@elizaos/plugin-local-inference 2.0.0-beta.1 → 2.0.11-beta.7

package/src/services/ffi-streaming-runner.ts

@@ -0,0 +1,354 @@
+/**
+ * In-process streaming-LLM runner.
+ *
+ * FFI streaming-LLM ABI declared in `ffi-streaming-llm.h`. The
+ * token-by-token loop hands `onTextChunk` accepted chunks and surfaces
+ * verifier events from native MTP.
+ *
+ * This file deliberately does not own the FFI context or the binding
+ * itself. It takes a narrow `LlmStreamingBinding` (see
+ * `services/llm-streaming-binding.ts`) + an opaque `LlmCtxHandle` as
+ * constructor arguments — that way it can be driven by libelizainference
+ * (via `wrapElizaInferenceFfi`) or any desktop libllama shim adapter without
+ * dragging in TTS/ASR surfaces. A single context can host concurrent generation
+ * sessions (one per pinned slot); the runner serialises with
+ * `slotInFlight`.
+ *
+ * Single-flight: lock map keyed by slot id, slot id `-1` unlocked. Two concurrent generates
+ * against the same pinned slot would interleave KV cache state, so the
+ * runner serializes them at the JS layer.
+ */
+
+import { performance } from "node:perf_hooks";
+
+import type {
+	LlmCtxHandle,
+	LlmStreamingBinding,
+} from "./llm-streaming-binding";
+import type { LlmStreamHandle, LlmStreamStep } from "./voice/ffi-bindings";
+import type { TextToken, VerifierStreamEvent } from "./voice/types";
+
+export interface FfiStreamingGenerateArgs {
+	/** Pre-tokenized prompt — the runner does not detokenize. */
+	promptTokens: Int32Array;
+	/** Pinned slot id; -1 disables pinning (any free slot). */
+	slotId: number;
+	/** Optional prompt cache key used to derive a slot when `slotId === -1`. */
+	cacheKey?: string;
+	maxTokens: number;
+	temperature: number;
+	topP: number;
+	topK: number;
+	repeatPenalty: number;
+	draftMin: number;
+	draftMax: number;
+	/** Reserved for separate draft-model speculation; null for Eliza-1 MTP. */
+	draftModelPath: string | null;
+	/**
+	 * Per-load GPU offload (ABI v8). Forwarded into the native session config
+	 * on `llmStreamOpen`. The fused libelizainference path loads the text model
+	 * once per ctx, so the FIRST session's value wins; later sessions reuse the
+	 * resident model. `undefined` selects the runtime default (all layers).
+	 * The desktop libllama path already applies gpuLayers at `loadModel()`, so
+	 * it ignores this field — it is load-time config, threaded here only so the
+	 * fused runner can mirror the libllama load decision.
+	 */
+	gpuLayers?: number;
+	/**
+	 * KV-cache K/V quant type names (ABI v8), e.g. "qjl1_256" / "q4_polar".
+	 * Same load-time semantics as `gpuLayers`: forwarded into the fused
+	 * session config so the first `llmStreamOpen` applies the quantized cache.
+	 */
+	cacheTypeK?: string | null;
+	cacheTypeV?: string | null;
+	/**
+	 * GBNF grammar source forcing the structured-reply envelope. Passed to
+	 * the native session's `llmStreamOpen` config so sampling is
+	 * grammar-constrained. `null` disables the constraint (free generation).
+	 */
+	gbnfGrammar?: string | null;
+	/** Cancellation signal — fires `llmStreamCancel` on the active session. */
+	signal?: AbortSignal;
+	/** Per-chunk text callback. */
+	onTextChunk?: (chunk: string) => void | Promise<void>;
+	/** Speculative accept/reject events from MTP verification. */
+	onVerifierEvent?: (event: VerifierStreamEvent) => void | Promise<void>;
+}
+
+export interface FfiStreamingGenerateResult {
+	text: string;
+	slotId: number;
+	firstTokenMs: number | null;
+	drafted: number;
+	accepted: number;
+}
+
+/** Default per-step caps. Match upstream llama-server's `n_predict` chunk size. */
+const DEFAULT_MAX_TOKENS_PER_STEP = 32;
+const DEFAULT_MAX_TEXT_BYTES = 1024;
+
+/**
+ * Backend used by the mobile and desktop FFI routes.
