Update app.py

app.py

@@ -296,40 +296,59 @@ if QWEN3_AVAILABLE:
 else:
     print("   ⚠️ Qwen3VL not in transformers version")
 
-# ── Model 3: CSM-DocExtract-VL-Q4KM (NEW, replaces Qwen3-2B) ──
-print("\n3️⃣  CSM-DocExtract-VL-Q4KM (8B Q4KM Quantized)...")
+# ── Model 3: CSM-DocExtract-VL-Q4KM (NEW — Qwen3VL, pre-quantized Q4KM) ──
+print("\n3️⃣  CSM-DocExtract-VL-Q4KM (8B Q4KM — Qwen3VL architecture)...")
 MODEL_ID_Q4KM = "Chhagan005/CSM-DocExtract-VL-Q4KM"
 CSM_Q4KM_AVAILABLE = False
 processor_q4km = model_q4km = None
 
 try:
-    processor_q4km = AutoProcessor.from_pretrained(MODEL_ID_Q4KM, trust_remote_code=True)
-    # Try loading as full quantized model first
-    try:
-        model_q4km = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+    # Processor load
+    processor_q4km = AutoProcessor.from_pretrained(
+        MODEL_ID_Q4KM, trust_remote_code=True
+    )
+
+    # Model is qwen3_vl type + ALREADY pre-quantized Q4KM
+    # So: use Qwen3VL class + NO extra quantization_config
+    if QWEN3_AVAILABLE:
+        model_q4km = Qwen3VLForConditionalGeneration.from_pretrained(
             MODEL_ID_Q4KM,
-            quantization_config=bnb_4bit_config,
-            torch_dtype=torch.float16,
+            torch_dtype="auto",       # model already has Q4KM weights
             device_map="auto",
             trust_remote_code=True,
         ).eval()
-    except:
-        # Fallback: try Qwen3VL architecture
-        if QWEN3_AVAILABLE:
-            model_q4km = Qwen3VLForConditionalGeneration.from_pretrained(
+        print("   ✅ Loaded! (Qwen3VL Q4KM pre-quantized)")
+        CSM_Q4KM_AVAILABLE = True
+    else:
+        # Qwen3VL not in transformers → use AutoModel fallback
+        from transformers import AutoModelForCausalLM
+        try:
+            from transformers import AutoModelForVisualQuestionAnswering
+            model_q4km = AutoModelForVisualQuestionAnswering.from_pretrained(
                 MODEL_ID_Q4KM,
-                quantization_config=bnb_4bit_config,
-                torch_dtype=torch.float16,
+                torch_dtype="auto",
                 device_map="auto",
                 trust_remote_code=True,
             ).eval()
-        else:
-            raise Exception("Neither Qwen2.5VL nor Qwen3VL architecture worked")
-    print("   ✅ Loaded! (~6-7GB VRAM)")
-    CSM_Q4KM_AVAILABLE = True
+        except:
+            # Last fallback: force load with Qwen2_5 but ignore arch warning
+            import warnings
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                model_q4km = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+                    MODEL_ID_Q4KM,
+                    torch_dtype="auto",
+                    device_map="auto",
+                    trust_remote_code=True,
+                    ignore_mismatched_sizes=True,
+                ).eval()
+        print("   ✅ Loaded! (fallback loader)")
+        CSM_Q4KM_AVAILABLE = True
+
 except Exception as e:
     print(f"   ❌ Failed: {e}")
 
+
 # ── Model 4: CSM-DocExtract-VL 4BNB (NEW, replaces Nanonets) ──
 print("\n4️⃣  CSM-DocExtract-VL 4BNB (BitsAndBytes 4-bit)...")
 MODEL_ID_4BNB = "Chhagan005/CSM-DocExtract-VL"