Upload folder using huggingface_hub

Dockerfile

@@ -8,7 +8,7 @@ ENV HF_HOME=/tmp/hf \
 WORKDIR /app
 
 RUN pip install --no-cache-dir \
-    "transformers>=4.52" \
+    "transformers>=5.13" \
     accelerate \
     "bitsandbytes>=0.43" \
     sentencepiece \

app.py

@@ -18,7 +18,7 @@ from fastapi import FastAPI, Request
 from fastapi.responses import JSONResponse
 from PIL import Image
 
-MODEL_ID = os.environ.get("MODEL_ID", "llmfan46/gemma-3-12b-it-ultra-uncensored-heretic")
+MODEL_ID = os.environ.get("MODEL_ID", "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic")
 
 app = FastAPI()
 _proc = None
@@ -32,7 +32,8 @@ def _load():
     if _model is not None or _load_error is not None:
         return
     try:
-        from transformers import AutoProcessor, BitsAndBytesConfig, Gemma3ForConditionalGeneration
+        # model-agnostic loader → works for gemma3 AND gemma4 (needs transformers>=5.13)
+        from transformers import AutoModelForImageTextToText, AutoProcessor, BitsAndBytesConfig
         _proc = AutoProcessor.from_pretrained(MODEL_ID)
         bnb = BitsAndBytesConfig(
             load_in_4bit=True,
@@ -40,7 +41,7 @@ def _load():
             bnb_4bit_compute_dtype=torch.float16,   # T4 has no bf16
             bnb_4bit_use_double_quant=True,
         )
-        _model = Gemma3ForConditionalGeneration.from_pretrained(
+        _model = AutoModelForImageTextToText.from_pretrained(
             MODEL_ID, quantization_config=bnb, device_map="auto", torch_dtype=torch.float16,
         )
         _model.eval()