wllama-service 2.0.5 → 2.0.6

package/dist/index.d.mts

@@ -13,6 +13,8 @@ interface GenerateRequest {
     temperature?: number;
     topK?: number;
     topP?: number;
+    stop?: string | string[];
+    abortSignal?: AbortSignal;
 }
 interface GenerateResult {
     success: boolean;
@@ -35,44 +37,38 @@ interface ConnectionResult {
 declare class WllamaService {
     private wllama;
     private modelName;
-    private config;
+    private readonly config;
     constructor(config?: WllamaServiceConfig);
-    /** Check browser compatibility and capabilities */
+    /**
+     * Check WebGPU without directly accessing navigator.gpu.
+     *
+     * This avoids requiring @webgpu/types merely for feature detection.
+     */
+    private hasWebGPU;
+    /** Check browser compatibility and capabilities. */
     checkEnvironment(): ConnectionResult;
-    /** Load a GGUF model file from the user's filesystem */
+    /** Load a GGUF model file from the user's filesystem. */
     loadModel(file: File, onProgress?: (progress: number) => void): Promise<LoadModelResult>;
     /**
-     * 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 a chat response using the model's embedded chat template.
+     *
+     * Suitable for instruction-tuned models such as Phi-3 Instruct,
+     * Llama Instruct, ChatML models, and similar chat models.
      */
     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.
+     * Generate a raw completion without applying a chat template.
      *
-     * 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).
+     * Suitable for base completion models, manually formatted prompts,
+     * few-shot completion, fill-in-the-middle, and custom templates.
      */
     generateCompletion(req: GenerateRequest): Promise<GenerateResult>;
-    /** Unload the current model and free memory */
+    /** Unload the current model and release its resources. */
     unload(): Promise<void>;
-    /** Whether a model is currently loaded */
+    /** Whether a model is currently loaded. */
     get isLoaded(): boolean;
-    /** Name of the currently loaded model */
+    /** Name of the currently loaded model. */
     get currentModel(): string;
-    private storeInCache;
-    private deleteFromCache;
 }
 
 export { type ConnectionResult, type GenerateRequest, type GenerateResult, type LoadModelResult, WllamaService, type WllamaServiceConfig };

package/dist/index.d.ts

@@ -13,6 +13,8 @@ interface GenerateRequest {
     temperature?: number;
     topK?: number;
     topP?: number;
+    stop?: string | string[];
+    abortSignal?: AbortSignal;
 }
 interface GenerateResult {
     success: boolean;
@@ -35,44 +37,38 @@ interface ConnectionResult {
 declare class WllamaService {
     private wllama;
     private modelName;
-    private config;
+    private readonly config;
     constructor(config?: WllamaServiceConfig);
-    /** Check browser compatibility and capabilities */
+    /**
+     * Check WebGPU without directly accessing navigator.gpu.
+     *
+     * This avoids requiring @webgpu/types merely for feature detection.
+     */
+    private hasWebGPU;
+    /** Check browser compatibility and capabilities. */
     checkEnvironment(): ConnectionResult;
-    /** Load a GGUF model file from the user's filesystem */
+    /** Load a GGUF model file from the user's filesystem. */
     loadModel(file: File, onProgress?: (progress: number) => void): Promise<LoadModelResult>;
     /**
-     * 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 a chat response using the model's embedded chat template.
+     *
+     * Suitable for instruction-tuned models such as Phi-3 Instruct,
+     * Llama Instruct, ChatML models, and similar chat models.
      */
     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.
+     * Generate a raw completion without applying a chat template.
      *
-     * 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).
+     * Suitable for base completion models, manually formatted prompts,
+     * few-shot completion, fill-in-the-middle, and custom templates.
      */
     generateCompletion(req: GenerateRequest): Promise<GenerateResult>;
-    /** Unload the current model and free memory */
+    /** Unload the current model and release its resources. */
     unload(): Promise<void>;
-    /** Whether a model is currently loaded */
+    /** Whether a model is currently loaded. */
     get isLoaded(): boolean;
-    /** Name of the currently loaded model */
+    /** Name of the currently loaded model. */
     get currentModel(): string;
-    private storeInCache;
-    private deleteFromCache;
 }
 
 export { type ConnectionResult, type GenerateRequest, type GenerateResult, type LoadModelResult, WllamaService, type WllamaServiceConfig };

package/dist/index.js

@@ -36,104 +36,136 @@ var WllamaService = class {
       nCtx: config.nCtx ?? 2048
     };
   }
-  /** Check browser compatibility and capabilities */
+  /**
+   * Check WebGPU without directly accessing navigator.gpu.
+   *
+   * This avoids requiring @webgpu/types merely for feature detection.
+   */
+  hasWebGPU() {
+    return typeof navigator !== "undefined" && "gpu" in navigator;
+  }
+  /** Check browser compatibility and capabilities. */
   checkEnvironment() {
-    if (typeof WebAssembly === "undefined") {
-      return { success: false, error: "WebAssembly is not supported in this browser." };
+    if (typeof window === "undefined" || typeof navigator === "undefined") {
+      return {
+        success: false,
+        error: "WllamaService can only run in a browser environment."
+      };
     }
-    if (!("caches" in window)) {
-      return { success: false, error: "Cache API is not supported in this browser." };
+    if (typeof WebAssembly === "undefined") {
+      return {
+        success: false,
+        error: "WebAssembly is not supported in this browser."
