composte-hermes.yml aktualisiert
This commit is contained in:
parent
bd4ccbe7df
commit
aa03f88031
|
|
@ -5,7 +5,6 @@
|
||||||
# Hermes spricht vLLM intern über http://vllm:8000/v1 an – kein extra
|
# Hermes spricht vLLM intern über http://vllm:8000/v1 an – kein extra
|
||||||
# Netzwerk-Routing nötig, alles lokal auf Cortex.
|
# Netzwerk-Routing nötig, alles lokal auf Cortex.
|
||||||
#
|
#
|
||||||
|
|
||||||
services:
|
services:
|
||||||
vllm:
|
vllm:
|
||||||
image: vllm/vllm-openai:latest
|
image: vllm/vllm-openai:latest
|
||||||
|
|
@ -18,6 +17,8 @@ services:
|
||||||
environment:
|
environment:
|
||||||
- NVIDIA_VISIBLE_DEVICES=0,1
|
- NVIDIA_VISIBLE_DEVICES=0,1
|
||||||
- HUGGING_FACE_HUB_TOKEN=${HF_TOKEN:-}
|
- HUGGING_FACE_HUB_TOKEN=${HF_TOKEN:-}
|
||||||
|
# Reduziert Fragmentierung im CUDA-Allocator – hilft bei knappem VRAM
|
||||||
|
- PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
|
||||||
volumes:
|
volumes:
|
||||||
- ${HOME:-/root}/.cache/huggingface:/root/.cache/huggingface
|
- ${HOME:-/root}/.cache/huggingface:/root/.cache/huggingface
|
||||||
deploy:
|
deploy:
|
||||||
|
|
@ -31,19 +32,34 @@ services:
|
||||||
- --model=Qwen/Qwen3-14B-AWQ
|
- --model=Qwen/Qwen3-14B-AWQ
|
||||||
- --quantization=awq
|
- --quantization=awq
|
||||||
- --tensor-parallel-size=2
|
- --tensor-parallel-size=2
|
||||||
- --gpu-memory-utilization=0.90
|
# Von 0.90 auf 0.82 gesenkt: lässt Platz für CUDA-Graph-Pools
|
||||||
|
# und NCCL all_reduce Buffer, die zusätzlich zum KV-Cache-Pool anfallen
|
||||||
|
- --gpu-memory-utilization=0.82
|
||||||
|
# 32768 ist die native Context-Länge von Qwen3-14B – YaRN entfernt,
|
||||||
|
# da hier gar keine Erweiterung über die native Länge hinaus stattfindet
|
||||||
- --max-model-len=32768
|
- --max-model-len=32768
|
||||||
- --hf-overrides={"rope_scaling":{"rope_type":"yarn","factor":2.0,"original_max_position_embeddings":32768}}
|
# Begrenzt parallele Sequenzen -> kleinerer KV-Cache-Bedarf,
|
||||||
|
# wichtig bei nur 12GB pro GPU. Bei Bedarf hochsetzen, wenn stabil.
|
||||||
|
- --max-num-seqs=16
|
||||||
|
# FP8 KV-Cache spart ca. 50% Speicher gegenüber FP16 bei minimalem
|
||||||
|
# Qualitätsverlust – gerade bei 32k Context relevant
|
||||||
|
- --kv-cache-dtype=fp8
|
||||||
- --served-model-name=qwen3-14b-awq
|
- --served-model-name=qwen3-14b-awq
|
||||||
|
healthcheck:
|
||||||
|
test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
|
||||||
|
interval: 30s
|
||||||
|
timeout: 10s
|
||||||
|
retries: 5
|
||||||
|
start_period: 180s
|
||||||
networks:
|
networks:
|
||||||
- cortex-net
|
- cortex-net
|
||||||
|
|
||||||
hermes-gateway:
|
hermes-gateway:
|
||||||
image: nousresearch/hermes-agent:latest
|
image: nousresearch/hermes-agent:latest
|
||||||
container_name: cortex-hermes-gateway
|
container_name: cortex-hermes-gateway
|
||||||
restart: unless-stopped
|
restart: unless-stopped
|
||||||
depends_on:
|
depends_on:
|
||||||
- vllm
|
vllm:
|
||||||
|
condition: service_healthy
|
||||||
command: ["gateway", "run"]
|
command: ["gateway", "run"]
|
||||||
shm_size: 1g
|
shm_size: 1g
|
||||||
ports:
|
ports:
|
||||||
|
|
@ -61,7 +77,6 @@ services:
|
||||||
# - API_SERVER_KEY=${API_SERVER_KEY}
|
# - API_SERVER_KEY=${API_SERVER_KEY}
|
||||||
networks:
|
networks:
|
||||||
- cortex-net
|
- cortex-net
|
||||||
|
|
||||||
hermes-dashboard:
|
hermes-dashboard:
|
||||||
image: nousresearch/hermes-agent:latest
|
image: nousresearch/hermes-agent:latest
|
||||||
container_name: cortex-hermes-dashboard
|
container_name: cortex-hermes-dashboard
|
||||||
|
|
@ -79,7 +94,6 @@ services:
|
||||||
- GATEWAY_HEALTH_URL=http://cortex-hermes-gateway:8642
|
- GATEWAY_HEALTH_URL=http://cortex-hermes-gateway:8642
|
||||||
networks:
|
networks:
|
||||||
- cortex-net
|
- cortex-net
|
||||||
|
|
||||||
networks:
|
networks:
|
||||||
cortex-net:
|
cortex-net:
|
||||||
driver: bridge
|
driver: bridge
|
||||||
Loading…
Reference in New Issue