[add]add skip-image-cache parameters

Author: SangChengC

lightllm/server/embed_cache/utils.py

@@ -32,7 +32,8 @@ def create_shm(name, data):
         mem_view = shared_memory.buf
         mem_view[:data_size] = data
     except FileExistsError:
-        print("Warning create shm {} failed because of FileExistsError!".format(name))
+        pass
+        # print("Warning create shm {} failed because of FileExistsError!".format(name))
 
 
 def read_shm(name):

lightllm/server/multimodal_params.py

@@ -137,9 +137,11 @@ def __init__(
         self,
         images: List[dict] = [],
         audios: List[dict] = [],
+        skip_image_cache: bool = False,
     ) -> None:
         self.images = [ImageItem(**i) for i in images]
         self.audios = [AudioItem(**a) for a in audios]
+        self.skip_image_cache = skip_image_cache
         return
 
     async def verify_and_preload(self, request: Request):

lightllm/server/visualserver/manager.py

@@ -123,10 +123,12 @@ async def loop_for_fwd(self):
 
                     multimodal_params = group_req_indexes.multimodal_params
 
-                    img_uuids = [img.uuid for img in multimodal_params.images]
-                    ready_image = obtain(self.cache_client.root.get_items_embed(img_uuids))
-
-                    for img, ready in zip(multimodal_params.images, ready_image):
+                    img_uuids = list(dict.fromkeys(img.uuid for img in multimodal_params.images))
+                    if multimodal_params.skip_image_cache:
+                        ready_image = [False] * len(img_uuids)
+                    else:
+                        ready_image = obtain(self.cache_client.root.get_items_embed(img_uuids))
+                    for img, ready in zip(img_uuids, ready_image):
                         if not ready:
                             images_need_infer.append(img)
 

lightllm/server/visualserver/model_infer/model_rpc.py

@@ -100,9 +100,8 @@ def exposed_encode(self, images: List[ImageItem]):
         all_img_embeds = all_img_embeds.to(torch.device("cpu"))
 
         if self.tp_rank_id == 0:
-            ready_flags = obtain(self.cache_client.root.get_items_embed(uuids))
             ids_to_set = []
-            for i, ready in enumerate(ready_flags):
+            for i, ready in uuids:
                 if ready:
                     continue
                 uid = uuids[i]