+      };
     }
-    const hasWebGPU = !!navigator.gpu;
-    const isolated = window.crossOriginIsolated;
     return {
       success: true,
-      hasWebGPU,
-      crossOriginIsolated: isolated
+      hasWebGPU: this.hasWebGPU(),
+      crossOriginIsolated: window.crossOriginIsolated
     };
   }
-  /** Load a GGUF model file from the user's filesystem */
+  /** Load a GGUF model file from the user's filesystem. */
   async loadModel(file, onProgress) {
-    const env = this.checkEnvironment();
-    if (!env.success) return { success: false, error: env.error };
+    const environment = this.checkEnvironment();
+    if (!environment.success) {
+      return {
+        success: false,
+        error: environment.error
+      };
+    }
     await this.unload();
     try {
       onProgress?.(10);
-      const configPaths = { default: this.config.wasmPath };
-      this.wllama = new import_esm.Wllama(configPaths);
+      this.wllama = new import_esm.Wllama({
+        default: this.config.wasmPath
+      });
       onProgress?.(30);
-      await this.storeInCache(file);
-      onProgress?.(50);
-      const hasWebGPU = !!navigator.gpu;
+      const useWebGPU = this.hasWebGPU() && this.config.nGpuLayers > 0;
       await this.wllama.loadModel([file], {
         n_ctx: this.config.nCtx,
-        ...hasWebGPU && this.config.nGpuLayers > 0 ? { n_gpu_layers: this.config.nGpuLayers } : {}
+        n_gpu_layers: useWebGPU ? this.config.nGpuLayers : 0,
+        jinja: true
       });
       this.modelName = file.name;
       onProgress?.(100);
-      return { success: true, usedWebGPU: hasWebGPU };
-    } catch (e) {
+      return {
+        success: true,
+        usedWebGPU: useWebGPU
+      };
+    } catch (error) {
       await this.unload();
-      const message = e instanceof Error ? e.message : "Failed to load model";
-      return { success: false, error: message };
+      return {
+        success: false,
+        error: error instanceof Error ? error.message : "Failed to load 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 a chat response using the model's embedded chat template.
+   *
+   * Suitable for instruction-tuned models such as Phi-3 Instruct,
+   * Llama Instruct, ChatML models, and similar chat models.
    */
   async generate(req) {
     if (!this.wllama) {
-      return { success: false, error: "No model loaded. Call loadModel() first." };
+      return {
+        success: false,
+        error: "No model loaded. Call loadModel() first."
+      };
     }
     try {
-      const t0 = performance.now();
+      const startedAt = performance.now();
       const messages = [];
-      if (req.system) messages.push({ role: "system", content: req.system });
-      messages.push({ role: "user", content: req.prompt });
+      if (req.system) {
+        messages.push({
+          role: "system",
+          content: req.system
+        });
+      }
+      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,
-        top_p: req.topP ?? 0.95
+        top_p: req.topP ?? 0.95,
+        stop: req.stop,
+        abortSignal: req.abortSignal
       });
-      const text = response?.choices?.[0]?.message?.content ?? "";
+      const text = response.choices?.[0]?.message?.content ?? "";
       return {
         success: true,
         text,
-        timeMs: Math.round(performance.now() - t0)
+        timeMs: Math.round(
+          performance.now() - startedAt
+        )
+      };
+    } catch (error) {
+      return {
+        success: false,
+        error: error instanceof Error ? error.message : "Generation failed"
       };
-    } catch (e) {
-      const message = e instanceof Error ? e.message : "Generation failed";
-      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.
+   * Generate a raw completion without applying a chat template.
    *
-   * 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).
+   * Suitable for base completion models, manually formatted prompts,
+   * few-shot completion, fill-in-the-middle, and custom templates.
    */
   async generateCompletion(req) {
     if (!this.wllama) {
-      return { success: false, error: "No model loaded. Call loadModel() first." };
+      return {
+        success: false,
+        error: "No model loaded. Call loadModel() first."
+      };
     }
     try {
-      const t0 = performance.now();
+      const startedAt = performance.now();
       const prompt = req.system ? `${req.system}
 
 ${req.prompt}` : req.prompt;
@@ -143,59 +175,50 @@ ${req.prompt}` : req.prompt;
         max_tokens: req.maxTokens ?? 512,
         temperature: req.temperature ?? 0.7,
         top_k: req.topK ?? 40,
-        top_p: req.topP ?? 0.95
+        top_p: req.topP ?? 0.95,
+        stop: req.stop,
+        abortSignal: req.abortSignal
+      });
+      const choice = response.choices?.[0];
+      console.debug("[wllama] raw completion", {
+        text: choice?.text,
+        finishReason: choice?.finish_reason,
+        usage: response.usage,
+        timings: response.timings
       });
-      const text = response?.choices?.[0]?.text ?? "";
       return {
         success: true,
-        text,
