Upload 2 files

app.py

@@ -2862,6 +2862,9 @@ def _chat_stream(history, version, ckpt_label, mode_key, use_custom,
         }
     else:
         cfg = dict(MODES[mode_key])
+        # Max new tokens slider always applies (independent of preset override)
+        cfg["max_new_tokens"] = int(max_tokens)
+        cfg["context_window"] = int(ctx_win)
 
     if prior_msgs:
         prompt = _build_conversation_prompt(prior_msgs, new_msg)
@@ -2903,6 +2906,9 @@ def _compare_fn(prompt, selected_versions, mode_key, use_custom,
         }
     else:
         cfg = dict(MODES[mode_key])
+        # Max new tokens slider always applies (independent of preset override)
+        cfg["max_new_tokens"] = int(max_tokens)
+        cfg["context_window"] = int(ctx_win)
 
     all_versions = _collection_versions()
     selected = set(selected_versions or [])
@@ -3047,8 +3053,8 @@ def _advanced_block():
             soft_cap = gr.Slider(0.0, 50.0, value=20.0, step=0.5, label="Logit soft cap")
             loop_pen = gr.Slider(0.0, 50.0, value=15.0, step=0.5, label="Loop penalty")
         with gr.Row():
-            max_tokens = gr.Slider(16, 2048, value=512, step=16, label="Max new tokens")
-            ctx_win = gr.Slider(128, 2048, value=2048, step=128, label="Context window")
+            max_tokens = gr.Slider(16, 4096, value=512, step=16, label="Max new tokens")
+            ctx_win = gr.Slider(128, 4096, value=2048, step=128, label="Context window")
     return use_custom, temperature, top_k, top_p, min_p, rep_penalty, ngram_size, soft_cap, loop_pen, max_tokens, ctx_win, raw_mode