Аудио‑эмбеддинги на GPU: поиск и классификация

				
					version: "3.9"

x-common-env: &env
  MODELS_DIR: /models
  CACHE_DIR: /var/cache/aemb
  SAMPLE_RATE: "32000"
  WINDOW_S: "0.96"
  HOP_S: "0.48"
  EMBED_PRECISION: "fp16"         # bf16|fp16
  INDEX_TYPE: "ivf_flat"          # flat|hnsw|ivf_flat|ivf_pq
  INDEX_PARTS: "8"                # для IVF
  EFSEARCH: "128"                 # для HNSW
  TOPK: "20"

services:
  aemb-realtime:
    image: cloudcompute/audio-embeddings:latest
    environment:
      <<: *env
      SERVICE_MODE: "realtime"
    deploy:
      resources:
        reservations:
          devices: [{ capabilities: ["gpu"] }]
    ports: ["8050:8050"]
    volumes:
      - /nvme/models:/models
      - /nvme/aemb-cache:/var/cache/aemb
      - /nvme/index:/index
    command: ["python", "serve_rt.py", "--host=0.0.0.0", "--port=8050", "--index=/index/main"]

  aemb-batch:
    image: cloudcompute/audio-embeddings:latest
    environment:
      <<: *env
      SERVICE_MODE: "batch"
      MAX_CHUNK_SECONDS: "120"
      OVERLAP_SECONDS: "0.3"
    deploy:
      resources:
        reservations:
          devices: [{ capabilities: ["gpu"] }]
    volumes:
      - /nvme/models:/models
      - /nvme/aemb-cache:/var/cache/aemb
      - /mnt/audio:/data/in
      - /nvme/emb:/data/emb
    command: ["python", "run_batch.py", "--input=/data/in", "--output=/data/emb"]

  index-builder:
    image: cloudcompute/vector-index:latest
    environment:
      INDEX_TYPE: "ivf_flat"
      INDEX_PARTS: "8"
      PQ_M: "16"
      PQ_BITS: "8"
    volumes:
      - /nvme/emb:/data/emb
      - /nvme/index:/index
    command: ["python", "build_index.py", "--emb=/data/emb", "--out=/index/main"]


				
			

				
					apiVersion: apps/v1
kind: Deployment
metadata:
  name: aemb-realtime
spec:
  replicas: 2
  selector: { matchLabels: { app: aemb-realtime } }
  template:
    metadata:
      labels: { app: aemb-realtime }
    spec:
      containers:
        - name: aemb
          image: cloudcompute/audio-embeddings:latest
          ports: [{ containerPort: 8050 }]
          env:
            - { name: EMBED_PRECISION, value: "fp16" }
            - { name: INDEX_TYPE, value: "ivf_flat" }
            - { name: TOPK, value: "20" }
          volumeMounts:
            - { name: models, mountPath: /models }
            - { name: cache,  mountPath: /var/cache/aemb }
            - { name: index,  mountPath: /index }
          resources:
            limits:
              nvidia.com/gpu: 1
              memory: "24Gi"
              cpu: "4"
      volumes:
        - name: models
          hostPath: { path: /nvme/models }
        - name: cache
          hostPath: { path: /nvme/aemb-cache }
        - name: index
          hostPath: { path: /nvme/index }


				
			

				
					from fastapi import FastAPI, Body
import numpy as np, base64, json

app = FastAPI()

# Заглушки: embed_gpu, search_gpu, load_audio
def embed_gpu(x_32khz_mono_fp32) -> np.ndarray:
    # Возвращает L2-нормированный вектор shape [dim]
    ...

def search_gpu(vec, topk=20):
    # Возвращает (ids, distances) из GPU-индекса
    ...

@app.post("/embed")
def embed_endpoint(payload: dict = Body(...)):
    audio = base64.b64decode(payload["audio_b64"])
    x = load_audio(audio, sr=32000)        # mono fp32
    v = embed_gpu(x).astype("float16")     # экономия канала
    return {"dim": int(v.shape[0]),
            "embedding_b64": base64.b64encode(v.tobytes()).decode("ascii")}

@app.post("/search")
def search_endpoint(payload: dict = Body(...)):
    v = np.frombuffer(base64.b64decode(payload["embedding_b64"]), dtype=np.float16).astype(np.float32)
    ids, dists = search_gpu(v, topk=int(payload.get("topk", 20)))
    return {"ids": ids[:], "distances": dists[:]}