-        timeMs: Math.round(performance.now() - t0)
+        text: choice?.text ?? "",
+        timeMs: Math.round(
+          performance.now() - startedAt
+        )
+      };
+    } catch (error) {
+      return {
+        success: false,
+        error: error instanceof Error ? error.message : "Generation failed"
       };
-    } catch (e) {
-      const message = e instanceof Error ? e.message : "Generation failed";
-      return { success: false, error: message };
     }
   }
-  /** Unload the current model and free memory */
+  /** Unload the current model and release its resources. */
   async unload() {
     if (this.wllama) {
       try {
         await this.wllama.exit();
-      } catch (_) {
+      } catch {
       }
       this.wllama = null;
     }
-    if (this.modelName) {
-      try {
-        await this.deleteFromCache(this.modelName);
-      } catch (_) {
-      }
-      this.modelName = "";
-    }
+    this.modelName = "";
   }
-  /** Whether a model is currently loaded */
+  /** Whether a model is currently loaded. */
   get isLoaded() {
     return this.wllama !== null;
   }
-  /** Name of the currently loaded model */
+  /** Name of the currently loaded model. */
   get currentModel() {
     return this.modelName;
   }
-  async storeInCache(file) {
-    const url = `${window.location.origin}/wllama-local-models/${file.name}`;
-    const cache = await caches.open("wllama-local-models");
-    await cache.put(
-      url,
-      new Response(file, {
-        headers: { "Content-Type": "application/octet-stream" }
-      })
-    );
-  }
-  async deleteFromCache(fileName) {
-    const url = `${window.location.origin}/wllama-local-models/${fileName}`;
-    const cache = await caches.open("wllama-local-models");
-    await cache.delete(url);
-  }
 };
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {

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  /**\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":[]}
+{"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';","// src/WllamaService.ts\n\nimport { Wllama } from \"@wllama/wllama/esm/index.js\";\n\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 = \"\";\n  private readonly 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  /**\n   * Check WebGPU without directly accessing navigator.gpu.\n   *\n   * This avoids requiring @webgpu/types merely for feature detection.\n   */\n  private hasWebGPU(): boolean {\n    return typeof navigator !== \"undefined\" && \"gpu\" in navigator;\n  }\n\n  /** Check browser compatibility and capabilities. */\n  checkEnvironment(): ConnectionResult {\n    if (\n      typeof window === \"undefined\" ||\n      typeof navigator === \"undefined\"\n    ) {\n      return {\n        success: false,\n        error: \"WllamaService can only run in a browser environment.\",\n      };\n    }\n\n    if (typeof WebAssembly === \"undefined\") {\n      return {\n        success: false,\n        error: \"WebAssembly is not supported in this browser.\",\n      };\n    }\n\n    return {\n      success: true,\n      hasWebGPU: this.hasWebGPU(),\n      crossOriginIsolated: window.crossOriginIsolated,\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 environment = this.checkEnvironment();\n\n    if (!environment.success) {\n      return {\n        success: false,\n        error: environment.error,\n      };\n    }\n\n    await this.unload();\n\n    try {\n      onProgress?.(10);\n\n      this.wllama = new Wllama({\n        default: this.config.wasmPath,\n      });\n\n      onProgress?.(30);\n\n      const useWebGPU =\n        this.hasWebGPU() && this.config.nGpuLayers > 0;\n\n      await this.wllama.loadModel([file], {\n        n_ctx: this.config.nCtx,\n        n_gpu_layers: useWebGPU\n          ? this.config.nGpuLayers\n          : 0,\n        jinja: true,\n      });\n\n      this.modelName = file.name;\n\n      onProgress?.(100);\n\n      return {\n        success: true,\n        usedWebGPU: useWebGPU,\n      };\n    } catch (error: unknown) {\n      await this.unload();\n\n      return {\n        success: false,\n        error:\n          error instanceof Error\n            ? error.message\n            : \"Failed to load model\",\n      };\n    }\n  }\n\n  /**\n   * Generate a chat response using the model's embedded chat template.\n   *\n   * Suitable for instruction-tuned models such as Phi-3 Instruct,\n   * Llama Instruct, ChatML models, and similar chat models.\n   */\n  async generate(\n    req: GenerateRequest,\n  ): Promise<GenerateResult> {\n    if (!this.wllama) {\n      return {\n        success: false,\n        error: \"No model loaded. Call loadModel() first.\",\n      };\n    }\n\n    try {\n      const startedAt = performance.now();\n\n      const messages: Array<{\n        role: \"system\" | \"user\";\n        content: string;\n      }> = [];\n\n      if (req.system) {\n        messages.push({\n          role: \"system\",\n          content: req.system,\n        });\n      }\n\n      messages.push({\n        role: \"user\",\n        content: req.prompt,\n      });\n\n      /*\n       * Keep the cast here if the installed @wllama/wllama declaration\n       * does not expose the exact createChatCompletion signature.