Upload folder using huggingface_hub

live.html

@@ -79,7 +79,7 @@ the rejection sampling below is the actual algorithm on actual model logits.
 <a href="index.html">← back to the interactive explainer</a></p>
 
 <div class="card">
-  <button id="loadBtn">Load models (~70 MB, cached after first visit)</button>
+  <button id="loadBtn">Load models (~215 MB, cached after first visit)</button>
   <span id="status"></span>
   <div id="loadWarn" class="warn"></div>
 </div>
@@ -123,9 +123,19 @@ a 70B cloud cross-check) you can run locally with <code>python app.py</code>.</p
 </main>
 
 <script type="module">
-import { AutoTokenizer, AutoModelForCausalLM, Tensor }
+import { AutoTokenizer, AutoModelForCausalLM, Tensor, env }
   from "https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.5.2";
 
+// ?local=1 → serve models from ./models/ (offline mirrors, CI testing)
+if (new URLSearchParams(location.search).get("local")) {
+  env.allowRemoteModels = false;
+  env.allowLocalModels = true;
+  env.localModelPath = "./models/";
+}
+// pin the small legacy quantized files — plain {dtype:"q8"} resolves to the
+// much larger model_quantized.onnx (237 MB) on these repos
+const MODEL_OPTS = { model_file_name: "decoder_model_merged", dtype: "q8" };
+
 const $ = id => document.getElementById(id);
 const esc = s => s.replace(/&/g,"&amp;").replace(/</g,"&lt;").replace(/>/g,"&gt;");
 const vis = t => t.trim() === "" ? t.replace(/ /g,"·") || "·" : t;
@@ -187,18 +197,18 @@ $("loadBtn").onclick = async () => {
     tok = await AutoTokenizer.from_pretrained("Xenova/distilgpt2");
     status("loading draft (distilgpt2)…");
     draft = await AutoModelForCausalLM.from_pretrained("Xenova/distilgpt2",
-      { dtype: "q8", device, progress_callback: progress });
+      { ...MODEL_OPTS, device, progress_callback: progress });
     status("loading target (gpt2)…");
     target = await AutoModelForCausalLM.from_pretrained("Xenova/gpt2",
-      { dtype: "q8", device, progress_callback: progress });
+      { ...MODEL_OPTS, device, progress_callback: progress });
   } catch (e) {
     if (device === "webgpu"){ // some GPUs fail on q8 — retry on wasm
       $("loadWarn").textContent = "WebGPU load failed, retrying on WASM…";
       device = "wasm";
       draft = await AutoModelForCausalLM.from_pretrained("Xenova/distilgpt2",
-        { dtype: "q8", device, progress_callback: progress });
+        { ...MODEL_OPTS, device, progress_callback: progress });
       target = await AutoModelForCausalLM.from_pretrained("Xenova/gpt2",
-        { dtype: "q8", device, progress_callback: progress });
+        { ...MODEL_OPTS, device, progress_callback: progress });
     } else { status("load failed: " + e.message); $("loadBtn").disabled = false; return; }
   }
   status(`ready · running on ${device}`);