diff --git a/config.json b/config.json index 3da0714..5a583a6 100644 --- a/config.json +++ b/config.json @@ -20,7 +20,7 @@ "revision": "", "models": { "default": "Qwen3.5-35B-A3B-GPTQ-Int4", - "selected": "Qwen3.5-35B-A3B-FP8", + "selected": "Qwen3.5-27B-FP8", "profiles": { "GLM-4.7-Flash-AWQ": { "local_path": "GLM-4.7-Flash-AWQ", @@ -39,13 +39,13 @@ "Qwen3.5-27B-FP8": { "local_path": "Qwen3.5-27B-FP8", "dtype": "auto", - "ctx": "65536", + "ctx": "131072", "trust_remote": true, "valid_tp": [1, 2], "max_num_seqs": 12, - "max_tokens": "8192", + "max_tokens": "16384", "enforce_eager": false, - "gpu_util": "0.9", + "gpu_util": "0.94", "kv_cache_dtype": "fp8", "enable_prefix_caching": true, "max_num_batched_tokens": 16384,