{
  "affected_packages": [],
  "configuration_checks": [
    {
      "condition": "equals",
      "context": "Applies when starting vLLM 0.34.0 with the old runner setting.",
      "effect": "startup_failure",
      "evidence_url": "https://github.com/vllm-project/vllm/releases/tag/v0.34.0",
      "setting": "VLLM_USE_V2_MODEL_RUNNER",
      "value": "0"
    },
    {
      "condition": "uses_default",
      "context": "The default changed from 16 to 64. ROCm paged attention stays at 16.",
      "effect": "default_changed",
      "evidence_url": "https://github.com/vllm-project/vllm/releases/tag/v0.34.0",
      "setting": "--block-size",
      "value": null
    }
  ],
  "editorial": {
    "audience": "Teams that serve models with vLLM and own their GPU deployment configuration.",
    "first_check": "Check staging deployment manifests for the old runner setting and explicit block-size overrides. Then compare installed dependencies with the release notes.",
    "scenario": "Review existing inference deployments before upgrading an environment to vLLM 0.34.0."
  },
  "event_id": "evt-eedebcdab0bdeccfc99fd72f",
  "event_kind": "breaking-change",
  "event_revision": 1,
  "schema_version": 1,
  "source_facts": [
    {
      "evidence_url": "https://github.com/vllm-project/vllm/releases/tag/v0.34.0",
      "text": "Model Runner V1 was removed. VLLM_USE_V2_MODEL_RUNNER=0 fails at startup."
    },
    {
      "evidence_url": "https://github.com/vllm-project/vllm/releases/tag/v0.34.0",
      "text": "The default --block-size changed from 16 to 64. ROCm paged attention still uses 16."
    },
    {
      "evidence_url": "https://github.com/vllm-project/vllm/releases/tag/v0.34.0",
      "text": "The release notes specify minimum versions of 2.0.0 for huggingface_hub, 0.23.3 for tokenizers, and 1.0.30 for timm."
    }
  ],
  "unknowns": [
    "Latency and memory impact for your workload",
    "Rollback behavior in your environment"
  ]
}
