AFD config reference: the additional_config["afd"] keys

Code version: afd-plugin @ 7452d52 · 2026-10-27 · vLLM @ ced6857

How the config is passed#

You give AFD config to vLLM under one key. Each vllm serve process gets --additional-config '{"afd": { ... }}'. vLLM stores the outer mapping on its additional_config field (vllm:vllm/config/vllm.py:434). The plugin reads the inner "afd" mapping and builds an AFDConfig dataclass (afd_plugin/config.py:42).

The two tables in the next section are generated from the source. The validation, read-site and example tables that follow are hand-written, each row with a citation.

The keys#

The build generates the tables below from AFDConfig and the connector source, so the names, types, defaults and lines stay correct at the pinned commit. Run python3 tools/gen_afd_config.py to refresh them.

Generated by tools/gen_afd_config.py from afd_plugin/config.py and the connector source. Do not edit between the markers.

The afd keys#

connector_extra_config by connector#

P2pNcclAFDConnector (GPU): no keys. A non-empty connector_extra_config raises ValueError (afd_plugin/connectors/gpu/p2p.py:155).

CAMP2pAFDConnector (NPU)

CAMAsyncAFDConnector (NPU)

The async connector accepts the wire key dynamicQuant for its typed field dynamic_quant.

Validation rules#

The plugin coerces each value to its type, then validates it. Coercion happens in _normalize_mapping (afd_plugin/config.py:120-176); validation happens in validate_afd_config (afd_plugin/config.py:311-350). An invalid value raises before the run starts.

Where each value is read, and what it changes#

The table lists the main GPU read site of each key. The NPU connectors read the same keys through their own code.

Where the final value is set#

Lesson from the fact checks: a default is not final until the code sets it. Trace each value through four steps. A missing key takes the dataclass default; a present key is coerced, stored, then validated.

1. Recipe or CLI. The --additional-config JSON gives the raw mapping. vLLM stores it on additional_config (vllm:vllm/config/vllm.py:434).
2. Normalize. _normalize_mapping rewrites aliases, splits out connector_extra_config, and coerces port, the rank counts, the timeout and the booleans (afd_plugin/config.py:147-174).
3. Construct. AFDConfig(**normalized) fills any absent key with the dataclass default (afd_plugin/config.py:189-190). So the defaults in the table apply only to keys the recipe omits.
4. Validate. validate_afd_config raises on a bad value (afd_plugin/config.py:311-350). The worker re-parses and re-validates through assert_compatible_afd_stack (afd_plugin/validation.py:328).

Worked example: the recipe's afd block#

This block is the attention server from the simplest GPU recipe, 2a2f_eager_dbo_dp1tp2.sh (recipe/gpu/P2pNcclAFDConnector/deepseek_v2_lite/prefill_decode_colocation/2a2f_eager_dbo_dp1tp2.sh:11-20). The FFN server uses the same block with "role": "ffn". The tour Tour: Life of a request follows a request through this recipe.

{
  "afd": {
    "role": "attention",
    "connector": "P2pNcclAFDConnector",
    "host": "127.0.0.1",
    "port": 6269,
    "num_attention_ranks": 2,
    "num_ffn_ranks": 2
  }
}

The recipe omits five keys, so they take their defaults: async_dp=false, compute_gate_on_attention=false, afd_process_group_timeout_s=120, an empty connector_extra_config, and the coerced port and rank integers. The A >= F rule holds because both counts are 2 (afd_plugin/distributed/topology.py:52-58).