forked from Blendman974/faceclaw
Support Apple Intelligence as an on-device AI model for the voice assistant/etc
This commit is contained in:
1 parent
3c3aa31542
commit
4973fd2aea
22 files changed
+1551
-40
No files matched your search
+4
-6
@@ -6,12 +6,14 @@ import type {
|
||||
LlmStreamOptions,
|
||||
} from "../assistant/llm-protocol";
|
||||
import { buildQwenPrompt } from "../assistant/qwen-prompt";
|
||||
import type { LocalModelState } from "./local-llm-types";
|
||||
|
||||
declare const com: any;
|
||||
declare const java: any;
|
||||
|
||||
/**
|
||||
* On-phone LLM backend: model download management plus a streaming client
|
||||
* Android's on-phone LLM backend (reached through local-llm.ts; iOS uses
|
||||
* local-llm.ios.ts instead): model download management plus a streaming client
|
||||
* with the same LlmStreamOptions surface as the Anthropic/OpenAI clients, so
|
||||
* the direct assistant loop can treat "local" as just another provider.
|
||||
* Inference is llama.cpp on the CPU (FaceclawLlamaRunner over JNI).
|
||||
@@ -38,11 +40,7 @@ const TOP_K = 20;
|
||||
/** Free the ~3GB of model + KV this long after the last generation. */
|
||||
const IDLE_UNLOAD_MS = 2 * 60 * 1000;
|
||||
|
||||
export type LocalModelState = {
|
||||
status: "absent" | "downloading" | "ready";
|
||||
bytesDownloaded: number;
|
||||
totalBytes: number;
|
||||
};
|
||||
export type { LocalModelState } from "./local-llm-types";
|
||||
|
||||
let runner: any = null;
|
||||
let downloader: any = null;
|
||||
|
||||
Reference in new issue
Block a user