Update app.py

app.py

@@ -20,6 +20,7 @@ import fitz
 
 from transformers import (
     Qwen2_5_VLForConditionalGeneration,
+    Blip2ForConditionalGeneration,
     AutoModelForCausalLM,
     AutoModelForVision2Seq,
     AutoModelForImageTextToText,
@@ -78,16 +79,16 @@ model_g = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     MODEL_ID_G, trust_remote_code=True, subfolder=SUBFOLDER, torch_dtype=torch.float16
 ).to(device).eval()
 
-MODEL_ID_I = "microsoft/Florence-2-large"
+MODEL_ID_I = "Salesforce/blip2-opt-2.7b"
 processor_i = AutoProcessor.from_pretrained(MODEL_ID_I, trust_remote_code=True)
-model_i = AutoModelForCausalLM.from_pretrained(
-    MODEL_ID_I, trust_remote_code=True, torch_dtype=torch.float16, _attn_implementation="flash_attention_2"
+model_i = Blip2ForConditionalGeneration.from_pretrained(
+    MODEL_ID_I, trust_remote_code=True, torch_dtype=torch.float16
 ).to(device).eval()
 
-MODEL_ID_A = "microsoft/Florence-2-base"
+MODEL_ID_A = "facebook/Perception-LM-1B"
 processor_a = AutoProcessor.from_pretrained(MODEL_ID_A, trust_remote_code=True)
-model_a = AutoModelForCausalLM.from_pretrained(
-    MODEL_ID_A, trust_remote_code=True, torch_dtype=torch.float16, _attn_implementation="flash_attention_2"
+model_a = AutoModelForImageTextToText.from_pretrained(
+    MODEL_ID_A, trust_remote_code=True, torch_dtype=torch.float16
 ).to(device).eval()
 
 # --- PDF Generation and Preview Utility Function ---
@@ -184,8 +185,8 @@ def process_document_stream(
     elif model_name == "LFM2-VL-1.6B": processor, model = processor_t, model_t
     elif model_name == "SmolVLM-Instruct-250M": processor, model = processor_c, model_c
     elif model_name == "MonkeyOCR-pro-1.2B": processor, model = processor_g, model_g
-    elif model_name == "Florence-2-large": processor, model = processor_i, model_i
-    elif model_name == "Florence-2-base": processor, model = processor_a, model_a
+    elif model_name == "blip2-opt-2.7b": processor, model = processor_i, model_i
+    elif model_name == "Perception-LM-1B": processor, model = processor_a, model_a
     else:
         yield "Invalid model selected.", ""
         return
@@ -242,7 +243,7 @@ def create_gradio_interface():
             # Left Column (Inputs)
             with gr.Column(scale=1):
                 model_choice = gr.Dropdown(
-                    choices=["LFM2-VL-1.6B", "LFM2-VL-450M", "SmolVLM-Instruct-250M", "Florence-2-large", "MonkeyOCR-pro-1.2B", "Florence-2-base"],
+                    choices=["LFM2-VL-1.6B", "LFM2-VL-450M", "SmolVLM-Instruct-250M", "blip2-opt-2.7b", "MonkeyOCR-pro-1.2B", "Perception-LM-1B"],
                     label="Select Model", value="LFM2-VL-1.6B"
                 )
                 prompt_input = gr.Textbox(label="Query Input", placeholder="✦︎ Enter your query")