wllama-service 2.0.3 → 2.0.5

package/dist/index.d.mts

@@ -41,8 +41,30 @@ declare class WllamaService {
     checkEnvironment(): ConnectionResult;
     /** Load a GGUF model file from the user's filesystem */
     loadModel(file: File, onProgress?: (progress: number) => void): Promise<LoadModelResult>;
-    /** Generate a response from the loaded model */
+    /**
+     * Generate a response using the model's chat template.
+     * Use this for normal conversational / instruction-following use cases —
+     * it accepts a `messages`-style request (system + user) and
+     * createChatCompletion formats it with the model's real chat template
+     * (ChatML, Llama-3, etc.) before running inference.
+     */
     generate(req: GenerateRequest): Promise<GenerateResult>;
+    /**
+     * Generate a raw completion with no chat-template formatting.
+     * Use this when you need to control the exact prompt string yourself —
+     * e.g. few-shot prompts, base-model completion, custom templates, or
+     * fill-in-the-middle style prompts.
+     *
+     * createCompletion (confirmed against @wllama/wllama@3.5.1's real type
+     * defs) takes a single options object with a raw `prompt` string. It does
+     * NOT accept a `messages` array and does NOT apply any chat template —
+     * whatever string you pass is exactly what the model sees. `req.system`
+     * (if provided) is prepended as plain text; it is your responsibility to
+     * format it however the target model expects (call
+     * `this.wllama.getChatTemplate()` if you want to match the model's
+     * native template manually).
+     */
+    generateCompletion(req: GenerateRequest): Promise<GenerateResult>;
     /** Unload the current model and free memory */
     unload(): Promise<void>;
     /** Whether a model is currently loaded */

package/dist/index.d.ts

@@ -41,8 +41,30 @@ declare class WllamaService {
     checkEnvironment(): ConnectionResult;
     /** Load a GGUF model file from the user's filesystem */
     loadModel(file: File, onProgress?: (progress: number) => void): Promise<LoadModelResult>;
-    /** Generate a response from the loaded model */
+    /**
+     * Generate a response using the model's chat template.
+     * Use this for normal conversational / instruction-following use cases —
+     * it accepts a `messages`-style request (system + user) and
+     * createChatCompletion formats it with the model's real chat template
+     * (ChatML, Llama-3, etc.) before running inference.
+     */
     generate(req: GenerateRequest): Promise<GenerateResult>;
+    /**
+     * Generate a raw completion with no chat-template formatting.
+     * Use this when you need to control the exact prompt string yourself —
+     * e.g. few-shot prompts, base-model completion, custom templates, or
+     * fill-in-the-middle style prompts.
+     *
+     * createCompletion (confirmed against @wllama/wllama@3.5.1's real type
+     * defs) takes a single options object with a raw `prompt` string. It does
+     * NOT accept a `messages` array and does NOT apply any chat template —
+     * whatever string you pass is exactly what the model sees. `req.system`
+     * (if provided) is prepended as plain text; it is your responsibility to
+     * format it however the target model expects (call
+     * `this.wllama.getChatTemplate()` if you want to match the model's
+     * native template manually).
+     */
+    generateCompletion(req: GenerateRequest): Promise<GenerateResult>;
     /** Unload the current model and free memory */
     unload(): Promise<void>;
     /** Whether a model is currently loaded */

package/dist/index.js

@@ -78,7 +78,13 @@ var WllamaService = class {
       return { success: false, error: message };
     }
   }
-  /** Generate a response from the loaded model */
+  /**
+   * Generate a response using the model's chat template.
+   * Use this for normal conversational / instruction-following use cases —
+   * it accepts a `messages`-style request (system + user) and
+   * createChatCompletion formats it with the model's real chat template
+   * (ChatML, Llama-3, etc.) before running inference.
+   */
   async generate(req) {
     if (!this.wllama) {
       return { success: false, error: "No model loaded. Call loadModel() first." };
@@ -90,6 +96,7 @@ var WllamaService = class {
       messages.push({ role: "user", content: req.prompt });
       const response = await this.wllama.createChatCompletion({
         messages,
+        stream: false,
         max_tokens: req.maxTokens ?? 512,
         temperature: req.temperature ?? 0.7,
         top_k: req.topK ?? 40,
@@ -106,6 +113,49 @@ var WllamaService = class {
       return { success: false, error: message };
     }
   }
+  /**
+   * Generate a raw completion with no chat-template formatting.
+   * Use this when you need to control the exact prompt string yourself —
+   * e.g. few-shot prompts, base-model completion, custom templates, or
+   * fill-in-the-middle style prompts.
+   *
+   * createCompletion (confirmed against @wllama/wllama@3.5.1's real type
+   * defs) takes a single options object with a raw `prompt` string. It does
+   * NOT accept a `messages` array and does NOT apply any chat template —
+   * whatever string you pass is exactly what the model sees. `req.system`
+   * (if provided) is prepended as plain text; it is your responsibility to
+   * format it however the target model expects (call
+   * `this.wllama.getChatTemplate()` if you want to match the model's
+   * native template manually).
+   */
+  async generateCompletion(req) {
+    if (!this.wllama) {
+      return { success: false, error: "No model loaded. Call loadModel() first." };
+    }
+    try {
+      const t0 = performance.now();
+      const prompt = req.system ? `${req.system}
+
+${req.prompt}` : req.prompt;
+      const response = await this.wllama.createCompletion({
+        prompt,
+        stream: false,
+        max_tokens: req.maxTokens ?? 512,
+        temperature: req.temperature ?? 0.7,
+        top_k: req.topK ?? 40,
+        top_p: req.topP ?? 0.95
+      });
+      const text = response?.choices?.[0]?.text ?? "";
+      return {
+        success: true,
+        text,
+        timeMs: Math.round(performance.now() - t0)
+      };
+    } catch (e) {
+      const message = e instanceof Error ? e.message : "Generation failed";
+      return { success: false, error: message };
+    }
+  }
