Fix LongCat n-gram embedding in PD-disaggregated scheduler loops (#31134)

Co-authored-by: whn09 <whn09@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

python/sglang/srt/disaggregation/decode.py

@@ -2001,6 +2001,9 @@ class SchedulerDisaggregationDecodeMixin:
             )
             self.running_batch = plan.running_batch
             batch = plan.batch_to_run
+            batch = self.ngram_embedding_manager.prepare_for_forward(
+                batch, chunked_req=self.chunked_req
+            )
             self.cur_batch_for_debug = batch
 
             # Launch the current batch
@@ -2039,6 +2042,9 @@ class SchedulerDisaggregationDecodeMixin:
             )
             self.running_batch = plan.running_batch
             batch = plan.batch_to_run
+            batch = self.ngram_embedding_manager.prepare_for_forward(
+                batch, chunked_req=self.chunked_req
+            )
             self.cur_batch_for_debug = batch
             # overlap + spec + grammar is unsupported (would desync DP ranks).
             disable_overlap_for_batch = self.is_disable_overlap_for_batch(

python/sglang/srt/disaggregation/prefill.py

@@ -521,6 +521,9 @@ class SchedulerDisaggregationPrefillMixin:
             )
             self.running_batch = plan.running_batch
             batch = plan.batch_to_run
+            batch = self.ngram_embedding_manager.prepare_for_forward(
+                batch, chunked_req=self.chunked_req
+            )
             self.cur_batch_for_debug = batch
 
             # Launch the current batch
@@ -559,6 +562,9 @@ class SchedulerDisaggregationPrefillMixin:
             )
             self.running_batch = plan.running_batch
             batch = plan.batch_to_run
+            batch = self.ngram_embedding_manager.prepare_for_forward(
+                batch, chunked_req=self.chunked_req
+            )
             self.cur_batch_for_debug = batch
 
             # Launch the current batch