some modifications

cli_demo.py

@@ -54,6 +54,9 @@ def _load_model_processor(args):
     model, vis_processors, _ = load_model_and_preprocess("minigpt4qwen", args.model_type)
     model.load_checkpoint(args.checkpoint_path)
 
+    model.llm_model.transformer.bfloat16()
+    model.llm_model.lm_head.bfloat16()
+
     generation_config = {
     "chat_format": "chatml",
     "eos_token_id": 151643,
@@ -233,8 +236,12 @@ def main():
                 if '<ImageHere>' not in query:
                     query = '<Img><ImageHere></Img> ' + query
                 first = False
-            with torch.autocast(device_type="cpu",enabled=True,dtype=torch.bfloat16) if args.cpu_only else torch.cuda.amp.autocast(enabled=True,dtype=torch.bfloat16):
-                response, history = model.chat(query, history=history, image_tensor=image_tensor, generation_config=generation_config)
+            if args.cpu_only:
+                model.bfloat16()
+                response, history = model.chat(query, history=history, image_tensor=image_tensor.bfloat16(), generation_config=generation_config)
+            else:
+                with torch.cuda.amp.autocast(enabled=True,dtype=torch.bfloat16):
+                    response, history = model.chat(query, history=history, image_tensor=image_tensor, generation_config=generation_config)
             _clear_screen()
             print(f"\nUser: {query}")
             print(f"\nQwen-Chat: {response}")

lavis/projects/pp_qwen14b/sft_pp.yaml

@@ -39,7 +39,7 @@ model:
   lora_dropout: 0.05
 
   # text length when training
-  max_txt_len: 256
+  max_txt_len: 512
 
   # enable autocast of vit
   enable_autocast: False
@@ -77,14 +77,14 @@ run:
   dist_url: "env://"
   distributed: True
 
-  max_epoch: 2
+  max_epoch: 1
   log_freq: 10
 
   lr_sched: "linear_warmup_cosine_lr_step-wise"
   warmup_lr: 0
   init_lr: 2e-5
   min_lr: 0
-  warmup_ratio: 0.3
+  warmup_ratio: 0.1
 
   deepspeed_config:
     # global batch = 128 = n_ranks * grad_acc_steps * micro_batch_size = (4//2) * 64 * 1

webui_demo.py

@@ -37,6 +37,9 @@ def _load_model_processor(args):
     model, vis_processors, _ = load_model_and_preprocess("minigpt4qwen", args.model_type)
     model.load_checkpoint(args.checkpoint_path)
 
+    model.llm_model.transformer.bfloat16()
+    model.llm_model.lm_head.bfloat16()
+
     generation_config = {
     "chat_format": "chatml",
     "eos_token_id": 151643,
@@ -148,8 +151,12 @@ def gradio_answer(chatbot, history, img_list, do_sample,num_beams, temperature,
     image_tensor =  img_list[0]  # 如果想支持多图情况：torch.stack(img_list).to(self.device)
     generation_config = GenerationConfig.from_dict(generation_config)
     global args
-    with torch.autocast(device_type="cpu",enabled=True,dtype=torch.bfloat16) if args.cpu_only else torch.cuda.amp.autocast(enabled=True,dtype=torch.bfloat16):
-        response, history = model.chat(query=chatbot[-1][0], history=history, image_tensor=image_tensor, generation_config=generation_config,verbose=True)
+    if args.cpu_only:
+        model.bfloat16()
+        response, history = model.chat(query=chatbot[-1][0], history=history, image_tensor=image_tensor.bfloat16(), generation_config=generation_config,verbose=True)
+    else:
+        with torch.cuda.amp.autocast(enabled=True,dtype=torch.bfloat16):
+            response, history = model.chat(query=chatbot[-1][0], history=history, image_tensor=image_tensor.bfloat16(), generation_config=generation_config,verbose=True)
     chatbot[-1][1] = response
     return chatbot, history, img_list
 
@@ -230,4 +237,4 @@ def gradio_answer(chatbot, history, img_list, do_sample,num_beams, temperature,
     )
     clear.click(gradio_reset, [history, img_list], [chatbot, image, text_input, upload_button, history, img_list], queue=False)
 
-demo.launch(share=True)
+demo.launch(share=True,inbrowser=True)