update

app.py

@@ -10,44 +10,83 @@ from vlm_inference import (
 )
 
 # =====================================================
-# Load model on CPU (ZeroGPU)
+# Load VLM on CPU (ZeroGPU)
 # =====================================================
+print("[DEBUG] Loading VLM model on CPU...")
 model = load_vlm_model()
 model.eval()
+print("[DEBUG] VLM model loaded.")
 
 
 # =====================================================
-# GPU inference (single-turn VLM)
+# message parser (multimodal=True 仕様準拠)
+# =====================================================
+def parse_message(message: dict):
+    """
+    message = {
+        "text": str,
+        "files": list  # PIL.Image が入る
+    }
+    """
+    print("[DEBUG] parse_message called")
+    print("[DEBUG] message type:", type(message))
+    print("[DEBUG] message content:", message)
+
+    text = message.get("text", "")
+    files = message.get("files", [])
+
+    print("[DEBUG] parsed text:", repr(text))
+    print("[DEBUG] parsed files:", files)
+
+    image = files[0] if files else None
+    print("[DEBUG] parsed image:", image)
+
+    return text, image
+
+
+# =====================================================
+# GPU inference (single-turn, VLM only)
 # =====================================================
 @spaces.GPU
-def infer_once(
-    image,
-    text,
+def chat_fn(
+    message,
+    history,        # unused (single-turn)
     temperature,
     top_p,
     top_k,
 ):
+    print("[DEBUG] chat_fn called")
+    print("[DEBUG] temperature:", temperature, "top_p:", top_p, "top_k:", top_k)
+
+    text, image = parse_message(message)
+
     if image is None:
-        yield "⚠️ Please upload an image."
-        return
+        print("[DEBUG] image is None -> returning error message")
+        return "Image input is required."
 
     device = "cuda"
+    print("[DEBUG] moving model to GPU")
     model_gpu = model.to(device)
 
-    # --- image tensor ---
+    print("[DEBUG] preprocessing image")
     image_tensor = image_processor(
         images=image.convert("RGB"),
         return_tensors="pt"
     )["pixel_values"].to(device)
 
-    # --- prompt (Colabと同一) ---
+    print("[DEBUG] image_tensor shape:", image_tensor.shape)
+
     prompt = (
         "<user>\n"
         f"{text}\n"
         "<assistant>\n"
     )
 
-    try:
+    print("[DEBUG] prompt:")
+    print(prompt)
+
+    def stream():
+        print("[DEBUG] stream generator started")
         for chunk in vlm_infer_stream(
             model=model_gpu,
             image_tensor=image_tensor,
@@ -57,72 +96,33 @@ def infer_once(
             top_p=top_p if top_p > 0 else None,
             top_k=top_k if top_k > 0 else None,
         ):
+            print("[DEBUG] yield chunk:", repr(chunk))
             yield chunk
-    finally:
+
+        print("[DEBUG] inference finished, cleaning up GPU")
         model_gpu.to("cpu")
         torch.cuda.empty_cache()
+        print("[DEBUG] GPU cleanup done")
 
-
-# =====================================================
-# UI logic (history is display-only)
-# =====================================================
-def submit(
-    image,
-    text,
-    history,
-    temperature,
-    top_p,
-    top_k,
-):
-    history = history or []
-    history.append((text, ""))
-
-    def stream():
-        acc = ""
-        for chunk in infer_once(image, text, temperature, top_p, top_k):
-            acc += chunk
-            history[-1] = (text, acc)
-            yield history
-
-    return history, stream()
+    return stream()
 
 
 # =====================================================
-# Gradio UI
+# UI (ChatInterface, multimodal)
 # =====================================================
-with gr.Blocks(title="EveryonesGPT Vision (Single-turn)") as demo:
-    gr.Markdown("## 🖼️ EveryonesGPT Vision\nSingle-turn VLM (Colab-compatible)")
-
-    with gr.Row():
-        with gr.Column(scale=1):
-            image_input = gr.Image(type="pil", label="Image")
-            text_input = gr.Textbox(
-                label="Prompt",
-                placeholder="Describe the image or ask a question",
-                lines=3,
-            )
-
-            temperature = gr.Slider(0.1, 2.0, value=0.5, step=0.05, label="Temperature")
-            top_p = gr.Slider(0.0, 1.0, value=0.9, step=0.05, label="Top-p")
-            top_k = gr.Slider(0, 200, value=0, step=1, label="Top-k")
-
-            submit_btn = gr.Button("Run")
-
-        with gr.Column(scale=1):
-            chatbot = gr.Chatbot(label="Output (history is display-only)")
-            state = gr.State([])
-
-    submit_btn.click(
-        fn=submit,
-        inputs=[
-            image_input,
-            text_input,
-            state,
-            temperature,
-            top_p,
-            top_k,
-        ],
-        outputs=[chatbot, chatbot],
-    )
+print("[DEBUG] Building Gradio UI")
+
+demo = gr.ChatInterface(
+    fn=chat_fn,
+    multimodal=True,
+    title="EveryonesGPT Vision (VLM only)",
+    description="Single-turn Vision-Language Model demo (CLIP ViT-L/14)",
+    additional_inputs=[
+        gr.Slider(0.1, 2.0, value=0.5, step=0.05, label="Temperature"),
+        gr.Slider(0.0, 1.0, value=0.9, step=0.05, label="Top-p"),
+        gr.Slider(0, 200, value=0, step=1, label="Top-k"),
+    ],
+)
 
+print("[DEBUG] Launching Gradio app")
 demo.launch()