fix(recipes): make the Solar Open2 250B benchmark and docs link usable

[snarravula-dl]

Fixes two defects found in post-deployment testing.

6786199 — benchmark manifests referenced an inaccessible AIPerf image

Both Jobs pinned nvcr.io/nvstaging/ai-dynamo/aiperf:0.13.0rc2, which returns 401 to an unauthenticated client, so the published figures could not be reproduced outside NVIDIA. The pin was inherited from the reference perf.yaml these were copied from.

Now nvcr.io/nvidia/ai-dynamo/aiperf:0.12.0, the latest public release (verified anonymously pullable). That release does not accept cache_bust.target: warmup_isolation_system, so the block is removed and the note describing warmup isolation is corrected.

Both profiles were re-measured on the public client, cold cache, full 15% trace:

Profile	tok/s/GPU	tok/s/user p50	TTFT p50	Requests
Aggregated	213.74	52.05	325 ms	1805/1805
Disaggregated 1P:2D	206.34	52.31	2,161 ms	1805/1805

Both clear the gates, and every figure sits inside the run-to-run spread already seen with the previous client, so the published table is unchanged.

6786216 — README linked to a nonexistent documentation route

The link used /dynamo/latest/recipes/model-recipes/..., which 404s. The published page is at /dynamo/dev/recipes/solar-open2-250b (verified 200), which is the route other recipes use. The .lycheeignore entry added while the page was unpublished is no longer needed.

Summary by CodeRabbit

• Documentation

  • Updated Solar Open2 250B performance notes to clarify KV cache behavior and cold-cache run requirements.
  • Updated the recipe documentation link.

• Benchmark Configuration

  • Updated benchmark tooling to version 0.12.0.
  • Removed warmup-isolation cache-busting settings.
  • Clarified that benchmark runs may inherit existing worker cache contents unless worker pods are restarted.

[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.

[github-actions]

🌿 Fern Docs Preview: https://nvidia-preview-01a0b054-cc1a-7141-bbf1-abab549d89af.docs.buildwithfern.com/dynamo/dev

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 8f6ca85c-895d-4457-8744-a211fa5ad113

📥 Commits

Reviewing files that changed from the base of the PR and between 0dbd9a8 and af040d6.

📒 Files selected for processing (5)

• .lycheeignore
• docs/fern/pages/recipes/model-recipes/solar-open2-250b.mdx
• recipes/solar-open2-250b/README.md
• recipes/solar-open2-250b/perf/perf-disagg.yaml
• recipes/solar-open2-250b/perf/perf.yaml

💤 Files with no reviewable changes (1)

• .lycheeignore

Included review availability: Your plan provides up to 12 included reviews per hour; 11 remain after this review.

---

Walkthrough

The Solar Open2 250B benchmark manifests now use a released aiperf image and omit warmup cache busting. The benchmark documentation describes KV cache behavior. Recipe documentation links now use the dev path.

Changes

Solar Open2 250B alignment

Layer / File(s)	Summary
Benchmark configuration and cache documentation recipes/solar-open2-250b/perf/*, docs/fern/pages/recipes/model-recipes/solar-open2-250b.mdx	Both performance manifests remove warmup cache-busting settings and use nvcr.io/nvidia/ai-dynamo/aiperf:0.12.0. The benchmark notes state that server-side KV cache remains and that worker pods must restart for cold-cache runs.
Recipe documentation references .lycheeignore, recipes/solar-open2-250b/README.md	The README links to the dev recipe path. The old URL is removed from .lycheeignore.

Priority: ⬇️ Low

Estimated code review effort: 1 (Trivial) | ~5 minutes

Merge Risk: ⚪ Minimal · up to af040

The benchmark configurations and documentation references are aligned, with no merge-blocking risk identified.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The description explains the defects, changes, validation results, and affected behavior. It does not follow the required template because it omits the reviewer-start section and the required Related …	Add the required template sections, including “Where should the reviewer start?” and “Related Issues.” In the Related Issues section, either link the relevant issue or confirm that no related issue exists.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the two main changes: fixing the Solar Open2 250B benchmark and documentation link.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 0…
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.

Full details: Description check

Explanation

The description explains the defects, changes, validation results, and affected behavior. It does not follow the required template because it omits the reviewer-start section and the required Related Issues declaration.