\n       */\n      const response = await (\n        this.wllama as unknown as {\n          createChatCompletion(options: {\n            messages: Array<{\n              role: \"system\" | \"user\";\n              content: string;\n            }>;\n            stream: false;\n            max_tokens: number;\n            temperature: number;\n            top_k: number;\n            top_p: number;\n            stop?: string | string[];\n            abortSignal?: AbortSignal;\n          }): Promise<{\n            choices?: Array<{\n              message?: {\n                content?: string;\n              };\n            }>;\n          }>;\n        }\n      ).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        stop: req.stop,\n        abortSignal: req.abortSignal,\n      });\n\n      const text =\n        response.choices?.[0]?.message?.content ?? \"\";\n\n      return {\n        success: true,\n        text,\n        timeMs: Math.round(\n          performance.now() - startedAt,\n        ),\n      };\n    } catch (error: unknown) {\n      return {\n        success: false,\n        error:\n          error instanceof Error\n            ? error.message\n            : \"Generation failed\",\n      };\n    }\n  }\n\n  /**\n   * Generate a raw completion without applying a chat template.\n   *\n   * Suitable for base completion models, manually formatted prompts,\n   * few-shot completion, fill-in-the-middle, and custom templates.\n   */\n  async generateCompletion(\n    req: GenerateRequest,\n  ): Promise<GenerateResult> {\n    if (!this.wllama) {\n      return {\n        success: false,\n        error: \"No model loaded. Call loadModel() first.\",\n      };\n    }\n\n    try {\n      const startedAt = performance.now();\n\n      const prompt = req.system\n        ? `${req.system}\\n\\n${req.prompt}`\n        : req.prompt;\n\n      const response =\n        await this.wllama.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          stop: req.stop,\n          abortSignal: req.abortSignal,\n        });\n\n      const choice = response.choices?.[0];\n\n      console.debug(\"[wllama] raw completion\", {\n        text: choice?.text,\n        finishReason: choice?.finish_reason,\n        usage: response.usage,\n        timings: response.timings,\n      });\n\n      return {\n        success: true,\n        text: choice?.text ?? \"\",\n        timeMs: Math.round(\n          performance.now() - startedAt,\n        ),\n      };\n    } catch (error: unknown) {\n      return {\n        success: false,\n        error:\n          error instanceof Error\n            ? error.message\n            : \"Generation failed\",\n      };\n    }\n  }\n\n  /** Unload the current model and release its resources. */\n  async unload(): Promise<void> {\n    if (this.wllama) {\n      try {\n        await this.wllama.exit();\n      } catch {\n        // Ignore shutdown errors.\n      }\n\n      this.wllama = null;\n    }\n\n    this.modelName = \"\";\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}"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACEA,iBAAuB;AAUhB,IAAM,gBAAN,MAAoB;AAAA,EAKzB,YAAY,SAA8B,CAAC,GAAG;AAJ9C,SAAQ,SAAwB;AAChC,SAAQ,YAAY;AAIlB,SAAK,SAAS;AAAA,MACZ,UAAU,OAAO,YAAY;AAAA,MAC7B,YAAY,OAAO,cAAc;AAAA,MACjC,MAAM,OAAO,QAAQ;AAAA,IACvB;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOQ,YAAqB;AAC3B,WAAO,OAAO,cAAc,eAAe,SAAS;AAAA,EACtD;AAAA;AAAA,EAGA,mBAAqC;AACnC,QACE,OAAO,WAAW,eAClB,OAAO,cAAc,aACrB;AACA,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OAAO;AAAA,MACT;AAAA,IACF;AAEA,QAAI,OAAO,gBAAgB,aAAa;AACtC,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OAAO;AAAA,MACT;AAAA,IACF;AAEA,WAAO;AAAA,MACL,SAAS;AAAA,MACT,WAAW,KAAK,UAAU;AAAA,MAC1B,qBAAqB,OAAO;AAAA,IAC9B;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,UACJ,MACA,YAC0B;AAC1B,UAAM,cAAc,KAAK,iBAAiB;AAE1C,QAAI,CAAC,YAAY,SAAS;AACxB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OAAO,YAAY;AAAA,MACrB;AAAA,IACF;AAEA,UAAM,KAAK,OAAO;AAElB,QAAI;AACF,mBAAa,EAAE;AAEf,WAAK,SAAS,IAAI,kBAAO;AAAA,QACvB,SAAS,KAAK,OAAO;AAAA,MACvB,CAAC;AAED,mBAAa,EAAE;AAEf,YAAM,YACJ,KAAK,UAAU,KAAK,KAAK,OAAO,aAAa;AAE/C,YAAM,KAAK,OAAO,UAAU,CAAC,IAAI,GAAG;AAAA,QAClC,OAAO,KAAK,OAAO;AAAA,QACnB,cAAc,YACV,KAAK,OAAO,aACZ;AAAA,QACJ,OAAO;AAAA,MACT,CAAC;AAED,WAAK,YAAY,KAAK;AAEtB,mBAAa,GAAG;AAEhB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,YAAY;AAAA,MACd;AAAA,IACF,SAAS,OAAgB;AACvB,YAAM,KAAK,OAAO;AAElB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OACE,iBAAiB,QACb,MAAM,UACN;AAAA,MACR;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,SACJ,KACyB;AACzB,QAAI,CAAC,KAAK,QAAQ;AAChB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OAAO;AAAA,MACT;AAAA,IACF;AAEA,QAAI;AACF,YAAM,YAAY,YAAY,IAAI;AAElC,YAAM,WAGD,CAAC;AAEN,UAAI,IAAI,QAAQ;AACd,iBAAS,KAAK;AAAA,UACZ,MAAM;AAAA,UACN,SAAS,IAAI;AAAA,QACf,CAAC;AAAA,MACH;AAEA,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN,SAAS,IAAI;AAAA,MACf,CAAC;AAMD,YAAM,WAAW,MACf,KAAK,OAqBL,qBAAqB;AAAA,QACrB;AAAA,QACA,QAAQ;AAAA,QACR,YAAY,IAAI,aAAa;AAAA,QAC7B,aAAa,IAAI,eAAe;AAAA,QAChC,OAAO,IAAI,QAAQ;AAAA,QACnB,OAAO,IAAI,QAAQ;AAAA,QACnB,MAAM,IAAI;AAAA,QACV,aAAa,IAAI;AAAA,MACnB,CAAC;AAED,YAAM,OACJ,SAAS,UAAU,CAAC,GAAG,SAAS,WAAW;AAE7C,aAAO;AAAA,QACL,SAAS;AAAA,QACT;AAAA,QACA,QAAQ,KAAK;AAAA,UACX,YAAY,IAAI,IAAI;AAAA,QACtB;AAAA,MACF;AAAA,IACF,SAAS,OAAgB;AACvB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OACE,iBAAiB,QACb,MAAM,UACN;AAAA,MACR;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,mBACJ,KACyB;AACzB,QAAI,CAAC,KAAK,QAAQ;AAChB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OAAO;AAAA,MACT;AAAA,IACF;AAEA,QAAI;AACF,YAAM,YAAY,YAAY,IAAI;AAElC,YAAM,SAAS,IAAI,SACf,GAAG,IAAI,MAAM;AAAA;AAAA,EAAO,IAAI,MAAM,KAC9B,IAAI;AAER,YAAM,WACJ,MAAM,KAAK,OAAO,iBAAiB;AAAA,QACjC;AAAA,QACA,QAAQ;AAAA,QACR,YAAY,IAAI,aAAa;AAAA,QAC7B,aAAa,IAAI,eAAe;AAAA,QAChC,OAAO,IAAI,QAAQ;AAAA,QACnB,OAAO,IAAI,QAAQ;AAAA,QACnB,MAAM,IAAI;AAAA,QACV,aAAa,IAAI;AAAA,MACnB,CAAC;AAEH,YAAM,SAAS,SAAS,UAAU,CAAC;AAEnC,cAAQ,MAAM,2BAA2B;AAAA,QACvC,MAAM,QAAQ;AAAA,QACd,cAAc,QAAQ;AAAA,QACtB,OAAO,SAAS;AAAA,QAChB,SAAS,SAAS;AAAA,MACpB,CAAC;AAED,aAAO;AAAA,QACL,SAAS;AAAA,QACT,MAAM,QAAQ,QAAQ;AAAA,QACtB,QAAQ,KAAK;AAAA,UACX,YAAY,IAAI,IAAI;AAAA,QACtB;AAAA,MACF;AAAA,IACF,SAAS,OAAgB;AACvB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OACE,iBAAiB,QACb,MAAM,UACN;AAAA,MACR;AAAA,IACF;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,SAAwB;AAC5B,QAAI,KAAK,QAAQ;AACf,UAAI;AACF,cAAM,KAAK,OAAO,KAAK;AAAA,MACzB,QAAQ;AAAA,MAER;AAEA,WAAK,SAAS;AAAA,IAChB;AAEA,SAAK,YAAY;AAAA,EACnB;AAAA;AAAA,EAGA,IAAI,WAAoB;AACtB,WAAO,KAAK,WAAW;AAAA,EACzB;AAAA;AAAA,EAGA,IAAI,eAAuB;AACzB,WAAO,KAAK;AAAA,EACd;AACF;","names":[]}

package/dist/index.mjs

@@ -10,104 +10,136 @@ var WllamaService = class {
       nCtx: config.nCtx ?? 2048
     };
   }
-  /** Check browser compatibility and capabilities */
+  /**
+   * Check WebGPU without directly accessing navigator.gpu.
+   *
+   * This avoids requiring @webgpu/types merely for feature detection.
+   */
+  hasWebGPU() {
+    return typeof navigator !== "undefined" && "gpu" in navigator;
+  }
+  /** Check browser compatibility and capabilities. */
   checkEnvironment() {
-    if (typeof WebAssembly === "undefined") {
-      return { success: false, error: "WebAssembly is not supported in this browser." };
+    if (typeof window === "undefined" || typeof navigator === "undefined") {
+      return {
+        success: false,
+        error: "WllamaService can only run in a browser environment."
+      };
     }
-    if (!("caches" in window)) {
-      return { success: false, error: "Cache API is not supported in this browser." };
+    if (typeof WebAssembly === "undefined") {
+      return {
+        success: false,
+        error: "WebAssembly is not supported in this browser."
+      };
     }
-    const hasWebGPU = !!navigator.gpu;
-    const isolated = window.crossOriginIsolated;
     return {
       success: true,
-      hasWebGPU,
-      crossOriginIsolated: isolated
+      hasWebGPU: this.hasWebGPU(),
+      crossOriginIsolated: window.crossOriginIsolated
     };
   }
-  /** Load a GGUF model file from the user's filesystem */
+  /** Load a GGUF model file from the user's filesystem. */
   async loadModel(file, onProgress) {
-    const env = this.checkEnvironment();
-    if (!env.success) return { success: false, error: env.error };
+    const environment = this.checkEnvironment();
+    if (!environment.success) {
+      return {
+        success: false,
+        error: environment.error
+      };
+    }
     await this.unload();
     try {
       onProgress?.(10);
-      const configPaths = { default: this.config.wasmPath };
-      this.wllama = new Wllama(configPaths);
+      this.wllama = new Wllama({
+        default: this.config.wasmPath
+      });
       onProgress?.(30);
-      await this.storeInCache(file);
-      onProgress?.(50);
-      const hasWebGPU = !!navigator.gpu;
+      const useWebGPU = this.hasWebGPU() && this.config.nGpuLayers > 0;
       await this.wllama.loadModel([file], {
         n_ctx: this.config.nCtx,
-        ...hasWebGPU && this.config.nGpuLayers > 0 ? { n_gpu_layers: this.config.nGpuLayers } : {}
+        n_gpu_layers: useWebGPU ? this.config.nGpuLayers : 0,
+        jinja: true
       });
       this.modelName = file.name;
       onProgress?.(100);
-      return { success: true, usedWebGPU: hasWebGPU };
-    } catch (e) {
+      return {
+        success: true,
+        usedWebGPU: useWebGPU
+      };
+    } catch (error) {
       await this.unload();
-      const message = e instanceof Error ? e.message : "Failed to load model";
-      return { success: false, error: message };
+      return {
+        success: false,
+        error: error instanceof Error ? error.message : "Failed to load 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 a chat response using the model's embedded chat template.
+   *
+   * Suitable for instruction-tuned models such as Phi-3 Instruct,
+   * Llama Instruct, ChatML models, and similar chat models.
    */
   async generate(req) {
     if (!this.wllama) {
-      return { success: false, error: "No model loaded. Call loadModel() first." };
+      return {
+        success: false,
+        error: "No model loaded. Call loadModel() first."
