Upload space/app.py with huggingface_hub

space/app.py

@@ -192,7 +192,7 @@ def load_model():
         cfg = mod.preset_config(
             run_args.get("model_size", "1.7b"),
             len(tok),
-            # Pretrain was 4096; chat SFT args say 2048 — serve at full pretrain window
+            # Pretrain was 4096; chat SFT now 4096 — match pretrain/Gradio window
             int(os.environ.get("BLOCK_SIZE", "4096")),
             0.0,
             run_args.get("pcs_a", 0.8309193524478643),