--- a/vllm/v1/attention/backends/mla/flashinfer_mla_sparse_sm120.py
+++ b/vllm/v1/attention/backends/mla/flashinfer_mla_sparse_sm120.py
@@ -32,6 +32,13 @@
 class FlashInferMLASparseSM120Impl(SparseMLAAttentionImpl[FlashInferMLASparseMetadata]):
     """SM120 FlashInfer sparse-MLA implementation."""
 
+    # glm52-dcp patch: FlashInfer's trtllm_batch_decode_with_kv_cache_mla
+    # accepts return_lse (verified in flashinfer 0.6.14), so this impl CAN
+    # return the decode LSE that DCP's cross-rank merge requires. Upstream
+    # never plumbed it through; see forward_mqa below. Mirrors the parent
+    # FlashInferMLASparseImpl's LSE handling.
+    can_return_lse_for_decode = True
+
     def __init__(
         self,
         num_heads: int,
@@ -126,8 +133,12 @@
             ),
         )
 
+        # glm52-dcp patch: size the out buffer from the query's actual head
+        # count, not self.num_heads. Under DCP the MLA decode all-gathers
+        # queries across ranks, so each rank sees the FULL head count (64)
+        # rather than the TP-local slice (16) — self.num_heads is stale here.
         output = q.new_empty(
-            (num_actual_toks, self.num_heads, self.kv_lora_rank),
+            (num_actual_toks, q.shape[1], self.kv_lora_rank),
             dtype=q.dtype,
         )
 
@@ -153,5 +164,27 @@
             bmm2_scale=1.0,
             sparse_mla_top_k=attn_metadata.topk_tokens,
             kv_scale_format=self.kv_scale_format,
+            # glm52-dcp patch: ask the kernel for the decode LSE when DCP
+            # needs it for the cross-rank merge (base class sets this flag
+            # only when dcp_world_size > 1).
+            return_lse=self.need_to_return_lse_for_decode,
         )
+        if self.need_to_return_lse_for_decode:
+            # With return_lse=True flashinfer returns (out, lse).
+            assert isinstance(out, tuple), (
+                "expected (out, lse) tuple from flashinfer when return_lse=True"
+            )
+            o, lse = out
+            o = o.view(-1, o.shape[-2], o.shape[-1])
+            # Reuse the parent impl's LSE normalization + empty-row masking
+            # conventions (the shared DCP reducer expects (num_tokens, num_heads)
+            # and -inf LSE on rows with no valid top-k indices).
+            from vllm.v1.attention.backends.mla.flashinfer_mla_sparse import (
+                FlashInferMLASparseImpl as _ParentImpl,
+            )
+            lse = _ParentImpl._normalize_lse(lse, o.shape[0], o.shape[1])
+            empty_rows = (topk_indices_physical == -1).all(dim=-1)
+            o.masked_fill_(empty_rows.view(-1, 1, 1), 0.0)
+            lse.masked_fill_(empty_rows.view(-1, 1), float("-inf"))
+            return o, lse
         return out.squeeze(1), None