[PD] optimize kv cache transfer directly using batch transfer (#9149)

Co-authored-by: Shangming Cai <csmthu@gmail.com>

python/sglang/srt/disaggregation/mooncake/conn.py

@@ -356,33 +356,49 @@ class MooncakeKVManager(BaseKVManager):
             ]
         assert layers_params is not None
 
-        # Worker function for processing a single layer
-        def process_layer(src_ptr: int, dst_ptr: int, item_len: int) -> int:
+        def set_transfer_blocks(
+            src_ptr: int, dst_ptr: int, item_len: int
+        ) -> List[Tuple[int, int, int]]:
             transfer_blocks = []
             for prefill_index, decode_index in zip(prefill_kv_blocks, dst_kv_blocks):
                 src_addr = src_ptr + int(prefill_index[0]) * item_len
                 dst_addr = dst_ptr + int(decode_index[0]) * item_len
                 length = item_len * len(prefill_index)
                 transfer_blocks.append((src_addr, dst_addr, length))
+            return transfer_blocks
 
+        # Worker function for processing a single layer
+        def process_layer(src_ptr: int, dst_ptr: int, item_len: int) -> int:
+            transfer_blocks = set_transfer_blocks(src_ptr, dst_ptr, item_len)
             return self._transfer_data(mooncake_session_id, transfer_blocks)
 
-        futures = [
-            executor.submit(
-                process_layer,
-                src_ptr,
-                dst_ptr,
-                item_len,
-            )
-            for (src_ptr, dst_ptr, item_len) in layers_params
-        ]
+        # Worker function for processing all layers in a batch
+        def process_layers(layers_params: List[Tuple[int, int, int]]) -> int:
+            transfer_blocks = []
+            for src_ptr, dst_ptr, item_len in layers_params:
+                transfer_blocks.extend(set_transfer_blocks(src_ptr, dst_ptr, item_len))
+            return self._transfer_data(mooncake_session_id, transfer_blocks)
 
-        for future in concurrent.futures.as_completed(futures):
-            status = future.result()
-            if status != 0:
-                for f in futures:
-                    f.cancel()
-                return status
+        if self.enable_custom_mem_pool:
+            futures = [
+                executor.submit(
+                    process_layer,
+                    src_ptr,
+                    dst_ptr,
+                    item_len,
+                )
+                for (src_ptr, dst_ptr, item_len) in layers_params
+            ]
+            for future in concurrent.futures.as_completed(futures):
+                status = future.result()
+                if status != 0:
+                    for f in futures:
+                        f.cancel()
+                    return status
+        else:
+            # Combining all layers' params in one batch transfer is more efficient
+            # compared to using multiple threads
+            return process_layers(layers_params)
 
         return 0