convert_req_index_to_global_index opt

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

@@ -302,6 +302,18 @@ def triton_convert_req_index_to_global_index(
     prefill_workspace_starts: int32 [num_prefills], 0-indexed workspace
         starts for each prefill request
     """
+    if (envs.USE_LIGHTOP_CONVERT_REQ_INDEX_TO_GLOBAL_INDEX):
+        from lightop import op
+        return op.convert_req_index_to_global_index(
+            req_id, 
+            block_table, 
+            token_indices, 
+            BLOCK_SIZE, 
+            NUM_TOPK_TOKENS,
+            HAS_PREFILL_WORKSPACE, 
+            prefill_workspace_request_ids, 
+            prefill_workspace_starts
+        )
     assert req_id.dtype == torch.int32
     assert block_table.dtype == torch.int32
     assert token_indices.dtype == torch.int32