Fix circular import in vision.py

backends/exllamav3/vision.py

@@ -1,6 +1,7 @@
 """Vision utilities for ExLlamaV3."""
 
-from common import model
+from typing import TYPE_CHECKING
+
 from common.optional_dependencies import dependencies
 from common.image_util import get_image
 from common.logger import xlogger
@@ -10,6 +11,9 @@ from common.logger import xlogger
 if dependencies.exllamav3:
     from exllamav3.tokenizer import MMEmbedding
 
+if TYPE_CHECKING:
+    from backends.exllamav3.model import ExllamaV3Container
+
 from collections import OrderedDict
 from hashlib import blake2b
 from typing import OrderedDict as OrderedDictType
@@ -22,7 +26,7 @@ def _image_key_128(s: str) -> bytes:
     return blake2b(s.encode("utf-8"), digest_size=16).digest()
 
 
-async def get_image_embedding_exl3(url: str) -> "MMEmbedding":
+async def get_image_embedding_exl3(container: "ExllamaV3Container", url: str) -> "MMEmbedding":
     key = _image_key_128(url)
 
     cached = _embedding_cache.get(key)
@@ -33,8 +37,8 @@ async def get_image_embedding_exl3(url: str) -> "MMEmbedding":
             return embedding
 
     image = await get_image(url)
-    embedding = model.container.vision_model.get_image_embeddings(
-        tokenizer=model.container.tokenizer,
+    embedding = container.vision_model.get_image_embeddings(
+        tokenizer=container.tokenizer,
         image=image,
         text_alias=None,
     )

common/multimodal.py

@@ -1,32 +1,20 @@
-from backends.exllamav3.vision import get_image_embedding_exl3
-from common import model
-from common.logger import xlogger
+from typing import List, TYPE_CHECKING
+
 from pydantic import BaseModel, Field
-from typing import List
 
-from common.optional_dependencies import dependencies
+from backends.exllamav3.vision import get_image_embedding_exl3
 
-if dependencies.exllamav3:
-    from exllamav3 import Model
+if TYPE_CHECKING:
+    from backends.exllamav3.model import ExllamaV3Container
 
 
 class MultimodalEmbeddingWrapper(BaseModel):
     """Common multimodal embedding wrapper"""
 
-    type: str = None
     content: list = Field(default_factory=list)
     text_alias: List[str] = Field(default_factory=list)
 
-    async def add(self, url: str):
-        # Determine the type of vision embedding to use
-        if not self.type:
-            if dependencies.exllamav3 and isinstance(model.container.vision_model, Model):
-                self.type = "MMEmbedding"
-
-        # Create the embedding
-        if self.type == "MMEmbedding":
-            embedding = await get_image_embedding_exl3(url)
-            self.content.append(embedding)
-            self.text_alias.append(embedding.text_alias)
-        else:
-            xlogger.error("No valid vision model to create embedding")
+    async def add(self, container: "ExllamaV3Container", url: str):
+        embedding = await get_image_embedding_exl3(container, url)
+        self.content.append(embedding)
+        self.text_alias.append(embedding.text_alias)

endpoints/OAI/utils/chat_completion.py

@@ -214,7 +214,7 @@ async def format_messages_with_template(
                 if content.type == "text":
                     concatenated_content += content.text
                 elif content.type == "image_url" and mm_embeddings:
-                    await mm_embeddings.add(content.image_url.url)
+                    await mm_embeddings.add(model.container, content.image_url.url)
                     concatenated_content += mm_embeddings.text_alias[-1]
 
             # Convert the message content into a concatenated string