Revert assistant to gemma-4-E4B (bf16) — 12B 8-bit was too slow

- VLM_MODEL_ID default -> prithivMLmods/gemma-4-E4B-it-Uncensored-MAX
(model_type gemma4), small/fast, loads bf16 (~8GB).
- VLM_LOAD_8BIT default -> 0 (bf16); set =1 only for the big 12B.
- requirements unchanged (transformers git main already supports gemma4),
so this is an app reload, not a full rebuild.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app.py

@@ -150,20 +150,18 @@ print("Pipelines loaded!")
 # describing a reference image. Loaded once, bf16 — fits alongside the two
 # diffusion pipelines on the shared GPU.
 # -----------------------------------------------------------------------------
-# Prior models, kept for easy revert (set VLM_MODEL_ID back to one of these):
-#   "rodrigomt/Qwen3.5-4B-Uncensored-Aggressive"   # no generation_config; needed eos pinning
-#   "prithivMLmods/gemma-4-E4B-it-Uncensored-MAX"  # model_type gemma4; under-rated explicit content (~2)
-#   "ccharnkij/Qwen3.5-9B-Uncensored"              # 9B Qwen3.5 VL, ~18.8 GB bf16 (the previous default)
-# Current: gemma-4-12B "unified" (text+image+audio+video), abliterated/uncensored.
-# model_type=gemma4_unified / Gemma4UnifiedForConditionalGeneration — NEWER than the
-# earlier gemma4, so it needs a very recent transformers (requirements pins git main).
-# It's ~24 GB in bf16, which does NOT fit on the zero-a10g (~24 GB) next to the two
-# diffusion pipelines, so we load it 8-bit (bitsandbytes, ~12 GB) by DEFAULT.
-# Flip VLM_LOAD_8BIT=0 to load full bf16 (only safe on a bigger GPU). Stop tokens are
-# handled model-agnostically by _resolve_vlm_eos_ids() (this model ships a proper
-# generation_config + chat_template.jinja).
-VLM_MODEL_ID = os.environ.get("VLM_MODEL_ID", "OpenYourMind/gemma-4-12B-it-abliterated-uncensored")
-VLM_LOAD_8BIT = os.environ.get("VLM_LOAD_8BIT", "1").lower() not in ("0", "false", "no", "")
+# Other models, kept for easy revert (set VLM_MODEL_ID, and VLM_LOAD_8BIT for big ones):
+#   "rodrigomt/Qwen3.5-4B-Uncensored-Aggressive"          # no generation_config; needed eos pinning
+#   "ccharnkij/Qwen3.5-9B-Uncensored"                     # 9B Qwen3.5 VL, ~18.8 GB bf16
+#   "OpenYourMind/gemma-4-12B-it-abliterated-uncensored"  # gemma4_unified, ~24GB; needs VLM_LOAD_8BIT=1 (slow)
+# Current: gemma-4-E4B (model_type=gemma4 / Gemma4ForConditionalGeneration), multimodal,
+# uncensored. Small/fast — loads full bf16 (~8 GB, fits the zero-a10g alongside the
+# diffusion pipelines). The 12B above was accurate but too slow at 8-bit; this is the
+# speed pick. Stop tokens are handled model-agnostically by _resolve_vlm_eos_ids()
+# (this model ships a proper generation_config). VLM_LOAD_8BIT=1 forces bitsandbytes
+# 8-bit (only needed for the 12B); default is bf16.
+VLM_MODEL_ID = os.environ.get("VLM_MODEL_ID", "prithivMLmods/gemma-4-E4B-it-Uncensored-MAX")
+VLM_LOAD_8BIT = os.environ.get("VLM_LOAD_8BIT", "0").lower() not in ("0", "false", "no", "")
 
 print(f"Loading assistant: {VLM_MODEL_ID} (8bit={VLM_LOAD_8BIT}) ...")
 vlm_processor = AutoProcessor.from_pretrained(