x

config.json

@@ -38,14 +38,17 @@
       },
       "Qwen3.5-27B-FP8": {
         "local_path": "Qwen3.5-27B-FP8",
-        "dtype": "bfloat16",
+        "dtype": "auto",
         "ctx": "65536",
         "trust_remote": true,
         "valid_tp": [1, 2],
-        "max_num_seqs": "4",
+        "max_num_seqs": 12,
         "max_tokens": "8192",
         "enforce_eager": false,
-        "gpu_util": "0.94",
+        "gpu_util": "0.9",
+        "kv_cache_dtype": "fp8",
+        "enable_prefix_caching": true,
+        "max_num_batched_tokens": 16384,
         "tool_call_parser": "qwen3_coder",
         "reasoning_parser": "qwen3",
         "enable_auto_tool_choice": true,

docker-compose.yml

@@ -17,6 +17,14 @@ services:
       VLLM_RPC_TIMEOUT: "300"
       VLLM_WORKER_MULTIPROC_METHOD: "spawn"
       PYTHONUNBUFFERED: "1"
+      # 启用 AITER 加速（R9700 属于 gfx1201 架构）
+      VLLM_ROCM_USE_AITER: "1"
+      VLLM_USE_AITER_UNIFIED_ATTENTION: "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"
+      # 使用 vLLM V1 引擎（性能更好）
+      VLLM_USE_V1: "1"
     devices:
       - /dev/kfd
       - /dev/dri