feat: switch assistant VLM Qwen3.5-4B -> gemma-4-E4B-it-Uncensored-MAX

Comment out the Qwen model (kept for revert) and load prithivMLmods/
gemma-4-E4B-it-Uncensored-MAX (Gemma4ForConditionalGeneration, multimodal).
Make the eos-token resolver model-agnostic: union the model's own
generation_config eos, the tokenizer eos, and known turn terminators across
families (Gemma <turn|>, Qwen <|im_end|>, Llama <|eot_id|>), skipping <unk>.
Gemma ships a generation_config so it stops natively; the union keeps the
manual-pin safety for models that don't. Update assistant UI copy; Gemma has
no thinking mode so the Reasoning toggle is now a documented no-op.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app.py

@@ -140,14 +140,17 @@ noobxl_pipe.to("cuda")
 print("Pipelines loaded!")
 
 # -----------------------------------------------------------------------------
-# Qwen3.5-4B multimodal assistant. Powers the "Prompt Assistant" tab: one-turn
-# question answering with an optional image, for turning rough ideas into rich
-# prompts or describing a reference image. Loaded once, bf16 (~9-10 GB) — fits
-# alongside the two diffusion pipelines on the shared GPU.
+# Multimodal assistant. Powers the "Prompt Assistant" tab: one-turn question
+# answering with an optional image, for turning rough ideas into rich prompts or
+# describing a reference image. Loaded once, bf16 — fits alongside the two
+# diffusion pipelines on the shared GPU.
 # -----------------------------------------------------------------------------
-VLM_MODEL_ID = "rodrigomt/Qwen3.5-4B-Uncensored-Aggressive"
+# Previous model (Qwen3.5-4B) — kept for easy revert. It shipped no
+# generation_config.json (needed manual eos pinning, see _resolve_vlm_eos_ids).
+# VLM_MODEL_ID = "rodrigomt/Qwen3.5-4B-Uncensored-Aggressive"
+VLM_MODEL_ID = "prithivMLmods/gemma-4-E4B-it-Uncensored-MAX"
 
-print("Loading Qwen3.5-4B assistant...")
+print(f"Loading assistant: {VLM_MODEL_ID} ...")
 vlm_processor = AutoProcessor.from_pretrained(
     VLM_MODEL_ID, token=HF_TOKEN, trust_remote_code=True
 )
@@ -159,17 +162,37 @@ vlm_model = AutoModelForImageTextToText.from_pretrained(
 )
 vlm_model.to("cuda").eval()
 
-# This finetune ships NO generation_config.json and its config carries no
-# eos_token_id, so generate() has no stop token and runs to max_new_tokens
-# (endless output). Pin the stop tokens explicitly from the tokenizer: the chat
-# template ends each assistant turn with <|im_end|>, so that's the real
-# terminator (plus <|endoftext|> as a fallback).
+
+def _resolve_vlm_eos_ids(tokenizer, model):
+    """Collect every stop token id for the active VLM, model-agnostically.
+
+    generate() stops on these. We union three sources so a model swap can't
+    silently lose its terminator: (1) the model's own generation_config (correct
+    for models that ship one, e.g. Gemma's <turn|>/<eos>), (2) the tokenizer's
+    eos, and (3) known chat-turn terminators across families (Gemma's <turn|>,
+    Qwen's <|im_end|>, Llama's <|eot_id|>). Tokens absent from the vocab resolve
+    to <unk> and are skipped. This is what stops a model like the old Qwen
+    finetune (no generation_config) from rambling to max_new_tokens.
+    """
+    ids = set()
+    gc = getattr(model, "generation_config", None)
+    gc_eos = getattr(gc, "eos_token_id", None) if gc is not None else None
+    if isinstance(gc_eos, int):
+        ids.add(gc_eos)
+    elif isinstance(gc_eos, (list, tuple)):
+        ids.update(int(x) for x in gc_eos)
+    if tokenizer.eos_token_id is not None:
+        ids.add(int(tokenizer.eos_token_id))
+    unk = tokenizer.unk_token_id
+    for t in ("<turn|>", "<end_of_turn>", "<|im_end|>", "<|eot_id|>", "<|endoftext|>"):
+        tid = tokenizer.convert_tokens_to_ids(t)
+        if isinstance(tid, int) and tid >= 0 and tid != unk:
+            ids.add(tid)
+    return sorted(ids)
+
+
 _vlm_tokenizer = getattr(vlm_processor, "tokenizer", vlm_processor)
-_VLM_EOS_IDS = sorted({
-    tid for tok in ("<|im_end|>", "<|endoftext|>")
-    for tid in (_vlm_tokenizer.convert_tokens_to_ids(tok),)
-    if isinstance(tid, int) and tid >= 0
-} | ({_vlm_tokenizer.eos_token_id} if _vlm_tokenizer.eos_token_id is not None else set()))
+_VLM_EOS_IDS = _resolve_vlm_eos_ids(_vlm_tokenizer, vlm_model)
 _VLM_PAD_ID = (
     _vlm_tokenizer.pad_token_id
     if _vlm_tokenizer.pad_token_id is not None
@@ -1306,7 +1329,7 @@ with gr.Blocks(fill_height=True) as demo:
 
         with gr.Tab("💬 Prompt Assistant"):
             gr.Markdown(
-                "Ask **Qwen3.5-4B** a single question — with or without an image. "
+                "Ask **Gemma 4 E4B** a single question — with or without an image. "
                 "Turn a rough idea into a rich prompt, or upload a reference image and "
                 "ask the model to describe it as a prompt."
             )
@@ -1327,7 +1350,7 @@ with gr.Blocks(fill_height=True) as demo:
                         choices=["Off", "On"],
                         value="Off",
                         label="🧠 Reasoning",
-                        info="Off: direct answer, best for prompts • On: think step-by-step first (slower, raise max tokens)",
+                        info="Direct answers are best for prompts. (Reasoning mode only applies to thinking-capable models; Gemma 4 answers directly.)",
                     )
                     with gr.Accordion("⚙️ Settings", open=False):
                         vlm_max_tokens = gr.Slider(