@elizaos/plugin-local-inference 2.0.0-beta.1 → 2.0.3-beta.3

package/src/services/vision/index.ts

@@ -0,0 +1,251 @@
+/**
+ * Vision-describe capability (WS2) — public entry point.
+ *
+ * This module is what plugin-vision (WS4), the IMAGE_DESCRIPTION
+ * handler in `provider.ts`, and computer-use (WS9) import to register
+ * vision capability with the WS1 MemoryArbiter.
+ *
+ * Wiring:
+ *
+ *   const arbiter = service.getMemoryArbiter();
+ *   const registration = createVisionCapabilityRegistration({
+ *     loader: createDefaultVisionLoader({ ... }),
+ *     arbiterCache: arbiter,
+ *   });
+ *   arbiter.registerCapability(registration);
+ *
+ * `createVisionCapabilityRegistration` wraps the underlying backend so
+ * the arbiter's `run(request)` path:
+ *
+ *   1. Hashes the request's image bytes (model-family-scoped).
+ *   2. Checks the arbiter's vision-embedding cache.
+ *   3. On miss: calls `backend.describe(request)`, lets the backend
+ *      run its own projector + decoder. Backends that cannot expose projected
+ *      tokens return decoder text only, so the cache stays empty for this hash.
+ *      The decoder text is what the caller wanted anyway.
+ *   4. On hit: calls `backend.describe(request, { projectedTokens })`.
+ *      Backends that support pre-projected token reuse skip the
+ *      projector entirely. Backends that don't ignore the hint; the
+ *      result is still correct but the projector cost is paid again.
+ */
+
+export {
+	type AospLlamaMtmdBinding,
+	type AospMtmdHandle,
+	type LoadAospVisionBackendOptions,
+	loadAospVisionBackend,
+} from "./aosp-unavailable";
+export {
+	type CapacitorLlamaMtmdBinding,
+	type CapacitorLlamaMtmdHandle,
+	type CapacitorLlamaVisionBackendOptions,
+	loadCapacitorLlamaVisionBackend,
+	VisionBackendUnavailableError,
+	type VisionManagerLike,
+} from "./capacitor-llama";
+export {
+	classifyLocalVisionError,
+	type LocalImageDescriptionHandler,
+	type LocalVisionOutcome,
+	type VisionCloudFallbackOptions,
+	type VisionFallbackReason,
+	type WrappedImageDescriptionHandler,
+	wrapImageDescriptionHandlerWithCloudFallback,
+} from "./cloud-fallback";
+export {
+	hashImageBytes,
+	hashRawPixels,
+	hashVisionInput,
+	resolveImageBytes,
+} from "./hash";
+export {
+	createLlamaServerVisionBackend,
+	type LlamaServerVisionBackendOptions,
+} from "./llama-server";
+export type {
+	VisionDescribeBackend,
+	VisionDescribeBackendLoader,
+	VisionDescribeBackendOptions,
+	VisionDescribeLoadArgs,
+	VisionDescribeRequest,
+	VisionDescribeResult,
+	VisionImageChannelOrder,
+	VisionImageInput,
+} from "./types";
+export {
+	type VisionVastFallbackOptions,
+	wrapImageDescriptionHandlerWithVastFallback,
+} from "./vast-fallback";
+
+import type {
+	ArbiterCapability,
+	CapabilityRegistration,
+} from "../memory-arbiter";
+import { hashVisionInput } from "./hash";
+import type {
+	VisionDescribeBackend,
+	VisionDescribeBackendLoader,
+	VisionDescribeRequest,
+	VisionDescribeResult,
+} from "./types";
+
+/**
+ * Minimal arbiter shape we need from the cache. Lets tests inject a
+ * fake cache without pulling in the whole MemoryArbiter.
