diff --git a/slime/backends/megatron_utils/model_provider.py b/slime/backends/megatron_utils/model_provider.py index 8174c7ac5..afc8ad0b4 100644 --- a/slime/backends/megatron_utils/model_provider.py +++ b/slime/backends/megatron_utils/model_provider.py @@ -33,6 +33,8 @@ def __init__( self.sequence_parallel = config.sequence_parallel if self.sequence_parallel: self.weight.sequence_parallel = True + if bias: + self.bias.sequence_parallel = True self.weight.data.normal_(mean=0.0, std=0.02) if bias: diff --git a/slime/backends/sglang_utils/arguments.py b/slime/backends/sglang_utils/arguments.py index 9755a43bb..c76946166 100644 --- a/slime/backends/sglang_utils/arguments.py +++ b/slime/backends/sglang_utils/arguments.py @@ -33,6 +33,7 @@ def add_sglang_arguments(parser): Add arguments to the parser for the SGLang server. """ parser = add_sglang_router_arguments(parser) + parser.set_defaults(router_balance_abs_threshold=10, router_balance_rel_threshold=1.2) parser.add_argument("--sglang-server-concurrency", type=int, default=512) old_add_argument = parser.add_argument