EPLB support for MTP (#7510)

python/sglang/srt/managers/expert_distribution.py

@@ -61,6 +61,10 @@ class ExpertDistributionRecorder(ABC):
     def with_debug_name(self, debug_name):
         yield
 
+    @contextmanager
+    def disable_this_region(self):
+        yield
+
     @contextmanager
     def with_forward_pass(self, forward_pass_id: int, forward_batch: ForwardBatch):
         yield
@@ -116,6 +120,7 @@ class _ExpertDistributionRecorderReal(ExpertDistributionRecorder):
         self._expert_location_metadata = expert_location_metadata
 
         self._recording = False
+        self._disable_all = False
         self._current_forward_pass_id = Withable()
         self._current_layer_idx = Withable()
         self._current_debug_name = Withable()
@@ -148,6 +153,16 @@ class _ExpertDistributionRecorderReal(ExpertDistributionRecorder):
             finally:
                 self._on_forward_pass_end(forward_pass_id)
 
+    @contextmanager
+    def disable_this_region(self):
+        """Context manager to temporarily disable recording."""
+        previous_disable_all = self._disable_all
+        self._disable_all = True
+        try:
+            yield
+        finally:
+            self._disable_all = previous_disable_all
+
     def _on_forward_pass_start(self, forward_batch: ForwardBatch):
         if not self._recording:
             return
@@ -189,6 +204,8 @@ class _ExpertDistributionRecorderReal(ExpertDistributionRecorder):
         )
 
     def _on_hook(self, hook_name: str, **kwargs):
+        if self._disable_all:
+            return
         if not (self._recording or torch.cuda.is_current_stream_capturing()):
             return
         gatherer = self._single_pass_gatherers[
@@ -462,6 +479,10 @@ class _SelectExpertsSinglePassGatherer(_LayerBasedGpuSinglePassGatherer):
     def on_select_experts(self, layer_idx: int, topk_ids: torch.Tensor):
         topk_ids = topk_ids.flatten()
         mask = topk_ids != -1
+        assert self._data[layer_idx, :].shape == topk_ids.shape, (
+            "Shape mismatch between data and topk_ids."
+            "Selecting expert is not supported for multiple token prediction at the moment."
+        )
         self._data[layer_idx, :].scatter_add_(
             dim=0, index=topk_ids.masked_fill(~mask, 0).long(), src=mask.int()
         )

python/sglang/srt/models/deepseek_nextn.py

@@ -28,6 +28,9 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
+from sglang.srt.managers.expert_distribution import (
+    get_global_expert_distribution_recorder,
+)
 from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.models.deepseek_v2 import DeepseekV2DecoderLayer, DeepseekV3ForCausalLM
@@ -82,7 +85,6 @@ class DeepseekModelNextN(nn.Module):
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
-
         zero_allocator = BumpAllocator(
             buffer_size=2,
             dtype=torch.float32,
@@ -108,9 +110,10 @@ class DeepseekModelNextN(nn.Module):
             )
 
         residual = None
-        hidden_states, residual = self.decoder(
-            positions, hidden_states, forward_batch, residual, zero_allocator
-        )
+        with get_global_expert_distribution_recorder().disable_this_region():
+            hidden_states, residual = self.decoder(
+                positions, hidden_states, forward_batch, residual, zero_allocator
+            )
 
         if not forward_batch.forward_mode.is_idle():
             if residual is not None: