x

docker-compose.yml

@@ -3,7 +3,7 @@ services:
     build:
       context: .
       dockerfile: Dockerfile
-    image: rocm-vllm-inference:nightly
+    image: vllm-openai-rocm:nightly
     container_name: rocm-vllm-openai
     entrypoint: ["python"]
     command: ["-m", "app.start_openai"]
@@ -23,10 +23,10 @@ services:
       # VLLM_WORKER_MULTIPROC_METHOD: "spawn"
       # PYTHONUNBUFFERED: "1"
       # # 启用 AITER 加速（R9700 属于 gfx1201 架构）
-      # VLLM_ROCM_USE_AITER: "0"
-      # VLLM_ROCM_USE_AITER_MHA: "0"
+      VLLM_ROCM_USE_AITER: "1"
+      VLLM_ROCM_USE_AITER_MHA: "0"
       # # 增强 All-Reduce 性能（双卡通信优化）
-      # VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: "INT4"
+      VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: "INT4"
     devices:
       - /dev/kfd
       - /dev/dri