   /** Unload the current model and free memory */
   async unload() {
     if (this.wllama) {

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/index.ts","../src/WllamaService.ts"],"sourcesContent":["export { WllamaService } from './WllamaService';\nexport type {\n  WllamaServiceConfig,\n  GenerateRequest,\n  GenerateResult,\n  LoadModelResult,\n  ConnectionResult,\n} from './types';","import { Wllama } from '@wllama/wllama/esm/index.js';\nimport type {\n  WllamaServiceConfig,\n  GenerateRequest,\n  GenerateResult,\n  LoadModelResult,\n  ConnectionResult,\n} from './types';\n\nexport class WllamaService {\n  private wllama: Wllama | null = null;\n  private modelName: string = '';\n  private config: Required<WllamaServiceConfig>;\n\n  constructor(config: WllamaServiceConfig = {}) {\n    this.config = {\n      wasmPath: config.wasmPath ?? '/wllama/wllama.wasm',\n      nGpuLayers: config.nGpuLayers ?? 999,\n      nCtx: config.nCtx ?? 2048,\n    };\n  }\n\n  /** Check browser compatibility and capabilities */\n  checkEnvironment(): ConnectionResult {\n    if (typeof WebAssembly === 'undefined') {\n      return { success: false, error: 'WebAssembly is not supported in this browser.' };\n    }\n    if (!('caches' in window)) {\n      return { success: false, error: 'Cache API is not supported in this browser.' };\n    }\n    const hasWebGPU = !!(navigator as any).gpu;\n    const isolated = window.crossOriginIsolated;\n    return {\n      success: true,\n      hasWebGPU,\n      crossOriginIsolated: isolated,\n    };\n  }\n\n  /** Load a GGUF model file from the user's filesystem */\n  async loadModel(\n    file: File,\n    onProgress?: (progress: number) => void,\n  ): Promise<LoadModelResult> {\n    const env = this.checkEnvironment();\n    if (!env.success) return { success: false, error: env.error };\n\n    await this.unload();\n\n    try {\n      onProgress?.(10);\n\n      // AssetsPathConfig requires `default` — this is the key the\n      // Wllama runtime actually reads (confirmed against the real\n      // @wllama/wllama@3.5.1 implementation, not just its docs).\n      // No `as any` needed: this shape is type-correct as written.\n      const configPaths = { default: this.config.wasmPath };\n      this.wllama = new Wllama(configPaths);\n      onProgress?.(30);\n\n      await this.storeInCache(file);\n      onProgress?.(50);\n\n      const hasWebGPU = !!(navigator as any).gpu;\n      await (this.wllama as any).loadModel([file], {\n        n_ctx: this.config.nCtx,\n        ...(hasWebGPU && this.config.nGpuLayers > 0\n          ? { n_gpu_layers: this.config.nGpuLayers }\n          : {}),\n      });\n\n      this.modelName = file.name;\n      onProgress?.(100);\n\n      return { success: true, usedWebGPU: hasWebGPU };\n    } catch (e: unknown) {\n      await this.unload();\n      const message = e instanceof Error ? e.message : 'Failed to load model';\n      return { success: false, error: message };\n    }\n  }\n\n  /** Generate a response from the loaded model */\n  async generate(req: GenerateRequest): Promise<GenerateResult> {\n    if (!this.wllama) {\n      return { success: false, error: 'No model loaded. Call loadModel() first.' };\n    }\n\n    try {\n      const t0 = performance.now();\n\n      const messages: { role: string; content: string }[] = [];\n      if (req.system) messages.push({ role: 'system', content: req.system });\n      messages.push({ role: 'user', content: req.prompt });\n\n      const response = await (this.wllama as any).createChatCompletion({\n        messages,\n        max_tokens: req.maxTokens ?? 512,\n        temperature: req.temperature ?? 0.7,\n        top_k: req.topK ?? 40,\n        top_p: req.topP ?? 0.95,\n      });\n\n      const text: string = response?.choices?.[0]?.message?.content ?? '';\n      return {\n        success: true,\n        text,\n        timeMs: Math.round(performance.now() - t0),\n      };\n    } catch (e: unknown) {\n      const message = e instanceof Error ? e.message : 'Generation failed';\n      return { success: false, error: message };\n    }\n  }\n\n  /** Unload the current model and free memory */\n  async unload(): Promise<void> {\n    if (this.wllama) {\n      try {\n        await (this.wllama as any).exit();\n      } catch (_) {\n        /* ignore */\n      }\n      this.wllama = null;\n    }\n    if (this.modelName) {\n      try {\n        await this.deleteFromCache(this.modelName);\n      } catch (_) {\n        /* ignore */\n      }\n      this.modelName = '';\n    }\n  }\n\n  /** Whether a model is currently loaded */\n  get isLoaded(): boolean {\n    return this.wllama !== null;\n  }\n\n  /** Name of the currently loaded model */\n  get currentModel(): string {\n    return this.modelName;\n  }\n\n  private async storeInCache(file: File): Promise<void> {\n    const url = `${window.location.origin}/wllama-local-models/${file.name}`;\n    const cache = await caches.open('wllama-local-models');\n    await cache.put(\n      url,\n      new Response(file, {\n        headers: { 'Content-Type': 'application/octet-stream' },\n      }),\n    );\n  }\n\n  private async deleteFromCache(fileName: string): Promise<void> {\n    const url = `${window.location.origin}/wllama-local-models/${fileName}`;\n    const cache = await caches.open('wllama-local-models');\n    await cache.delete(url);\n  }\n}"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACAA,iBAAuB;AAShB,IAAM,gBAAN,MAAoB;AAAA,EAKzB,YAAY,SAA8B,CAAC,GAAG;AAJ9C,SAAQ,SAAwB;AAChC,SAAQ,YAAoB;AAI1B,SAAK,SAAS;AAAA,MACZ,UAAU,OAAO,YAAY;AAAA,MAC7B,YAAY,OAAO,cAAc;AAAA,MACjC,MAAM,OAAO,QAAQ;AAAA,IACvB;AAAA,EACF;AAAA;AAAA,EAGA,mBAAqC;AACnC,QAAI,OAAO,gBAAgB,aAAa;AACtC,aAAO,EAAE,SAAS,OAAO,OAAO,gDAAgD;AAAA,IAClF;AACA,QAAI,EAAE,YAAY,SAAS;AACzB,aAAO,EAAE,SAAS,OAAO,OAAO,8CAA8C;AAAA,IAChF;AACA,UAAM,YAAY,CAAC,CAAE,UAAkB;AACvC,UAAM,WAAW,OAAO;AACxB,WAAO;AAAA,MACL,SAAS;AAAA,MACT;AAAA,MACA,qBAAqB;AAAA,IACvB;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,UACJ,MACA,YAC0B;AAC1B,UAAM,MAAM,KAAK,iBAAiB;AAClC,QAAI,CAAC,IAAI,QAAS,QAAO,EAAE,SAAS,OAAO,OAAO,IAAI,MAAM;AAE5D,UAAM,KAAK,OAAO;AAElB,QAAI;AACF,mBAAa,EAAE;AAMf,YAAM,cAAc,EAAE,SAAS,KAAK,OAAO,SAAS;AACpD,WAAK,SAAS,IAAI,kBAAO,WAAW;AACpC,mBAAa,EAAE;AAEf,YAAM,KAAK,aAAa,IAAI;AAC5B,mBAAa,EAAE;AAEf,YAAM,YAAY,CAAC,CAAE,UAAkB;AACvC,YAAO,KAAK,OAAe,UAAU,CAAC,IAAI,GAAG;AAAA,QAC3C,OAAO,KAAK,OAAO;AAAA,QACnB,GAAI,aAAa,KAAK,OAAO,aAAa,IACtC,EAAE,cAAc,KAAK,OAAO,WAAW,IACvC,CAAC;AAAA,MACP,CAAC;AAED,WAAK,YAAY,KAAK;AACtB,mBAAa,GAAG;AAEhB,aAAO,EAAE,SAAS,MAAM,YAAY,UAAU;AAAA,IAChD,SAAS,GAAY;AACnB,YAAM,KAAK,OAAO;AAClB,YAAM,UAAU,aAAa,QAAQ,EAAE,UAAU;AACjD,aAAO,EAAE,SAAS,OAAO,OAAO,QAAQ;AAAA,IAC1C;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,SAAS,KAA+C;AAC5D,QAAI,CAAC,KAAK,QAAQ;AAChB,aAAO,EAAE,SAAS,OAAO,OAAO,2CAA2C;AAAA,IAC7E;AAEA,QAAI;AACF,YAAM,KAAK,YAAY,IAAI;AAE3B,YAAM,WAAgD,CAAC;AACvD,UAAI,IAAI,OAAQ,UAAS,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,OAAO,CAAC;AACrE,eAAS,KAAK,EAAE,MAAM,QAAQ,SAAS,IAAI,OAAO,CAAC;AAEnD,YAAM,WAAW,MAAO,KAAK,OAAe,qBAAqB;AAAA,QAC/D;AAAA,QACA,YAAY,IAAI,aAAa;AAAA,QAC7B,aAAa,IAAI,eAAe;AAAA,QAChC,OAAO,IAAI,QAAQ;AAAA,QACnB,OAAO,IAAI,QAAQ;AAAA,MACrB,CAAC;AAED,YAAM,OAAe,UAAU,UAAU,CAAC,GAAG,SAAS,WAAW;AACjE,aAAO;AAAA,QACL,SAAS;AAAA,QACT;AAAA,QACA,QAAQ,KAAK,MAAM,YAAY,IAAI,IAAI,EAAE;AAAA,MAC3C;AAAA,IACF,SAAS,GAAY;AACnB,YAAM,UAAU,aAAa,QAAQ,EAAE,UAAU;AACjD,aAAO,EAAE,SAAS,OAAO,OAAO,QAAQ;AAAA,IAC1C;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,SAAwB;AAC5B,QAAI,KAAK,QAAQ;AACf,UAAI;AACF,cAAO,KAAK,OAAe,KAAK;AAAA,MAClC,SAAS,GAAG;AAAA,MAEZ;AACA,WAAK,SAAS;AAAA,IAChB;AACA,QAAI,KAAK,WAAW;AAClB,UAAI;AACF,cAAM,KAAK,gBAAgB,KAAK,SAAS;AAAA,MAC3C,SAAS,GAAG;AAAA,MAEZ;AACA,WAAK,YAAY;AAAA,IACnB;AAAA,EACF;AAAA;AAAA,EAGA,IAAI,WAAoB;AACtB,WAAO,KAAK,WAAW;AAAA,EACzB;AAAA;AAAA,EAGA,IAAI,eAAuB;AACzB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAc,aAAa,MAA2B;AACpD,UAAM,MAAM,GAAG,OAAO,SAAS,MAAM,wBAAwB,KAAK,IAAI;AACtE,UAAM,QAAQ,MAAM,OAAO,KAAK,qBAAqB;AACrD,UAAM,MAAM;AAAA,MACV;AAAA,MACA,IAAI,SAAS,MAAM;AAAA,QACjB,SAAS,EAAE,gBAAgB,2BAA2B;AAAA,MACxD,CAAC;AAAA,IACH;AAAA,EACF;AAAA,EAEA,MAAc,gBAAgB,UAAiC;AAC7D,UAAM,MAAM,GAAG,OAAO,SAAS,MAAM,wBAAwB,QAAQ;AACrE,UAAM,QAAQ,MAAM,OAAO,KAAK,qBAAqB;AACrD,UAAM,MAAM,OAAO,GAAG;AAAA,EACxB;AACF;","names":[]}
+{"version":3,"sources":["../src/index.ts","../src/WllamaService.ts"],"sourcesContent":["export { WllamaService } from './WllamaService';\nexport type {\n  WllamaServiceConfig,\n  GenerateRequest,\n  GenerateResult,\n  LoadModelResult,\n  ConnectionResult,\n} from './types';","import { Wllama } from '@wllama/wllama/esm/index.js';\nimport type {\n  WllamaServiceConfig,\n  GenerateRequest,\n  GenerateResult,\n  LoadModelResult,\n  ConnectionResult,\n} from './types';\n\nexport class WllamaService {\n  private wllama: Wllama | null = null;\n  private modelName: string = '';\n  private config: Required<WllamaServiceConfig>;\n\n  constructor(config: WllamaServiceConfig = {}) {\n    this.config = {\n      wasmPath: config.wasmPath ?? '/wllama/wllama.wasm',\n      nGpuLayers: config.nGpuLayers ?? 999,\n      nCtx: config.nCtx ?? 2048,\n    };\n  }\n\n  /** Check browser compatibility and capabilities */\n  checkEnvironment(): ConnectionResult {\n    if (typeof WebAssembly === 'undefined') {\n      return { success: false, error: 'WebAssembly is not supported in this browser.' };\n    }\n    if (!('caches' in window)) {\n      return { success: false, error: 'Cache API is not supported in this browser.' };\n    }\n    const hasWebGPU = !!(navigator as any).gpu;\n    const isolated = window.crossOriginIsolated;\n    return {\n      success: true,\n      hasWebGPU,\n      crossOriginIsolated: isolated,\n    };\n  }\n\n  /** Load a GGUF model file from the user's filesystem */\n  async loadModel(\n    file: File,\n    onProgress?: (progress: number) => void,\n  ): Promise<LoadModelResult> {\n    const env = this.checkEnvironment();\n    if (!env.success) return { success: false, error: env.error };\n\n    await this.unload();\n\n    try {\n      onProgress?.(10);\n\n      // AssetsPathConfig requires `default` — this is the key the\n      // Wllama runtime actually reads (confirmed against the real\n      // @wllama/wllama@3.5.1 implementation, not just its docs).\n      // No `as any` needed: this shape is type-correct as written.\n      const configPaths = { default: this.config.wasmPath };\n      this.wllama = new Wllama(configPaths);\n      onProgress?.(30);\n\n      await this.storeInCache(file);\n      onProgress?.(50);\n\n      const hasWebGPU = !!(navigator as any).gpu;\n      await (this.wllama as any).loadModel([file], {\n        n_ctx: this.config.nCtx,\n        ...(hasWebGPU && this.config.nGpuLayers > 0\n          ? { n_gpu_layers: this.config.nGpuLayers }\n          : {}),\n      });\n\n      this.modelName = file.name;\n      onProgress?.(100);\n\n      return { success: true, usedWebGPU: hasWebGPU };\n    } catch (e: unknown) {\n      await this.unload();\n      const message = e instanceof Error ? e.message : 'Failed to load model';\n      return { success: false, error: message };\n    }\n  }\n\n  /**\n   * Generate a response using the model's chat template.\n   * Use this for normal conversational / instruction-following use cases —\n   * it accepts a `messages`-style request (system + user) and\n   * createChatCompletion formats it with the model's real chat template\n   * (ChatML, Llama-3, etc.) before running inference.\n   */\n  async generate(req: GenerateRequest): Promise<GenerateResult> {\n    if (!this.wllama) {\n      return { success: false, error: 'No model loaded. Call loadModel() first.' };\n    }\n\n    try {\n      const t0 = performance.now();\n\n      const messages: { role: string; content: string }[] = [];\n      if (req.system) messages.push({ role: 'system', content: req.system });\n      messages.push({ role: 'user', content: req.prompt });\n\n      const response = await (this.wllama as any).createChatCompletion({\n        messages,\n        stream: false,\n        max_tokens: req.maxTokens ?? 512,\n        temperature: req.temperature ?? 0.7,\n        top_k: req.topK ?? 40,\n        top_p: req.topP ?? 0.95,\n      });\n\n      const text: string = response?.choices?.[0]?.message?.content ?? '';\n      return {\n        success: true,\n        text,\n        timeMs: Math.round(performance.now() - t0),\n      };\n    } catch (e: unknown) {\n      const message = e instanceof Error ? e.message : 'Generation failed';\n      return { success: false, error: message };\n    }\n  }\n\n  /**\n   * Generate a raw completion with no chat-template formatting.\n   * Use this when you need to control the exact prompt string yourself —\n   * e.g. few-shot prompts, base-model completion, custom templates, or\n   * fill-in-the-middle style prompts.