prompt-api-polyfill 1.2.0 → 1.2.1

package/README.md

@@ -140,6 +140,17 @@ npm install prompt-api-polyfill
     apiKey: 'dummy', // Required for now by the loader
     device: 'webgpu', // 'webgpu' or 'cpu'
     dtype: 'q4f16', // Quantization level
+    env: {
+      // Optional: Pass low-level Transformers.js environment overrides
+      allowRemoteModels: false,
+      backends: {
+        onnx: {
+          wasm: {
+            wasmPaths: 'https://cdn.example.com/wasm-assets/',
+          },
+        },
+      },
+    },
   };
 
   if (!('LanguageModel' in window)) {
@@ -225,6 +236,18 @@ This repo ships with a template file:
   // For Transformers.js:
   "device": "webgpu",
   "dtype": "q4f16",
+
+  // Optional library-level overrides:
+  "env": {
+    "allowRemoteModels": false,
+    "backends": {
+      "onnx": {
+        "wasm": {
+          "wasmPaths": "https://cdn.example.com/wasm-assets/"
+        }
+      }
+    }
+  }
 }
 ```
 
@@ -281,7 +304,17 @@ Then open `.env.json` and fill in the values.
   "apiKey": "dummy",
   "modelName": "onnx-community/gemma-3-1b-it-ONNX-GQA",
   "device": "webgpu",
-  "dtype": "q4f16"
+  "dtype": "q4f16",
+  "env": {
+    "allowRemoteModels": false,
+    "backends": {
+      "onnx": {
+        "wasm": {
+          "wasmPaths": "https://cdn.example.com/wasm-assets/"
+        }
+      }
+    }
+  }
 }
 ```
 
@@ -304,6 +337,9 @@ Then open `.env.json` and fill in the values.
 
 - `device`: **Transformers.js only**. Either `"webgpu"` or `"cpu"`.
 - `dtype`: **Transformers.js only**. Quantization level (e.g., `"q4f16"`).
+- `env` (optional): **Transformers.js only**. A flexible object to override
+  [Transformers.js environment variables](https://huggingface.co/docs/transformers.js/api/env).
+  This is useful for specifying local `wasmPaths` or proxy settings.
 
 - `modelName` (optional): The model ID to use. If not provided, the polyfill
   uses the defaults defined in [`backends/defaults.js`](backends/defaults.js).

