fix(pd): fix kv pools without end_layer (#25476)

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

@@ -56,7 +56,7 @@ class KVArgs:
     # reconstruct PP sub-ranges when kv_data_ptrs does not use a flat
     # layer-indexed layout (e.g. DeepSeek V4's buffer-type-organized flat
     # list).
-    prefill_end_layer: int
+    prefill_end_layer: Optional[int]
     # For DeepSeek V4 (and other compressed-MLA) memory pools only.
     # Full-model compression ratio per layer (entries are 0/4/128). Used by
     # the connection layer to slice the buffer-type-organized flat list in a

python/sglang/srt/disaggregation/prefill.py

@@ -147,7 +147,7 @@ class PrefillBootstrapQueue:
         kv_args.pp_rank = self.pp_rank
         kv_args.system_dp_rank = self.scheduler.ps.dp_rank
         kv_args.prefill_start_layer = self.token_to_kv_pool.start_layer
-        kv_args.prefill_end_layer = self.token_to_kv_pool.end_layer
+        kv_args.prefill_end_layer = getattr(self.token_to_kv_pool, "end_layer", None)
         kv_args.mla_compression_ratios = None
         kv_data_ptrs, kv_data_lens, kv_item_lens = (
             self.token_to_kv_pool.get_contiguous_buf_infos()