vllm-project/vllm · error · ValueError

Either vllm_config must be provided, or all of model_config,

Error message

Either vllm_config must be provided, or all of model_config, parallel_config, and cache_config must be provided.

What it means

ValueError from lmcache_get_or_create_engine_metadata-style config assembly in vLLM's LMCache integration utils: the caller passed neither a full vllm_config nor the complete trio of model_config, parallel_config, and cache_config. The function needs all three configs to build LMCacheEngineMetadata (KV dtype, MLA detection, parallelism), so it refuses partial input rather than guessing. It is an API-contract error, almost always a programming mistake in caller code, not a runtime/environment failure.

Source

Thrown at vllm/distributed/kv_transfer/kv_connector/v1/lmcache_integration/utils.py:134

                                          to vllm_config)
        cache_config (CacheConfig): Cache configuration (alternative to
                                    vllm_config)
    """
    # Third Party
    # First Party
    from lmcache.config import LMCacheEngineMetadata

    from vllm.utils.torch_utils import get_kv_cache_torch_dtype

    config = lmcache_get_or_create_config()
    # Support both vllm_config object and individual config parameters
    if vllm_config is not None:
        model_cfg = vllm_config.model_config
        parallel_cfg = vllm_config.parallel_config
        cache_cfg = vllm_config.cache_config
    else:
        if model_config is None or parallel_config is None or cache_config is None:
            raise ValueError(
                "Either vllm_config must be provided, or all of "
                "model_config, parallel_config, and cache_config must be provided."
            )
        model_cfg = model_config
        parallel_cfg = parallel_config
        cache_cfg = cache_config

    # Get KV cache dtype
    kv_dtype = get_kv_cache_torch_dtype(cache_cfg.cache_dtype, model_cfg.dtype)

    # Check if MLA is enabled
    use_mla = mla_enabled(model_cfg)

    # Construct KV shape (for memory pool)
    num_layer = model_cfg.get_num_layers(parallel_cfg)
    chunk_size = config.chunk_size
    num_kv_head = model_cfg.get_num_kv_heads(parallel_cfg)
    head_size = model_cfg.get_head_size()

View on GitHub (pinned to c794754062)

Solutions

1. Pass the complete VllmConfig object — the preferred and simplest path.
2. If you must pass individual configs, supply all three: model_config, parallel_config, and cache_config.
3. In tests, build a minimal VllmConfig via the standard mock/vllm_config factory instead of hand-assembling configs.

Example fix

# before
meta = build_metadata(model_config=mc)
# after
meta = build_metadata(vllm_config=vllm_config)

from vllm.config import VllmConfig
assert isinstance(vllm_config, VllmConfig) or all(
    c is not None for c in (model_config, parallel_config, cache_config)
)

def has_full_config(vllm_config, model_config, parallel_config, cache_config) -> bool:
    return vllm_config is not None or (
        model_config is not None
        and parallel_config is not None
        and cache_config is not None
    )

When it happens

Trigger: Calling the metadata/config helper with vllm_config=None and one of model_config/parallel_config/cache_config also None; passing only model_config; constructing the adapter manually in tests without a full VllmConfig.

Common situations: Custom scripts or tests that instantiate the LMCache integration directly instead of going through the engine; refactors that changed a function signature from three configs to a single vllm_config and callers were not all updated.

Related errors

• MLA only works with naive serde mode..
• layerwise MLA connector is not supported yet
• LMCacheMPConnector only works without hybrid kv cache manage
• Unknown KVConnectorRole: {self.role}
• The model is not multimodal.

AI-assisted analysis of vllm-project/vllm@c794754062 (2026-08-14). Data as JSON: /api/errors/fb1357cc044c7edc. Report an issue: GitHub.