The vLLM v1 map

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

The idea#

vLLM v1 runs a request through a fixed chain of parts. The API server process builds the config and owns HTTP. The EngineCore process holds the scheduler and the executor. The executor starts the workers. Each worker owns one device and builds a model runner, which runs the model.

afd-plugin does not replace this chain. It changes six small points along it: the config, the EngineCore start, the worker class, the model runner class and the model class. This page names each box once, gives its vLLM line, and points to the AFD hook.

An everyday comparison. Think of a factory line with six stations. AFD does not rebuild the line; it changes the tool at a few stations. The comparison stops here: one station (the FFN EngineCore) stops the line early and runs its own loop.

The picture#

Read the boxes top to bottom. Each box runs in a process or is built inside one. The violet row on each box is the one place AFD changes it. The afd-plugin loads once in each process, so the hooks take effect everywhere they are needed.

In the code#

The table names each box, its vLLM line at commit ced6857, and the AFD hook. The vLLM lines are in the pinned vLLM checkout; the hook lines are in afd-plugin.

The executor chooses between separate worker processes and one in-process worker.

        elif distributed_executor_backend == "mp":
            from vllm.v1.executor.multiproc_executor import MultiprocExecutor

            executor_class = MultiprocExecutor
        elif distributed_executor_backend == "uni":
            from vllm.v1.executor.uniproc_executor import UniProcExecutor

            executor_class = UniProcExecutor

The worker then imports the class named in worker_cls. AFD has already set this field to its role worker, so vLLM builds the AFD worker here.

        parallel_config = vllm_config.parallel_config
        if isinstance(parallel_config.worker_cls, str):
            worker_class: type[WorkerBase] = resolve_obj_by_qualname(
                parallel_config.worker_cls
            )

Deep dive#

In afd-plugin#

Each hook has its own page. Start here:

• The config and the worker rewrite: AFD configuration: additional_config["afd"] and Worker selection: choosing the role worker.
• The model wrappers: Model registration: the AFD<arch> wrappers.
• The two roles and the FFN EngineCore: Attention and FFN roles: what each server owns.
• The module that applies the hooks at load time: afd_plugin/__init__.py: the plugin entry module.