+ */
+export interface VisionEmbeddingCacheLike {
+	getCachedVisionEmbedding(hash: string): {
+		tokens: Float32Array;
+		tokenCount: number;
+		hiddenSize: number;
+		live?: boolean;
+	} | null;
+	setCachedVisionEmbedding(
+		hash: string,
+		entry: {
+			tokens: Float32Array;
+			tokenCount: number;
+			hiddenSize: number;
+		},
+		ttlMs?: number,
+	): void;
+}
+
+export interface CreateVisionCapabilityRegistrationOptions {
+	/**
+	 * The arbiter (or any object with the cache passthroughs). When
+	 * provided the wrapper performs hash → cache lookup before calling
+	 * the backend's `describe`.
+	 */
+	arbiterCache?: VisionEmbeddingCacheLike;
+	loader: VisionDescribeBackendLoader;
+	/** Default model family for the cache key. Defaults to `gemma-vl`. */
+	modelFamily?: string;
+	estimatedMb?: number;
+}
+
+/**
+ * Build a `CapabilityRegistration` ready to feed to
+ * `arbiter.registerCapability()`. The wrapper plumbs the cache hint
+ * into the backend's describe call so backends that support
+ * pre-projected tokens skip the projector.
+ */
+export function createVisionCapabilityRegistration(
+	opts: CreateVisionCapabilityRegistrationOptions,
+): CapabilityRegistration<
+	VisionDescribeBackend,
+	VisionDescribeRequest,
+	VisionDescribeResult
+> {
+	const capability: ArbiterCapability = "vision-describe";
+	const family = opts.modelFamily ?? "gemma-vl";
+	const cache = opts.arbiterCache;
+	const loader = opts.loader;
+	return {
+		capability,
+		residentRole: "vision",
+		estimatedMb: opts.estimatedMb ?? 600,
+		async load(modelKey) {
+			return await loader(modelKey);
+		},
+		async unload(backend) {
+			await backend.dispose();
+		},
+		async run(backend, request) {
+			const effectiveFamily = request.modelFamily ?? family;
+			const cached = (() => {
+				if (!cache) return null;
+				if (request.image.kind === "url") {
+					// URL inputs can't be hashed without first fetching; skip
+					// the cache lookup rather than paying the fetch cost twice.
+					return null;
+				}
+				try {
+					const hash = hashVisionInput(request.image, effectiveFamily);
+					const hit = cache.getCachedVisionEmbedding(hash);
+					if (hit && hit.live !== false) return { hash, hit };
+				} catch {
+					// Hashing failed (malformed data URL etc.); proceed without
+					// cache rather than failing the request.
+				}
+				return null;
+			})();
+			const projected = cached?.hit
+				? {
+						tokens: cached.hit.tokens,
+						tokenCount: cached.hit.tokenCount,
+						hiddenSize: cached.hit.hiddenSize,
+					}
+				: undefined;
+			const result = await backend.describe(request, {
+				projectedTokens: projected,
+			});
+			return {
+				...result,
+				cacheHit: Boolean(projected),
+			};
+		},
+	};
+}
+
+import type {
+	IAgentRuntime,
+	ImageDescriptionParams,
+	ImageDescriptionResult,
+} from "@elizaos/core";
+import {
+	type LocalImageDescriptionHandler,
+	type VisionCloudFallbackOptions,
+	wrapImageDescriptionHandlerWithCloudFallback,
+} from "./cloud-fallback";
+import {
+	type VisionVastFallbackOptions,
+	wrapImageDescriptionHandlerWithVastFallback,
+} from "./vast-fallback";
+
+/**
+ * Compose the full local → cloud → vast IMAGE_DESCRIPTION chain and
+ * terminate it as a runtime-shaped `ImageDescriptionHandler`. When all
+ * three paths return `{ kind: "fallback" }`, the terminator throws the
+ * underlying cause (or a structured upstream-fail message) so the runtime
+ * surfaces the failure cleanly rather than serving a sentinel result.
+ *
+ * This is the single entry point `ensure-local-inference-handler.ts`
+ * uses at the IMAGE_DESCRIPTION model registration site. Tests
+ * exercise the composition via the individual `wrap*` helpers; this
+ * function is the production wiring.
