{
  "agent": {
    "identifiers": [],
    "kind": "breaking-change",
    "review": "editorial",
    "source_fingerprint": "a5377a6e16874ed70fb8ed505e4d85dda65bd05bea5270514a91141a5bb9b446",
    "source_id": "vllm",
    "source_key": "447739190",
    "version": "v0.34.0"
  },
  "category": "local-ai",
  "commercial_opportunities": null,
  "compatibility_and_migration": "Check runner flags, block size overrides, and dependencies against the upgrade table.",
  "confidence": "high",
  "developer_impact": "GPU serving teams need to check existing launch configurations before rollout.",
  "discovery_date": "2026-11-21",
  "event_date": "2026-11-16",
  "evidence": [
    {
      "basis": "Maintainer release notes, including the Before you upgrade table.",
      "fields": [
        "title",
        "event_date",
        "technical_description",
        "what_changed",
        "previous_behavior",
        "new_behavior",
        "developer_impact",
        "known_limitations",
        "compatibility_and_migration",
        "summary"
      ],
      "url": "https://github.com/vllm-project/vllm/releases/tag/v0.34.0"
    }
  ],
  "id": "evt-eedebcdab0bdeccfc99fd72f",
  "known_limitations": "ROCm paged attention keeps a block size of 16.",
  "new_behavior": "vLLM sizes the extensible KV cache after warmup. The option VLLM_USE_V2_MODEL_RUNNER=0 now fails at startup.",
  "organization": "vLLM",
  "previous_behavior": "The previous default block size was 16. Configurations could select Model Runner V1.",
  "priority": {
    "band": "P1",
    "dimensions": {
      "adoption_potential": {
        "reason": "Estimate based on the scope of the documented change.",
        "value": 4
      },
      "agent_relevance": {
        "reason": "The score reflects relevance to agent and developer workflows.",
        "value": 5
      },
      "commercial_opportunity": {
        "reason": "This score reflects a possible workflow opportunity. Demand is unverified.",
        "value": 4
      },
      "commoditization_risk": {
        "reason": "The score reflects possible overlap with existing tools. Market effects are uncertain.",
        "value": 2
      },
      "security_impact": {
        "reason": "The record makes no security severity claim.",
        "value": 0
      },
      "source_credibility": {
        "reason": "The source is the official maintainer release or a reviewed advisory.",
        "value": 5
      },
      "technical_significance": {
        "reason": "The migration table documents startup and GPU configuration changes.",
        "value": 5
      },
      "time_sensitivity": {
        "reason": "The score reflects the age of the change and the need to review it.",
        "value": 4
      }
    },
    "reason": "Weighted evidence score of 75. The score meets the P1 threshold.",
    "score": 75
  },
  "related_event_ids": [],
  "revision": 1,
  "schema_version": 1,
  "security_implications": null,
  "security_severity": null,
  "source_authority": "maintainer",
  "source_urls": [
    "https://github.com/vllm-project/vllm/releases/tag/v0.34.0"
  ],
  "status": "active",
  "summary": "vLLM v0.34.0 removes Model Runner V1 and changes KV cache defaults. Review the upgrade table before deployment.",
  "tags": [
    "inference-migration",
    "open-source",
    "release"
  ],
  "technical_description": "The vLLM maintainers published v0.34.0. Model Runner V2 is now the only GPU runner.",
  "technology": "vLLM",
  "title": "vLLM v0.34.0 changes the GPU runner and KV cache defaults",
  "verification_date": "2026-11-21",
  "what_changed": "vLLM enables the extensible KV cache by default. The default block size is 64."
}
