From 67084267289333faf784e895a88cf39d559922bd Mon Sep 17 00:00:00 2001 From: Liangsheng Yin Date: Sun, 18 Oct 2026 16:41:21 -0700 Subject: [PATCH] [Fix] spec: Remove `--speculative-skip-dp-mlp-sync` (#45043) --- docs/docs/advanced_features/server_arguments.mdx | 6 ------ .../ascend-npus/reference/support_features.mdx | 6 ------ python/sglang/srt/arg_groups/field_order.py | 1 - python/sglang/srt/arg_groups/fields/spec.py | 4 ---- python/sglang/srt/arg_groups/speculative_hook.py | 6 ------ python/sglang/srt/managers/scheduler.py | 1 - .../srt/speculative/dspark_components/dspark_planner.py | 1 - .../unit/managers/test_scheduler_chunked_req_gate.py | 2 +- test/registered/unit/spec/test_spec_registry.py | 1 - 9 files changed, 1 insertion(+), 27 deletions(-) diff --git a/docs/docs/advanced_features/server_arguments.mdx b/docs/docs/advanced_features/server_arguments.mdx index ecb39fb7af..8d30cb6c51 100644 --- a/docs/docs/advanced_features/server_arguments.mdx +++ b/docs/docs/advanced_features/server_arguments.mdx @@ -1758,12 +1758,6 @@ Combining `--enable-response-store` with `--disaggregation-mode=prefill` or `dec `None` Type: str - - `--speculative-skip-dp-mlp-sync` - Skip the extra MLP sync that the scheduler performs before merging a new batch when speculative decoding + DP attention are both enabled. - `False` - bool flag (set to enable) - diff --git a/docs/docs/hardware-platforms/ascend-npus/reference/support_features.mdx b/docs/docs/hardware-platforms/ascend-npus/reference/support_features.mdx index fe30fc3ac9..a4d554402b 100644 --- a/docs/docs/hardware-platforms/ascend-npus/reference/support_features.mdx +++ b/docs/docs/hardware-platforms/ascend-npus/reference/support_features.mdx @@ -1592,12 +1592,6 @@ non-default speculative acceptance thresholds or deterministic inference. Type: str A2, A3, 950PR/DT Series - - `--speculative-skip-dp-mlp-sync` - `False` - bool flag (set to enable) - A2, A3, 950PR/DT Series - diff --git a/python/sglang/srt/arg_groups/field_order.py b/python/sglang/srt/arg_groups/field_order.py index 0e80eaf887..43d21dd4b1 100644 --- a/python/sglang/srt/arg_groups/field_order.py +++ b/python/sglang/srt/arg_groups/field_order.py @@ -288,7 +288,6 @@ POSITIONAL_FIELD_ORDER = ( "speculative_enable_w4a4_mxfp4_megamoe", "speculative_draft_model_quantization", "_speculative_draft_quantization_explicitly_set", - "speculative_skip_dp_mlp_sync", "enable_multi_layer_eagle", "speculative_adaptive", "speculative_adaptive_config", diff --git a/python/sglang/srt/arg_groups/fields/spec.py b/python/sglang/srt/arg_groups/fields/spec.py index b7367a447a..d86620d28a 100644 --- a/python/sglang/srt/arg_groups/fields/spec.py +++ b/python/sglang/srt/arg_groups/fields/spec.py @@ -226,10 +226,6 @@ class Spec(msgspec.Struct): Optional[bool], Arg(no_cli=True), ] = None - speculative_skip_dp_mlp_sync: A[ - bool, - "Skip the extra MLP sync that the scheduler performs before merging a new batch when speculative decoding + DP attention are both enabled.", - ] = False enable_multi_layer_eagle: A[ bool, Arg( diff --git a/python/sglang/srt/arg_groups/speculative_hook.py b/python/sglang/srt/arg_groups/speculative_hook.py index 4ab139b1d4..40995d45ea 100644 --- a/python/sglang/srt/arg_groups/speculative_hook.py +++ b/python/sglang/srt/arg_groups/speculative_hook.py @@ -248,12 +248,6 @@ def handle_speculative_decoding(server_args: ServerArgs) -> None: algo.validate_server_args, ) - if cfg.speculative_skip_dp_mlp_sync: - assert cfg.speculative_algorithm == "EAGLE", ( - "--speculative-skip-dp-mlp-sync is only supported with " - f"speculative_algorithm == EAGLE, got {cfg.speculative_algorithm}." - ) - if envs.SGLANG_ENABLE_DP_SPEC_PREFILL_COORDINATION.get(): if ( cfg.speculative_algorithm not in ("EAGLE", "EAGLE3", "DSPARK") diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py index 6f783d667b..644c3aac6e 100644 --- a/python/sglang/srt/managers/scheduler.py +++ b/python/sglang/srt/managers/scheduler.py @@ -3914,7 +3914,6 @@ class Scheduler( if ( need_mlp_sync and not self.spec_algorithm.is_none() - and not get_spec().speculative_skip_dp_mlp_sync and not envs.SGLANG_ENABLE_DP_SPEC_PREFILL_COORDINATION.get() ): # NOTE: This branch makes sure prefill and decode batches will not be mixed when spec and dp-attn is enabled. diff --git a/python/sglang/srt/speculative/dspark_components/dspark_planner.py b/python/sglang/srt/speculative/dspark_components/dspark_planner.py index 34198b4de5..17a342c721 100644 --- a/python/sglang/srt/speculative/dspark_components/dspark_planner.py +++ b/python/sglang/srt/speculative/dspark_components/dspark_planner.py @@ -171,7 +171,6 @@ class DSparkVerifyPlanner: and get_parallel().attn_cp_size == 1 and require_mlp_tp_gather() and not get_schedule().disable_overlap_schedule - and not get_spec().speculative_skip_dp_mlp_sync and get_disagg().disaggregation_mode == "null" and get_parallel().pp_size == 1 and not envs.SGLANG_SCHEDULER_SKIP_ALL_GATHER.get() diff --git a/test/registered/unit/managers/test_scheduler_chunked_req_gate.py b/test/registered/unit/managers/test_scheduler_chunked_req_gate.py index 2f0c1cf4c2..97e9442996 100644 --- a/test/registered/unit/managers/test_scheduler_chunked_req_gate.py +++ b/test/registered/unit/managers/test_scheduler_chunked_req_gate.py @@ -70,7 +70,7 @@ def _scheduler_for_get_next_batch(*, tree_cache, chunked_req) -> Scheduler: s.last_batch = None s.require_mlp_sync = False s.spec_algorithm = MagicMock() - s.server_args = MagicMock(speculative_skip_dp_mlp_sync=True) + s.server_args = MagicMock() s.running_batch = MagicMock() s.running_batch.is_empty.return_value = True s.running_batch.batch_size.return_value = 0 diff --git a/test/registered/unit/spec/test_spec_registry.py b/test/registered/unit/spec/test_spec_registry.py index 1fd2996bd2..bc0771f16e 100644 --- a/test/registered/unit/spec/test_spec_registry.py +++ b/test/registered/unit/spec/test_spec_registry.py @@ -221,7 +221,6 @@ class TestServerArgsHook(_RegistryIsolated): trust_remote_code=False, speculative_draft_window_size=None, speculative_draft_sink_size=None, - speculative_skip_dp_mlp_sync=False, speculative_adaptive=False, )