fix(inference): reject wrong-platform NEMOCLAW_VLLM_MODEL overrides before downloads

[andrewwhitecdw]

Summary

NEMOCLAW_VLLM_MODEL overrides for platform-restricted models without a runtime field could start image and model downloads on unsupported hardware. The installer now rejects every model whose platform list excludes the detected profile before callbacks or Docker side effects. Product scope verdict: PASS — #7358 fixes existing managed-vLLM platform enforcement and does not add a supported integration or product surface; this verdict is independent of GitHub mergeStateStatus.

Related Issue

Closes #7358

Changes

• Remove the runtime-field condition from the platform gate in runVllmInstall.
• Document that runVllmInstall enforces every model's platform list before downloads.
• Add regressions for Spark-only and Station-only overrides on generic Linux and a supported Spark override.

Type of Change

• Code change (feature, bug fix, or refactor)
• Code change with doc updates
• Doc only (prose changes, no code sample modifications)
• Doc only (includes code sample changes)

Quality Gates

• Tests added or updated for changed behavior
• Existing tests cover changed behavior — justification:
• Tests not applicable — justification:
• Docs updated for user-facing behavior changes
• Docs not applicable — justification: canonical docs already restrict managed-vLLM overrides to supported platforms and associate the affected models with DGX Spark or DGX Station.
• Sensitive paths changed (security, policy, credentials, preflight, onboarding, inference, runner, sandbox, or messaging)
• Sensitive-path review completed or maintainer-approved waiver recorded — reviewer/approval link/justification: exact-head security review PASS at fix(inference): reject wrong-platform NEMOCLAW_VLLM_MODEL overrides before downloads #7359 (review); no findings across the nine required categories.
• Non-success, skipped, or missing CI check accepted by maintainer — check name, approval link, and follow-up issue:

Documentation Writer Review

• Documentation writer subagent reviewed the completed changes
• Result: no-docs-needed
• Evidence: Exact-head review found only src/lib/inference/vllm-models.ts, src/lib/inference/vllm.test.ts, and src/lib/inference/vllm.ts. All PR product paths are unchanged from reviewed head b835f6049; the new commit is a merge-only update from main. The fix enforces documented platform support for managed-vLLM model overrides and adds no command, environment variable, or workflow. git diff --check 0670d2035...95498cf89 passed.
• Agent: Codex Desktop

DGX Station Hardware Evidence

• Tested on DGX Station
• Tested commit:
• Station profile/scenario:
• Result:
• Supporting evidence:

Verification

• PR description includes a Signed-off-by: line and every commit appears as Verified in GitHub
• Normal pre-commit, commit-msg, and pre-push hooks passed, or npm run check:diff passed when hooks were skipped or unavailable
• Targeted behavior tests pass for the current change set, or tests are marked not applicable above — npx vitest run --project cli src/lib/inference/vllm.test.ts src/lib/inference/vllm-models.test.ts (100/100 passed)
• Applicable broad gate passed — npm test for broad runtime/test-harness changes; npm run check for repo-wide validation/coverage changes — command/result:
• Quality Gates section completed with required justifications or waivers
• No secrets, API keys, or credentials committed
• npm run docs builds without warnings (doc changes only)
• Doc pages follow the style guide (doc changes only)
• New doc pages include SPDX header and frontmatter (new pages only)

---

Signed-off-by: Andrew White andrewh@cdw.com

[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.

[coderabbitai]

📝 Walkthrough

Walkthrough

vLLM installation now validates platform compatibility for all selected models, including NEMOCLAW_VLLM_MODEL overrides, before pulling images or downloading weights. Regression tests cover incompatible Linux overrides and a supported Spark override, and the model documentation reflects the validation.

Changes

vLLM platform compatibility

Layer / File(s)	Summary
Override validation and regression coverage src/lib/inference/vllm.ts, src/lib/inference/vllm-models.ts, src/lib/inference/vllm.test.ts	runVllmInstall rejects models whose platforms exclude the target profile before downloads, with tests covering Spark-only and Station-only overrides on generic Linux and a Spark-compatible override. The platforms documentation describes this behavior.

Estimated code review effort: 2 (Simple) | ~10 minutes

Suggested labels: provider: vllm, platform: dgx-spark, platform: dgx-station

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	The changes satisfy #7358 by enforcing platform checks without relying on runtime and adding regression tests for the affected overrides.
Out of Scope Changes check	✅ Passed	The PR stays in scope with a small validation fix, a JSDoc clarification, and targeted regression tests.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main change: rejecting wrong-platform NEMOCLAW_VLLM_MODEL overrides before downloads.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[github-actions]

PR Review Advisor — Informational

Advisor assessment: Informational / high confidence
Next action: No advisor follow-up needed.
Findings: 0 blockers · 0 warnings · 0 suggestions
Status: No actionable findings remain in the canonical review ledger.

Model lanes

• GPT-5.6 Terra (primary): Completed · high confidence · 0 blockers · 0 warnings · 0 suggestions
• Nemotron 3 Ultra (second opinion): Completed · high confidence · 0 blockers · 3 warnings · 1 suggestion
• Model comparison: normalized findings differ; normalized E2E selections match; Nemotron reported the same number of blockers, 3 more warnings, 1 more suggestion.

