diff --git a/python/sglang/srt/arg_groups/field_order.py b/python/sglang/srt/arg_groups/field_order.py index 43d21dd4b1..03bb58d317 100644 --- a/python/sglang/srt/arg_groups/field_order.py +++ b/python/sglang/srt/arg_groups/field_order.py @@ -1,6 +1,7 @@ """Field order preserving the ServerArgs positional constructor signature. -``collect_input_fields`` appends unlisted fields after these entries. +Append new built-in fields to this tuple; inserting or regrouping changes existing +calls. ``collect_input_fields`` appends unlisted extension fields after these entries. """ # fmt: off @@ -100,7 +101,6 @@ POSITIONAL_FIELD_ORDER = ( "shared_experts_tp_size", "enable_dense_mlp_attn_tp", "enable_layernorm_sp", - "enable_megamoe_for_layernorm_sp", "disable_attn_tp_gather", "enable_p2p_check", "device", @@ -521,5 +521,30 @@ POSITIONAL_FIELD_ORDER = ( "enable_decoder_swa_bounded_replay", "enable_response_store", "disaggregation_decode_host_receive_threshold", + "qsa_indexer_dtype", + "disaggregation_decode_allocation_policy", + "speculative_boundary_reduction", + "speculative_use_block_verification", + "speculative_draft_sink_size", + "disable_aiter_allreduce_fusion_in_prefill", + "disable_aiter_allreduce_fusion_in_decode", + "monokernel_backend", + "dsa_triton_union", + "enable_dsa_fused_indexer", + "enable_cp_tp_group_sharing", + "enable_watermark", + "watermark_config", + "watermark_key", + "watermark_key_b", + "watermark_context_window", + "watermark_mixing_probability", + "watermark_max_probability", + "watermark_default_enabled", + "watermark_enforce_all", + "grpc_response_timeout_secs", + "enable_paged_experts", + "paged_experts_num_resident", + "enable_nvidia_megakernel", + "enable_megamoe_for_layernorm_sp", ) # fmt: on diff --git a/test/registered/unit/server_args/test_server_args.py b/test/registered/unit/server_args/test_server_args.py index e2155722e5..6e02dc78ec 100644 --- a/test/registered/unit/server_args/test_server_args.py +++ b/test/registered/unit/server_args/test_server_args.py @@ -28,6 +28,7 @@ from sglang.srt.arg_groups.cuda_graph_hook import ( finalize_cuda_graph_prefill_max_context, handle_cuda_graph_config, ) +from sglang.srt.arg_groups.field_order import POSITIONAL_FIELD_ORDER from sglang.srt.arg_groups.hicache_hook import ( handle_hicache, handle_hicache_ratio_default, @@ -127,6 +128,61 @@ _mock_device = patch( _mock_device.start() +class TestServerArgsPositionalOrder(CustomTestCase): + def test_new_parallel_flag_preserves_existing_positional_slots(self): + """New options must not reinterpret existing positional constructor calls.""" + fields = msgspec.structs.fields(ServerArgs) + for index, name in ( + (95, "disable_attn_tp_gather"), + (96, "enable_p2p_check"), + (97, "device"), + (535, "grpc_response_timeout_secs"), + (536, "enable_paged_experts"), + (537, "paged_experts_num_resident"), + (538, "enable_nvidia_megakernel"), + (539, "enable_megamoe_for_layernorm_sp"), + ): + with self.subTest(index=index): + self.assertEqual(fields[index].name, name) + defaults = ServerArgs(model_path="dummy") + # Surviving legacy inputs retain their slots after an intentional removal. + positional = [getattr(defaults, field.name) for field in fields[:536]] + positional[95] = True + positional[96] = False + positional[97] = "cpu" + positional[535] = 123.0 + + args = ServerArgs(*positional) + + self.assertTrue(args.disable_attn_tp_gather) + self.assertFalse(args.enable_p2p_check) + self.assertEqual(args.device, "cpu") + self.assertEqual(args.grpc_response_timeout_secs, 123.0) + self.assertFalse(args.enable_megamoe_for_layernorm_sp) + + def test_builtin_fields_have_explicit_positional_slots(self): + """New namespace fields cannot shift previously appended built-in options.""" + self.assertEqual( + tuple(field.name for field in msgspec.structs.fields(ServerArgs)), + POSITIONAL_FIELD_ORDER, + ) + + def test_new_parallel_flag_keeps_keyword_and_cli_inputs(self): + parser = argparse.ArgumentParser() + ServerArgs.add_cli_args(parser) + for enabled in (False, True): + with self.subTest(enabled=enabled): + keyword_args = ServerArgs( + model_path="dummy", enable_megamoe_for_layernorm_sp=enabled + ) + argv = ["--model-path", "dummy"] + if enabled: + argv.append("--enable-megamoe-for-layernorm-sp") + cli_args = ServerArgs.from_cli_args(parser.parse_args(argv)) + self.assertEqual(keyword_args.enable_megamoe_for_layernorm_sp, enabled) + self.assertEqual(cli_args.enable_megamoe_for_layernorm_sp, enabled) + + class TestPrepareServerArgs(CustomTestCase): def test_optimistic_prefill_allows_l2_write_through_only(self): for policy, expected in (