src/triton_inference/deploy/compose/docker-compose.yml

Anuvani-Diarization / src /triton_inference /deploy /compose /docker-compose.yml

Anuvani-Diarization 1.0.0

63dd946 12 days ago

3.77 kB

# The serving deployment: the FastAPI service (the Rust engine: decoding, block scheduling over every GPU, masks,
# clustering, the decoder, the RTTM) and, from the file `diar up` writes beside its state ($DIAR_STATE/tritons.yml),
# one Triton per GPU holding the two networks, all on a private bridge network. Only the service's port is published
# (127.0.0.1 by default). Driven by ../../diar (up, down, status, logs), which also writes the .env this file reads:
#   ../../diar up            # every GPU of this machine, the default lane (ffn16)
#   ../../diar up 2 --exact  # the first two GPUs, the exact lane
# Each Triton sees only its own GPU (NVIDIA_VISIBLE_DEVICES=UUID) and holds that GPU's model diar_nn_g<i> (device 0
# inside the container) and resample_cpu; it checks every batch shape against the release's outputs before it serves.
# The service finds every diar_nn_g<i> model on the Tritons it is given (--gpus auto), on this machine or others.

services:
  # one process, one engine: it owns every Triton it is given, on this machine and on others (diar up --remote)
  service:
    image: ${SERVICE_IMAGE}
    networks: [diar]
    command: ["--listen", "0.0.0.0:8180", "--triton", "${TRITON_URLS}", "--gpus", "auto",
              "--slots-per-gpu", "${SLOTS_PER_GPU}", "--schedule", "${SCHEDULE}",
              "--spool", "/var/lib/anuvani-diar/spool", "--triton-wait-s", "1200",
              "--max-jobs", "${MAX_JOBS}", "--max-uploads", "${MAX_UPLOADS}", "--max-upload-mb", "${MAX_UPLOAD_MB}",
              "--max-spool-gb", "${MAX_SPOOL_GB}", "--audio-budget-gb", "${AUDIO_BUDGET_GB}",
              "--decode-threads", "${DECODE_THREADS}", "--post-threads", "${POST_THREADS}"]
    environment: ["ANUVANI_DIAR_TOKEN=${TOKEN}"]
    ports: ["${SERVICE_BIND}:${SERVICE_PORT}:8180"]
    volumes: ["spool:/var/lib/anuvani-diar"]
    # the service drains for 30 s (--drain-s), then stops
    stop_grace_period: 60s
    restart: unless-stopped
    logging: { driver: json-file, options: { max-size: "50m", max-file: "5" } }

  # runs as the account that started the deployment, so that it can read the service's token file (mode 0600)
  prometheus:
    image: prom/prometheus:v3.11.0
    profiles: [obs]
    networks: [diar]
    user: "${HOST_UID}:${HOST_GID}"
    command: ["--config.file=/etc/prometheus/prometheus.yml", "--storage.tsdb.path=/prometheus",
              "--storage.tsdb.retention.time=15d"]
    volumes: ["${PROM_CONFIG}:/etc/prometheus/prometheus.yml:ro", "${TOKEN_FILE}:/etc/prometheus/service_token:ro",
              "${PROM_DATA}:/prometheus"]
    ports: ["127.0.0.1:${PROM_PORT}:9090"]
    restart: unless-stopped

  grafana:
    image: grafana/grafana:12.4.2
    profiles: [obs]
    networks: [diar]
    environment:
      - GF_SECURITY_ADMIN_PASSWORD=${GRAFANA_PASSWORD}
      - GF_AUTH_ANONYMOUS_ENABLED=false
      - PROM_URL=http://prometheus:9090
    volumes: ["./grafana/provisioning:/etc/grafana/provisioning:ro", "./grafana/dashboards:/var/lib/grafana/dashboards:ro",
              "grafana-data:/var/lib/grafana"]
    ports: ["127.0.0.1:${GRAFANA_PORT}:3000"]
    restart: unless-stopped

  # GPU counters; without its PROF_* fields, which DCGM serves to one client per GPU (a host that already runs an
  # exporter with them: point Prometheus there instead, DCGM_TARGET)
  dcgm-exporter:
    image: nvcr.io/nvidia/k8s/dcgm-exporter:3.3.9-3.6.1-ubuntu22.04
    profiles: [dcgm]
    networks: [diar]
    runtime: nvidia
    cap_add: [SYS_ADMIN]
    environment: [NVIDIA_VISIBLE_DEVICES=all]
    command: ["-f", "/etc/dcgm-exporter/counters.csv", "-c", "1000"]
    volumes: ["./dcgm_counters.csv:/etc/dcgm-exporter/counters.csv:ro"]
    restart: unless-stopped

