v0.4

app.py

@@ -13,14 +13,14 @@ device = "cuda" if torch.cuda.is_available() else "cpu"
 print(f"Using {device}" if device != "cpu" else "Using CPU")
 
 def _load_model():
-  tokenizer = AutoTokenizer.from_pretrained("vikhyatk/moondream2", trust_remote_code=True, revision="2024-05-08")
+  tokenizer = AutoTokenizer.from_pretrained("vikhyatk/moondream2", trust_remote_code=True, revision="2024-05-08", torch_dtype=(torch.bfloat16 if device == 'cuda' else torch.float32))
   model = AutoModelForCausalLM.from_pretrained("vikhyatk/moondream2", device_map=device, trust_remote_code=True, revision="2024-05-08")
   return (model, tokenizer)
 
 class MoonDream():
   def __init__(self, model=None, tokenizer=None):
     self.model, self.tokenizer = (model, tokenizer)
-    if not model or not tokenizer:
+    if not model or model is None or not tokenizer or tokenizer is None:
       self.model, self.tokenizer = _load_model()
     self.device = device
     self.model.to(self.device)
@@ -89,12 +89,13 @@ Complete Description:
   return res.choices[0].text.split("```")[0]
 
 def xform_image_description(img, inst):
-  from together import Together
-  desc = dual_images(img)
-  tog = Together(api_key=os.getenv("TOGETHER_KEY"))
-  prompt=f"""Describe the image in aggressively verbose detail. I must know every freckle upon a man's brow and each blade of the grass intimately.\nDescription: ```text\n{desc}\n```\nInstructions:\n```text\n{inst}\n```\n\n\n---\nDetailed Description:\n```text"""
-  res = tog.completions.create(prompt=prompt, model="meta-llama/Meta-Llama-3-70B", stop=["```"], max_tokens=1024)
-  return res.choices[0].text[len(prompt):].split("```")[0]
+    #md = MoonDream()
+    from together import Together
+    desc = dual_images(img)
+    tog = Together(api_key=os.getenv("TOGETHER_KEY"))
+    prompt=f"""Describe the image in aggressively verbose detail. I must know every freckle upon a man's brow and each blade of the grass intimately.\nDescription: ```text\n{desc}\n```\nInstructions:\n```text\n{inst}\n```\n\n\n---\nDetailed Description:\n```text"""
+    res = tog.completions.create(prompt=prompt, model="meta-llama/Meta-Llama-3-70B", stop=["```"], max_tokens=1024)
+    return res.choices[0].text[len(prompt):].split("```")[0]
 
 with gr.Blocks() as demo:
   with gr.Row(visible=True):
@@ -110,6 +111,8 @@ with gr.Blocks() as demo:
           im1 = gr.Image(label="image 1", type='pil')
         with gr.Column():
           im2 = gr.Image(label="image 2", type='pil')
+        with gr.Row():
+            minst = gr.Textbox(label="Merge Instructions")
       with gr.Row():
         btn2 = gr.Button("submit batch")
       with gr.Row():
@@ -117,8 +120,6 @@ with gr.Blocks() as demo:
           otp2 = gr.Textbox(label="individual batch output (left)", interactive=True)
         with gr.Column():
           otp3 = gr.Textbox(label="individual batch output (right)", interactive=True)
-      with gr.Row():
-          minst = gr.Textbox(label="Merge Instructions")
       with gr.Row():
         btn_scd = gr.Button("Merge Descriptions to Single Combined Description")
       with gr.Row():