+ */
+export function withVisionFallbackChain(
+	local: LocalImageDescriptionHandler,
+	options: {
+		cloud?: VisionCloudFallbackOptions;
+		vast?: VisionVastFallbackOptions;
+	} = {},
+): (
+	runtime: IAgentRuntime,
+	params: ImageDescriptionParams | string,
+) => Promise<ImageDescriptionResult> {
+	const wrapped = wrapImageDescriptionHandlerWithVastFallback(
+		wrapImageDescriptionHandlerWithCloudFallback(local, options.cloud),
+		options.vast,
+	);
+	return async (_runtime, params) => {
+		const outcome = await wrapped(params);
+		if (
+			outcome &&
+			typeof outcome === "object" &&
+			"kind" in outcome &&
+			outcome.kind === "fallback"
+		) {
+			const causeMsg = outcome.cause?.message ?? outcome.reason;
+			const err = new Error(
+				`[VisionFallback] all IMAGE_DESCRIPTION providers exhausted (reason=${outcome.reason}): ${causeMsg}`,
+			);
+			if (outcome.cause) {
+				(err as Error & { cause?: unknown }).cause = outcome.cause;
+			}
+			throw err;
+		}
+		return outcome as ImageDescriptionResult;
+	};
+}

package/src/services/vision/llama-server.ts

@@ -0,0 +1,177 @@
+/**
+ * llama-server vision-describe backend (WS2).
+ *
+ * Wraps the out-of-process llama-server's `/completion` endpoint with
+ * the `image_data` array (base64-encoded payloads) and shapes the
+ * response to the WS2 `VisionDescribeBackend` contract.
+ *
+ * llama-server image-data API recap (verified against llama.cpp commit
+ * b8198+, May 2026):
+ *
+ *   POST /completion
+ *   { "prompt": "<...>USER: [img-12] What's in this image?\nASSISTANT:",
+ *     "image_data": [
+ *       { "data": "<base64 png/jpeg>", "id": 12 }
+ *     ],
+ *     "n_predict": 256,
+ *     "temperature": 0.2,
+ *     "stream": false }
+ *
+ *   Response:
+ *   { "content": "A photo of a cat.", "stop": true,
+ *     "timings": { "prompt_ms": 180.4, "predicted_ms": 423.1 } }
+ *
+ * Server-side mmproj is loaded via the `--mmproj <path>` flag on
+ * llama-server startup. The FFI runtime wrapper passes this flag
+ * already for tiers with vision enabled; this backend assumes the
+ * server has been started with the right mmproj for the active model.
+ *
+ * Backend responsibility:
+ *   - Encode the image as base64 (when not already).
+ *   - Build the prompt with the `[img-N]` placeholder convention.
+ *   - POST to `/completion`, parse the text + timings.
+ *   - Honour AbortSignal by passing it through to the fetch call.
+ *
+ * Backend explicitly does NOT:
+ *   - Start / stop the server. That's the FFI runtime wrapper's job.
+ *   - Resolve the mmproj path — the server already has it. The arbiter's
+ *     `--mmproj` was set when the text model loaded.
+ *   - Implement projector-token reuse. llama-server has no API to
+ *     accept pre-projected tokens; if the WS1 cache hit happens, this
+ *     backend ignores the hint and re-runs the projector. The cache
+ *     is more useful with the in-process node-llama-cpp backend.
+ *
+ * Metal / CUDA validation:
+ *   The llama-server build embeds the same mtmd_encode path the
+ *   in-process binding will eventually expose. On a Metal build the
+ *   image encode dispatches through the Metal compute encoder; on a
+ *   CUDA build through cuBLAS. We have no GPU on this host — see the
+ *   `__tests__/vision-describe.test.ts` notes for the GPU smoke check.