\n   *\n   * createCompletion (confirmed against @wllama/wllama@3.5.1's real type\n   * defs) takes a single options object with a raw `prompt` string. It does\n   * NOT accept a `messages` array and does NOT apply any chat template —\n   * whatever string you pass is exactly what the model sees. `req.system`\n   * (if provided) is prepended as plain text; it is your responsibility to\n   * format it however the target model expects (call\n   * `this.wllama.getChatTemplate()` if you want to match the model's\n   * native template manually).\n   */\n  async generateCompletion(req: GenerateRequest): Promise<GenerateResult> {\n    if (!this.wllama) {\n      return { success: false, error: 'No model loaded. Call loadModel() first.' };\n    }\n\n    try {\n      const t0 = performance.now();\n\n      const prompt = req.system ? `${req.system}\\n\\n${req.prompt}` : req.prompt;\n\n      const response = await (this.wllama as any).createCompletion({\n        prompt,\n        stream: false,\n        max_tokens: req.maxTokens ?? 512,\n        temperature: req.temperature ?? 0.7,\n        top_k: req.topK ?? 40,\n        top_p: req.topP ?? 0.95,\n      });\n\n      const text: string = response?.choices?.[0]?.text ?? '';\n      return {\n        success: true,\n        text,\n        timeMs: Math.round(performance.now() - t0),\n      };\n    } catch (e: unknown) {\n      const message = e instanceof Error ? e.message : 'Generation failed';\n      return { success: false, error: message };\n    }\n  }\n\n  /** Unload the current model and free memory */\n  async unload(): Promise<void> {\n    if (this.wllama) {\n      try {\n        await (this.wllama as any).exit();\n      } catch (_) {\n        /* ignore */\n      }\n      this.wllama = null;\n    }\n    if (this.modelName) {\n      try {\n        await this.deleteFromCache(this.modelName);\n      } catch (_) {\n        /* ignore */\n      }\n      this.modelName = '';\n    }\n  }\n\n  /** Whether a model is currently loaded */\n  get isLoaded(): boolean {\n    return this.wllama !== null;\n  }\n\n  /** Name of the currently loaded model */\n  get currentModel(): string {\n    return this.modelName;\n  }\n\n  private async storeInCache(file: File): Promise<void> {\n    const url = `${window.location.origin}/wllama-local-models/${file.name}`;\n    const cache = await caches.open('wllama-local-models');\n    await cache.put(\n      url,\n      new Response(file, {\n        headers: { 'Content-Type': 'application/octet-stream' },\n      }),\n    );\n  }\n\n  private async deleteFromCache(fileName: string): Promise<void> {\n    const url = `${window.location.origin}/wllama-local-models/${fileName}`;\n    const cache = await caches.open('wllama-local-models');\n    await cache.delete(url);\n  }\n}"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACAA,iBAAuB;AAShB,IAAM,gBAAN,MAAoB;AAAA,EAKzB,YAAY,SAA8B,CAAC,GAAG;AAJ9C,SAAQ,SAAwB;AAChC,SAAQ,YAAoB;AAI1B,SAAK,SAAS;AAAA,MACZ,UAAU,OAAO,YAAY;AAAA,MAC7B,YAAY,OAAO,cAAc;AAAA,MACjC,MAAM,OAAO,QAAQ;AAAA,IACvB;AAAA,EACF;AAAA;AAAA,EAGA,mBAAqC;AACnC,QAAI,OAAO,gBAAgB,aAAa;AACtC,aAAO,EAAE,SAAS,OAAO,OAAO,gDAAgD;AAAA,IAClF;AACA,QAAI,EAAE,YAAY,SAAS;AACzB,aAAO,EAAE,SAAS,OAAO,OAAO,8CAA8C;AAAA,IAChF;AACA,UAAM,YAAY,CAAC,CAAE,UAAkB;AACvC,UAAM,WAAW,OAAO;AACxB,WAAO;AAAA,MACL,SAAS;AAAA,MACT;AAAA,MACA,qBAAqB;AAAA,IACvB;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,UACJ,MACA,YAC0B;AAC1B,UAAM,MAAM,KAAK,iBAAiB;AAClC,QAAI,CAAC,IAAI,QAAS,QAAO,EAAE,SAAS,OAAO,OAAO,IAAI,MAAM;AAE5D,UAAM,KAAK,OAAO;AAElB,QAAI;AACF,mBAAa,EAAE;AAMf,YAAM,cAAc,EAAE,SAAS,KAAK,OAAO,SAAS;AACpD,WAAK,SAAS,IAAI,kBAAO,WAAW;AACpC,mBAAa,EAAE;AAEf,YAAM,KAAK,aAAa,IAAI;AAC5B,mBAAa,EAAE;AAEf,YAAM,YAAY,CAAC,CAAE,UAAkB;AACvC,YAAO,KAAK,OAAe,UAAU,CAAC,IAAI,GAAG;AAAA,QAC3C,OAAO,KAAK,OAAO;AAAA,QACnB,GAAI,aAAa,KAAK,OAAO,aAAa,IACtC,EAAE,cAAc,KAAK,OAAO,WAAW,IACvC,CAAC;AAAA,MACP,CAAC;AAED,WAAK,YAAY,KAAK;AACtB,mBAAa,GAAG;AAEhB,aAAO,EAAE,SAAS,MAAM,YAAY,UAAU;AAAA,IAChD,SAAS,GAAY;AACnB,YAAM,KAAK,OAAO;AAClB,YAAM,UAAU,aAAa,QAAQ,EAAE,UAAU;AACjD,aAAO,EAAE,SAAS,OAAO,OAAO,QAAQ;AAAA,IAC1C;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,SAAS,KAA+C;AAC5D,QAAI,CAAC,KAAK,QAAQ;AAChB,aAAO,EAAE,SAAS,OAAO,OAAO,2CAA2C;AAAA,IAC7E;AAEA,QAAI;AACF,YAAM,KAAK,YAAY,IAAI;AAE3B,YAAM,WAAgD,CAAC;AACvD,UAAI,IAAI,OAAQ,UAAS,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,OAAO,CAAC;AACrE,eAAS,KAAK,EAAE,MAAM,QAAQ,SAAS,IAAI,OAAO,CAAC;AAEnD,YAAM,WAAW,MAAO,KAAK,OAAe,qBAAqB;AAAA,QAC/D;AAAA,QACA,QAAQ;AAAA,QACR,YAAY,IAAI,aAAa;AAAA,QAC7B,aAAa,IAAI,eAAe;AAAA,QAChC,OAAO,IAAI,QAAQ;AAAA,QACnB,OAAO,IAAI,QAAQ;AAAA,MACrB,CAAC;AAED,YAAM,OAAe,UAAU,UAAU,CAAC,GAAG,SAAS,WAAW;AACjE,aAAO;AAAA,QACL,SAAS;AAAA,QACT;AAAA,QACA,QAAQ,KAAK,MAAM,YAAY,IAAI,IAAI,EAAE;AAAA,MAC3C;AAAA,IACF,SAAS,GAAY;AACnB,YAAM,UAAU,aAAa,QAAQ,EAAE,UAAU;AACjD,aAAO,EAAE,SAAS,OAAO,OAAO,QAAQ;AAAA,IAC1C;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAiBA,MAAM,mBAAmB,KAA+C;AACtE,QAAI,CAAC,KAAK,QAAQ;AAChB,aAAO,EAAE,SAAS,OAAO,OAAO,2CAA2C;AAAA,IAC7E;AAEA,QAAI;AACF,YAAM,KAAK,YAAY,IAAI;AAE3B,YAAM,SAAS,IAAI,SAAS,GAAG,IAAI,MAAM;AAAA;AAAA,EAAO,IAAI,MAAM,KAAK,IAAI;AAEnE,YAAM,WAAW,MAAO,KAAK,OAAe,iBAAiB;AAAA,QAC3D;AAAA,QACA,QAAQ;AAAA,QACR,YAAY,IAAI,aAAa;AAAA,QAC7B,aAAa,IAAI,eAAe;AAAA,QAChC,OAAO,IAAI,QAAQ;AAAA,QACnB,OAAO,IAAI,QAAQ;AAAA,MACrB,CAAC;AAED,YAAM,OAAe,UAAU,UAAU,CAAC,GAAG,QAAQ;AACrD,aAAO;AAAA,QACL,SAAS;AAAA,QACT;AAAA,QACA,QAAQ,KAAK,MAAM,YAAY,IAAI,IAAI,EAAE;AAAA,MAC3C;AAAA,IACF,SAAS,GAAY;AACnB,YAAM,UAAU,aAAa,QAAQ,EAAE,UAAU;AACjD,aAAO,EAAE,SAAS,OAAO,OAAO,QAAQ;AAAA,IAC1C;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,SAAwB;AAC5B,QAAI,KAAK,QAAQ;AACf,UAAI;AACF,cAAO,KAAK,OAAe,KAAK;AAAA,MAClC,SAAS,GAAG;AAAA,MAEZ;AACA,WAAK,SAAS;AAAA,IAChB;AACA,QAAI,KAAK,WAAW;AAClB,UAAI;AACF,cAAM,KAAK,gBAAgB,KAAK,SAAS;AAAA,MAC3C,SAAS,GAAG;AAAA,MAEZ;AACA,WAAK,YAAY;AAAA,IACnB;AAAA,EACF;AAAA;AAAA,EAGA,IAAI,WAAoB;AACtB,WAAO,KAAK,WAAW;AAAA,EACzB;AAAA;AAAA,EAGA,IAAI,eAAuB;AACzB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAc,aAAa,MAA2B;AACpD,UAAM,MAAM,GAAG,OAAO,SAAS,MAAM,wBAAwB,KAAK,IAAI;AACtE,UAAM,QAAQ,MAAM,OAAO,KAAK,qBAAqB;AACrD,UAAM,MAAM;AAAA,MACV;AAAA,MACA,IAAI,SAAS,MAAM;AAAA,QACjB,SAAS,EAAE,gBAAgB,2BAA2B;AAAA,MACxD,CAAC;AAAA,IACH;AAAA,EACF;AAAA,EAEA,MAAc,gBAAgB,UAAiC;AAC7D,UAAM,MAAM,GAAG,OAAO,SAAS,MAAM,wBAAwB,QAAQ;AACrE,UAAM,QAAQ,MAAM,OAAO,KAAK,qBAAqB;AACrD,UAAM,MAAM,OAAO,GAAG;AAAA,EACxB;AACF;","names":[]}

package/dist/index.mjs

@@ -52,7 +52,13 @@ var WllamaService = class {
       return { success: false, error: message };
     }
   }
-  /** Generate a response from the loaded model */
+  /**
+   * Generate a response using the model's chat template.
