x

2026-03-29 19:49:29 +08:00
parent eef84a2199
commit 757a29d7ee
8 changed files with 4 additions and 4 deletions
@@ -17,7 +17,7 @@
  "revision": "",
  "models": {
    "default": "Qwen3.5-35B-A3B-GPTQ-Int4",
-    "selected": "Qwen3.5-27B-FP8",
+    "selected": "Qwen3.5-35B-A3B-GPTQ-Int4",
    "profiles": {
      "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit": {
        "local_path": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
@@ -52,11 +52,11 @@
      "Qwen3.5-27B-FP8": {
        "local_path": "Qwen3.5-27B-FP8",
        "dtype": "bfloat16",
-        "ctx": "32768",
+        "ctx": "65536",
        "trust_remote": true,
        "valid_tp": [1, 2],
        "max_num_seqs": "4",
-        "max_tokens": "16384",
+        "max_tokens": "8192",
        "gpu_util": "0.94",
        "tool_call_parser": "qwen3_xml",
        "enable_auto_tool_choice": true,
@@ -71,7 +71,7 @@
        "trust_remote": true,
        "valid_tp": [1, 2],
        "max_num_seqs": "8",
-        "max_tokens": "4096",
+        "max_tokens": "8192",
        "gpu_util": "0.92",
        "tool_call_parser": "qwen3_xml",
        "enable_auto_tool_choice": true,