Update app.py

app.py

@@ -9,6 +9,7 @@ import requests
 from transformers import AutoModelForImageTextToText, AutoProcessor
 from transformers.models.qwen2_vl.image_processing_qwen2_vl import smart_resize
 import torch
+from torch.ao.quantization import quantize_dynamic
 import re
 import traceback
 
@@ -17,6 +18,32 @@ MODEL_ID = "Hcompany/Holo1-3B"
 
 # --- Helpers (robust across different transformers versions) ---
 
+def locate_text_backbone(model):
+    """
+    Tries common attribute names used by VLMs to find the LLM/text stack.
+    Falls back to the whole model if unknown.
+    """
+    # common in Qwen-like / custom repos
+    for name in [
+        "language_model",   # e.g., model.language_model
+        "text_model",       # e.g., model.text_model
+        "model",            # sometimes the text core is 'model'
+        "llm",              # generic
+        "transformer",      # some repos expose raw transformer as 'transformer'
+    ]:
+        m = getattr(model, name, None)
+        if m is not None:
+            return m, name
+
+    # last resort: look for a child that has an lm_head or tied weights
+    for name, child in model.named_children():
+        if hasattr(child, "lm_head") or hasattr(child, "get_input_embeddings"):
+            return child, name
+
+    # if still not found, return the model itself
+    return model, None
+
+
 def pick_device() -> str:
     # Force CPU per request
     return "cpu"
@@ -239,7 +266,7 @@ except Exception as e:
         pass
 
 # --- Gradio UI ---
-title = "Holo1-7B: Action VLM Localization Demo (CPU)"
+title = "Holo1-3B: Action VLM Localization Demo (CPU)"
 article = f"""
 <p style='text-align: center'>
 Model: <a href='https://huggingface.co/{MODEL_ID}' target='_blank'>{MODEL_ID}</a> by HCompany |