[DCP] Enable decode context parallel for Kimi K2.5 NVFP4 (#31514)

Co-authored-by: kpham-sgl <kpham-sgl@users.noreply.github.com>

python/sglang/srt/layers/dcp/comm.py

@@ -181,7 +181,7 @@ def all_gather_kv_cache_for_mha_extend(
     k_pe: torch.Tensor,
 ):
     prefix_kv_a, prefix_k_pe = token_to_kv_pool.get_mla_kv_buffer(
-        attn_mqa, dcp_local_prefix_kv_indices
+        attn_mqa, dcp_local_prefix_kv_indices, dst_dtype=kv_a.dtype
     )
     extend_prefix_lens_cpu = torch.tensor(extend_prefix_lens_cpu)
     gathered_kv_cache = all_gather_kv_cache_for_dcp(

python/sglang/srt/models/kimi_k25.py

@@ -786,6 +786,35 @@ class KimiK25ForConditionalGeneration(nn.Module):
             else {}
         )
 
+    def prepare_context_parallel_metadata_for_dcp(
+        self,
+        seq_lens: torch.Tensor,
+        extend_prefix_lens: torch.Tensor,
+        extend_prefix_lens_cpu: torch.Tensor,
+        extend_seq_lens: torch.Tensor,
+        req_pool_indices: torch.Tensor,
+        req_to_token: torch.Tensor,
+        seq_lens_sum: int,
+        kv_buffer_shape: torch.Size,
+        kv_cache_dtype,
+        kv_cache_device,
+        create_chunked_prefix_cache_kv_indices_fn,
+    ):
+        # DCP metadata is built on the inner DeepSeek-V3 language model.
+        return self.language_model.prepare_context_parallel_metadata_for_dcp(
+            seq_lens=seq_lens,
+            extend_prefix_lens=extend_prefix_lens,
+            extend_prefix_lens_cpu=extend_prefix_lens_cpu,
+            extend_seq_lens=extend_seq_lens,
+            req_pool_indices=req_pool_indices,
+            req_to_token=req_to_token,
+            seq_lens_sum=seq_lens_sum,
+            kv_buffer_shape=kv_buffer_shape,
+            kv_cache_dtype=kv_cache_dtype,
+            kv_cache_device=kv_cache_device,
+            create_chunked_prefix_cache_kv_indices_fn=create_chunked_prefix_cache_kv_indices_fn,
+        )
+
     def forward(
         self,
         input_ids: torch.Tensor,