x

2026-03-29 05:47:30 +08:00
parent dd5afbe6d0
commit 23909c205e
5 changed files with 46 additions and 7 deletions
@@ -12,12 +12,15 @@
  "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": "Qwen3.5-35B-A3B-GPTQ-Int4",
    "profiles": {
      "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit": {
+        "local_path": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
        "ctx": "24576",
        "trust_remote": true,
        "valid_tp": [2],
@@ -34,6 +37,7 @@
        "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],
@@ -46,6 +50,7 @@
        "hf_model_id": "THUDM/GLM-4.7-Flash-AWQ"
      },
      "Qwen3.5-27B-FP8": {
+        "local_path": "Qwen3.5-27B-FP8",
        "ctx": "32768",
        "trust_remote": true,
        "valid_tp": [1, 2],
@@ -58,6 +63,7 @@
        "hf_model_id": "RedHatAI/Qwen3.5-27B-FP8-dynamic"
      },
      "Qwen3.5-35B-A3B-GPTQ-Int4": {
+        "local_path": "Qwen3.5-35B-A3B-GPTQ-Int4",
        "ctx": "32768",
        "trust_remote": true,
        "valid_tp": [1, 2],