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,
)