From 80a549bdcd380c48e939b34da09c89445ec7f748 Mon Sep 17 00:00:00 2001
From: chenjw28 <792430652@qq.com>
Date: Fri, 17 Apr 2026 09:38:30 +0800
Subject: [PATCH] x

---
 config.json | 118 +++++++++++++++++++++++++++++++++++++++-------------
 1 file changed, 88 insertions(+), 30 deletions(-)

diff --git a/config.json b/config.json
index ebd41ce..9b90be6 100644
--- a/config.json
+++ b/config.json
@@ -20,40 +20,70 @@
   "revision": "",
   "models": {
     "default": "Qwen3.5-35B-A3B-GPTQ-Int4",
-    "selected": "Qwen3.5-27B-FP8",
+    "selected": "Qwen3.6-35B-A3B-FP8",
     "profiles": {
-      "GLM-4.7-Flash-AWQ": {
-        "local_path": "GLM-4.7-Flash-AWQ",
-        "ctx": "32768",
+      "Qwen3.6-35B-A3B": {
+        "local_path": "Qwen3.6-35B-A3B",
+        "dtype": "float16",
+        "quantization": "auto",
+        "ctx": "131072",
+        "max_tokens": "8192",
         "trust_remote": true,
-        "valid_tp": [1, 2],
-        "max_num_seqs": "16",
-        "max_tokens": "32768",
-        "gpu_util": "0.98",
+        "enforce_eager": false,
+        "valid_tp": [
+          2
+        ],
+        "max_num_seqs": "12",
+        "gpu_util": "0.90",
         "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"
+        "served_model_name": "Qwen3.6-35B-A3B",
+        "hf_model_id": "Qwen/Qwen3.6-35B-A3B",
+        "env": {
+          "VLLM_RPC_TIMEOUT": "300"
+        }
       },
-      "Qwen3.5-27B-FP8": {
-        "local_path": "Qwen3.5-27B-FP8",
-        "dtype": "auto",
+      "Qwen3.6-35B-A3B-FP8": {
+        "local_path": "Qwen3.6-35B-A3B-FP8",
+        "dtype": "float16",
+        "quantization": "auto",
         "ctx": "131072",
+        "max_tokens": "8192",
         "trust_remote": true,
-        "valid_tp": [1, 2],
-        "max_num_seqs": 32,
-        "max_tokens": "16384",
         "enforce_eager": false,
+        "valid_tp": [
+          2
+        ],
+        "max_num_seqs": "12",
         "gpu_util": "0.90",
+        "tool_call_parser": "qwen3_coder",
+        "reasoning_parser": "qwen3",
+        "enable_auto_tool_choice": true,
+        "served_model_name": "Qwen3.6-35B-A3B-FP8",
+        "hf_model_id": "Qwen/Qwen3.6-35B-A3B-FP8"
+      },
+      "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": "Qwen3.5-27B-FP8",
-        "hf_model_id": "RedHatAI/Qwen3.5-27B-FP8-dynamic"
+        "served_model_name": "Qwen3.5-35B-A3B-FP8",
+        "hf_model_id": "Qwen/Qwen3.5-35B-A3B-FP8"
       },
       "Qwen3.5-35B-A3B-GPTQ-Int4": {
         "local_path": "Qwen3.5-35B-A3B-GPTQ-Int4",
@@ -61,7 +91,9 @@
         "quantization": "gptq",
         "ctx": "131072",
         "trust_remote": true,
-        "valid_tp": [2],
+        "valid_tp": [
+          2
+        ],
         "max_num_seqs": "6",
         "max_tokens": "4096",
         "gpu_util": "0.90",
@@ -75,31 +107,37 @@
           "VLLM_RPC_TIMEOUT": "300"
         }
       },
-      "Qwen3.5-35B-A3B-FP8": {
-        "local_path": "Qwen3.5-35B-A3B-FP8",
-        "dtype": "float16",
+      "Qwen3.5-27B-FP8": {
+        "local_path": "Qwen3.5-27B-FP8",
+        "dtype": "auto",
         "ctx": "131072",
-        "enforce_eager": false,
         "trust_remote": true,
-        "valid_tp": [1, 2],
-        "max_num_seqs": "8",
+        "valid_tp": [
+          1,
+          2
+        ],
+        "max_num_seqs": 32,
         "max_tokens": "16384",
-        "gpu_util": "0.94",
+        "enforce_eager": false,
+        "gpu_util": "0.90",
         "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"
+        "served_model_name": "Qwen3.5-27B-FP8",
+        "hf_model_id": "RedHatAI/Qwen3.5-27B-FP8-dynamic"
       },
       "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],
+        "valid_tp": [
+          1,
+          2
+        ],
         "max_num_seqs": "8",
         "max_tokens": "65536",
         "max_num_batched_tokens": 16384,
@@ -112,12 +150,32 @@
         "served_model_name": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
         "hf_model_id": "cpatonn/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit"
       },
+      "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"
+      },
       "Gemma-4-26B-A4B": {
         "local_path": "gemma-4-26B-A4B",
         "dtype": "auto",
         "ctx": "65536",
         "trust_remote": true,
-        "valid_tp": [1, 2],
+        "valid_tp": [
+          1,
+          2
+        ],
         "max_num_seqs": 16,
         "max_tokens": "8192",
         "enforce_eager": false,