release kind rather than on the data happening to be clean, so a future preview entry cannot leak in unnoticed.

Note on Duplication

The accordion and the <llms-only> twin now emit the same 60 rows. Keeping both is deliberate. The twin exists precisely because component-rendered output may be dropped from Fern's agent-facing markdown exports, and <Accordion> is a component. The local dev server does not serve the .md export (returns 400), so this could not be verified here, and losing the matrix for agents is a worse failure than repeating it. A comment in the generator records the intent so it is not pruned without checking; the redundancy can be removed once the deployed .md export is confirmed to carry Accordion children.

Verification

• fern check — 0 errors
• scripts/check_reference.sh — gates 1-4 pass (agent twins fresh, custom.js parses, no stale matrix links, absolute reference hrefs match the nav)
• Generator idempotency confirmed with --check
• Local preview: accordion renders data-state="closed", 60 data rows, both wizard headings gone, and the link out to the machine-readable page resolves

Pre-Existing Failure, Not From This PR

check_reference.sh gate 5 reports two fern docs broken-links errors inside reference/, both from a single dead link definition:

[kv-routing]: ../user-guides/kv-cache-aware-routing.md

That line is byte-identical on main and is untouched by this diff — no user-guides/ directory exists anywhere under docs/fern. main fails this gate today without these changes. Flagging rather than fixing, since the correct target is a guess and the link belongs to a neighbouring section.

Data Accuracy

Spot-checked the matrix against the build system and the pre-refactor authoritative table. No drift found. container/render.py pins cuda_version to 13.0 for vllm, 13.1 for trtllm and 13.0 for sglang, matching the v1.3.0 rows exactly, and every row agrees with the old docs/reference/support-matrix.md including the Experimental notes on the v0.8.x CUDA 13.0 rows.

---

Summary by CodeRabbit

• Documentation
  • Replaced the compatibility wizard with a release support matrix.
  • Added an expandable table showing CUDA toolkit and minimum driver requirements by release and backend.
  • Added notes covering patch versions, early-access images, and experimental CUDA 13 image availability.
  • Improved generation of multiple independently updated tables within documentation pages.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[coderabbitai]

Walkthrough

The generator now supports multiple marked spans per page and produces a released-version CUDA support matrix. The compatibility page replaces its release wizard with a generated accordion containing CUDA, driver, patch-version, and early-access information.

Changes

Compatibility documentation generation

Layer / File(s)	Summary
Multi-span page generation docs/fern/scripts/gen_llms_tables.py	Adds configurable Block definitions, marker-specific matching, replacement or append behavior, and sequential rendering of multiple spans per page.
Support matrix rendering docs/fern/scripts/gen_llms_tables.py	Filters CUDA history to released stable and patch versions, renders the support matrix, and raises an error when no matching rows exist.
Compatibility page integration docs/fern/reference/compatibility.mdx	Replaces the compatibility wizard with a generated release support matrix accordion and supporting release notes.

Estimated code review effort: 4 (Complex) | ~45 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The PR has solid overview and verification details, but it omits the required 'Where should reviewer start?' and 'Related Issues' sections.	Add a 'Where should reviewer start?' section and the required Related Issues block (e.g. Closes #XXXX or the no-issue checkbox).

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title is concise and accurately describes the main change: replacing the compatibility CUDA picker with a generated support matrix.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

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

[coderabbitai]

🧹 Nitpick comments (2)

docs/fern/scripts/gen_llms_tables.py (2)

955-982: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Consider keyword arguments for Block's boolean fields to avoid the boolean-trap.

Block("support-matrix", render_support_matrix, False, False) and the other Block(...) call sites rely on positional booleans (llms_only, append_if_missing) that are indistinguishable at the call site without checking the class definition. Since Block is a NamedTuple, callers can pass keyword args for clarity.

♻️ Proposed refactor

 PAGES: dict[str, tuple[Block, ...]] = {
     "compatibility.mdx": (
-        Block("support-matrix", render_support_matrix, False, False),
-        Block("llms-tables", render_compatibility, True, True),
+        Block("support-matrix", render_support_matrix, llms_only=False, append_if_missing=False),
+        Block("llms-tables", render_compatibility, llms_only=True, append_if_missing=True),
     ),
     "release-artifacts.mdx": (
-        Block("llms-tables", render_release_artifacts, True, True),
+        Block("llms-tables", render_release_artifacts, llms_only=True, append_if_missing=True),
     ),
     "model-early-access-builds.mdx": (
-        Block("llms-tables", render_model_ea_builds, True, True),
+        Block("llms-tables", render_model_ea_builds, llms_only=True, append_if_missing=True),
     ),
-    "releases-data.mdx": (Block("llms-tables", render_releases_data, False, True),),
+    "releases-data.mdx": (Block("llms-tables", render_releases_data, llms_only=False, append_if_missing=True),),
 }

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/fern/scripts/gen_llms_tables.py` around lines 955 - 982, Update the
Block constructions in PAGES to pass llms_only and append_if_missing as keyword
arguments rather than positional booleans, preserving each existing value and
renderer.

---

991-999: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

No accompanying test for the newly parameterized critical regex.

block_re is now built from a variable marker name (previously a hard-coded pattern) and is central to correctly locating/replacing generated spans across multiple pages/markers. As per coding guidelines, "Be careful with regex escaping in raw strings... when changing critical regexes, add a one-line test to prove it matches." No test evidence for this regex change is included in this diff.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/fern/scripts/gen_llms_tables.py` around lines 991 - 999, Add a focused
one-line test for block_re that verifies a parameterized marker is correctly
matched and escaped when locating a generated span, including the expected
begin/end markers. Keep the test scoped to block_re and preserve its support for
edited note text between the markers.

Sources: Coding guidelines, Path instructions

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@docs/fern/scripts/gen_llms_tables.py`:
- Around line 955-982: Update the Block constructions in PAGES to pass llms_only
and append_if_missing as keyword arguments rather than positional booleans,
preserving each existing value and renderer.
- Around line 991-999: Add a focused one-line test for block_re that verifies a
parameterized marker is correctly matched and escaped when locating a generated
span, including the expected begin/end markers. Keep the test scoped to block_re
and preserve its support for edited note text between the markers.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 78a67533-89e9-4ebc-afdb-478d3ea88c85

📥 Commits

Reviewing files that changed from the base of the PR and between 0fe93a1 and 912dc86.

📒 Files selected for processing (2)

• docs/fern/reference/compatibility.mdx
• docs/fern/scripts/gen_llms_tables.py

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 912dc86875

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[github-actions]

🌿 Fern Docs Preview: https://nvidia-preview-019fbea0-0882-723a-a1cb-e636c456a3b8.docs.buildwithfern.com/dynamo/dev

[dagil-nvidia]

/ok to test 99614d7

[dagil-nvidia]

/ok to test b5ff7ab

[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.

[dagil-nvidia]

/ok to test 6d8df6a

[dagil-nvidia]

/ok to test 9934563

[dagil-nvidia]

/ok to test 2f71a70