[Spec] Clamp multimodal pad sentinels in spec-v2 draft prefill embedding (#27512)

python/sglang/srt/models/mimo_v2_nextn.py

@@ -201,7 +201,12 @@ class MiMoV2ModelNextN(nn.Module):
         input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
         if input_embeds is None:
-            hidden_states = self.embed_tokens(input_ids)
+            # Multimodal pad sentinels (MM_PAD_SHIFT_VALUE + hash) sit out of vocab;
+            # clamp to avoid an OOB gather. The draft gets visual semantics from target
+            # hidden_states, so the embedding at these positions is unused anyway.
+            hidden_states = self.embed_tokens(
+                input_ids.clamp(min=0, max=self.vocab_size - 1)
+            )
         else:
             hidden_states = input_embeds
         if hidden_states.shape[0] > 0:

python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py

@@ -383,6 +383,16 @@ class MultiLayerEagleDraftWorker(BaseDraftWorker):
             target_hidden_states: Hidden states from the target model forward
             next_token_ids: Next token ids generated from the target forward.
         """
+        # The draft embed clamps unconditionally (to tolerate multimodal pad
+        # sentinels), so probe next_token_ids here first -- otherwise a corrupted id
+        # would be clamped away instead of surfacing.
+        maybe_detect_oob(
+            next_token_ids,
+            0,
+            self.model_config.vocab_size,
+            "draft_extend_for_prefill: next_token_ids before draft embed",
+        )
+
         # Construct spec_info
         next_draft_input = EagleDraftInput(
             hidden_states=target_hidden_states,