bugfix: llava-hf/llava-interleave-qwen-7b-hf (#2497)

- fix init raise exception because tie_word_embeddings config
- max_batch_size option for start

lmdeploy/serve/openai/api_server.py

@@ -15,7 +15,7 @@
 
 from lmdeploy.archs import get_task
 from lmdeploy.messages import (GenerationConfig, LogitsProcessor,
-                               PytorchEngineConfig, TurbomindEngineConfig)
+                               PytorchEngineConfig, TurbomindEngineConfig, VisionConfig)
 from lmdeploy.model import ChatTemplateConfig
 from lmdeploy.serve.async_engine import AsyncEngine
 from lmdeploy.serve.openai.protocol import (  # noqa: E501
@@ -1054,13 +1054,15 @@ def serve(model_path: str,
 
     _, pipeline_class = get_task(model_path)
 
+    vision_config = VisionConfig(kwargs.get("vision_max_batch_size", 1))
     VariableInterface.async_engine = pipeline_class(
         model_path=model_path,
         model_name=model_name,
         backend=backend,
         backend_config=backend_config,
         chat_template_config=chat_template_config,
         max_log_len=max_log_len,
+        vision_config=vision_config,
         **kwargs)
 
     if proxy_url is not None:

lmdeploy/vl/model/llava_hf.py

@@ -24,21 +24,40 @@ def build_model(self):
             warnings.simplefilter('ignore')
             from transformers import LlavaForConditionalGeneration
             model = LlavaForConditionalGeneration._from_config(self.hf_config)
+
+        if not getattr(model.config, "tie_word_embeddings", False):
+            if not self.with_llm:
+                del model.language_model
+                for key in ['language_model']:
+                    setattr(model, key, None)
+            else:
+                self.vl_model = model
+            with disable_logging():
+                load_checkpoint_and_dispatch(
+                    model=model,
+                    max_memory=self.max_memory,
+                    checkpoint=self.model_path,
+                    device_map='auto' if not self.with_llm else {'': 'cpu'},
+                    no_split_module_classes=['CLIPEncoderLayer', 'SiglipEncoderLayer'],
+                    dtype=torch.half)
+        else:
+            # fix for llava-hf/llava-interleave-qwen-7b-hf
+            # we have to remove llm after init model for using call llm.get_output_embedding()
+            with disable_logging():
+                load_checkpoint_and_dispatch(
+                    model=model,
+                    max_memory=self.max_memory,
+                    checkpoint=self.model_path,
+                    device_map='auto' if not self.with_llm else {'': 'cpu'},
+                    no_split_module_classes=['CLIPEncoderLayer', 'SiglipEncoderLayer'],
+                    dtype=torch.half)
             if not self.with_llm:
                 del model.language_model
                 for key in ['language_model']:
                     setattr(model, key, None)
             else:
                 self.vl_model = model
 
-        with disable_logging():
-            load_checkpoint_and_dispatch(
-                model=model,
-                max_memory=self.max_memory,
-                checkpoint=self.model_path,
-                device_map='auto' if not self.with_llm else {'': 'cpu'},
-                no_split_module_classes=['CLIPEncoderLayer'],
-                dtype=torch.half)
         model.eval()
         self.model = model
         # processor