perf: run both models in parallel with ThreadPoolExecutor

Whichever finishes first shows first. No more sequential waiting.
If Buleyean is faster, it appears before the base model.

app.py

@@ -6,6 +6,7 @@ Base model vs Void-trained model. Live inference. Nothing hardcoded.
 import gradio as gr
 import torch
 import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 print("[Void] Loading base model...", flush=True)
@@ -53,13 +54,42 @@ def compare(prompt):
         yield "", "", "", ""
         return
 
-    base_text, base_time, base_toks, base_ms = gen(prompt, base_model, base_tokenizer)
-    base_stats = f"{base_toks} tokens in {base_time:.1f}s ({base_ms:.0f}ms/tok)"
-    yield base_text, "generating...", base_stats, "running..."
-
-    bule_text, bule_time, bule_toks, bule_ms = gen(prompt, bule_model, bule_tokenizer)
-    bule_stats = f"{bule_toks} tokens in {bule_time:.1f}s ({bule_ms:.0f}ms/tok)"
-    yield base_text, bule_text, base_stats, bule_stats
+    # Run both in parallel -- whichever finishes first shows first
+    base_result = [None]
+    bule_result = [None]
+
+    def run_base():
+        base_result[0] = gen(prompt, base_model, base_tokenizer)
+
+    def run_bule():
+        bule_result[0] = gen(prompt, bule_model, bule_tokenizer)
+
+    with ThreadPoolExecutor(max_workers=2) as pool:
+        futures = {
+            pool.submit(run_base): "base",
+            pool.submit(run_bule): "bule",
+        }
+        for future in as_completed(futures):
+            name = futures[future]
+            future.result()  # raise if error
+            if name == "base" and base_result[0]:
+                text, t, toks, ms = base_result[0]
+                stats = f"{toks} tokens in {t:.1f}s ({ms:.0f}ms/tok)"
+                bule_text = bule_result[0][0] if bule_result[0] else "generating..."
+                bule_s = f"{bule_result[0][2]} tokens in {bule_result[0][1]:.1f}s ({bule_result[0][3]:.0f}ms/tok)" if bule_result[0] else "running..."
+                yield text, bule_text, stats, bule_s
+            elif name == "bule" and bule_result[0]:
+                text, t, toks, ms = bule_result[0]
+                stats = f"{toks} tokens in {t:.1f}s ({ms:.0f}ms/tok)"
+                base_text = base_result[0][0] if base_result[0] else "generating..."
+                base_s = f"{base_result[0][2]} tokens in {base_result[0][1]:.1f}s ({base_result[0][3]:.0f}ms/tok)" if base_result[0] else "running..."
+                yield base_text, text, base_s, stats
+
+    # Final yield with both results
+    if base_result[0] and bule_result[0]:
+        bt, b_t, b_toks, b_ms = base_result[0]
+        vt, v_t, v_toks, v_ms = bule_result[0]
+        yield bt, vt, f"{b_toks} tokens in {b_t:.1f}s ({b_ms:.0f}ms/tok)", f"{v_toks} tokens in {v_t:.1f}s ({v_ms:.0f}ms/tok)"
 
 
 CSS = """