package/dist/backends/transformers.js

@@ -1,68 +1,74 @@
-import { pipeline as z, TextStreamer as C } from "@huggingface/transformers";
-import { P as j, D as y } from "../chunks/defaults-CNQngzSd.js";
-class D extends j {
+import { env as j, pipeline as z, TextStreamer as C } from "@huggingface/transformers";
+import { P as $, D as y } from "../chunks/defaults-CNQngzSd.js";
+class I extends $ {
   #e;
   #t;
-  #a;
-  #n;
+  #i;
   #o;
-  constructor(n = {}) {
-    super(n.modelName || y.transformers.modelName), this.#a = n.device || y.transformers.device, this.#n = n.dtype || y.transformers.dtype;
+  #n;
+  constructor(o = {}) {
+    if (super(o.modelName || y.transformers.modelName), this.#i = o.device || y.transformers.device, this.#o = o.dtype || y.transformers.dtype, o.env) {
+      const t = (n, a) => {
+        for (const [l, e] of Object.entries(a))
+          e && typeof e == "object" && !Array.isArray(e) && n[l] && typeof n[l] == "object" ? t(n[l], e) : n[l] = e;
+      };
+      t(j, o.env);
+    }
   }
   /**
    * Loaded models can be large, so we initialize them lazily.
    * @param {EventTarget} [monitorTarget] - The event target to dispatch download progress events to.
    * @returns {Promise<Object>} The generator.
    */
-  async #s(n) {
+  async #s(o) {
     if (!this.#e) {
-      const t = /* @__PURE__ */ new Map(), a = await $(this.modelName, {
-        dtype: this.#n
+      const t = /* @__PURE__ */ new Map(), n = await M(this.modelName, {
+        dtype: this.#o
       });
-      for (const { path: e, size: o } of a)
-        t.set(e, { loaded: 0, total: o });
-      const i = (e) => {
-        if (!n)
+      for (const { path: e, size: r } of n)
+        t.set(e, { loaded: 0, total: r });
+      const a = (e) => {
+        if (!o)
           return;
-        const o = 1 / 65536, c = Math.floor(e / o) * o;
-        c <= n.__lastProgressLoaded || (n.dispatchEvent(
+        const r = 1 / 65536, c = Math.floor(e / r) * r;
+        c <= o.__lastProgressLoaded || (o.dispatchEvent(
           new ProgressEvent("downloadprogress", {
             loaded: c,
             total: 1,
             lengthComputable: !0
           })
-        ), n.__lastProgressLoaded = c);
+        ), o.__lastProgressLoaded = c);
       }, l = (e) => {
         if (e.status === "initiate")
           if (t.has(e.file)) {
-            const o = t.get(e.file);
-            e.total && (o.total = e.total);
+            const r = t.get(e.file);
+            e.total && (r.total = e.total);
           } else
             t.set(e.file, { loaded: 0, total: e.total || 0 });
         else if (e.status === "progress")
           t.has(e.file) && (t.get(e.file).loaded = e.loaded);
         else if (e.status === "done") {
           if (t.has(e.file)) {
-            const o = t.get(e.file);
-            o.loaded = o.total;
+            const r = t.get(e.file);
+            r.loaded = r.total;
           }
         } else if (e.status === "ready") {
-          i(1);
+          a(1);
           return;
         }
         if (e.status === "progress" || e.status === "done") {
-          let o = 0, c = 0;
+          let r = 0, c = 0;
           for (const { loaded: f, total: d } of t.values())
-            o += f, c += d;
+            r += f, c += d;
           if (c > 0) {
-            const f = o / c;
-            i(Math.min(f, 0.9999));
+            const f = r / c;
+            a(Math.min(f, 0.9999));
           }
         }
       };
-      i(0), this.#e = await z("text-generation", this.modelName, {
-        device: this.#a,
-        dtype: this.#n,
+      a(0), this.#e = await z("text-generation", this.modelName, {
+        device: this.#i,
+        dtype: this.#o,
         progress_callback: l
       }), this.#t = this.#e.tokenizer;
     }
@@ -73,9 +79,9 @@ class D extends j {
    * @param {Object} options - LanguageModel options.
    * @returns {string} 'available' or 'unavailable'.
    */
-  static availability(n) {
-    if (n?.expectedInputs && Array.isArray(n.expectedInputs)) {
-      for (const t of n.expectedInputs)
+  static availability(o) {
+    if (o?.expectedInputs && Array.isArray(o.expectedInputs)) {
+      for (const t of o.expectedInputs)
         if (t.type === "audio" || t.type === "image")
           return "unavailable";
     }
@@ -88,55 +94,55 @@ class D extends j {
    * @param {EventTarget} [monitorTarget] - The event target to dispatch download progress events to.