+ */
+export class FfiStreamingRunner {
+	private readonly slotInFlight = new Map<number, Promise<void>>();
+
+	/**
+	 * Constructor takes the narrow `LlmStreamingBinding` (see
+	 * `services/llm-streaming-binding.ts`) so both libelizainference (via
+	 * `wrapElizaInferenceFfi`) and desktop libllama adapters can
+	 * satisfy it. The runner never touches TTS/ASR/mmap surfaces.
+	 */
+	constructor(
+		private readonly ffi: LlmStreamingBinding,
+		private readonly ctx: LlmCtxHandle,
+	) {}
+
+	/**
+	 * Run one generation. Mirrors `MtpLlamaServer.generateWithUsage()`
+	 * — same single-flight rule, same callback shape, same result block
+	 * minus the metrics scrape (FFI does not have a `/metrics` endpoint).
+	 */
+	async generateWithUsage(
+		args: FfiStreamingGenerateArgs,
+	): Promise<FfiStreamingGenerateResult> {
+		if (args.slotId < 0) {
+			return this.runGenerate(args);
+		}
+		const prior = this.slotInFlight.get(args.slotId);
+		const run = (prior ?? Promise.resolve())
+			.catch(() => {})
+			.then(() => this.runGenerate(args));
+		const tail = run.then(
+			() => {},
+			() => {},
+		);
+		this.slotInFlight.set(args.slotId, tail);
+		try {
+			return await run;
+		} finally {
+			if (this.slotInFlight.get(args.slotId) === tail) {
+				this.slotInFlight.delete(args.slotId);
+			}
+		}
+	}
+
+	/**
+	 * Async-iterable variant. Yields each accepted-token batch as it lands
+	 * so callers that want token-grained control (e.g. the voice scheduler
+	 * driving phrase-chunking off accept/reject events) don't have to
+	 * register a callback. Internally still routes through `generateWithUsage`
+	 * via a pump so the single-flight rule applies.
+	 */
+	async *generateStream(
+		args: FfiStreamingGenerateArgs,
+	): AsyncIterable<LlmStreamStep> {
+		// Queue accumulates steps the inner callback produces; the iterator
+		// drains it. Using a plain array + resolver is simpler than wiring a
+		// real async-queue for the single-consumer case here.
+		const queue: LlmStreamStep[] = [];
+		let resume: (() => void) | null = null;
+		let finished = false;
+		let failure: Error | null = null;
+
+		const wakeConsumer = () => {
+			const wake = resume;
+			resume = null;
+			if (wake) wake();
+		};
+
+		const onStep = (step: LlmStreamStep) => {
+			queue.push(step);
+			wakeConsumer();
+		};
+
+		const work = (async () => {
+			try {
+				await this.runGenerateInner(args, onStep);
+			} catch (err) {
+				failure = err instanceof Error ? err : new Error(String(err));
+			} finally {
+				finished = true;
+				wakeConsumer();
+			}
+		})();
+
+		try {
+			while (true) {
+				if (queue.length > 0) {
+					const next = queue.shift();
+					if (next === undefined) continue;
+					yield next;
+					if (next.done) return;
+					continue;
+				}
+				if (failure) throw failure;
+				if (finished) return;
+				await new Promise<void>((resolve) => {
+					resume = resolve;
+				});
+			}
+		} finally {
+			await work;
+		}
+	}
+
+	/**
+	 * Save the streaming slot KV state to disk. Best called between turns
+	 * — calling mid-stream is racy and the FFI side is allowed to refuse.