+   * Use this for normal conversational / instruction-following use cases —
+   * it accepts a `messages`-style request (system + user) and
+   * createChatCompletion formats it with the model's real chat template
+   * (ChatML, Llama-3, etc.) before running inference.
+   */
   async generate(req) {
     if (!this.wllama) {
       return { success: false, error: "No model loaded. Call loadModel() first." };
@@ -64,6 +70,7 @@ var WllamaService = class {
       messages.push({ role: "user", content: req.prompt });
       const response = await this.wllama.createChatCompletion({
         messages,
+        stream: false,
         max_tokens: req.maxTokens ?? 512,
         temperature: req.temperature ?? 0.7,
         top_k: req.topK ?? 40,
@@ -80,6 +87,49 @@ var WllamaService = class {
       return { success: false, error: message };
     }
   }
+  /**
+   * Generate a raw completion with no chat-template formatting.
+   * Use this when you need to control the exact prompt string yourself —
+   * e.g. few-shot prompts, base-model completion, custom templates, or
+   * fill-in-the-middle style prompts.
+   *
+   * createCompletion (confirmed against @wllama/wllama@3.5.1's real type
+   * defs) takes a single options object with a raw `prompt` string. It does
+   * NOT accept a `messages` array and does NOT apply any chat template —
+   * whatever string you pass is exactly what the model sees. `req.system`
+   * (if provided) is prepended as plain text; it is your responsibility to
+   * format it however the target model expects (call
+   * `this.wllama.getChatTemplate()` if you want to match the model's
+   * native template manually).
+   */
+  async generateCompletion(req) {
+    if (!this.wllama) {
+      return { success: false, error: "No model loaded. Call loadModel() first." };
+    }
+    try {
+      const t0 = performance.now();
+      const prompt = req.system ? `${req.system}
+
+${req.prompt}` : req.prompt;
+      const response = await this.wllama.createCompletion({
+        prompt,
+        stream: false,
+        max_tokens: req.maxTokens ?? 512,
+        temperature: req.temperature ?? 0.7,
+        top_k: req.topK ?? 40,
+        top_p: req.topP ?? 0.95
+      });
+      const text = response?.choices?.[0]?.text ?? "";
+      return {
+        success: true,
+        text,
+        timeMs: Math.round(performance.now() - t0)
+      };
+    } catch (e) {
+      const message = e instanceof Error ? e.message : "Generation failed";
+      return { success: false, error: message };
+    }
+  }
   /** Unload the current model and free memory */
   async unload() {
     if (this.wllama) {

package/dist/index.mjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/WllamaService.ts"],"sourcesContent":["import { Wllama } from '@wllama/wllama/esm/index.js';\nimport type {\n  WllamaServiceConfig,\n  GenerateRequest,\n  GenerateResult,\n  LoadModelResult,\n  ConnectionResult,\n} from './types';\n\nexport class WllamaService {\n  private wllama: Wllama | null = null;\n  private modelName: string = '';\n  private config: Required<WllamaServiceConfig>;\n\n  constructor(config: WllamaServiceConfig = {}) {\n    this.config = {\n      wasmPath: config.wasmPath ?? '/wllama/wllama.wasm',\n      nGpuLayers: config.nGpuLayers ?? 999,\n      nCtx: config.nCtx ?? 2048,\n    };\n  }\n\n  /** Check browser compatibility and capabilities */\n  checkEnvironment(): ConnectionResult {\n    if (typeof WebAssembly === 'undefined') {\n      return { success: false, error: 'WebAssembly is not supported in this browser.' };\n    }\n    if (!('caches' in window)) {\n      return { success: false, error: 'Cache API is not supported in this browser.' };\n    }\n    const hasWebGPU = !!(navigator as any).gpu;\n    const isolated = window.crossOriginIsolated;\n    return {\n      success: true,\n      hasWebGPU,\n      crossOriginIsolated: isolated,\n    };\n  }\n\n  /** Load a GGUF model file from the user's filesystem */\n  async loadModel(\n    file: File,\n    onProgress?: (progress: number) => void,\n  ): Promise<LoadModelResult> {\n    const env = this.checkEnvironment();\n    if (!env.success) return { success: false, error: env.error };\n\n    await this.unload();\n\n    try {\n      onProgress?.(10);\n\n      // AssetsPathConfig requires `default` — this is the key the\n      // Wllama runtime actually reads (confirmed against the real\n      // @wllama/wllama@3.5.1 implementation, not just its docs).\n      // No `as any` needed: this shape is type-correct as written.\n      const configPaths = { default: this.config.wasmPath };\n      this.wllama = new Wllama(configPaths);\n      onProgress?.(30);\n\n      await this.storeInCache(file);\n      onProgress?.(50);\n\n      const hasWebGPU = !!(navigator as any).gpu;\n      await (this.wllama as any).loadModel([file], {\n        n_ctx: this.config.nCtx,\n        ...(hasWebGPU && this.config.nGpuLayers > 0\n          ? { n_gpu_layers: this.config.nGpuLayers }\n          : {}),\n      });\n\n      this.modelName = file.name;\n      onProgress?.(100);\n\n      return { success: true, usedWebGPU: hasWebGPU };\n    } catch (e: unknown) {\n      await this.unload();\n      const message = e instanceof Error ? e.message : 'Failed to load model';\n      return { success: false, error: message };\n    }\n  }\n\n  /** Generate a response from the loaded model */\n  async generate(req: GenerateRequest): Promise<GenerateResult> {\n    if (!this.wllama) {\n      return { success: false, error: 'No model loaded. Call loadModel() first.' };\n    }\n\n    try {\n      const t0 = performance.now();\n\n      const messages: { role: string; content: string }[] = [];\n      if (req.system) messages.push({ role: 'system', content: req.system });\n      messages.push({ role: 'user', content: req.prompt });\n\n      const response = await (this.wllama as any).createChatCompletion({\n        messages,\n        max_tokens: req.maxTokens ?? 