Samuele E. Locatelli
51 lines
1.5 KiB
Plaintext
51 lines
1.5 KiB
Plaintext
services:
|
|
# Primo modello: GPT-OSS (più veloce da caricare)
|
|
gpt-oss:
|
|
image: nvcr.io/nvidia/vllm:26.03.post1-py3
|
|
container_name: vllm-gpt-oss
|
|
restart: always
|
|
runtime: nvidia
|
|
ipc: host
|
|
environment:
|
|
- NVIDIA_VISIBLE_DEVICES=all
|
|
ports:
|
|
- "8000:8000"
|
|
volumes:
|
|
- /root/.cache/huggingface:/root/.cache/huggingface
|
|
command: >
|
|
vllm serve openai/gpt-oss-20b
|
|
--gpu-memory-utilization 0.40
|
|
--served-model-name gpt-oss-20b
|
|
healthcheck:
|
|
test: ["CMD-SHELL", "curl -f http://localhost:8000/health || exit 1"]
|
|
interval: 30s
|
|
timeout: 10s
|
|
retries: 10
|
|
start_period: 60s
|
|
|
|
# Secondo modello: Qwen (parte solo quando GPT-OSS è pronto)
|
|
qwen-36:
|
|
image: nvcr.io/nvidia/vllm:26.03.post1-py3
|
|
container_name: vllm-qwen
|
|
restart: always
|
|
runtime: nvidia
|
|
ipc: host
|
|
environment:
|
|
- NVIDIA_VISIBLE_DEVICES=all
|
|
- HF_HUB_ENABLE_HF_TRANSFER=1 # Accelera il caricamento se scarica
|
|
depends_on:
|
|
gpt-oss:
|
|
condition: service_healthy
|
|
ports:
|
|
- "8001:8000"
|
|
volumes:
|
|
- /root/.cache/huggingface:/root/.cache/huggingface
|
|
command: >
|
|
vllm serve /root/.cache/huggingface/hub/models--Qwen--Qwen3.6-35B-A3B-FP8/snapshots/61a5771f218894aaacf97551e24a25b866750fc2
|
|
--quantization fp8
|
|
--kv-cache-dtype fp8
|
|
--max-model-len 32768
|
|
--gpu-memory-utilization 0.40
|
|
--trust-remote-code
|
|
--served-model-name qwen-3.6-blackwell
|