diff --git a/config.json b/config.json index ca9c0ca..633a462 100644 --- a/config.json +++ b/config.json @@ -69,19 +69,18 @@ "local_path": "Qwen3.6-27B-FP8", "dtype": "auto", "quantization": "fp8", - "ctx": "262144", - "max_tokens": "65536", - "max_num_batched_tokens": 32768, + "ctx": "32768", + "max_tokens": "16384", + "max_num_batched_tokens": 8192, "trust_remote": true, "enforce_eager": false, "valid_tp": [ - 1, 2 ], - "max_num_seqs": 256, - "gpu_util": "0.92", + "max_num_seqs": 16, + "gpu_util": "0.80", "kv_cache_dtype": "fp8", - "enable_prefix_caching": true, + "enable_prefix_caching": false, "tool_call_parser": "qwen3_coder", "reasoning_parser": "qwen3", "enable_auto_tool_choice": true,