• Fix all pre-merge checks with AI

---

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

[ynpandey-nv]

/ok to test af040d6

[dmitry-tokarev-nv]

Approved

Both fixes verified by execution. One P3 on the performance table, which does not block.

What I ran, and where verification stopped

The docs link. Resolved with curl, following redirects:

URL	status
docs.nvidia.com/dynamo/dev/recipes/solar-open2-250b (this PR)	200
docs.nvidia.com/dynamo/latest/recipes/model-recipes/solar-open2-250b (merge base)	404

The old path was wrong, not moved. The Fern navigation at docs/fern/index.yml:804 gives the page the slug solar-open2-250b inside a section with skip-slug: true, so recipes/model-recipes/ was never part of the route. The new form matches every sibling recipe README: kimi-k3, deepseek-v4-1-flash, gemma-4-31b, ax-k2.

The link checker could not have caught the old link. .github/workflows/docs-link-check.yml runs lychee over the whole tree, so it does read recipes/*/README.md, but the exact 404 URL was listed in .lycheeignore. The same change that added the link added the entry that hid it. Removing the entry restores the coverage, and lychee is green on this PR with no entry.

The benchmark. The blocking defect was the image. Anonymous manifest fetch from nvcr.io:

image	status
nvcr.io/nvstaging/ai-dynamo/aiperf:0.13.0rc2 (merge base)	401
nvcr.io/nvidia/ai-dynamo/aiperf:0.12.0 (this PR)	200

I extracted the ConfigMap body from all four manifests, base and head, and ran aiperf config validate against a real aiperf 0.12.0 install. The base pair is rejected on cache_bust.target, and the head pair is valid. warmup_isolation_system is a 0.13.0 enum value. The removal is required, not cosmetic.

Where verification stopped. I did not run the benchmark. That needs a B200 cluster. I could not check the claim that the new figures sit inside the run-to-run spread, because the spread is not stated. Both points are unverified.

Siblings. No other recipe carries either defect. Every other perf.yaml already points at nvcr.io/nvidia/ai-dynamo/aiperf, and no other manifest in the tree uses cache_bust. The /latest/recipes/model-recipes/ link form appeared only here.

Style. The page keeps its SPDX header and Fern frontmatter, adds no body H1, and keeps the vLLM casing.

Out of scope, for a separate change. This recipe has no entry under docs/fern/pages/recipes/_catalog/recipes/, unlike the other 26. That gap predates this PR.

[dmitry-tokarev-nv]

[P3] The published table was not measured with these manifests. docs/fern/pages/recipes/model-recipes/solar-open2-250b.mdx:253. The numbers in the table above come from the old image and the removed cache_bust block. The PR body reports 213.74 tok/s/GPU and 325 ms for aggregated. Please update the table, or say which client produced the figures.

Measured: aiperf 0.12.0 rejects the old configuration, accepts the new one

Both ConfigMap bodies extracted from the manifests and checked with aiperf config validate on a real 0.12.0 install (pip, macOS ARM64):

tree	perf.yaml	perf-disagg.yaml
merge base	rejected	rejected
this PR	valid	valid

The rejection on the merge base is the same on both files:

benchmark.datasets.0.file.cache_bust.target
  Input should be 'none', 'system_prefix', 'system_suffix',
  'first_turn_prefix' or 'first_turn_suffix'
  [type=enum, input_value='warmup_isolation_system', input_type=str]

So the change is necessary and correct: warmup_isolation_system is a 0.13.0 enum value and does not exist in 0.12.0.

The point of this comment is narrower. The table at lines 250-251 still reads 216.88 / 52.02 / 308 ms and 206.20 / 51.69 / 2277 ms, and the sentence at line 253 tells the reader those figures come from perf.yaml and perf-disagg.yaml. Those two files no longer configure warmup isolation and no longer use the client that produced the figures. A reader who reproduces the run gets the PR body numbers, not the table numbers.

The PR body says every figure sits inside the run-to-run spread. I could not check that claim. The spread is not stated, and I have no B200 cluster. Unverified.

Registry access, checked anonymously with a token from nvcr.io/proxy_auth:

image	manifest HTTP status
nvcr.io/nvstaging/ai-dynamo/aiperf:0.13.0rc2 (merge base)	401
nvcr.io/nvidia/ai-dynamo/aiperf:0.12.0 (this PR)	200