PD disaggregation, isolated transfer, prefill OOM fixed. (#38094)

Co-authored-by: sglang-npu-bot <sglangnpu@163.com>

python/sglang/srt/managers/schedule_batch.py

@@ -971,6 +971,7 @@ class Req(ReqDllmMixin):
         multi_item_delimiter_indices: Optional[List[int]] = None,
         session_id: Optional[str] = None,
         cache_salt: Optional[str] = None,
+        disable_radix_cache: bool = False,
     ):
         # Input and output info
         self.rid = rid
@@ -1273,7 +1274,9 @@ class Req(ReqDllmMixin):
         # retracted request is rebootstrapped. Set in pause_generation(retract)
         # and consumed in the decode transfer commit; never plumbed to prefill.
         self.pd_rebootstrap_forced_output_id: Optional[int] = None
-        self.skip_radix_cache_insert = bootstrap_host == FAKE_BOOTSTRAP_HOST
+        self.skip_radix_cache_insert = (
+            bootstrap_host == FAKE_BOOTSTRAP_HOST and disable_radix_cache
+        )
         self.disagg_kv_sender: Optional[BaseKVSender] = None
 
         self.routed_dp_rank: Optional[int] = routed_dp_rank

python/sglang/srt/managers/scheduler.py

@@ -3321,6 +3321,7 @@ class Scheduler(
             time_stats=recv_req.time_stats,
             return_pooled_hidden_states=recv_req.return_pooled_hidden_states,
             multi_item_delimiter_indices=recv_req.multi_item_delimiter_indices,
+            disable_radix_cache=self.disable_radix_cache,
         )
         req.tokenizer = self.tokenizer
         self._maybe_namespace_elastic_radix_cache(req)