Merge pull request hiyouga#4706 from T-Atlas/main

chore: Update vllm_engine.py to support vllm version >= 0.5.1

src/llamafactory/chat/vllm_engine.py

@@ -18,7 +18,7 @@
 from ..data import get_template_and_fix_tokenizer
 from ..extras.logging import get_logger
 from ..extras.misc import get_device_count
-from ..extras.packages import is_vllm_available, is_vllm_version_greater_than_0_5
+from ..extras.packages import is_vllm_available, is_vllm_version_greater_than_0_5, is_vllm_version_greater_than_0_5_1
 from ..model import load_config, load_tokenizer
 from ..model.model_utils.quantization import QuantizationMethod
 from ..model.model_utils.visual import LlavaMultiModalProjectorForYiVLForVLLM
@@ -29,7 +29,9 @@
     from vllm import AsyncEngineArgs, AsyncLLMEngine, RequestOutput, SamplingParams
     from vllm.lora.request import LoRARequest
 
-    if is_vllm_version_greater_than_0_5():
+    if is_vllm_version_greater_than_0_5_1():
+        pass
+    elif is_vllm_version_greater_than_0_5():
         from vllm.multimodal.image import ImagePixelData
     else:
         from vllm.sequence import MultiModalData
@@ -130,8 +132,10 @@ async def _generate(
         if self.processor is not None and image is not None:  # add image features
             image_processor: "BaseImageProcessor" = getattr(self.processor, "image_processor")
             pixel_values = image_processor(image, return_tensors="pt")["pixel_values"]
-            if is_vllm_version_greater_than_0_5():
-                multi_modal_data = ImagePixelData(image=pixel_values)
+            if is_vllm_version_greater_than_0_5_1():
+                multi_modal_data = {"image": pixel_values}
+            elif is_vllm_version_greater_than_0_5():
+                multi_modal_data = ImagePixelData(image=pixel_values)   
             else:  # TODO: remove vllm 0.4.3 support
                 multi_modal_data = MultiModalData(type=MultiModalData.Type.IMAGE, data=pixel_values)
         else:

src/llamafactory/extras/packages.py

@@ -81,3 +81,8 @@ def is_vllm_available():
 @lru_cache
 def is_vllm_version_greater_than_0_5():
     return _get_package_version("vllm") >= version.parse("0.5.0")
+
+
+@lru_cache
+def is_vllm_version_greater_than_0_5_1():
+    return _get_package_version("vllm") >= version.parse("0.5.1")