Create localmodel.mjs

core/localmodel.mjs

@@ -0,0 +1,54 @@
+/* Modelo local MiMo en el navegador (wllama 3.5.1, WASM/CPU).
+   El chat de este Space no depende de ningún servicio externo de modelos:
+   el GGUF se descarga desde Hugging Face (aprox. 5,8 GB la primera vez) y se
+   reutiliza desde la caché del navegador mientras esta se conserve. */
+const WLLAMA_ESM = 'https://cdn.jsdelivr.net/npm/@wllama/wllama@3.5.1/esm/index.js';
+const WLLAMA_WASM = 'https://cdn.jsdelivr.net/npm/@wllama/wllama@3.5.1/esm/wasm/wllama.wasm';
+const MODEL_URL = 'https://huggingface.co/bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF/resolve/main/MiMo-V2.6-Distill-Qwen-9B-Q4_K_M.gguf';
+export const APLocalModel = {
+  state: 'idle', // idle | downloading | ready | error
+  error: '',
+  wllama: null,
+  label: 'MiMo-V2.6-Distill-Qwen-9B (local en este navegador)',
+  async cleanupCache() {
+    try { const keys = await caches.keys(); for (const k of keys) await caches.delete(k); } catch (_) {}
+  },
+  async load(onProgress) {
+    if (this.state === 'ready' || this.state === 'downloading') return;
+    this.state = 'downloading'; this.error = '';
+    try {
+      const { Wllama } = await import(WLLAMA_ESM);
+      if (this.wllama) { try { await this.wllama.exit(); } catch (_) {} }
+      this.wllama = new Wllama({ default: WLLAMA_WASM }, { parallelDownloads: 5 });
+      const opts = {
+        n_gpu_layers: 0, // WASM/CPU: sin aceleración garantizada
+        n_ctx: 4096, n_batch: 512,
+        n_threads: Math.max(1, Math.min(8, navigator.hardwareConcurrency || 4)),
+        progressCallback: ({ loaded, total }) => { if (onProgress && total > 0) onProgress(loaded, total); },
+      };
+      try {
+        await this.wllama.loadModelFromUrl(MODEL_URL, opts);
+      } catch (first) {
+        // Caché incompleta o corrupta: limpiar y reintentar una vez.
+        await this.cleanupCache();
+        await this.wllama.loadModelFromUrl(MODEL_URL, opts);
+      }
+      this.state = 'ready';
+    } catch (e) {
+      this.state = 'error'; this.error = String(e && e.message || e).slice(0, 200);
+      throw e;
+    }
+  },
+  async chat(messages, { maxTokens = 640, temperature = 0.2, onToken } = {}) {
+    if (this.state !== 'ready') throw new Error('modelo local no cargado');
+    const stream = await this.wllama.createChatCompletion({ messages, max_tokens: maxTokens, temperature, stream: true });
+    let answer = '';
+    for await (const chunk of stream) {
+      const delta = (chunk.choices && chunk.choices[0] && chunk.choices[0].delta) || {};
+      const tk = delta.content || '';
+      if (tk) { answer += tk; if (onToken) onToken(answer); }
+    }
+    return answer;
+  },
+};
+globalThis.APLocalModel = APLocalModel;