512,\n        temperature: req.temperature ?? 0.7,\n        top_k: req.topK ?? 40,\n        top_p: req.topP ?? 0.95,\n      });\n\n      const text: string = response?.choices?.[0]?.message?.content ?? '';\n      return {\n        success: true,\n        text,\n        timeMs: Math.round(performance.now() - t0),\n      };\n    } catch (e: unknown) {\n      const message = e instanceof Error ? e.message : 'Generation failed';\n      return { success: false, error: message };\n    }\n  }\n\n  /** Unload the current model and free memory */\n  async unload(): Promise<void> {\n    if (this.wllama) {\n      try {\n        await (this.wllama as any).exit();\n      } catch (_) {\n        /* ignore */\n      }\n      this.wllama = null;\n    }\n    if (this.modelName) {\n      try {\n        await this.deleteFromCache(this.modelName);\n      } catch (_) {\n        /* ignore */\n      }\n      this.modelName = '';\n    }\n  }\n\n  /** Whether a model is currently loaded */\n  get isLoaded(): boolean {\n    return this.wllama !== null;\n  }\n\n  /** Name of the currently loaded model */\n  get currentModel(): string {\n    return this.modelName;\n  }\n\n  private async storeInCache(file: File): Promise<void> {\n    const url = `${window.location.origin}/wllama-local-models/${file.name}`;\n    const cache = await caches.open('wllama-local-models');\n    await cache.put(\n      url,\n      new Response(file, {\n        headers: { 'Content-Type': 'application/octet-stream' },\n      }),\n    );\n  }\n\n  private async deleteFromCache(fileName: string): Promise<void> {\n    const url = `${window.location.origin}/wllama-local-models/${fileName}`;\n    const cache = await caches.open('wllama-local-models');\n    await cache.delete(url);\n  }\n}"],"mappings":";AAAA,SAAS,cAAc;AAShB,IAAM,gBAAN,MAAoB;AAAA,EAKzB,YAAY,SAA8B,CAAC,GAAG;AAJ9C,SAAQ,SAAwB;AAChC,SAAQ,YAAoB;AAI1B,SAAK,SAAS;AAAA,MACZ,UAAU,OAAO,YAAY;AAAA,MAC7B,YAAY,OAAO,cAAc;AAAA,MACjC,MAAM,OAAO,QAAQ;AAAA,IACvB;AAAA,EACF;AAAA;AAAA,EAGA,mBAAqC;AACnC,QAAI,OAAO,gBAAgB,aAAa;AACtC,aAAO,EAAE,SAAS,OAAO,OAAO,gDAAgD;AAAA,IAClF;AACA,QAAI,EAAE,YAAY,SAAS;AACzB,aAAO,EAAE,SAAS,OAAO,OAAO,8CAA8C;AAAA,IAChF;AACA,UAAM,YAAY,CAAC,CAAE,UAAkB;AACvC,UAAM,WAAW,OAAO;AACxB,WAAO;AAAA,MACL,SAAS;AAAA,MACT;AAAA,MACA,qBAAqB;AAAA,IACvB;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,UACJ,MACA,YAC0B;AAC1B,UAAM,MAAM,KAAK,iBAAiB;AAClC,QAAI,CAAC,IAAI,QAAS,QAAO,EAAE,SAAS,OAAO,OAAO,IAAI,MAAM;AAE5D,UAAM,KAAK,OAAO;AAElB,QAAI;AACF,mBAAa,EAAE;AAMf,YAAM,cAAc,EAAE,SAAS,KAAK,OAAO,SAAS;AACpD,WAAK,SAAS,IAAI,OAAO,WAAW;AACpC,mBAAa,EAAE;AAEf,YAAM,KAAK,aAAa,IAAI;AAC5B,mBAAa,EAAE;AAEf,YAAM,YAAY,CAAC,CAAE,UAAkB;AACvC,YAAO,KAAK,OAAe,UAAU,CAAC,IAAI,GAAG;AAAA,QAC3C,OAAO,KAAK,OAAO;AAAA,QACnB,GAAI,aAAa,KAAK,OAAO,aAAa,IACtC,EAAE,cAAc,KAAK,OAAO,WAAW,IACvC,CAAC;AAAA,MACP,CAAC;AAED,WAAK,YAAY,KAAK;AACtB,mBAAa,GAAG;AAEhB,aAAO,EAAE,SAAS,MAAM,YAAY,UAAU;AAAA,IAChD,SAAS,GAAY;AACnB,YAAM,KAAK,OAAO;AAClB,YAAM,UAAU,aAAa,QAAQ,EAAE,UAAU;AACjD,aAAO,EAAE,SAAS,OAAO,OAAO,QAAQ;AAAA,IAC1C;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,SAAS,KAA+C;AAC5D,QAAI,CAAC,KAAK,QAAQ;AAChB,aAAO,EAAE,SAAS,OAAO,OAAO,2CAA2C;AAAA,IAC7E;AAEA,QAAI;AACF,YAAM,KAAK,YAAY,IAAI;AAE3B,YAAM,WAAgD,CAAC;AACvD,UAAI,IAAI,OAAQ,UAAS,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,OAAO,CAAC;AACrE,eAAS,KAAK,EAAE,MAAM,QAAQ,SAAS,IAAI,OAAO,CAAC;AAEnD,YAAM,WAAW,MAAO,KAAK,OAAe,qBAAqB;AAAA,QAC/D;AAAA,QACA,YAAY,IAAI,aAAa;AAAA,QAC7B,aAAa,IAAI,eAAe;AAAA,QAChC,OAAO,IAAI,QAAQ;AAAA,QACnB,OAAO,IAAI,QAAQ;AAAA,MACrB,CAAC;AAED,YAAM,OAAe,UAAU,UAAU,CAAC,GAAG,SAAS,WAAW;AACjE,aAAO;AAAA,QACL,SAAS;AAAA,QACT;AAAA,QACA,QAAQ,KAAK,MAAM,YAAY,IAAI,IAAI,EAAE;AAAA,MAC3C;AAAA,IACF,SAAS,GAAY;AACnB,YAAM,UAAU,aAAa,QAAQ,EAAE,UAAU;AACjD,aAAO,EAAE,SAAS,OAAO,OAAO,QAAQ;AAAA,IAC1C;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,SAAwB;AAC5B,QAAI,KAAK,QAAQ;AACf,UAAI;AACF,cAAO,KAAK,OAAe,KAAK;AAAA,MAClC,SAAS,GAAG;AAAA,MAEZ;AACA,WAAK,SAAS;AAAA,IAChB;AACA,QAAI,KAAK,WAAW;AAClB,UAAI;AACF,cAAM,KAAK,gBAAgB,KAAK,SAAS;AAAA,MAC3C,SAAS,GAAG;AAAA,MAEZ;AACA,WAAK,YAAY;AAAA,IACnB;AAAA,EACF;AAAA;AAAA,EAGA,IAAI,WAAoB;AACtB,WAAO,KAAK,WAAW;AAAA,EACzB;AAAA;AAAA,EAGA,IAAI,eAAuB;AACzB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAc,aAAa,MAA2B;AACpD,UAAM,MAAM,GAAG,OAAO,SAAS,MAAM,wBAAwB,KAAK,IAAI;AACtE,UAAM,QAAQ,MAAM,OAAO,KAAK,qBAAqB;AACrD,UAAM,MAAM;AAAA,MACV;AAAA,MACA,IAAI,SAAS,MAAM;AAAA,QACjB,SAAS,EAAE,gBAAgB,2BAA2B;AAAA,MACxD,CAAC;AAAA,IACH;AAAA,EACF;AAAA,EAEA,MAAc,gBAAgB,UAAiC;AAC7D,UAAM,MAAM,GAAG,OAAO,SAAS,MAAM,wBAAwB,QAAQ;AACrE,UAAM,QAAQ,MAAM,OAAO,KAAK,qBAAqB;AACrD,UAAM,MAAM,OAAO,GAAG;AAAA,EACxB;AACF;","names":[]}
+{"version":3,"sources":["../src/WllamaService.ts"],"sourcesContent":["import { Wllama } from '@wllama/wllama/esm/index.js';\nimport type {\n  WllamaServiceConfig,\n  GenerateRequest,\n  GenerateResult,\n  LoadModelResult,\n  ConnectionResult,\n} from './types';\n\nexport class WllamaService {\n  private wllama: Wllama | null = null;\n  private modelName: string = '';\n  private config: Required<WllamaServiceConfig>;\n\n  constructor(config: WllamaServiceConfig = {}) {\n    this.config = {\n      wasmPath: config.wasmPath ?? '/wllama/wllama.wasm',\n      nGpuLayers: config.nGpuLayers ?? 999,\n      nCtx: config.nCtx ?? 2048,\n    };\n  }\n\n  /** Check browser compatibility and capabilities */\n  checkEnvironment(): ConnectionResult {\n    if (typeof WebAssembly === 'undefined') {\n      return { success: false, error: 'WebAssembly is not supported in this browser.' };\n    }\n    if (!('caches' in window)) {\n      return { success: false, error: 'Cache API is not supported in this browser.' };\n    }\n    const hasWebGPU = !!