+      };
     }
     try {
-      const t0 = performance.now();
+      const startedAt = performance.now();
       const messages = [];
-      if (req.system) messages.push({ role: "system", content: req.system });
-      messages.push({ role: "user", content: req.prompt });
+      if (req.system) {
+        messages.push({
+          role: "system",
+          content: req.system
+        });
+      }
+      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,
-        top_p: req.topP ?? 0.95
+        top_p: req.topP ?? 0.95,
+        stop: req.stop,
+        abortSignal: req.abortSignal
       });
-      const text = response?.choices?.[0]?.message?.content ?? "";
+      const text = response.choices?.[0]?.message?.content ?? "";
       return {
         success: true,
         text,
-        timeMs: Math.round(performance.now() - t0)
+        timeMs: Math.round(
+          performance.now() - startedAt
+        )
+      };
+    } catch (error) {
+      return {
+        success: false,
+        error: error instanceof Error ? error.message : "Generation failed"
       };
-    } catch (e) {
-      const message = e instanceof Error ? e.message : "Generation failed";
-      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.
+   * Generate a raw completion without applying a chat template.
    *
-   * 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).
+   * Suitable for base completion models, manually formatted prompts,
+   * few-shot completion, fill-in-the-middle, and custom templates.
    */
   async generateCompletion(req) {
     if (!this.wllama) {
-      return { success: false, error: "No model loaded. Call loadModel() first." };
+      return {
+        success: false,
+        error: "No model loaded. Call loadModel() first."
+      };
     }
     try {
-      const t0 = performance.now();
+      const startedAt = performance.now();
       const prompt = req.system ? `${req.system}
 
 ${req.prompt}` : req.prompt;
@@ -117,59 +149,50 @@ ${req.prompt}` : req.prompt;
         max_tokens: req.maxTokens ?? 512,
         temperature: req.temperature ?? 0.7,
         top_k: req.topK ?? 40,
-        top_p: req.topP ?? 0.95
+        top_p: req.topP ?? 0.95,
+        stop: req.stop,
+        abortSignal: req.abortSignal
+      });
+      const choice = response.choices?.[0];
+      console.debug("[wllama] raw completion", {
+        text: choice?.text,
+        finishReason: choice?.finish_reason,
+        usage: response.usage,
+        timings: response.timings
       });
-      const text = response?.choices?.[0]?.text ?? "";
       return {
         success: true,
-        text,
-        timeMs: Math.round(performance.now() - t0)
+        text: choice?.text ?? "",
+        timeMs: Math.round(
+          performance.now() - startedAt
+        )
+      };
+    } catch (error) {
+      return {
+        success: false,
+        error: error instanceof Error ? error.message : "Generation failed"
       };
-    } catch (e) {
-      const message = e instanceof Error ? e.message : "Generation failed";
-      return { success: false, error: message };
     }
   }
-  /** Unload the current model and free memory */
+  /** Unload the current model and release its resources. */
   async unload() {
     if (this.wllama) {
       try {
         await this.wllama.exit();
-      } catch (_) {
+      } catch {
       }
       this.wllama = null;
     }
-    if (this.modelName) {
-      try {
-        await this.deleteFromCache(this.modelName);
-      } catch (_) {
-      }
-      this.modelName = "";
-    }
+    this.modelName = "";
   }
-  /** Whether a model is currently loaded */
+  /** Whether a model is currently loaded. */
   get isLoaded() {
     return this.wllama !== null;
   }
-  /** Name of the currently loaded model */
+  /** Name of the currently loaded model. */
   get currentModel() {
     return this.modelName;
   }
-  async storeInCache(file) {
-    const url = `${window.location.origin}/wllama-local-models/${file.name}`;
-    const cache = await caches.open("wllama-local-models");
-    await cache.put(
-      url,
-      new Response(file, {
-        headers: { "Content-Type": "application/octet-stream" }
-      })
-    );
-  }
-  async deleteFromCache(fileName) {
-    const url = `${window.location.origin}/wllama-local-models/${fileName}`;
-    const cache = await caches.open("wllama-local-models");
-    await cache.delete(url);
-  }
 };
 export {
   WllamaService

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  /**\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":[]}
