diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py index 64ea2c185..81debbb9b 100644 --- a/python/sglang/srt/server_args.py +++ b/python/sglang/srt/server_args.py @@ -6565,6 +6565,12 @@ class ServerArgs: f"Please use --nsa-{label}-backend=flashmla_sparse or omit it." ) + if self.kv_cache_dtype != "bfloat16": + raise ValueError( + f"HiSparse requires bfloat16 KV cache, but got --kv-cache-dtype={self.kv_cache_dtype}. " + f"Please use --kv-cache-dtype=bfloat16." + ) + assert ( self.schedule_conservativeness >= 0 ), "schedule_conservativeness must be non-negative"