+	 * Surfaced here so the conversation registry can persist between
+	 * mobile backgrounds the same way `MtpLlamaServer.persistSlot` does.
+	 */
+	saveSlot(stream: LlmStreamHandle, filename: string): void {
+		if (this.ffi.llmStreamSaveSlot === undefined) {
+			throw new Error(
+				"[ffi-streaming-runner] llmStreamSaveSlot is not exported by this build",
+			);
+		}
+		this.ffi.llmStreamSaveSlot({ stream, filename });
+	}
+
+	/** Restore a previously-saved slot KV file into a fresh session. */
+	restoreSlot(stream: LlmStreamHandle, filename: string): void {
+		if (this.ffi.llmStreamRestoreSlot === undefined) {
+			throw new Error(
+				"[ffi-streaming-runner] llmStreamRestoreSlot is not exported by this build",
+			);
+		}
+		this.ffi.llmStreamRestoreSlot({ stream, filename });
+	}
+
+	/* ----- internals -------------------------------------------------- */
+
+	private async runGenerate(
+		args: FfiStreamingGenerateArgs,
+	): Promise<FfiStreamingGenerateResult> {
+		const aggregated: string[] = [];
+		let totalDrafted = 0;
+		let totalAccepted = 0;
+		let firstTokenMs: number | null = null;
+		const startedAt = performance.now();
+
+		await this.runGenerateInner(args, (step) => {
+			if (step.text.length > 0 && firstTokenMs === null) {
+				firstTokenMs = performance.now() - startedAt;
+			}
+			aggregated.push(step.text);
+			totalDrafted += step.drafterDrafted;
+			totalAccepted += step.drafterAccepted;
+		});
+
+		return {
+			text: aggregated.join(""),
+			slotId: args.slotId,
+			firstTokenMs,
+			drafted: totalDrafted,
+			accepted: totalAccepted,
+		};
+	}
+
+	/**
+	 * Shared inner loop. Opens the session, runs the prefill + next pump,
+	 * forwards each step through `onStep` plus the optional caller
+	 * callbacks, and wires abort + cancel.
+	 */
+	private async runGenerateInner(
+		args: FfiStreamingGenerateArgs,
+		onStep: (step: LlmStreamStep) => void,
+	): Promise<void> {
+		if (
+			this.ffi.llmStreamOpen === undefined ||
+			this.ffi.llmStreamPrefill === undefined ||
+			this.ffi.llmStreamNext === undefined ||
+			this.ffi.llmStreamClose === undefined
+		) {
+			throw new Error(
+				"[ffi-streaming-runner] libelizainference is missing streaming-LLM symbols. " +
+					"Rebuild against the current eliza-inference-ffi.h.",
+			);
+		}
+
+		const stream = this.ffi.llmStreamOpen({
+			ctx: this.ctx,
+			config: {
+				maxTokens: args.maxTokens,
+				temperature: args.temperature,
+				topP: args.topP,
+				topK: args.topK,
+				repeatPenalty: args.repeatPenalty,
+				slotId: args.slotId,
+				promptCacheKey: args.cacheKey ?? null,
+				draftMin: args.draftMin,
+				draftMax: args.draftMax,
+				draftModelPath: args.draftModelPath,
+				gbnfGrammar: args.gbnfGrammar ?? null,
+				gpuLayers: args.gpuLayers,
+				cacheTypeK: args.cacheTypeK,
+				cacheTypeV: args.cacheTypeV,
+			},
+		});
+
+		let abortListener: (() => void) | null = null;
+		if (args.signal) {
+			if (args.signal.aborted) {
+				this.ffi.llmStreamCancel?.(stream);
+				this.ffi.llmStreamClose(stream);
+				throw new Error("[ffi-streaming-runner] aborted before start");
+			}
+			abortListener = () => {
+				this.ffi.llmStreamCancel?.(stream);
+			};
+			args.signal.addEventListener("abort", abortListener, { once: true });
+		}
+
+		try {
+			this.ffi.llmStreamPrefill({ stream, tokens: args.promptTokens });
+			if (args.maxTokens <= 0) {
+				return;
+			}
+
+			let tokenIndex = 0;
+			while (true) {
+				if (args.signal?.aborted) {
+					this.ffi.llmStreamCancel?.(stream);
+					throw new Error("[ffi-streaming-runner] aborted");
+				}
+				const step = this.ffi.llmStreamNext({
+					stream,
+					maxTokensPerStep: DEFAULT_MAX_TOKENS_PER_STEP,
+					maxTextBytes: DEFAULT_MAX_TEXT_BYTES,
+				});
+				onStep(step);
+
+				if (args.onTextChunk && step.text.length > 0) {
+					await args.onTextChunk(step.text);
+				}
+				if (args.onVerifierEvent) {
+					const tokens: TextToken[] = step.tokens.map((id, i) => ({
+						index: tokenIndex + i,
+						text: i === 0 ? step.text : "",
+						id,
+					}));
+					// The FFI ABI commits accepted tokens per step (the drafter
+					// accept/reject decomposition is delivered through the
+					// separate `setVerifierCallback` channel — see ffi.h §v2).