+{"version":3,"sources":["../src/WllamaService.ts"],"sourcesContent":["// src/WllamaService.ts\n\nimport { Wllama } from \"@wllama/wllama/esm/index.js\";\n\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 = \"\";\n  private readonly 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  /**\n   * Check WebGPU without directly accessing navigator.gpu.\n   *\n   * This avoids requiring @webgpu/types merely for feature detection.\n   */\n  private hasWebGPU(): boolean {\n    return typeof navigator !== \"undefined\" && \"gpu\" in navigator;\n  }\n\n  /** Check browser compatibility and capabilities. */\n  checkEnvironment(): ConnectionResult {\n    if (\n      typeof window === \"undefined\" ||\n      typeof navigator === \"undefined\"\n    ) {\n      return {\n        success: false,\n        error: \"WllamaService can only run in a browser environment.\",\n      };\n    }\n\n    if (typeof WebAssembly === \"undefined\") {\n      return {\n        success: false,\n        error: \"WebAssembly is not supported in this browser.\",\n      };\n    }\n\n    return {\n      success: true,\n      hasWebGPU: this.hasWebGPU(),\n      crossOriginIsolated: window.crossOriginIsolated,\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 environment = this.checkEnvironment();\n\n    if (!environment.success) {\n      return {\n        success: false,\n        error: environment.error,\n      };\n    }\n\n    await this.unload();\n\n    try {\n      onProgress?.(10);\n\n      this.wllama = new Wllama({\n        default: this.config.wasmPath,\n      });\n\n      onProgress?.(30);\n\n      const useWebGPU =\n        this.hasWebGPU() && this.config.nGpuLayers > 0;\n\n      await this.wllama.loadModel([file], {\n        n_ctx: this.config.nCtx,\n        n_gpu_layers: useWebGPU\n          ? this.config.nGpuLayers\n          : 0,\n        jinja: true,\n      });\n\n      this.modelName = file.name;\n\n      onProgress?.(100);\n\n      return {\n        success: true,\n        usedWebGPU: useWebGPU,\n      };\n    } catch (error: unknown) {\n      await this.unload();\n\n      return {\n        success: false,\n        error:\n          error instanceof Error\n            ? error.message\n            : \"Failed to load model\",\n      };\n    }\n  }\n\n  /**\n   * Generate a chat response using the model's embedded chat template.\n   *\n   * Suitable for instruction-tuned models such as Phi-3 Instruct,\n   * Llama Instruct, ChatML models, and similar chat models.\n   */\n  async generate(\n    req: GenerateRequest,\n  ): Promise<GenerateResult> {\n    if (!this.wllama) {\n      return {\n        success: false,\n        error: \"No model loaded. Call loadModel() first.\",\n      };\n    }\n\n    try {\n      const startedAt = performance.now();\n\n      const messages: Array<{\n        role: \"system\" | \"user\";\n        content: string;\n      }> = [];\n\n      if (req.system) {\n        messages.push({\n          role: \"system\",\n          content: req.system,\n        });\n      }\n\n      messages.push({\n        role: \"user\",\n        content: req.prompt,\n      });\n\n      /*\n       * Keep the cast here if the installed @wllama/wllama declaration\n       * does not expose the exact createChatCompletion signature.\n       */\n      const response = await (\n        this.wllama as unknown as {\n          createChatCompletion(options: {\n            messages: Array<{\n              role: \"system\" | \"user\";\n              content: string;\n            }>;\n            stream: false;\n            max_tokens: number;\n            temperature: number;\n            top_k: number;\n            top_p: number;\n            stop?: string | string[];\n            abortSignal?: AbortSignal;\n          }): Promise<{\n            choices?: Array<{\n              message?: {\n                content?: string;\n              };\n            }>;\n          }>;\n        }\n      ).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        stop: req.stop,\n        abortSignal: req.abortSignal,\n      });\n\n      const text =\n        response.choices?.[0]?.message?.content ?? \"\";\n\n      return {\n        success: true,\n        text,\n        timeMs: Math.round(\n          performance.now() - startedAt,\n        ),\n      };\n    } catch (error: unknown) {\n      return {\n        success: false,\n        error:\n          error instanceof Error\n            ? error.message\n            : \"Generation failed\",\n      };\n    }\n  }\n\n  /**\n   * Generate a raw completion without applying a chat template.\n   *\n   * Suitable for base completion models, manually formatted prompts,\n   * few-shot completion, fill-in-the-middle, and custom templates.\n   */\n  async generateCompletion(\n    req: GenerateRequest,\n  ): Promise<GenerateResult> {\n    if (!this.wllama) {\n      return {\n        success: false,\n        error: \"No model loaded. Call loadModel() first.\",\n      };\n    }\n\n    try {\n      const startedAt = performance.now();\n\n      const prompt = req.system\n        ? `${req.system}\\n\\n${req.prompt}`\n        : req.prompt;\n\n      const response =\n        await this.wllama.