[Fix] [NPU] dspark: Fix attn_tp > 1 draft wo_b all_reduce group

python/sglang/srt/hardware_backend/npu/attention/ascend_dsv4_backend.py

@@ -1021,6 +1021,18 @@ def _init_dspark_sparse_metadata(self, forward_batch: ForwardBatch) -> None:
             out_cache_loc=out_cache_loc,
             block_size=block_size,
         )
+        # The query is padded to an attn_tp multiple (prepare_mlp_sync_batch),
+        # but seq_lens is not, so pad the tail with -1 ("skip") to match.
+        num_padded_tokens = out_cache_loc.shape[0]
+        if ori_sparse_indices.shape[0] < num_padded_tokens:
+            pad_rows = num_padded_tokens - ori_sparse_indices.shape[0]
+            pad = torch.full(
+                (pad_rows, ori_sparse_indices.shape[1]),
+                -1,
+                dtype=ori_sparse_indices.dtype,
+                device=ori_sparse_indices.device,
+            )
+            ori_sparse_indices = torch.cat([ori_sparse_indices, pad], dim=0)
         ori_sparse_indices = ori_sparse_indices.unsqueeze(1).contiguous()
 
         fm.ori_sparse_indices = ori_sparse_indices

python/sglang/srt/models/deepseek_v4.py

@@ -1015,6 +1015,7 @@ def __init__(
             bias=False,
             quant_config=quant_config,
             reduce_results=reduce_results,
+            use_dp_attention_reduce=is_dp_attention_enabled(),
             prefix=add_prefix("wo_b", prefix),
             parallel_group="attn_tp",
         )

python/sglang/srt/speculative/dspark_components/dspark_worker_v2.py

@@ -175,6 +175,7 @@ def __init__(
         )
         if (
             get_parallel().attn_dp_enabled
+            and not _is_npu
             and self._draft_is_moe
             and get_parallel().attn_tp_size > 1
         ):