m3_server.py

import asyncio
import os
import time
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
from textwrap import dedent
from typing import List, Tuple, Union
from uuid import uuid4

import torch
from fastapi import FastAPI, HTTPException, Request
from fastapi.responses import JSONResponse
from FlagEmbedding import BGEM3FlagModel
from pydantic import BaseModel
from starlette.status import HTTP_504_GATEWAY_TIMEOUT

Path("/tmp/cache").mkdir(exist_ok=True)
os.environ[
    "HF_HOME"
] = "/tmp/cache"  # does not quite work, need Path("/tmp/cache").mkdir(exist_ok=True)?

batch_size = 2  # gpu batch_size in order of your available vram
max_request = 10  # max request for future improvements on api calls / gpu batches (for now is pretty basic)
max_length = 5000  # max context length for embeddings and passages in re-ranker
max_q_length = 256  # max context lenght for questions in re-ranker
request_flush_timeout = 0.1  # flush time out for future improvements on api calls / gpu batches (for now is pretty basic)
rerank_weights = [0.4, 0.2, 0.4]  # re-rank score weights
request_time_out = 30  # Timeout threshold
gpu_time_out = 5  # gpu processing timeout threshold
port = 3000
port = 7860
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"


class m3Wrapper:
    def __init__(self, model_name: str, device: str = DEVICE):
        """Init."""
        self.model = BGEM3FlagModel(
            model_name, device=device, use_fp16=True if device != "cpu" else False
        )

    def embed(self, sentences: List[str]) -> List[List[float]]:
        embeddings = self.model.encode(
            sentences, batch_size=batch_size, max_length=max_length
        )["dense_vecs"]
        embeddings = embeddings.tolist()
        return embeddings

    def rerank(self, sentence_pairs: List[Tuple[str, str]]) -> List[float]:
        scores = self.model.compute_score(
            sentence_pairs,
            batch_size=batch_size,
            max_query_length=max_q_length,
            max_passage_length=max_length,
            weights_for_different_modes=rerank_weights,
        )["colbert+sparse+dense"]
        return scores


class EmbedRequest(BaseModel):
    sentences: List[str]


class RerankRequest(BaseModel):
    sentence_pairs: List[Tuple[str, str]]


class EmbedResponse(BaseModel):
    embeddings: List[List[float]]


class RerankResponse(BaseModel):
    scores: List[float]


class RequestProcessor:
    def __init__(
        self, model: m3Wrapper, max_request_to_flush: int, accumulation_timeout: float
    ):
        """Init."""
        self.model = model
        self.max_batch_size = max_request_to_flush
        self.accumulation_timeout = accumulation_timeout
        self.queue = asyncio.Queue()
        self.response_futures = {}
        self.processing_loop_task = None
        self.processing_loop_started = False  # Processing pool flag lazy init state
        self.executor = ThreadPoolExecutor()  # Thread pool
        self.gpu_lock = asyncio.Semaphore(1)  # Sem for gpu sync usage

    async def ensure_processing_loop_started(self):
        if not self.processing_loop_started:
            print("starting processing_loop")
            self.processing_loop_task = asyncio.create_task(self.processing_loop())
            self.processing_loop_started = True

    async def processing_loop(self):
        while True:
            requests, request_types, request_ids = [], [], []
            start_time = asyncio.get_event_loop().time()

            while len(requests) < self.max_batch_size:
                timeout = self.accumulation_timeout - (
                    asyncio.get_event_loop().time() - start_time
                )
                if timeout <= 0:
                    break

                try:
                    req_data, req_type, req_id = await asyncio.wait_for(
                        self.queue.get(), timeout=timeout
                    )
                    requests.append(req_data)
                    request_types.append(req_type)
                    request_ids.append(req_id)
                except asyncio.TimeoutError:
                    break

            if requests:
                await self.process_requests_by_type(
                    requests, request_types, request_ids
                )

    async def process_requests_by_type(self, requests, request_types, request_ids):
        tasks = []
        for request_data, request_type, request_id in zip(
            requests, request_types, request_ids
        ):
            if request_type == "embed":
                task = asyncio.create_task(
                    self.run_with_semaphore(
                        self.model.embed, request_data.sentences, request_id
                    )
                )
            else:  # 'rerank'
                task = asyncio.create_task(
                    self.run_with_semaphore(
                        self.model.rerank, request_data.sentence_pairs, request_id
                    )
                )
            tasks.append(task)
        await asyncio.gather(*tasks)

    async def run_with_semaphore(self, func, data, request_id):
        async with self.gpu_lock:  # Wait for sem
            future = self.executor.submit(func, data)
            try:
                result = await asyncio.wait_for(
                    asyncio.wrap_future(future), timeout=gpu_time_out
                )
                self.response_futures[request_id].set_result(result)
            except asyncio.TimeoutError:
                self.response_futures[request_id].set_exception(
                    TimeoutError("GPU processing timeout")
                )
            except Exception as e:
                self.response_futures[request_id].set_exception(e)

    async def process_request(
        self, request_data: Union[EmbedRequest, RerankRequest], request_type: str
    ):
        try:
            await self.ensure_processing_loop_started()
            request_id = str(uuid4())
            self.response_futures[request_id] = asyncio.Future()
            await self.queue.put((request_data, request_type, request_id))
            return await self.response_futures[request_id]
        except Exception as e:
            raise HTTPException(status_code=500, detail=f"Internal Server Error {e}")


description = dedent(
    """\
    ```bash
    curl -X 'POST' \
      'https://mikeee-baai-m3.hf.space/embeddings/' \
      -H 'accept: application/json' \
      -H 'Content-Type: application/json' \
      -d '{
      "sentences": [
        "string", "string1"
      ]
    }'
    ```"""
)

app = FastAPI(
    title="baai m3, serving embed and rerank",
    # description="Swagger UI at https://mikeee-baai-m3.hf.space/docs",
    description=description,
    version="0.1.0a0",
)

# Initialize the model and request processor
model = m3Wrapper("BAAI/bge-m3")
processor = RequestProcessor(
    model, accumulation_timeout=request_flush_timeout, max_request_to_flush=max_request
)

# Adding a middleware returning a 504 error if the request processing time is above a certain threshold
@app.middleware("http")
async def timeout_middleware(request: Request, call_next):
    try:
        start_time = time.time()
        return await asyncio.wait_for(call_next(request), timeout=request_time_out)

    except asyncio.TimeoutError:
        process_time = time.time() - start_time
        return JSONResponse(
            {
                "detail": "Request processing time excedeed limit",
                "processing_time": process_time,
            },
            status_code=HTTP_504_GATEWAY_TIMEOUT,
        )


@app.get("/")
async def landing():
    """Define landing page."""
    return "Swagger UI at https://mikeee-baai-m3.hf.space/docs"


@app.post("/embeddings/", response_model=EmbedResponse)
async def get_embeddings(request: EmbedRequest):
    embeddings = await processor.process_request(request, "embed")
    return EmbedResponse(embeddings=embeddings)


@app.post("/rerank/", response_model=RerankResponse)
async def rerank(request: RerankRequest):
    scores = await processor.process_request(request, "rerank")
    return RerankResponse(scores=scores)


if __name__ == "__main__":
    import uvicorn

    uvicorn.run(app, host="0.0.0.0", port=port)