x

2026-04-23 15:42:35 +08:00
parent b534ceac58
commit 719c593c82
2 changed files with 30 additions and 14 deletions
@@ -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",