+					// Surface the batched accept here so HTTP-path callers see a
+					// matching event shape.
+					if (tokens.length > 0) {
+						await args.onVerifierEvent({
+							kind: "accept",
+							tokens,
+						});
+					}
+				}
+				tokenIndex += step.tokens.length;
+				if (step.done) break;
+			}
+		} finally {
+			if (abortListener && args.signal) {
+				args.signal.removeEventListener("abort", abortListener);
+			}
+			this.ffi.llmStreamClose(stream);
+		}
+	}
+}

package/src/services/ffi-unload-ordering.test.ts

@@ -0,0 +1,162 @@
+import { fileURLToPath } from "node:url";
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+import type { BackendPlan } from "./backend";
+import {
+	type FfiBackendRuntime,
+	type FfiBackendSession,
+	FfiStreamingBackend,
+} from "./ffi-streaming-backend";
+
+// Hoisted spy shared between the mock factory (hoisted) and the test so we can
+// assert the fused lib's native free was attempted exactly once.
+const ffiCloseMock = vi.hoisted(() =>
+	vi.fn(() => {
+		throw new Error("ov_free segfault surrogate");
+	}),
+);
+
+// Replace the fused-lib FFI loader so importing the runtime never pulls bun:ffi
+// or dlopens a native library. `loadElizaInferenceFfi` is the only value the
+// desktop fused runtime imports from the bindings module; the fake exposes the
+// v9 surface the runtime touches during acquire()/release().
+vi.mock("./voice/ffi-bindings", () => ({
+	loadElizaInferenceFfi: vi.fn(() => ({
+		create: () => 1n,
+		destroy: vi.fn(),
+		close: ffiCloseMock,
+		tokenizeSupported: () => true,
+		tokenize: () => new Int32Array(),
+		llmStreamSupported: () => true,
+		llmStreamOpen: () => 0n,
+		llmStreamPrefill: () => 0,
+		llmStreamNext: () => 0,
+		llmStreamCancel: () => 0,
+		llmStreamClose: () => undefined,
+	})),
+}));
+
+/**
+ * Tests for #14: unload() must await the native release BEFORE nulling the
+ * session refs, otherwise a throwing release leaves the backend wedged —
+ * session === null while the runtime still holds a live session, so the next
+ * load() skips unload(), calls acquire(), and acquire()'s live-session guard
+ * throws forever.
+ */
+
+const PLAN: BackendPlan = {
+	modelPath: "/fake/model.gguf",
+} as unknown as BackendPlan;
+
+function fakeSession(): FfiBackendSession {
+	return {
+		binding: {} as never,
+		ctx: {} as never,
+		runner: {} as never,
+		tokenize: () => new Int32Array(),
+		mtp: null,
+		draftModelPath: null,
+		mmprojPath: null,
+	};
+}
+
+/**
+ * Minimal runtime that mirrors the real acquire/release live-session guard:
+ * acquire() throws if a session is already live (exactly like
+ * DesktopFusedFfiBackendRuntime). release() can be made to throw to simulate a
+ * native bun:ffi free rejecting.