networks:
  diar: {}

volumes:
  spool: {}
  grafana-data: {}

	# The serving deployment: the FastAPI service (the Rust engine: decoding, block scheduling over every GPU, masks,
	# clustering, the decoder, the RTTM) and, from the file `diar up` writes beside its state ($DIAR_STATE/tritons.yml),
	# one Triton per GPU holding the two networks, all on a private bridge network. Only the service's port is published
	# (127.0.0.1 by default). Driven by ../../diar (up, down, status, logs), which also writes the .env this file reads:
	# ../../diar up # every GPU of this machine, the default lane (ffn16)
	# ../../diar up 2 --exact # the first two GPUs, the exact lane
	# Each Triton sees only its own GPU (NVIDIA_VISIBLE_DEVICES=UUID) and holds that GPU's model diar_nn_g<i> (device 0
	# inside the container) and resample_cpu; it checks every batch shape against the release's outputs before it serves.
	# The service finds every diar_nn_g<i> model on the Tritons it is given (--gpus auto), on this machine or others.

	services:
	# one process, one engine: it owns every Triton it is given, on this machine and on others (diar up --remote)
	service:
	image: ${SERVICE_IMAGE}
	networks: [diar]
	command: ["--listen", "0.0.0.0:8180", "--triton", "${TRITON_URLS}", "--gpus", "auto",
	"--slots-per-gpu", "${SLOTS_PER_GPU}", "--schedule", "${SCHEDULE}",
	"--spool", "/var/lib/anuvani-diar/spool", "--triton-wait-s", "1200",
	"--max-jobs", "${MAX_JOBS}", "--max-uploads", "${MAX_UPLOADS}", "--max-upload-mb", "${MAX_UPLOAD_MB}",
	"--max-spool-gb", "${MAX_SPOOL_GB}", "--audio-budget-gb", "${AUDIO_BUDGET_GB}",
	"--decode-threads", "${DECODE_THREADS}", "--post-threads", "${POST_THREADS}"]
	environment: ["ANUVANI_DIAR_TOKEN=${TOKEN}"]
	ports: ["${SERVICE_BIND}:${SERVICE_PORT}:8180"]
	volumes: ["spool:/var/lib/anuvani-diar"]
	# the service drains for 30 s (--drain-s), then stops
	stop_grace_period: 60s
	restart: unless-stopped
	logging: { driver: json-file, options: { max-size: "50m", max-file: "5" } }

	# runs as the account that started the deployment, so that it can read the service's token file (mode 0600)
	prometheus:
	image: prom/prometheus:v3.11.0
	profiles: [obs]
	networks: [diar]
	user: "${HOST_UID}:${HOST_GID}"
	command: ["--config.file=/etc/prometheus/prometheus.yml", "--storage.tsdb.path=/prometheus",
	"--storage.tsdb.retention.time=15d"]
	volumes: ["${PROM_CONFIG}:/etc/prometheus/prometheus.yml:ro", "${TOKEN_FILE}:/etc/prometheus/service_token:ro",
	"${PROM_DATA}:/prometheus"]
	ports: ["127.0.0.1:${PROM_PORT}:9090"]
	restart: unless-stopped

	grafana:
	image: grafana/grafana:12.4.2
	profiles: [obs]
	networks: [diar]
	environment:
	- GF_SECURITY_ADMIN_PASSWORD=${GRAFANA_PASSWORD}
	- GF_AUTH_ANONYMOUS_ENABLED=false
	- PROM_URL=http://prometheus:9090
	volumes: ["./grafana/provisioning:/etc/grafana/provisioning:ro", "./grafana/dashboards:/var/lib/grafana/dashboards:ro",
	"grafana-data:/var/lib/grafana"]
	ports: ["127.0.0.1:${GRAFANA_PORT}:3000"]
	restart: unless-stopped

	# GPU counters; without its PROF_* fields, which DCGM serves to one client per GPU (a host that already runs an
	# exporter with them: point Prometheus there instead, DCGM_TARGET)
	dcgm-exporter:
	image: nvcr.io/nvidia/k8s/dcgm-exporter:3.3.9-3.6.1-ubuntu22.04
	profiles: [dcgm]
	networks: [diar]
	runtime: nvidia
	cap_add: [SYS_ADMIN]
	environment: [NVIDIA_VISIBLE_DEVICES=all]
	command: ["-f", "/etc/dcgm-exporter/counters.csv", "-c", "1000"]
	volumes: ["./dcgm_counters.csv:/etc/dcgm-exporter/counters.csv:ro"]
	restart: unless-stopped

	networks:
	diar: {}

	volumes:
	spool: {}
	grafana-data: {}