Update app.py

app.py

@@ -68,17 +68,28 @@ ps3_pkg.PS3VisionModel = _PS3VisionModel
 sys.modules["ps3"] = ps3_pkg
 
 # ===============================
-# Quantization stub to avoid Triton path
-# VILA falls back to "from FloatPointQuantizeTorch import *" if Triton import fails.
-# Provide a tiny no-op module so imports succeed.
-# ===============================
-fpqt = types.ModuleType("FloatPointQuantizeTorch")
-def _id(x, *a, **k): return x  # identity
-# names used by llava.model.qfunction
-fpqt.block_cut = _id
-fpqt.block_quant = _id
-fpqt.block_reshape = _id
-sys.modules["FloatPointQuantizeTorch"] = fpqt
+# Quantization stubs to avoid Triton/Torch custom kernels
+# VILA sometimes imports:
+#  - from .FloatPointQuantizeTriton import *
+#  - from FloatPointQuantizeTriton import *
+#  - from FloatPointQuantizeTorch import *
+# Provide both names (absolute and package-qualified) with no-op funcs.
+# ===============================
+def _mk_fpq_module(mod_name: str):
+    mod = types.ModuleType(mod_name)
+    # Provide the APIs qfunction expects
+    def _id(x, *a, **k): return x
+    mod.block_cut = _id
+    mod.block_quant = _id
+    mod.block_reshape = _id
+    return mod
+
+# Absolute names
+sys.modules["FloatPointQuantizeTorch"] = _mk_fpq_module("FloatPointQuantizeTorch")
+sys.modules["FloatPointQuantizeTriton"] = _mk_fpq_module("FloatPointQuantizeTriton")
+# Package-qualified under llava.model
+sys.modules["llava.model.FloatPointQuantizeTorch"] = sys.modules["FloatPointQuantizeTorch"]
+sys.modules["llava.model.FloatPointQuantizeTriton"] = sys.modules["FloatPointQuantizeTriton"]
 
 # ===============================
 # Load VILA
@@ -112,13 +123,15 @@ if getattr(tokenizer, "chat_template", None) is None:
 # ===============================
 # Inference
 # ===============================
+from PIL import Image as _PILImage
+
 def vila_infer(image, prompt):
     if image is None:
         return "Please upload an image."
     if not prompt or not str(prompt).strip():
         prompt = "Please describe the image."
 
-    pil = Image.fromarray(image).convert("RGB")
+    pil = _PILImage.fromarray(image).convert("RGB")
 
     try:
         out = model.generate_content(
@@ -129,7 +142,7 @@ def vila_infer(image, prompt):
                     {"type": "text", "value": prompt}
                 ]
             }],
-            generation_config=None
+            generation_config=None  # default decoding
         )
         return str(out).strip()
     except Exception as e: