OAI: Initial vision support in OAI chat completions

* Support image_url inputs containing URLs or base64 strings following OAI vision spec
* Use async lru cache for image embeddings
* Add generic wrapper class for multimodal embeddings

backends/exllamav2/model.py

@@ -6,6 +6,8 @@ import gc
 import math
 import pathlib
 import traceback
+from backends.exllamav2.vision import clear_image_embedding_cache
+from common.multimodal import MultimodalEmbeddingWrapper
 import torch
 import uuid
 from copy import deepcopy
@@ -816,6 +818,9 @@ class ExllamaV2Container:
             # Delete references held in the grammar module
             clear_grammar_func_cache()
 
+            # Clear the image embedding cache
+            clear_image_embedding_cache()
+
             # Unload LoRAs
             if self.generator and self.generator.generator.current_loras:
                 for lora in self.generator.generator.current_loras:
@@ -908,12 +913,17 @@ class ExllamaV2Container:
         return dict(zip_longest(top_tokens, cleaned_values))
 
     async def generate(
-        self, prompt: str, request_id: str, abort_event: asyncio.Event = None, **kwargs
+        self,
+        prompt: str,
+        embeddings: MultimodalEmbeddingWrapper,
+        request_id: str,
+        abort_event: asyncio.Event = None,
+        **kwargs,
     ):
         """Generate a response to a prompt."""
         generations = []
         async for generation in self.generate_gen(
-            prompt, request_id, abort_event, **kwargs
+            prompt, embeddings, request_id, abort_event, **kwargs
         ):
             generations.append(generation)
 
@@ -979,6 +989,7 @@ class ExllamaV2Container:
     async def generate_gen(
         self,
         prompt: str,
+        embeddings: MultimodalEmbeddingWrapper,
         request_id: str,
         abort_event: Optional[asyncio.Event] = None,
         **kwargs,
@@ -1246,7 +1257,10 @@ class ExllamaV2Container:
         # Encode both positive and negative prompts
         input_ids = [
             self.tokenizer.encode(
-                prompt, add_bos=add_bos_token, encode_special_tokens=True
+                prompt,
+                add_bos=add_bos_token,
+                encode_special_tokens=True,
+                embeddings=embeddings.content,
             )
             for prompt in prompts
         ]
@@ -1297,6 +1311,7 @@ class ExllamaV2Container:
             banned_strings=banned_strings,
             token_healing=token_healing,
             identifier=job_id,
+            embeddings=embeddings.content,
         )
 
         # Save generated tokens and full response

backends/exllamav2/vision.py

@@ -4,18 +4,14 @@ import io
 import base64
 import re
 from PIL import Image
+from common import model
 import aiohttp
 from common.networking import (
     handle_request_error,
 )
 from fastapi import HTTPException
-from exllamav2 import (
-    ExLlamaV2,
-    ExLlamaV2Tokenizer,
-    ExLlamaV2VisionTower,
-    ExLlamaV2MMEmbedding,
-)
-from functools import lru_cache
+from exllamav2.generator import ExLlamaV2MMEmbedding
+from async_lru import alru_cache
 
 
 async def get_image(url: str) -> Image:
@@ -50,14 +46,16 @@ async def get_image(url: str) -> Image:
     return Image.open(io.BytesIO(bytes_image))
 
 
-@lru_cache(20)
-async def get_image_embedding(
-    model: ExLlamaV2,
-    tokenizer: ExLlamaV2Tokenizer,
-    vision_model: ExLlamaV2VisionTower,
-    url: str,
-) -> ExLlamaV2MMEmbedding:
+@alru_cache(20)
+async def get_image_embedding(url: str) -> ExLlamaV2MMEmbedding:
     image = await get_image(url)
-    return vision_model.get_image_embeddings(
-        model=model, tokenizer=tokenizer, image=image
+    return model.container.vision_model.get_image_embeddings(
+        model=model.container.model,
+        tokenizer=model.container.tokenizer,
+        image=image,
+        text_alias=None,
     )
+
+
+def clear_image_embedding_cache():
+    get_image_embedding.cache_clear()

common/multimodal.py

@@ -0,0 +1,36 @@
+from typing import List
+from backends.exllamav2.vision import get_image_embedding
+from common import model
+from pydantic import BaseModel
+from loguru import logger
+
+from common.optional_dependencies import dependencies
+
+if dependencies.exllamav2:
+    from exllamav2 import ExLlamaV2VisionTower
+
+
+class MultimodalEmbeddingWrapper(BaseModel):
+    """Common multimodal embedding wrapper"""
+
+    type: str = None
+    content: List = []
+    text_alias: List[str] = []
+
+
+async def add_image_embedding(
+    embeddings: MultimodalEmbeddingWrapper, url: str
+) -> MultimodalEmbeddingWrapper:
+    # Determine the type of vision embedding to use
+    if not embeddings.type:
+        if isinstance(model.container.vision_model, ExLlamaV2VisionTower):
+            embeddings.type = "ExLlamaV2MMEmbedding"
+
+    if embeddings.type == "ExLlamaV2MMEmbedding":
+        embedding = await get_image_embedding(url)
+        embeddings.content.append(embedding)
+        embeddings.text_alias.append(embedding.text_alias)
+    else:
+        logger.error("No valid vision model to create embedding")
+
+    return embeddings

endpoints/OAI/router.py

@@ -17,6 +17,7 @@ from endpoints.OAI.types.embedding import EmbeddingsRequest, EmbeddingsResponse
 from endpoints.OAI.utils.chat_completion import (
     format_prompt_with_template,
     generate_chat_completion,
+    preprocess_vision_request,
     stream_generate_chat_completion,
 )
 from endpoints.OAI.utils.completion import (
@@ -126,6 +127,8 @@ async def chat_completion_request(
     if isinstance(data.messages, str):
         prompt = data.messages
     else:
+        if model.container.use_vision:
+            data.messages, embeddings = await preprocess_vision_request(data.messages)
         prompt = await format_prompt_with_template(data)
 