(navigator as any).gpu;\n    const isolated = window.crossOriginIsolated;\n    return {\n      success: true,\n      hasWebGPU,\n      crossOriginIsolated: isolated,\n    };\n  }\n\n  /** Load a GGUF model file from the user's filesystem */\n  async loadModel(\n    file: File,\n    onProgress?: (progress: number) => void,\n  ): Promise<LoadModelResult> {\n    const env = this.checkEnvironment();\n    if (!env.success) return { success: false, error: env.error };\n\n    await this.unload();\n\n    try {\n      onProgress?.(10);\n\n      // AssetsPathConfig requires `default` — this is the key the\n      // Wllama runtime actually reads (confirmed against the real\n      // @wllama/wllama@3.5.1 implementation, not just its docs).\n      // No `as any` needed: this shape is type-correct as written.\n      const configPaths = { default: this.config.wasmPath };\n      this.wllama = new Wllama(configPaths);\n      onProgress?.(30);\n\n      await this.storeInCache(file);\n      onProgress?.(50);\n\n      const hasWebGPU = !!(navigator as any).gpu;\n      await (this.wllama as any).loadModel([file], {\n        n_ctx: this.config.nCtx,\n        ...(hasWebGPU && this.config.nGpuLayers > 0\n          ? { n_gpu_layers: this.config.nGpuLayers }\n          : {}),\n      });\n\n      this.modelName = file.name;\n      onProgress?.(100);\n\n      return { success: true, usedWebGPU: hasWebGPU };\n    } catch (e: unknown) {\n      await this.unload();\n      const message = e instanceof Error ? e.message : 'Failed to load model';\n      return { success: false, error: message };\n    }\n  }\n\n  /**\n   * Generate a response using the model's chat template.\n   * Use this for normal conversational / instruction-following use cases —\n   * it accepts a `messages`-style request (system + user) and\n   * createChatCompletion formats it with the model's real chat template\n   * (ChatML, Llama-3, etc.) before running inference.\n   */\n  async generate(req: GenerateRequest): Promise<GenerateResult> {\n    if (!this.wllama) {\n      return { success: false, error: 'No model loaded. Call loadModel() first.' };\n    }\n\n    try {\n      const t0 = performance.now();\n\n      const messages: { role: string; content: string }[] = [];\n      if (req.system) messages.push({ role: 'system', content: req.system });\n      messages.push({ role: 'user', content: req.prompt });\n\n      const response = await (this.wllama as any).createChatCompletion({\n        messages,\n        stream: false,\n        max_tokens: req.maxTokens ?? 512,\n        temperature: req.temperature ?? 0.7,\n        top_k: req.topK ?? 40,\n        top_p: req.topP ?? 0.95,\n      });\n\n      const text: string = response?.choices?.[0]?.message?.content ?? '';\n      return {\n        success: true,\n        text,\n        timeMs: Math.round(performance.now() - t0),\n      };\n    } catch (e: unknown) {\n      const message = e instanceof Error ? e.message : 'Generation failed';\n      return { success: false, error: message };\n    }\n  }\n\n  /**\n   * Generate a raw completion with no chat-template formatting.\n   * Use this when you need to control the exact prompt string yourself —\n   * e.g. few-shot prompts, base-model completion, custom templates, or\n   * fill-in-the-middle style prompts.\n   *\n   * createCompletion (confirmed against @wllama/wllama@3.5.1's real type\n   * defs) takes a single options object with a raw `prompt` string. It does\n   * NOT accept a `messages` array and does NOT apply any chat template —\n   * whatever string you pass is exactly what the model sees. `req.system`\n   * (if provided) is prepended as plain text; it is your responsibility to\n   * format it however the target model expects (call\n   * `this.wllama.getChatTemplate()` if you want to match the model's\n   * native template manually).\n   */\n  async generateCompletion(req: GenerateRequest): Promise<GenerateResult> {\n    if (!this.wllama) {\n      return { success: false, error: 'No model loaded. Call loadModel() first.' };\n    }\n\n    try {\n      const t0 = performance.now();\n\n      const prompt = req.system ? `${req.system}\\n\\n${req.prompt}` : req.prompt;\n\n      const response = await (this.wllama as any).createCompletion({\n        prompt,\n        stream: false,\n        max_tokens: req.maxTokens ?? 512,\n        temperature: req.temperature ?? 0.7,\n        top_k: req.topK ?? 40,\n        top_p: req.topP ?? 0.95,\n      });\n\n      const text: string = response?.choices?.[0]?.text ?? '';\n      return {\n        success: true,\n        text,\n        timeMs: Math.round(performance.now() - t0),\n      };\n    } catch (e: unknown) {\n      const message = e instanceof Error ? e.message : 'Generation failed';\n      return { success: false, error: message };\n    }\n  }\n\n  /** Unload the current model and free memory */\n  async unload(): Promise<void> {\n    if (this.wllama) {\n      try {\n        await (this.wllama as any).exit();\n      } catch (_) {\n        /* ignore */\n      }\n      this.wllama = null;\n    }\n    if (this.modelName) {\n      try {\n        await this.deleteFromCache(this.modelName);\n      } catch (_) {\n        /* ignore */\n      }\n      this.modelName = '';\n    }\n  }\n\n  /** Whether a model is currently loaded */\n  get isLoaded(): boolean {\n    return this.wllama !