+ */
+class GuardedRuntime implements FfiBackendRuntime {
+	private active = false;
+	releaseShouldThrow = false;
+	releaseCalls = 0;
+
+	supported(): boolean {
+		return true;
+	}
+
+	async acquire(): Promise<FfiBackendSession> {
+		if (this.active) {
+			throw new Error("acquire() called with a live session; release() first");
+		}
+		this.active = true;
+		return fakeSession();
+	}
+
+	async release(): Promise<void> {
+		this.releaseCalls += 1;
+		if (this.releaseShouldThrow) {
+			// The runtime still has a live session — a real release that throws
+			// mid-free leaves `active` set (the runtime's own finally is what
+			// clears it; here we model the throw-before-clear case).
+			throw new Error("native free rejected");
+		}
+		this.active = false;
+	}
+}
+
+describe("FfiStreamingBackend.unload() ordering (#14)", () => {
+	it("nulls session refs even when release() throws", async () => {
+		const runtime = new GuardedRuntime();
+		const backend = new FfiStreamingBackend(runtime);
+		await backend.load(PLAN);
+		expect(backend.hasLoadedModel()).toBe(true);
+
+		runtime.releaseShouldThrow = true;
+		await expect(backend.unload()).rejects.toThrow("native free rejected");
+
+		// The finally must have cleared our refs despite the throw, so the
+		// backend doesn't report a phantom loaded model.
+		expect(backend.hasLoadedModel()).toBe(false);
+		expect(backend.currentModelPath()).toBeNull();
+	});
+
+	it("awaits release before nulling refs (release observed first)", async () => {
+		const order: string[] = [];
+		const runtime: FfiBackendRuntime = {
+			supported: () => true,
+			acquire: async () => fakeSession(),
+			release: vi.fn(async () => {
+				order.push("release");
+			}),
+		};
+		const backend = new FfiStreamingBackend(runtime);
+		await backend.load(PLAN);
+		await backend.unload();
+		// hasLoadedModel reads session, which is nulled only after release.
+		order.push(backend.hasLoadedModel() ? "still-loaded" : "cleared");
+		expect(order).toEqual(["release", "cleared"]);
+	});
+});
+
+describe("DesktopFusedFfiBackendRuntime.release() ordering (#14)", () => {
+	beforeEach(() => {
+		// resolveFusedLibraryPath() returns the first existing candidate; point it
+		// at a real file so acquire() resolves a lib path (the FFI loader itself is
+		// mocked, so the path's contents are irrelevant).
+		process.env.ELIZA_INFERENCE_LIBRARY = fileURLToPath(import.meta.url);
+		ffiCloseMock.mockClear();
+	});
+
+	afterEach(() => {
+		process.env.ELIZA_INFERENCE_LIBRARY = undefined;
+	});
+
+	it("clears the active session even when the fused close() throws", async () => {
+		const { DesktopFusedFfiBackendRuntime } = await import(
+			"./desktop-fused-ffi-backend-runtime"
+		);
+		const runtime = new DesktopFusedFfiBackendRuntime();
+		await runtime.acquire(PLAN);
+
+		// close() throws, but release() must still clear `active` via its finally.
+		await expect(runtime.release()).rejects.toThrow(
+			"ov_free segfault surrogate",
+		);
+		expect(ffiCloseMock).toHaveBeenCalledTimes(1);
+
+		// The runtime is not hidden-wedged on the old live-session guard, but it
+		// is explicitly poisoned so a new native model is not allocated over a
+		// failed cleanup state.
+		await expect(runtime.acquire(PLAN)).rejects.toThrow(/restart required/i);
+	});
+});