     # Set an empty JSON schema if the request wants a JSON response
@@ -136,12 +139,14 @@ async def chat_completion_request(
 
     if data.stream and not disable_request_streaming:
         return EventSourceResponse(
-            stream_generate_chat_completion(prompt, data, request, model_path),
+            stream_generate_chat_completion(
+                prompt, embeddings, data, request, model_path
+            ),
             ping=maxsize,
         )
     else:
         generate_task = asyncio.create_task(
-            generate_chat_completion(prompt, data, request, model_path)
+            generate_chat_completion(prompt, embeddings, data, request, model_path)
         )
 
         response = await run_with_request_disconnect(

endpoints/OAI/utils/chat_completion.py

@@ -3,9 +3,10 @@
 import asyncio
 import pathlib
 from asyncio import CancelledError
-from typing import List, Optional
+from typing import Dict, List, Optional
 import json
 
+from common.multimodal import MultimodalEmbeddingWrapper, add_image_embedding
 from fastapi import HTTPException, Request
 from jinja2 import TemplateError
 from loguru import logger
@@ -279,7 +280,11 @@ async def format_prompt_with_template(
 
 
 async def stream_generate_chat_completion(
-    prompt: str, data: ChatCompletionRequest, request: Request, model_path: pathlib.Path
+    prompt: str,
+    embeddings: MultimodalEmbeddingWrapper,
+    data: ChatCompletionRequest,
+    request: Request,
+    model_path: pathlib.Path,
 ):
     """Generator for the generation process."""
     abort_event = asyncio.Event()
@@ -298,6 +303,7 @@ async def stream_generate_chat_completion(
                     n,
                     gen_queue,
                     prompt,
+                    embeddings,
                     request.state.id,
                     abort_event,
                     **task_gen_params.model_dump(exclude={"prompt"}),
@@ -372,7 +378,11 @@ async def stream_generate_chat_completion(
 
 
 async def generate_chat_completion(
-    prompt: str, data: ChatCompletionRequest, request: Request, model_path: pathlib.Path
+    prompt: str,
+    embeddings: MultimodalEmbeddingWrapper,
+    data: ChatCompletionRequest,
+    request: Request,
+    model_path: pathlib.Path,
 ):
     gen_tasks: List[asyncio.Task] = []
 
@@ -381,7 +391,10 @@ async def generate_chat_completion(
             gen_tasks.append(
                 asyncio.create_task(
                     model.container.generate(
-                        prompt, request.state.id, **data.model_dump(exclude={"prompt"})
+                        prompt,
+                        embeddings,
+                        request.state.id,
+                        **data.model_dump(exclude={"prompt"}),
                     )
                 )
             )
@@ -459,3 +472,22 @@ def postprocess_tool_call(call_str: str) -> List[ToolCall]:
             tool_call["function"]["arguments"]
         )
     return [ToolCall(**tool_call) for tool_call in tool_calls]
+
+
+async def preprocess_vision_request(messages: List[Dict]):
+    embeddings = MultimodalEmbeddingWrapper()
+    for message in messages:
+        if isinstance(message["content"], list):
+            concatenated_content = ""
+            for content in message["content"]:
+                if content["type"] == "text":
+                    concatenated_content += content["text"]
+                elif content["type"] == "image_url":
+                    embeddings = await add_image_embedding(
+                        embeddings, content["image_url"]["url"]
+                    )
+                    concatenated_content += embeddings.text_alias[-1]
+
+            message["content"] = concatenated_content
+
+    return messages, embeddings

endpoints/OAI/utils/completion.py

@@ -7,6 +7,7 @@ Also serves as a common module for completions and chat completions.
 import asyncio
 import pathlib
 from asyncio import CancelledError
+from common.multimodal import MultimodalEmbeddingWrapper
 from fastapi import HTTPException, Request
 from typing import List, Union
 
@@ -87,6 +88,7 @@ async def _stream_collector(
     task_idx: int,
     gen_queue: asyncio.Queue,
     prompt: str,
+    embeddings: MultimodalEmbeddingWrapper,
     request_id: str,
     abort_event: asyncio.Event,
     **kwargs,
@@ -95,7 +97,7 @@ async def _stream_collector(
 
     try:
         new_generation = model.container.generate_gen(
-            prompt, request_id, abort_event, **kwargs
+            prompt, embeddings, request_id, abort_event, **kwargs
         )
         async for generation in new_generation:
             generation["index"] = task_idx

pyproject.toml

@@ -29,6 +29,7 @@ dependencies = [
     "lm-format-enforcer >= 0.9.6",
     "aiofiles",
     "aiohttp",
+    "async_lru",
     "huggingface_hub",
     "psutil",
     "httptools>=0.5.0",