Engine configuration

Steering has two configuration levels. Engine arguments declare the algorithms, execution mode, and capacity for the lifetime of an LLM or vllm serve process. Each request supplies its vectors and token selection through a SteeringSpec.

Python and CLI arguments

Pass Python arguments to LLM(...); use the corresponding flags with vllm serve MODEL ....

Python argument	CLI flag	Default	Purpose
enable_steer_vector	--enable-steer-vector	False	Enable steering. A startup steering_config also enables it.
steer_algorithms	--steer-algorithms	None	Algorithm list, comma-separated names, or "all". Required when enabling steering without a startup default.
steer_multi_vector	--steer-multi-vector	False	Allow more than one vector in a request. Implied by "all" or a multi-vector startup default.
steer_graph_mode	--steer-graph-mode	"auto"	Select in_graph or split; see graph modes.
steer_graph_max_rank	--steer-graph-max-rank	32	Maximum LoReFT/LM-Steer rank accepted by an in_graph engine.
max_steer_vectors	--max-steer-vectors	min(256, max_num_seqs)	Maximum distinct steering configurations active in a running batch.
steer_vector_dtype	--steer-vector-dtype	"auto"	Payload dtype: model dtype for auto, or float16, bfloat16, float32.
steer_require_preload	--steer-require-preload	False	Require file-backed payloads to be preloaded before generation requests.
steering_config	--steering-config	None	Startup default as JSON text or the path to a spec JSON file.

The declaration can include several algorithms while each request uses just one. For example, steer_algorithms=["direct", "attention_add"] declares two algorithms, while steer_multi_vector=True permits combining vectors in a single spec. These are separate choices.

When a startup default is supplied, its algorithms are added to the declaration. The engine can inspect that default's payload to choose its graph mode. Later default updates and per-request overrides must fit the resolved engine settings.

Common configurations

One additive intervention per requestLoReFT with a known rank boundInteractive multi-algorithm server

from vllm import LLM

llm = LLM(
    model="Qwen/Qwen2.5-1.5B-Instruct",
    enable_steer_vector=True,
    steer_algorithms=["direct"],
)

Auto selects in_graph. Requests can use different files, scales, layers, selectors, and normalization settings within that declaration.

from vllm import LLM

llm = LLM(
    model="Qwen/Qwen2.5-1.5B-Instruct",
    enable_steer_vector=True,
    steer_algorithms=["loreft"],
    steer_graph_mode="in_graph",
    steer_graph_max_rank=32,
)

Each request must supply a LoReFT payload of rank at most 32. Use auto to select split when the declaration must accept arbitrary ranks.

vllm serve Qwen/Qwen2.5-1.5B-Instruct \
  --enable-steer-vector --steer-algorithms all \
  --max-steer-vectors 16 --port 8017

all implies multi-vector support and selects split. Component-specific requests still require a compatible model; declaring moe_router, for example, does not add MoE gates to a dense model.

Capacity and defaults

max_steer_vectors counts active configurations, not the number of weight files or submitted requests. Requests with identical configurations share a slot. If all slots are in use, requests needing a new configuration wait for one to become available. This does not require the client to resubmit them.

Payload storage is content-addressed separately: several configurations that use the same weights can share a materialized payload. A larger slot capacity permits more distinct concurrent configurations, but increases persistent graph-table memory. Low-rank graph buffers also grow with steer_graph_max_rank; use a bound suited to the checkpoints you serve.

Changing the engine default with llm.set_default_steering(...) affects new requests. It does not reserve a slot permanently or change the graph mode. See request inheritance and HTTP management endpoints.

Standard vLLM settings

Model loading, scheduling, sampling, memory allocation, and CUDA graph capture sizes remain vLLM settings. The fork uses vLLM 0.29.0 and its V2 GPU runner for steering and capture; leave the runner selection at its default.

Use the upstream engine argument reference for general options and this page for EasySteer's additions. Match upstream documentation to the installed vLLM version when defaults change.