Walkthrough

The vLLM model registry entry for nvidia/Qwen3.6-35B-A3B-NVFP4 doubles its maxModelLen from 65536 to 131072, increasing the --max-model-len serve command flag. The test is updated to verify the new expected value.

Changes

Model Configuration and Test Update

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~2 minutes

Possibly related PRs

• NVIDIA/NemoClaw#4619: Both PRs update the vLLM model registry/test for the same qwen3.6-35b-a3b-nvfp4 model (in src/lib/inference/vllm-models.ts and vllm-models.test.ts), with the main PR specifically adjusting its maxModelLen (i.e., --max-model-len) to a new value.
• NVIDIA/NemoClaw#4689: The main PR increases the vLLM --max-model-len for qwen3.6-35b-a3b-nvfp4 (thereby changing what /v1/models.max_model_len reports), and the retrieved PR reads that runtime max_model_len and applies it to NEMOCLAW_CONTEXT_WINDOW, so they directly connect through the same max_model_len value.

Suggested labels

Platform: DGX Spark, Provider: vLLM

Suggested reviewers

• cv

Poem

🐰 A model grows twice as tall,
From 65K to twice that, all!
The test now cheers, 131072 bright,
Qwen's context expanded to new height. ✨

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

E2E Recommendation Advisor

Base: origin/main
Head: HEAD
Confidence: medium

Required E2E

• inference-routing-e2e (medium): The PR changes inference-serving configuration. This is the closest existing merge-blocking E2E signal for assistant inference paths: it validates onboarding-created inference routes, inference.local access from the sandbox, OpenAI-compatible chat completion routing, and error classification.

Optional E2E

• gpu-e2e (high): Useful adjacent confidence for local GPU inference onboarding and sandbox inference behavior, although it exercises Ollama rather than managed vLLM and will not verify the changed --max-model-len value.
• onboard-inference-smoke-e2e (low): Adjacent regression coverage that setupInference fails unless the configured route serves a real chat completion. It does not exercise managed vLLM startup but is relevant to onboarding/inference success semantics.

New E2E recommendations

• managed vLLM DGX Spark install (high): No existing E2E appears to start the managed vLLM container with NEMOCLAW_VLLM_MODEL=qwen3.6-35b-a3b-nvfp4 and assert the generated vllm serve command includes --max-model-len 131072, reaches /v1/models, and serves a sandbox chat completion through inference.local. The current change is only directly covered by unit tests.
  • Suggested test: Add a DGX Spark/NVIDIA GPU managed-vLLM E2E job that runs non-interactive onboarding with NEMOCLAW_PROVIDER=install-vllm and NEMOCLAW_VLLM_MODEL=qwen3.6-35b-a3b-nvfp4, then verifies container readiness, command flags, and a real OpenAI-compatible chat completion from inside the sandbox.

Dispatch hint

• Workflow: .github/workflows/nightly-e2e.yaml
• jobs input: inference-routing-e2e

E2E Scenario Advisor

Base: origin/main
Head: HEAD
Confidence: high

Required scenario E2E

• None. The PR changes the vLLM model registry and its unit tests outside test/e2e-scenario/. The current dispatchable scenario ROUTES do not include a managed-vLLM or DGX Spark scenario that would exercise NEMOCLAW_VLLM_MODEL or the updated max-model-len for the NVFP4 checkpoint; existing scenario inference coverage is cloud NVIDIA/OpenAI-compatible or local Ollama.

Optional scenario E2E

• None.

Relevant changed files

• None.