Fix: load OCR model as LoRA adapter on Qwen2-VL-2B-Instruct base using PEFT

app.py

@@ -1,8 +1,8 @@
 """
 Chhagan DocVL AI - Document Intelligence Demo
 Models:
-  - Chhagan005/Chhagan-DocVL-Qwen3  (Qwen3-VL architecture)
-  - Chhagan005/Chhagan_ML-VL-OCR-v1 (Qwen2-VL architecture)
+  - Chhagan005/Chhagan-DocVL-Qwen3  (Qwen3-VL LoRA adapter)
+  - Chhagan005/Chhagan_ML-VL-OCR-v1 (Qwen2-VL LoRA adapter on Qwen/Qwen2-VL-2B-Instruct)
 Based on: Qwen3-VL-Outpost reference pattern
 """
 import os
@@ -14,6 +14,7 @@ import gradio as gr
 import spaces
 import torch
 from PIL import Image
+from peft import PeftModel
 
 from transformers import (
     Qwen3VLForConditionalGeneration,
@@ -26,7 +27,8 @@ logging.getLogger("transformers").setLevel(logging.ERROR)
 
 # ---------- CONFIG ----------
 MODEL_DOCVL = "Chhagan005/Chhagan-DocVL-Qwen3"
-MODEL_OCR   = "Chhagan005/Chhagan_ML-VL-OCR-v1"
+MODEL_OCR_ADAPTER = "Chhagan005/Chhagan_ML-VL-OCR-v1"
+MODEL_OCR_BASE = "Qwen/Qwen2-VL-2B-Instruct"
 HF_TOKEN = os.getenv("HF_TOKEN")
 
 MAX_MAX_NEW_TOKENS = 4096
@@ -35,7 +37,7 @@ DEFAULT_MAX_NEW_TOKENS = 1024
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 print("Using device:", device)
 
-# ---------- LOAD MODEL 1: DocVL (Qwen3-VL) ----------
+# ---------- LOAD MODEL 1: DocVL (Qwen3-VL LoRA) ----------
 print("=" * 60)
 print("🚀 Loading Chhagan-DocVL-Qwen3 ...")
 print("=" * 60)
@@ -48,18 +50,25 @@ model_docvl = Qwen3VLForConditionalGeneration.from_pretrained(
 ).to(device).eval()
 print("  ✅ Chhagan-DocVL-Qwen3 loaded")
 
-# ---------- LOAD MODEL 2: OCR-v1 (Qwen2-VL) ----------
+# ---------- LOAD MODEL 2: OCR-v1 (Qwen2-VL + LoRA adapter) ----------
 print("=" * 60)
-print("🚀 Loading Chhagan_ML-VL-OCR-v1 ...")
+print("🚀 Loading Chhagan_ML-VL-OCR-v1 (base + LoRA) ...")
 print("=" * 60)
 
 processor_ocr = AutoProcessor.from_pretrained(
-    MODEL_OCR, trust_remote_code=True, token=HF_TOKEN,
+    MODEL_OCR_BASE, trust_remote_code=True, token=HF_TOKEN,
 )
-model_ocr = Qwen2VLForConditionalGeneration.from_pretrained(
-    MODEL_OCR, trust_remote_code=True, torch_dtype=torch.float16, token=HF_TOKEN,
-).to(device).eval()
-print("  ✅ Chhagan_ML-VL-OCR-v1 loaded")
+print("  ✅ OCR Processor loaded (from base)")
+
+base_model_ocr = Qwen2VLForConditionalGeneration.from_pretrained(
+    MODEL_OCR_BASE, trust_remote_code=True, torch_dtype=torch.float16, token=HF_TOKEN,
+).to(device)
+print("  ✅ OCR Base model loaded")
+
+model_ocr = PeftModel.from_pretrained(
+    base_model_ocr, MODEL_OCR_ADAPTER, token=HF_TOKEN,
+).eval()
+print("  ✅ OCR LoRA adapter applied")
 print("=" * 60)