Upgrade to vllm 0.17.0 corex v4.1 overlay

vllm/v1/attention/backends/mla/flashmla.py

@@ -156,6 +156,8 @@ class FlashMLAMetadataBuilder(MLACommonMetadataBuilder[FlashMLAMetadata]):
         query_start_loc_device: torch.Tensor,
         num_decode_tokens: int,
         dcp_tot_seq_lens_device: torch.Tensor | None,
+        max_decode_seq_len: int = 0,
+        use_cuda_graph: bool = False,
     ) -> FlashMLADecodeMetadata:
         query_lens_cpu = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]
         # we use the max but all should be the same due to uniform length requirement
@@ -179,8 +181,10 @@ class FlashMLAMetadataBuilder(MLACommonMetadataBuilder[FlashMLAMetadata]):
         return FlashMLADecodeMetadata(
             block_table=block_table_tensor,
             seq_lens=seq_lens_device,
-            scheduler_metadata=scheduler_metadata,
             dcp_tot_seq_lens=dcp_tot_seq_lens_device,
+            max_decode_seq_len=max_decode_seq_len,
+            use_cuda_graph=use_cuda_graph,
+            scheduler_metadata=scheduler_metadata,
         )