feat(recipes): add Qwen3.5-0.8B vLLM aggregated recipe

[MatejKosec]

Summary

• Adds an aggregated single-GPU vLLM recipe for Qwen/Qwen3.5-0.8B under recipes/qwen3.5-0.8b/.
• Targets nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.1.0. The recipe will not work against 1.0.0: that image's vLLM and Transformers pre-date the qwen3_5 model type.
• Wires --dyn-reasoning-parser qwen3 and --dyn-tool-call-parser qwen3_coder on the worker. qwen3_coder is Dynamo's registered Qwen3-family XML tool-call parser.

Closes #8988.

Notes

• Uses --no-enable-log-requests, the renamed flag introduced in fix(recipes): rename --disable-log-requests to --no-enable-log-requests for vLLM #8693. Older recipes still using --disable-log-requests will fail with unrecognized arguments on 1.1.0.
• Directory is named qwen3.5-0.8b/ (with the dot) to track the official model name -- matches the precedent set by kimi-k2.5/.

[coderabbitai]

Walkthrough

This PR adds a complete Qwen3.5-0.8B aggregated model deployment recipe. It includes a DynamoGraphDeployment manifest (vLLM with Frontend/VllmWorker services), Kubernetes PVC definitions for model and compilation caches, a model-download Job for HuggingFace artifact preparation, and comprehensive README documentation with quick-start instructions and operational guidance.

Changes

Qwen3.5-0.8B Aggregated Deployment Recipe

Layer / File(s)	Summary
Deployment Definition recipes/qwen3.5-0.8b/vllm/agg/deploy.yaml	DynamoGraphDeployment qwen3-5-0-8b-agg specifies Frontend and VllmWorker services. VllmWorker configures vLLM with Qwen-specific reasoning (qwen3) and tool-call (qwen3_coder) parsers, multimodal support, tensor-parallel size 1, GPU memory 0.85, max-length 32768, and prefix caching.
Storage Infrastructure recipes/qwen3.5-0.8b/model-cache/model-cache.yaml	Two PersistentVolumeClaim manifests: model-cache (20Gi) and compilation-cache (10Gi), both with ReadWriteOnce access and placeholder storageClassName.
Model Download Job recipes/qwen3.5-0.8b/model-cache/model-download.yaml	Kubernetes batch Job downloads Qwen/Qwen3.5-0.8B model via huggingface_hub 1.11.0, pulls HF token from hf-token-secret, writes to model-cache PVC at /home/dynamo/.cache/huggingface, with backoffLimit: 3 and single completion/parallelism.
Documentation & Quick-Start recipes/qwen3.5-0.8b/README.md	README documents configuration options, prerequisites (1x H100 80GB, Dynamo, HF token secret), quick-start commands (create secrets, apply PVCs, deploy graph, test /v1/chat/completions), parser wiring details, architecture notes on vLLM model registry and EOS workaround, and validation notes (vLLM runtime 1.1.0, CLI flag --no-enable-log-requests, Qwen3.5 hybrid-attention caching behavior).

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title 'feat(recipes): add Qwen3.5-0.8B vLLM aggregated recipe' clearly describes the main change: adding a new recipe for Qwen3.5-0.8B with vLLM aggregated configuration.
Linked Issues check	✅ Passed	The PR successfully addresses both requirements from issue #8988: it pins vLLM to version 1.1.0 (which supports qwen3_5 model type) and uses qwen3_coder as the tool-call parser, with test confirmation of proper architecture resolution and tool-call functionality.
Out of Scope Changes check	✅ Passed	All changes are directly within scope: README documentation, Kubernetes manifests for model caching, job configuration, and vLLM deployment configuration are all necessary for the single-GPU Qwen3.5-0.8B aggregated recipe implementation.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Description check	✅ Passed	The PR description provides a clear summary, links a closing issue, and includes helpful notes about prerequisites and naming conventions.

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

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[coderabbitai]

Actionable comments posted: 2

🤖 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.

Inline comments:
In `@recipes/qwen3.5-0.8b/model-cache/model-download.yaml`:
- Around line 17-44: The container "model-download" is missing a securityContext
and runs as root; add a securityContext block under the container spec for
model-download to harden defaults: set runAsNonRoot: true and a non-root
runAsUser (e.g. 1000), set allowPrivilegeEscalation: false, and consider adding
readOnlyRootFilesystem: true and seccompProfile/runtimeClass if desired; update
the container spec (name: model-download) to include this securityContext so the
pod no longer runs as root and privilege escalation is disabled.

In `@recipes/qwen3.5-0.8b/README.md`:
- Around line 73-78: The fenced log snippet in recipes/qwen3.5-0.8b/README.md
lacks a language identifier (triggering markdownlint MD040); update the fence
for that block (the triple backticks surrounding the log lines) to include a
language token such as "text" so it becomes ```text, ensuring the block around
the lines like "2026-05-07T21:16:07  INFO model.__post_init__: Resolved
architecture: Qwen3_5ForConditionalGeneration" is closed with matching
backticks.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 95d0a12e-9c04-458b-94d3-36f62406f148

📥 Commits

Reviewing files that changed from the base of the PR and between 91516b0 and f03f250.

📒 Files selected for processing (4)

• recipes/qwen3.5-0.8b/README.md
• recipes/qwen3.5-0.8b/model-cache/model-cache.yaml
• recipes/qwen3.5-0.8b/model-cache/model-download.yaml
• recipes/qwen3.5-0.8b/vllm/agg/deploy.yaml