    * @returns {Promise<Object>} The generator.
    */
-  async createSession(n, t, a) {
-    return n.responseConstraint && console.warn(
+  async createSession(o, t, n) {
+    return o.responseConstraint && console.warn(
       "The `responseConstraint` flag isn't supported by the Transformers.js backend and was ignored."
-    ), await this.#s(a), this.generationConfig = {
+    ), await this.#s(n), this.generationConfig = {
       max_new_tokens: 512,
       // Default limit
       temperature: t.generationConfig?.temperature ?? 1,
       top_p: 1,
       do_sample: t.generationConfig?.temperature !== 0,
       return_full_text: !1
-    }, this.#o = t.systemInstruction, this.#e;
+    }, this.#n = t.systemInstruction, this.#e;
   }
-  async generateContent(n) {
-    const t = await this.#s(), a = this.#r(n), i = this.#t.apply_chat_template(a, {
+  async generateContent(o) {
+    const t = await this.#s(), n = this.#r(o), a = this.#t.apply_chat_template(n, {
       tokenize: !1,
       add_generation_prompt: !0
-    }), e = (await t(i, {
+    }), e = (await t(a, {
       ...this.generationConfig,
       add_special_tokens: !1
-    }))[0].generated_text, o = await this.countTokens(n);
-    return { text: e, usage: o };
+    }))[0].generated_text, r = await this.countTokens(o);
+    return { text: e, usage: r };
   }
-  async generateContentStream(n) {
-    const t = await this.#s(), a = this.#r(n), i = this.#t.apply_chat_template(a, {
+  async generateContentStream(o) {
+    const t = await this.#s(), n = this.#r(o), a = this.#t.apply_chat_template(n, {
       tokenize: !1,
       add_generation_prompt: !0
     }), l = [];
-    let e, o = new Promise((r) => e = r), c = !1;
-    const f = (r) => {
-      l.push(r), e && (e(), e = null);
+    let e, r = new Promise((i) => e = i), c = !1;
+    const f = (i) => {
+      l.push(i), e && (e(), e = null);
     }, d = new C(this.#t, {
       skip_prompt: !0,
       skip_special_tokens: !0,
       callback_function: f
     });
-    return t(i, {
+    return t(a, {
       ...this.generationConfig,
       add_special_tokens: !1,
       streamer: d
     }).then(() => {
       c = !0, e && (e(), e = null);
-    }).catch((r) => {
-      console.error("[Transformers.js] Generation error:", r), c = !0, e && (e(), e = null);
+    }).catch((i) => {
+      console.error("[Transformers.js] Generation error:", i), c = !0, e && (e(), e = null);
     }), (async function* () {
       for (; ; ) {
-        for (l.length === 0 && !c && (e || (o = new Promise((r) => e = r)), await o); l.length > 0; ) {
-          const r = l.shift();
+        for (l.length === 0 && !c && (e || (r = new Promise((i) => e = i)), await r); l.length > 0; ) {
+          const i = l.shift();
           yield {
-            text: () => r,
+            text: () => i,
             usageMetadata: { totalTokenCount: 0 }
           };
         }
@@ -145,67 +151,67 @@ class D extends j {
       }
     })();
   }
-  async countTokens(n) {
+  async countTokens(o) {
     await this.#s();
-    const t = this.#r(n);
+    const t = this.#r(o);
     return this.#t.apply_chat_template(t, {
       tokenize: !0,
       add_generation_prompt: !1,
       return_tensor: !1
     }).length;
   }
-  #r(n) {
-    const t = n.map((a) => {
-      let i = a.role === "model" ? "assistant" : a.role === "system" ? "system" : "user";
-      const l = a.parts.map((e) => e.text).join("");
-      return { role: i, content: l };
+  #r(o) {
+    const t = o.map((n) => {
+      let a = n.role === "model" ? "assistant" : n.role === "system" ? "system" : "user";
+      const l = n.parts.map((e) => e.text).join("");
+      return { role: a, content: l };
     });
-    if (this.#o && !t.some((a) => a.role === "system") && t.unshift({ role: "system", content: this.#o }), this.modelName.toLowerCase().includes("gemma")) {
-      const a = t.findIndex((i) => i.role === "system");
-      if (a !== -1) {
-        const i = t[a], l = t.findIndex(
-          (e, o) => e.role === "user" && o > a
+    if (this.#n && !t.some((n) => n.role === "system") && t.unshift({ role: "system", content: this.#n }), this.modelName.toLowerCase().includes("gemma")) {
+      const n = t.findIndex((a) => a.role === "system");
+      if (n !== -1) {
+        const a = t[n], l = t.findIndex(
+          (e, r) => e.role === "user" && r > n
         );
-        l !== -1 ? (t[l].content = i.content + `
+        l !== -1 ? (t[l].content = a.content + `
 