+ */
+
+import { resolveImageBytes } from "./hash";
+import type {
+	VisionDescribeBackend,
+	VisionDescribeRequest,
+	VisionDescribeResult,
+} from "./types";
+
+export interface LlamaServerVisionBackendOptions {
+	/**
+	 * Base URL of the llama-server. The FFI runtime wrapper exposes
+	 * this via `currentBaseUrl()`; pass the resolved URL here at load
+	 * time. The backend keeps it as-is across calls.
+	 */
+	baseUrl: string;
+	/**
+	 * Optional fetch override. Tests inject a fake fetch; production
+	 * uses global fetch. The signature mirrors `fetch` so the test
+	 * surface is the same as the real one.
+	 */
+	fetch?: typeof fetch;
+	/**
+	 * Default `n_predict` budget when the caller doesn't specify
+	 * `maxTokens`. 256 matches the description-length budget the
+	 * Florence-2 / VisionManager path uses today.
+	 */
+	defaultMaxTokens?: number;
+}
+
+export function createLlamaServerVisionBackend(
+	opts: LlamaServerVisionBackendOptions,
+): VisionDescribeBackend {
+	const fetchImpl = opts.fetch ?? globalThis.fetch;
+	const defaultMaxTokens = opts.defaultMaxTokens ?? 256;
+	let baseUrl = opts.baseUrl.replace(/\/$/, "");
+
+	if (!baseUrl) {
+		throw new Error(
+			"[vision/llama-server] baseUrl is required; pass FFI runtime's currentBaseUrl()",
+		);
+	}
+
+	return {
+		id: "llama-server",
+		async describe(
+			request: VisionDescribeRequest,
+		): Promise<VisionDescribeResult> {
+			const { bytes, mimeType } = resolveImageBytes(request.image);
+			const base64 = Buffer.from(bytes).toString("base64");
+			const prompt = buildVisionPrompt(request.prompt);
+			const startMs = Date.now();
+			const body = JSON.stringify({
+				prompt,
+				image_data: [{ data: base64, id: 12 }],
+				n_predict: request.maxTokens ?? defaultMaxTokens,
+				temperature: request.temperature ?? 0.2,
+				stream: false,
+				// `cache_prompt: false` here so each describe call gets a
+				// fresh slot; the WS1 vision-embedding cache handles repeat-
+				// frame reuse on the JS side, and the server-side prompt
+				// cache would only conflict with that (different KV state
+				// for the same projector tokens).
+				cache_prompt: false,
+			});
+			const res = await fetchImpl(`${baseUrl}/completion`, {
+				method: "POST",
+				headers: {
+					"content-type": "application/json",
+					...(mimeType ? { "x-image-mime": mimeType } : {}),
+				},
+				body,
+				signal: request.signal,
+			});
+			if (!res.ok) {
+				const text = await res.text().catch(() => "<unreadable>");
+				throw new Error(
+					`[vision/llama-server] /completion returned ${res.status}: ${text.slice(0, 200)}`,
+				);
+			}
+			const payload = (await res.json()) as {
+				content?: unknown;
+				timings?: { prompt_ms?: number; predicted_ms?: number };
+			};
+			if (typeof payload.content !== "string") {
+				throw new Error(
+					"[vision/llama-server] /completion response missing string `content`",
+				);
+			}
+			const elapsed = Date.now() - startMs;
+			return shape(payload.content, {
+				projectorMs: payload.timings?.prompt_ms,
+				decodeMs: payload.timings?.predicted_ms ?? elapsed,
+				cacheHit: false,
+			});
+		},
+		async dispose() {
+			// llama-server lifetime is owned by the FFI runtime wrapper.
+			// This backend just holds the baseUrl; nothing to free.
+			baseUrl = "";
+		},
+	};
+}
+
+function buildVisionPrompt(userPrompt?: string): string {
+	const ask = userPrompt?.trim() || "Describe what is in this image.";
+	// `[img-N]` is the placeholder llama-server's mtmd path replaces with
+	// the encoded image tokens. The `N` must match the `image_data[*].id`
+	// we send in the body; we use 12 because llama-server's stock
+	// example uses small integer ids — any positive integer works.
+	return `<start_of_turn>user\n[img-12]\n${ask}<end_of_turn>\n<start_of_turn>model\n`;
+}
+
+function shape(
+	text: string,
+	telemetry: { projectorMs?: number; decodeMs?: number; cacheHit?: boolean },
+): VisionDescribeResult {
+	const trimmed = text.trim();
+	if (!trimmed) {
+		throw new Error("[vision/llama-server] empty text from /completion");
+	}
+	const title = trimmed.split(/[.!?]/, 1)[0]?.trim() || "Image";
+	return {
+		title,
+		description: trimmed,
+		...telemetry,
+	};
+}

package/src/services/vision/types.ts

@@ -0,0 +1,163 @@
+/**
+ * Vision-language describe-image types (WS2).
+ *
+ * Two layers live here:
+ *
+ *   1. The **request/result** contract every WS2 backend implements
+ *      (`VisionDescribeRequest`, `VisionDescribeResult`). Callers pass
+ *      raw image bytes + a prompt; backends return a title+description.
+ *
+ *   2. The **backend** interface (`VisionDescribeBackend`) that the
+ *      `MemoryArbiter` registers as a capability handler. One backend
+ *      per binding family (node-llama-cpp / llama-server / AOSP libllama
+ *      shim). All three implement the same `load → describe → unload`
+ *      shape so the arbiter can swap between them without caring how
+ *      the projector is wired underneath.
+ *
+ * Why a separate file: the arbiter's `CapabilityRegistration<TBackend,
+ * TRequest, TResult>` is generic; pinning concrete shapes here keeps
+ * the registration sites short and removes a dozen casts at the
+ * call-site.
+ */
+
+/**
+ * Channel order for the raw pixel buffer. Most platforms hand us RGBA
+ * (HTMLCanvasElement, Capacitor `Camera`, the desktop `puppeteer-core`
+ * screenshot pipeline). The encoder normalizes internally; this enum
+ * stays so the hash step can pick a stable byte layout that doesn't
+ * depend on the platform-provided buffer order.
+ */
+export type VisionImageChannelOrder = "rgba" | "rgb" | "bgra" | "bgr";
+
+/**
+ * The raw image data the backend will encode. The arbiter does not see
+ * this — it gets handed straight to the backend's `run()`. The reason
+ * we accept multiple wrappers (URL / base64 / bytes) is that the three
+ * upstream entry points (HTTP route, agent runtime model handler,
+ * computer-use frame loop) each prefer a different shape. The backend
+ * resolves to bytes once.
+ */
+export type VisionImageInput =
+	| { kind: "bytes"; bytes: Uint8Array; mimeType?: string }
+	| { kind: "base64"; base64: string; mimeType?: string }
+	| { kind: "dataUrl"; dataUrl: string }
+	| { kind: "url"; url: string; mimeType?: string };
+
+/**
+ * Caller request to `describeImage`. The `modelFamily` distinguishes
+ * projected-token cache entries from different VL families that share
+ * the same hash space — Gemma-VL tokens are not interchangeable with
+ * Florence-2 tokens. Default is `gemma-vl` (the WS2 deliverable);
+ * each additional family registers under its own identifier.
+ */
+export interface VisionDescribeRequest {
+	image: VisionImageInput;
+	prompt?: string;
+	/**
+	 * The model family identifier. Used to namespace the projector cache
+	 * so swapping the backend's model family invalidates cached tokens.
+	 * Defaults to `"gemma-vl"` when omitted.
+	 */
+	modelFamily?: string;
+	/** Max output tokens; defaults to 256 (description-length budget). */
+	maxTokens?: number;
+	/** 0..1, default 0.2 (descriptions should be deterministic-ish). */
+	temperature?: number;
+	signal?: AbortSignal;
+	/**
+	 * Per-token callback. When set and the backend exposes streaming vision
+	 * (the fused ABI-v13 path), the description is decoded token-by-token and
+	 * each piece is delivered here as it generates — the same pipe as chat text.
