[Qwen3.5] Enable MTP spec_v2 and add test for nvidia/Qwen3.5-397B-A17B-NVFP4 (#19391)

python/sglang/srt/speculative/eagle_worker_v2.py

@@ -465,6 +465,7 @@ class EagleDraftWorker(BaseDraftWorker):
         batch: ModelWorkerBatch,
         target_hidden_states: torch.Tensor,
         next_token_ids: torch.Tensor,
+        mm_input_embeds: Optional[torch.Tensor] = None,
     ):
         """
         Run draft model extend to correctly fill the KV cache.
@@ -498,6 +499,8 @@ class EagleDraftWorker(BaseDraftWorker):
 
         # Run forward
         forward_batch = ForwardBatch.init_new(batch, self.draft_runner)
+        if mm_input_embeds is not None:
+            forward_batch.mm_input_embeds = mm_input_embeds
         logits_output = self.draft_runner.forward(forward_batch).logits_output
 
         # Update spec_info for the next draft step
@@ -668,6 +671,7 @@ class EAGLEWorkerV2(BaseSpecWorker):
                         model_worker_batch,
                         batch_output.logits_output.hidden_states,
                         batch_output.next_token_ids,
+                        batch_output.logits_output.mm_input_embeds,
                     )
                 )
                 return batch_output