From 91a506cff84a79bc73be358f79d1094ca52dc5ae Mon Sep 17 00:00:00 2001
From: SZCJW <792430652@qq.com>
Date: Wed, 15 Apr 2026 23:21:01 +0800
Subject: [PATCH] x

---
 config.json | 133 ++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 133 insertions(+)
 create mode 100644 config.json

diff --git a/config.json b/config.json
new file mode 100644
index 0000000..43efba6
--- /dev/null
+++ b/config.json
@@ -0,0 +1,133 @@
+{
+  "services": {
+    "api": {
+      "host": "0.0.0.0",
+      "port": 8002
+    },
+    "openai": {
+      "host": "0.0.0.0",
+      "port": 8001
+    }
+  },
+  "public_model_name": "Qwen_local_model",
+  "default_enable_thinking": false,
+  "reasoning_enabled": false,
+  "api_key": "sk-szcjw",
+  "tensor_parallel_size": 2,
+  "dtype": "bfloat16",
+  "model_root": "/opt/model",
+  "offline_mode": true,
+  "revision": "",
+  "models": {
+    "default": "Qwen3.5-35B-A3B-GPTQ-Int4",
+    "selected": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
+    "profiles": {
+      "GLM-4.7-Flash-AWQ": {
+        "local_path": "GLM-4.7-Flash-AWQ",
+        "ctx": "32768",
+        "trust_remote": true,
+        "valid_tp": [1, 2],
+        "max_num_seqs": "16",
+        "max_tokens": "32768",
+        "gpu_util": "0.98",
+        "tool_call_parser": "qwen3_coder",
+        "reasoning_parser": "qwen3",
+        "enable_auto_tool_choice": true,
+        "served_model_name": "GLM-4.7-Flash-AWQ",
+        "hf_model_id": "THUDM/GLM-4.7-Flash-AWQ"
+      },
+      "Qwen3.5-27B-FP8": {
+        "local_path": "Qwen3.5-27B-FP8",
+        "dtype": "auto",
+        "ctx": "65536",
+        "trust_remote": true,
+        "valid_tp": [1, 2],
+        "max_num_seqs": 12,
+        "max_tokens": "8192",
+        "enforce_eager": false,
+        "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,
+        "served_model_name": "Qwen3.5-27B-FP8",
+        "hf_model_id": "RedHatAI/Qwen3.5-27B-FP8-dynamic"
+      },
+      "Qwen3.5-35B-A3B-GPTQ-Int4": {
+        "local_path": "Qwen3.5-35B-A3B-GPTQ-Int4",
+        "dtype": "float16",
+        "quantization": "gptq",
+        "ctx": "131072",
+        "trust_remote": true,
+        "valid_tp": [2],
+        "max_num_seqs": "6",
+        "max_tokens": "4096",
+        "gpu_util": "0.90",
+        "enforce_eager": false,
+        "tool_call_parser": "qwen3_coder",
+        "reasoning_parser": "qwen3",
+        "enable_auto_tool_choice": true,
+        "served_model_name": "Qwen3.5-35B-A3B-GPTQ-Int4",
+        "hf_model_id": "Qwen/Qwen3.5-35B-A3B-GPTQ-Int4",
+        "env": {
+          "VLLM_RPC_TIMEOUT": "300"
+        }
+      },
+      "Qwen3.5-35B-A3B-FP8": {
+        "local_path": "Qwen3.5-35B-A3B-FP8",
+        "dtype": "float16",
+        "ctx": "131072",
+        "enforce_eager": false,
+        "trust_remote": true,
+        "valid_tp": [1, 2],
+        "max_num_seqs": "8",
+        "max_tokens": "16384",
+        "gpu_util": "0.94",
+        "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,
+        "served_model_name": "Qwen3.5-35B-A3B-FP8",
+        "hf_model_id": "Qwen/Qwen3.5-35B-A3B-FP8"
+      },
+      "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit": {
+        "local_path": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
+        "dtype": "auto",
+        "ctx": "131072",
+        "trust_remote": true,
+        "valid_tp": [1, 2],
+        "max_num_seqs": "8",
+        "max_tokens": "65536",
+        "max_num_batched_tokens": 16384,
+        "gpu_util": "0.94",
+        "enforce_eager": false,
+        "enable_prefix_caching": true,
+        "tool_call_parser": "qwen3_coder",
+        "reasoning_parser": "qwen3",
+        "enable_auto_tool_choice": true,
+        "served_model_name": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
+        "hf_model_id": "cpatonn/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit"
+      },
+      "Gemma-4-26B-A4B": {
+        "local_path": "gemma-4-26B-A4B",
+        "dtype": "auto",
+        "ctx": "65536",
+        "trust_remote": true,
+        "valid_tp": [1, 2],
+        "max_num_seqs": 16,
+        "max_tokens": "8192",
+        "enforce_eager": false,
+        "gpu_util": "0.85",
+        "kv_cache_dtype": "fp8",
+        "enable_prefix_caching": true,
+        "max_num_batched_tokens": 16384,
+        "served_model_name": "Gemma-4-26B-A4B",
+        "hf_model_id": "google/gemma-4-26B-A4B"
+      }
+    }
+  }
+}