wassemgtk's picture
Switch local vision stack to Qwen Omni and add weights bundle manifest
c4b8a5d verified
|
Raw
History Blame Contribute Delete
1.06 kB

Standard vLLM Deployment

This directory contains launch files for the local model services used by glm-5.2-visual-runtime.

Direct Launch

chmod +x vllm/serve_glm52_fp8.sh
PORT=8001 TENSOR_PARALLEL_SIZE=8 ./vllm/serve_glm52_fp8.sh

Qwen Omni Vision / Audio / Video

chmod +x vllm/serve_qwen3_omni.sh
PORT=8002 TENSOR_PARALLEL_SIZE=2 ./vllm/serve_qwen3_omni.sh

Optional Qwen3.6-27B Reasoning

Use this if you want Qwen3.6-27B as the reasoning engine instead of GLM-5.2:

chmod +x vllm/serve_qwen36_27b.sh
PORT=8001 TENSOR_PARALLEL_SIZE=2 ./vllm/serve_qwen36_27b.sh

Docker Compose

docker compose -f vllm/docker-compose.vllm.yml up

Smoke Test

pip install openai
VLLM_BASE_URL=http://localhost:8001/v1 VLLM_MODEL=glm-5.2 python vllm/openai_smoke_test.py

Gateway Connection

Configure the visual runtime gateway:

GLM_BASE_URL=http://localhost:8001/v1
GLM_MODEL=glm-5.2
VISION_BASE_URL=http://localhost:8002/v1
VISION_MODEL=qwen3-omni
VISUAL_RUNTIME_MODE=cloud