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          stop: req.stop,\n          abortSignal: req.abortSignal,\n        });\n\n      const choice = response.choices?.[0];\n\n      console.debug(\"[wllama] raw completion\", {\n        text: choice?.text,\n        finishReason: choice?.finish_reason,\n        usage: response.usage,\n        timings: response.timings,\n      });\n\n      return {\n        success: true,\n        text: choice?.text ?? \"\",\n        timeMs: Math.round(\n          performance.now() - startedAt,\n        ),\n      };\n    } catch (error: unknown) {\n      return {\n        success: false,\n        error:\n          error instanceof Error\n            ? error.message\n            : \"Generation failed\",\n      };\n    }\n  }\n\n  /** Unload the current model and release its resources. */\n  async unload(): Promise<void> {\n    if (this.wllama) {\n      try {\n        await this.wllama.exit();\n      } catch {\n        // Ignore shutdown errors.\n      }\n\n      this.wllama = null;\n    }\n\n    this.modelName = \"\";\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}"],"mappings":";AAEA,SAAS,cAAc;AAUhB,IAAM,gBAAN,MAAoB;AAAA,EAKzB,YAAY,SAA8B,CAAC,GAAG;AAJ9C,SAAQ,SAAwB;AAChC,SAAQ,YAAY;AAIlB,SAAK,SAAS;AAAA,MACZ,UAAU,OAAO,YAAY;AAAA,MAC7B,YAAY,OAAO,cAAc;AAAA,MACjC,MAAM,OAAO,QAAQ;AAAA,IACvB;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOQ,YAAqB;AAC3B,WAAO,OAAO,cAAc,eAAe,SAAS;AAAA,EACtD;AAAA;AAAA,EAGA,mBAAqC;AACnC,QACE,OAAO,WAAW,eAClB,OAAO,cAAc,aACrB;AACA,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OAAO;AAAA,MACT;AAAA,IACF;AAEA,QAAI,OAAO,gBAAgB,aAAa;AACtC,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OAAO;AAAA,MACT;AAAA,IACF;AAEA,WAAO;AAAA,MACL,SAAS;AAAA,MACT,WAAW,KAAK,UAAU;AAAA,MAC1B,qBAAqB,OAAO;AAAA,IAC9B;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,UACJ,MACA,YAC0B;AAC1B,UAAM,cAAc,KAAK,iBAAiB;AAE1C,QAAI,CAAC,YAAY,SAAS;AACxB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OAAO,YAAY;AAAA,MACrB;AAAA,IACF;AAEA,UAAM,KAAK,OAAO;AAElB,QAAI;AACF,mBAAa,EAAE;AAEf,WAAK,SAAS,IAAI,OAAO;AAAA,QACvB,SAAS,KAAK,OAAO;AAAA,MACvB,CAAC;AAED,mBAAa,EAAE;AAEf,YAAM,YACJ,KAAK,UAAU,KAAK,KAAK,OAAO,aAAa;AAE/C,YAAM,KAAK,OAAO,UAAU,CAAC,IAAI,GAAG;AAAA,QAClC,OAAO,KAAK,OAAO;AAAA,QACnB,cAAc,YACV,KAAK,OAAO,aACZ;AAAA,QACJ,OAAO;AAAA,MACT,CAAC;AAED,WAAK,YAAY,KAAK;AAEtB,mBAAa,GAAG;AAEhB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,YAAY;AAAA,MACd;AAAA,IACF,SAAS,OAAgB;AACvB,YAAM,KAAK,OAAO;AAElB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OACE,iBAAiB,QACb,MAAM,UACN;AAAA,MACR;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,SACJ,KACyB;AACzB,QAAI,CAAC,KAAK,QAAQ;AAChB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OAAO;AAAA,MACT;AAAA,IACF;AAEA,QAAI;AACF,YAAM,YAAY,YAAY,IAAI;AAElC,YAAM,WAGD,CAAC;AAEN,UAAI,IAAI,QAAQ;AACd,iBAAS,KAAK;AAAA,UACZ,MAAM;AAAA,UACN,SAAS,IAAI;AAAA,QACf,CAAC;AAAA,MACH;AAEA,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN,SAAS,IAAI;AAAA,MACf,CAAC;AAMD,YAAM,WAAW,MACf,KAAK,OAqBL,qBAAqB;AAAA,QACrB;AAAA,QACA,QAAQ;AAAA,QACR,YAAY,IAAI,aAAa;AAAA,QAC7B,aAAa,IAAI,eAAe;AAAA,QAChC,OAAO,IAAI,QAAQ;AAAA,QACnB,OAAO,IAAI,QAAQ;AAAA,QACnB,MAAM,IAAI;AAAA,QACV,aAAa,IAAI;AAAA,MACnB,CAAC;AAED,YAAM,OACJ,SAAS,UAAU,CAAC,GAAG,SAAS,WAAW;AAE7C,aAAO;AAAA,QACL,SAAS;AAAA,QACT;AAAA,QACA,QAAQ,KAAK;AAAA,UACX,YAAY,IAAI,IAAI;AAAA,QACtB;AAAA,MACF;AAAA,IACF,SAAS,OAAgB;AACvB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OACE,iBAAiB,QACb,MAAM,UACN;AAAA,MACR;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,mBACJ,KACyB;AACzB,QAAI,CAAC,KAAK,QAAQ;AAChB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OAAO;AAAA,MACT;AAAA,IACF;AAEA,QAAI;AACF,YAAM,YAAY,YAAY,IAAI;AAElC,YAAM,SAAS,IAAI,SACf,GAAG,IAAI,MAAM;AAAA;AAAA,EAAO,IAAI,MAAM,KAC9B,IAAI;AAER,YAAM,WACJ,MAAM,KAAK,OAAO,iBAAiB;AAAA,QACjC;AAAA,QACA,QAAQ;AAAA,QACR,YAAY,IAAI,aAAa;AAAA,QAC7B,aAAa,IAAI,eAAe;AAAA,QAChC,OAAO,IAAI,QAAQ;AAAA,QACnB,OAAO,IAAI,QAAQ;AAAA,QACnB,MAAM,IAAI;AAAA,QACV,aAAa,IAAI;AAAA,MACnB,CAAC;AAEH,YAAM,SAAS,SAAS,UAAU,CAAC;AAEnC,cAAQ,MAAM,2BAA2B;AAAA,QACvC,MAAM,QAAQ;AAAA,QACd,cAAc,QAAQ;AAAA,QACtB,OAAO,SAAS;AAAA,QAChB,SAAS,SAAS;AAAA,MACpB,CAAC;AAED,aAAO;AAAA,QACL,SAAS;AAAA,QACT,MAAM,QAAQ,QAAQ;AAAA,QACtB,QAAQ,KAAK;AAAA,UACX,YAAY,IAAI,IAAI;AAAA,QACtB;AAAA,MACF;AAAA,IACF,SAAS,OAAgB;AACvB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,OACE,iBAAiB,QACb,MAAM,UACN;AAAA,MACR;AAAA,IACF;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,SAAwB;AAC5B,QAAI,KAAK,QAAQ;AACf,UAAI;AACF,cAAM,KAAK,OAAO,KAAK;AAAA,MACzB,QAAQ;AAAA,MAER;AAEA,WAAK,SAAS;AAAA,IAChB;AAEA,SAAK,YAAY;AAAA,EACnB;AAAA;AAAA,EAGA,IAAI,WAAoB;AACtB,WAAO,KAAK,WAAW;AAAA,EACzB;AAAA;AAAA,EAGA,IAAI,eAAuB;AACzB,WAAO,KAAK;AAAA,EACd;AACF;","names":[]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "wllama-service",
-  "version": "2.0.5",
+  "version": "2.0.6",
   "description": "Framework-agnostic browser LLM service wrapper with WebGPU and WebAssembly support",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",