== null;\n  }\n\n  /** Name of the currently loaded model */\n  get currentModel(): string {\n    return this.modelName;\n  }\n\n  private async storeInCache(file: File): Promise<void> {\n    const url = `${window.location.origin}/wllama-local-models/${file.name}`;\n    const cache = await caches.open('wllama-local-models');\n    await cache.put(\n      url,\n      new Response(file, {\n        headers: { 'Content-Type': 'application/octet-stream' },\n      }),\n    );\n  }\n\n  private async deleteFromCache(fileName: string): Promise<void> {\n    const url = `${window.location.origin}/wllama-local-models/${fileName}`;\n    const cache = await caches.open('wllama-local-models');\n    await cache.delete(url);\n  }\n}"],"mappings":";AAAA,SAAS,cAAc;AAShB,IAAM,gBAAN,MAAoB;AAAA,EAKzB,YAAY,SAA8B,CAAC,GAAG;AAJ9C,SAAQ,SAAwB;AAChC,SAAQ,YAAoB;AAI1B,SAAK,SAAS;AAAA,MACZ,UAAU,OAAO,YAAY;AAAA,MAC7B,YAAY,OAAO,cAAc;AAAA,MACjC,MAAM,OAAO,QAAQ;AAAA,IACvB;AAAA,EACF;AAAA;AAAA,EAGA,mBAAqC;AACnC,QAAI,OAAO,gBAAgB,aAAa;AACtC,aAAO,EAAE,SAAS,OAAO,OAAO,gDAAgD;AAAA,IAClF;AACA,QAAI,EAAE,YAAY,SAAS;AACzB,aAAO,EAAE,SAAS,OAAO,OAAO,8CAA8C;AAAA,IAChF;AACA,UAAM,YAAY,CAAC,CAAE,UAAkB;AACvC,UAAM,WAAW,OAAO;AACxB,WAAO;AAAA,MACL,SAAS;AAAA,MACT;AAAA,MACA,qBAAqB;AAAA,IACvB;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,UACJ,MACA,YAC0B;AAC1B,UAAM,MAAM,KAAK,iBAAiB;AAClC,QAAI,CAAC,IAAI,QAAS,QAAO,EAAE,SAAS,OAAO,OAAO,IAAI,MAAM;AAE5D,UAAM,KAAK,OAAO;AAElB,QAAI;AACF,mBAAa,EAAE;AAMf,YAAM,cAAc,EAAE,SAAS,KAAK,OAAO,SAAS;AACpD,WAAK,SAAS,IAAI,OAAO,WAAW;AACpC,mBAAa,EAAE;AAEf,YAAM,KAAK,aAAa,IAAI;AAC5B,mBAAa,EAAE;AAEf,YAAM,YAAY,CAAC,CAAE,UAAkB;AACvC,YAAO,KAAK,OAAe,UAAU,CAAC,IAAI,GAAG;AAAA,QAC3C,OAAO,KAAK,OAAO;AAAA,QACnB,GAAI,aAAa,KAAK,OAAO,aAAa,IACtC,EAAE,cAAc,KAAK,OAAO,WAAW,IACvC,CAAC;AAAA,MACP,CAAC;AAED,WAAK,YAAY,KAAK;AACtB,mBAAa,GAAG;AAEhB,aAAO,EAAE,SAAS,MAAM,YAAY,UAAU;AAAA,IAChD,SAAS,GAAY;AACnB,YAAM,KAAK,OAAO;AAClB,YAAM,UAAU,aAAa,QAAQ,EAAE,UAAU;AACjD,aAAO,EAAE,SAAS,OAAO,OAAO,QAAQ;AAAA,IAC1C;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,SAAS,KAA+C;AAC5D,QAAI,CAAC,KAAK,QAAQ;AAChB,aAAO,EAAE,SAAS,OAAO,OAAO,2CAA2C;AAAA,IAC7E;AAEA,QAAI;AACF,YAAM,KAAK,YAAY,IAAI;AAE3B,YAAM,WAAgD,CAAC;AACvD,UAAI,IAAI,OAAQ,UAAS,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,OAAO,CAAC;AACrE,eAAS,KAAK,EAAE,MAAM,QAAQ,SAAS,IAAI,OAAO,CAAC;AAEnD,YAAM,WAAW,MAAO,KAAK,OAAe,qBAAqB;AAAA,QAC/D;AAAA,QACA,QAAQ;AAAA,QACR,YAAY,IAAI,aAAa;AAAA,QAC7B,aAAa,IAAI,eAAe;AAAA,QAChC,OAAO,IAAI,QAAQ;AAAA,QACnB,OAAO,IAAI,QAAQ;AAAA,MACrB,CAAC;AAED,YAAM,OAAe,UAAU,UAAU,CAAC,GAAG,SAAS,WAAW;AACjE,aAAO;AAAA,QACL,SAAS;AAAA,QACT;AAAA,QACA,QAAQ,KAAK,MAAM,YAAY,IAAI,IAAI,EAAE;AAAA,MAC3C;AAAA,IACF,SAAS,GAAY;AACnB,YAAM,UAAU,aAAa,QAAQ,EAAE,UAAU;AACjD,aAAO,EAAE,SAAS,OAAO,OAAO,QAAQ;AAAA,IAC1C;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAiBA,MAAM,mBAAmB,KAA+C;AACtE,QAAI,CAAC,KAAK,QAAQ;AAChB,aAAO,EAAE,SAAS,OAAO,OAAO,2CAA2C;AAAA,IAC7E;AAEA,QAAI;AACF,YAAM,KAAK,YAAY,IAAI;AAE3B,YAAM,SAAS,IAAI,SAAS,GAAG,IAAI,MAAM;AAAA;AAAA,EAAO,IAAI,MAAM,KAAK,IAAI;AAEnE,YAAM,WAAW,MAAO,KAAK,OAAe,iBAAiB;AAAA,QAC3D;AAAA,QACA,QAAQ;AAAA,QACR,YAAY,IAAI,aAAa;AAAA,QAC7B,aAAa,IAAI,eAAe;AAAA,QAChC,OAAO,IAAI,QAAQ;AAAA,QACnB,OAAO,IAAI,QAAQ;AAAA,MACrB,CAAC;AAED,YAAM,OAAe,UAAU,UAAU,CAAC,GAAG,QAAQ;AACrD,aAAO;AAAA,QACL,SAAS;AAAA,QACT;AAAA,QACA,QAAQ,KAAK,MAAM,YAAY,IAAI,IAAI,EAAE;AAAA,MAC3C;AAAA,IACF,SAAS,GAAY;AACnB,YAAM,UAAU,aAAa,QAAQ,EAAE,UAAU;AACjD,aAAO,EAAE,SAAS,OAAO,OAAO,QAAQ;AAAA,IAC1C;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,SAAwB;AAC5B,QAAI,KAAK,QAAQ;AACf,UAAI;AACF,cAAO,KAAK,OAAe,KAAK;AAAA,MAClC,SAAS,GAAG;AAAA,MAEZ;AACA,WAAK,SAAS;AAAA,IAChB;AACA,QAAI,KAAK,WAAW;AAClB,UAAI;AACF,cAAM,KAAK,gBAAgB,KAAK,SAAS;AAAA,MAC3C,SAAS,GAAG;AAAA,MAEZ;AACA,WAAK,YAAY;AAAA,IACnB;AAAA,EACF;AAAA;AAAA,EAGA,IAAI,WAAoB;AACtB,WAAO,KAAK,WAAW;AAAA,EACzB;AAAA;AAAA,EAGA,IAAI,eAAuB;AACzB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAc,aAAa,MAA2B;AACpD,UAAM,MAAM,GAAG,OAAO,SAAS,MAAM,wBAAwB,KAAK,IAAI;AACtE,UAAM,QAAQ,MAAM,OAAO,KAAK,qBAAqB;AACrD,UAAM,MAAM;AAAA,MACV;AAAA,MACA,IAAI,SAAS,MAAM;AAAA,QACjB,SAAS,EAAE,gBAAgB,2BAA2B;AAAA,MACxD,CAAC;AAAA,IACH;AAAA,EACF;AAAA,EAEA,MAAc,gBAAgB,UAAiC;AAC7D,UAAM,MAAM,GAAG,OAAO,SAAS,MAAM,wBAAwB,QAAQ;AACrE,UAAM,QAAQ,MAAM,OAAO,KAAK,qBAAqB;AACrD,UAAM,MAAM,OAAO,GAAG;AAAA,EACxB;AACF;","names":[]}

package/package.json

@@ -1,11 +1,8 @@
 {
   "name": "wllama-service",
-  "version": "2.0.3",
+  "version": "2.0.5",
   "description": "Framework-agnostic browser LLM service wrapper with WebGPU and WebAssembly support",
   "main": "./dist/index.js",
-  "bin": {
-    "copy-wllama-wasm": "./scripts/postinstall.js"
-  },
   "module": "./dist/index.mjs",
   "types": "./dist/index.d.ts",
   "exports": {
@@ -20,7 +17,6 @@
     "README.md"
   ],
   "scripts": {
-    "postinstall": "node scripts/postinstall.js",
     "build": "tsup",
     "dev": "tsup --watch",
     "typecheck": "tsc --noEmit",

package/scripts/postinstall.js

@@ -1,53 +0,0 @@
-#!/usr/bin/env node
-'use strict';
-
-const fs = require('fs');
-const path = require('path');
-
-/**
- * Resolve @wllama/wllama's wasm file using Node's real module resolution
- * instead of a guessed relative path. This works correctly regardless of
- * where in the yarn workspaces hoisting tree @wllama/wllama actually
- * landed (monorepo root, nested under wllama-service, etc) — Node's own
- * require.resolve algorithm finds it the same way `require()` would.
- */
-function resolveWasmSrc() {
-  const pkgJsonPath = require.resolve('@wllama/wllama/package.json', {
-    paths: [process.cwd()],
-  });
-  const pkgRoot = path.dirname(pkgJsonPath);
-
-  // Different @wllama/wllama versions ship different layouts:
-  //   older: esm/wasm/wllama.wasm
-  //   newer: esm/single-thread/wllama.wasm (+ esm/multi-thread/wllama.wasm)
-  const candidates = [
-    path.join(pkgRoot, 'esm', 'wasm', 'wllama.wasm'),
-    path.join(pkgRoot, 'esm', 'single-thread', 'wllama.wasm'),
-  ];
-  const found = candidates.find((p) => fs.existsSync(p));
-  if (!found) {
-    throw new Error(
-      `wllama.wasm not found in any known location under ${pkgRoot}\n` +
-        `Checked: ${candidates.join(', ')}`,
-    );
-  }
-  return found;
-}
-
-// Destination: this app's own public folder. Adjust this path if your
-// app's public folder lives somewhere other than apps/<this-app>/public
-// relative to where this script sits.
-const destDir = path.resolve(__dirname, '..', 'public', 'wllama');
-
-let wasmSrc;
-try {
-  wasmSrc = resolveWasmSrc();
-} catch (err) {
-  console.warn('[copy-wllama-wasm] ' + err.message);
-  process.exit(0);
-}
-
-fs.mkdirSync(destDir, { recursive: true });
-const dest = path.join(destDir, 'wllama.wasm');
-fs.copyFileSync(wasmSrc, dest);
-console.log(`[copy-wllama-wasm] ✓ Copied ${wasmSrc} -> ${dest}`);