Nemotron output stays in workflow artifacts and does not change the assessment above.

E2E guidance

Advisory only. E2E / PR Gate selects and runs jobs independently.

Recommended E2E: inference-routing, network-policy

Workflow run details

This automated review informs maintainers. Warnings and suggestions do not require a response. A maintainer decides whether to merge.

[wscurran]

✨ Thanks for the fix. This closes the platform gate gap for wrong-platform env overrides. Maintainers will review the vLLM platform enforcement changes.

---

Related open issues:

• #7358 vLLM platform gate misses wrong-platform NEMOCLAW_VLLM_MODEL overrides without a runtime field

[senthilr-nv]

Reviewed PR #7359 at head bfd9f10 against base fe56943. Exact-pair non-secret CI passed. Approve the audited skip for credential-bearing inference-routing and network-policy E2E; no fork code may receive repository secrets.

Approval was submitted while attempting to approve the protected E2E-skip environment. PR approval was premature and referenced a stale SHA/base pair; exact-head required CI and E2E gates remain pending.

[prekshivyas]

Sensitive-path security review — PASS

Exact pair reviewed: head b835f60490e0a8fd3b7b6863099a0bb37502c3bf against base f0848373b6866c7a4071491a6d054767a7a5968b.

Verdict: No security findings. This is safe from a code-security perspective. It strengthens the existing managed-vLLM platform boundary by applying the registry allowlist to every selected model before callbacks, prerequisite probes, Docker inspection, image pulls, or model downloads. Product scope is established by #7358 and the linked maintainer acknowledgment that #7359 addresses the reported existing-surface defect. This security PASS does not waive CI or required exact-head E2E.

Findings: none.

1. Secrets and credentials — PASS: no secret-handling code or credential material changed.
2. Input validation and sanitization — PASS: the user-controlled model override must resolve through the existing registry and now must match profile.platform before side effects.
3. Authentication and authorization — PASS: no auth boundary changes; the new check restricts rather than expands behavior.
4. Dependencies and third-party libraries — PASS: no dependency changes.
5. Error handling and logging — PASS: rejection reports only the model label and profile name and exposes no sensitive data.
6. Cryptography and data protection — PASS: no cryptographic or stored-data behavior changes.
7. Configuration and security headers — PASS: no container flags, ports, image references, network policy, or HTTP headers change.
8. Security testing — PASS: regressions cover Spark-only and Station-only overrides on generic Linux before callbacks/Docker work, plus the valid Spark path.
9. System security — PASS: the check is fail-closed, uses the registry's explicit platform allowlist, runs before all relevant side effects, and introduces no TOCTOU or privilege expansion.

Files reviewed in full: src/lib/inference/vllm-models.ts, src/lib/inference/vllm.ts, and src/lib/inference/vllm.test.ts.

[andrewwhitecdw]

@prekshivyas do you prefer I use the update branch button when the branch gets behind main or leave it alone and let a maintainer do it?

[prekshivyas]

Sensitive-path maintainer review refreshed for exact head 95498cf89caecee6b88cd3c9d80c3e5c93b75c59 on base 0670d20352298dd4686d6efe6abc5118ab8f27d2: PASS.

The three product blobs reviewed previously (src/lib/inference/vllm-models.ts, src/lib/inference/vllm.ts, and src/lib/inference/vllm.test.ts) are byte-identical to reviewed head b835f60490e0a8fd3b7b6863099a0bb37502c3bf; the new commit is a GitHub-Verified merge-only refresh from main. The prior nine-category security review therefore remains exact-content applicable: the platform gate still fails before callbacks, Docker probing, or other side effects on unsupported hosts; tests cover both unsupported platform paths and the supported path; no credentials, policy broadening, dependency, logging, or cryptographic changes are introduced. git diff --check 0670d2035...95498cf89 passes.

Approval still requires fresh exact-head CI/advisors and non-skipped exact-head E2E evidence.

[prekshivyas]

@prekshivyas do you prefer I use the update branch button when the branch gets behind main or leave it alone and let a maintainer do it?

@andrewwhitecdw generally if there are no conflicts it should not matter, but if we change ci and e2e tests on main it can affect the PR, so if the contributor frequently can, that is well and good or we take care of it!

[prekshivyas]

Exact-head gate status for 95498cf89caecee6b88cd3c9d80c3e5c93b75c59 against 0670d20352298dd4686d6efe6abc5118ab8f27d2: ordinary CI, DCO/Verified, documentation receipt, advisor replacement, mergeability, and the sensitive-path review are green. Approval remains blocked because the fork gate selected inference-routing and network-policy but did not run them; it offers only a credentialed-E2E skip. This review requires current non-skipped exact-head evidence, so I am not approving the skip or the PR. The remaining path is an internal trusted branch/superseding PR for the same reviewed change, or an explicit maintainer policy decision changing that requirement.

[prekshivyas]

Approved at exact head 95498cf. Security review passed with no blockers; focused platform-allowlist validation fails before side effects, targeted tests pass, and all CI checks are satisfied. Credentialed E2E was skipped through the audited fork-PR gate as required by repository policy.