+	 * Backends without streaming describe ignore it and return the final result.
+	 */
+	onTextChunk?: (chunk: string) => void | Promise<void>;
+	/** Per-step token cap for streaming describe (smaller = finer-grained UI). */
+	maxTokensPerStep?: number;
+}
+
+/** Backend response — same shape that ImageDescriptionResult expects. */
+export interface VisionDescribeResult {
+	title: string;
+	description: string;
+	/** Best-effort: ms spent in the projector (for arbiter telemetry). */
+	projectorMs?: number;
+	/** Best-effort: ms spent in the decoder. */
+	decodeMs?: number;
+	/** Whether the projected tokens came from the WS1 vision cache. */
+	cacheHit?: boolean;
+}
+
+/**
+ * Per-load arguments for a vision-describe backend. The arbiter's
+ * `load(modelKey)` only carries an opaque key; the binding resolves
+ * that key to real model+mmproj paths through this struct, which
+ * `createVisionCapabilityRegistration` populates from the catalog.
+ */
+export interface VisionDescribeLoadArgs {
+	/** Absolute path to the text decoder GGUF (the "main" model). */
+	modelPath: string;
+	/** Absolute path to the matching mmproj projector GGUF. */
+	mmprojPath: string;
+	/**
+	 * GPU offload preference. The backend translates this to its native
+	 * knob: node-llama-cpp `gpuLayers`, llama-server `--n-gpu-layers`,
+	 * AOSP libllama shim `eliza_llama_model_params_set_n_gpu_layers`.
+	 * `"auto"` lets the binding decide; numeric is honoured verbatim.
+	 */
+	gpuLayers?: number | "auto" | "max";
+	/** Max sampled context window in tokens. Defaults to 4096. */
+	contextSize?: number;
+}
+
+/**
+ * The contract every WS2 backend implements. The shape is intentionally
+ * narrow: the arbiter only ever calls `describe`. `dispose` is wrapped
+ * by the arbiter's `unload` so the backend can free GPU/VRAM and drop
+ * file descriptors on eviction.
+ */
+export interface VisionDescribeBackend {
+	/** Stable identifier — `"capacitor-llama"`, `"llama-server"`, `"aosp"`, or `"fake"` (tests). */
+	readonly id: "capacitor-llama" | "llama-server" | "aosp" | "fake";
+	/**
+	 * Run a describe pass. Backends MAY consult an injected projector cache
+	 * via `args.projectedTokens` (when the caller's hash already produced
+	 * a cache hit) instead of running the projector again; backends that
+	 * don't implement projector-token reuse ignore the field.
+	 */
+	describe(
+		request: VisionDescribeRequest,
+		args?: VisionDescribeBackendOptions,
+	): Promise<VisionDescribeResult>;
+	/** Release the loaded weights. Idempotent. */
+	dispose(): Promise<void>;
+}
+
+/**
+ * Per-call options the arbiter wrapper passes into the backend. Lives
+ * here (rather than on `VisionDescribeRequest`) so the caller-facing
+ * request type stays free of arbiter implementation details.
+ */
+export interface VisionDescribeBackendOptions {
+	/**
+	 * Pre-computed projected tokens from the WS1 vision-embedding cache.
+	 * When present the backend SHOULD skip its own projector step and
+	 * decode against these tokens directly. Backends that can't do this
+	 * still produce a correct result by ignoring the field; the arbiter's
+	 * wrapper will measure `cacheHit: false` in that case.
+	 */
+	projectedTokens?: {
+		tokens: Float32Array;
+		tokenCount: number;
+		hiddenSize: number;
+	};
+}
+
+/**
+ * Capability handler load function. The arbiter calls it with a model
+ * key (e.g. `"gemma-vl-4b"`); the implementation resolves to a real
+ * `(modelPath, mmprojPath)` pair from the catalog + installed registry
+ * and returns a live backend.
+ */
+export type VisionDescribeBackendLoader = (
+	modelKey: string,
+) => Promise<VisionDescribeBackend>;