x

config.json

@@ -27,7 +27,8 @@
         "dtype": "float16",
         "quantization": "fp8",
         "ctx": "262144",
-        "max_tokens": "32768",
+        "max_tokens": "65536",
+        "max_num_batched_tokens": 16384,
         "trust_remote": true,
         "enforce_eager": false,
         "valid_tp": [
@@ -54,7 +55,7 @@
           2
         ],
         "max_num_seqs": 32,
-        "gpu_util": "0.92",
+        "gpu_util": "0.96",
         "kv_cache_dtype": "fp8",
         "enable_prefix_caching": true,
         "tool_call_parser": "qwen3_coder",
@@ -64,6 +65,29 @@
         "served_model_name": "Qwen3.6-27B",
         "hf_model_id": "Qwen/Qwen3.6-27B"
       },
+      "Qwen3.6-27B-FP8": {
+        "local_path": "Qwen3.6-27B-FP8",
+        "dtype": "auto",
+        "ctx": "262144",
+        "max_tokens": "65536",
+        "max_num_batched_tokens": 16384,
+        "trust_remote": true,
+        "enforce_eager": false,
+        "valid_tp": [
+          1,
+          2
+        ],
+        "max_num_seqs": 32,
+        "gpu_util": "0.92",
+        "kv_cache_dtype": "fp8",
+        "enable_prefix_caching": true,
+        "tool_call_parser": "qwen3_coder",
+        "reasoning_parser": "qwen3",
+        "enable_auto_tool_choice": true,
+        "language_model_only": true,
+        "served_model_name": "Qwen3.6-27B-FP8",
+        "hf_model_id": "Qwen/Qwen3.6-27B-FP8"
+      },
       "Qwen3.5-35B-A3B-FP8": {
         "local_path": "Qwen3.5-35B-A3B-FP8",
         "dtype": "float16",

docker-compose.yml

@@ -19,18 +19,10 @@ services:
       HF_HUB_OFFLINE: "1"
       TRANSFORMERS_OFFLINE: "1"
       HF_DATASETS_OFFLINE: "1"
-      # vLLM 相关配置
+      VLLM_ROCM_USE_AITER: "1"  # 打开总开关
-      # VLLM_RPC_TIMEOUT: "300"
+      # 明确禁用不稳定的MoE后端，让其他AITER优化（如MHA）生效
-      # VLLM_WORKER_MULTIPROC_METHOD: "spawn"
+      VLLM_ROCM_USE_AITER_MOE: "0"
-      # PYTHONUNBUFFERED: "1"
+      VLLM_ROCM_MOE_BACKEND: "TRITON"  # 强制MoE使用Triton
-      # # 启用 AITER 加速（R9700 属于 gfx1201 架构）
-      # 核心修复: 禁用 AITER 的 FP8 MoE 后端，回退到 Triton 内核
-      VLLM_ROCM_USE_AITER: "0"
-      # 可选: 显式指定使用更稳定的 MoE 后端
-      VLLM_ROCM_MOE_BACKEND: "TRITON"
-      VLLM_ROCM_USE_AITER_MHA: "1"
-      VLLM_V1_USE_PREFILL_DECODE_ATTENTION: "0"
-      NCCL_MIN_NCHANNELS: "112"
     devices:
       - /dev/kfd
       - /dev/dri