-` + t[l].content, t.splice(a, 1)) : (i.content += `
+` + t[l].content, t.splice(n, 1)) : (a.content += `
 
-`, i.role = "user");
+`, a.role = "user");
       }
     }
     return t;
   }
 }
-async function $(g, n = {}) {
-  const { dtype: t = "q8", branch: a = "main" } = n;
-  let i = null;
-  const l = `transformers_model_files_${g}_${t}_${a}`;
+async function M(g, o = {}) {
+  const { dtype: t = "q8", branch: n = "main" } = o;
+  let a = null;
+  const l = `transformers_model_files_${g}_${t}_${n}`;
   try {
     const s = localStorage.getItem(l);
     if (s) {
-      i = JSON.parse(s);
-      const { timestamp: p, files: u } = i, v = 1440 * 60 * 1e3;
+      a = JSON.parse(s);
+      const { timestamp: p, files: u } = a, v = 1440 * 60 * 1e3;
       if (Date.now() - p < v)
         return u;
     }
   } catch (s) {
     console.warn("Failed to read from localStorage cache:", s);
   }
-  const e = `https://huggingface.co/api/models/${g}/tree/${a}?recursive=true`;
-  let o;
+  const e = `https://huggingface.co/api/models/${g}/tree/${n}?recursive=true`;
+  let r;
   try {
-    if (o = await fetch(e), !o.ok)
-      throw new Error(`Manifest fetch failed: ${o.status}`);
+    if (r = await fetch(e), !r.ok)
+      throw new Error(`Manifest fetch failed: ${r.status}`);
   } catch (s) {
-    if (i)
+    if (a)
       return console.warn(
         "Failed to fetch manifest from network, falling back to cached data (expired):",
         s
-      ), i.files;
+      ), a.files;
     throw s;
   }
-  const c = await o.json(), f = new Map(c.map((s) => [s.path, s.size])), d = [], h = (s) => f.has(s), r = (s) => h(s) ? (d.push({ path: s, size: f.get(s) }), !0) : !1;
-  r("config.json"), r("generation_config.json"), r("preprocessor_config.json"), h("tokenizer.json") ? (r("tokenizer.json"), r("tokenizer_config.json")) : (r("tokenizer_config.json"), r("special_tokens_map.json"), r("vocab.json"), r("merges.txt"), r("vocab.txt"));
+  const c = await r.json(), f = new Map(c.map((s) => [s.path, s.size])), d = [], h = (s) => f.has(s), i = (s) => h(s) ? (d.push({ path: s, size: f.get(s) }), !0) : !1;
+  i("config.json"), i("generation_config.json"), i("preprocessor_config.json"), h("tokenizer.json") ? (i("tokenizer.json"), i("tokenizer_config.json")) : (i("tokenizer_config.json"), i("special_tokens_map.json"), i("vocab.json"), i("merges.txt"), i("vocab.txt"));
   const w = "onnx";
   let m = [];
   t === "fp32" ? m = [""] : t === "quantized" ? m = ["_quantized"] : (m = [`_${t}`], t === "q8" && m.push(""));
@@ -228,11 +234,11 @@ async function $(g, n = {}) {
     (s) => s.includes("decoder_model_merged")
   ), b = _.filter((s) => !(x && s.includes("decoder_model") && !s.includes("merged")));
   for (const s of b) {
-    r(s);
+    i(s);
     const p = `${s}_data`;
-    if (r(p)) {
+    if (i(p)) {
       let u = 1;
-      for (; r(`${p}_${u}`); )
+      for (; i(`${p}_${u}`); )
         u++;
     }
   }
@@ -250,5 +256,5 @@ async function $(g, n = {}) {
   return d;
 }
 export {
-  D as default
+  I as default
 };

package/dot_env.json

@@ -7,5 +7,15 @@
   "reCaptchaSiteKey": "",
   "useLimitedUseAppCheckTokens": true,
   "device": "webgpu",
-  "dtype": "q4f16"
+  "dtype": "q4f16",
+  "env": {
+    "allowRemoteModels": true,
+    "backends": {
+      "onnx": {
+        "wasm": {
+          "wasmPaths": ""
+        }
+      }
+    }
+  }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "prompt-api-polyfill",
-  "version": "1.2.0",
+  "version": "1.2.1",
   "description": "Polyfill for the Prompt API (`LanguageModel`) backed by Firebase AI Logic, Gemini API, OpenAI API, or Transformers.js.",
   "type": "module",
   "main": "./dist/prompt-api-polyfill.js",
@@ -49,7 +49,7 @@
     "vite": "^7.3.1"
   },
   "dependencies": {
-    "@google/genai": "^1.41.0",
+    "@google/genai": "^1.42.0",
     "@huggingface/transformers": "^3.8.1",
     "firebase": "^12.9.0",
     "openai": "^6.22.0"