minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

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

Browse Source 
Co-authored-by: sglang-npu-bot <sglangnpu@163.com>
This commit is contained in:
Liwansi
2026-09-08 11:44:35 +08:00
committed by GitHub
co-authored by sglang-npu-bot
parent dfd9b5c2a4
commit 92371e9887
2 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/schedule_batch.py
+4 -1
View File
@@ -971,6 +971,7 @@ class Req(ReqDllmMixin):
multi_item_delimiter_indices: Optional[List[int]] = None, multi_item_delimiter_indices: Optional[List[int]] = None,
session_id: Optional[str] = None, session_id: Optional[str] = None,
cache_salt: Optional[str] = None, cache_salt: Optional[str] = None,
disable_radix_cache: bool = False,
): ):
# Input and output info # Input and output info
self.rid = rid self.rid = rid
@@ -1273,7 +1274,9 @@ class Req(ReqDllmMixin):
# retracted request is rebootstrapped. Set in pause_generation(retract) # retracted request is rebootstrapped. Set in pause_generation(retract)
# and consumed in the decode transfer commit; never plumbed to prefill. # and consumed in the decode transfer commit; never plumbed to prefill.
self.pd_rebootstrap_forced_output_id: Optional[int] = None 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.disagg_kv_sender: Optional[BaseKVSender] = None
self.routed_dp_rank: Optional[int] = routed_dp_rank self.routed_dp_rank: Optional[int] = routed_dp_rank
python/sglang/srt/managers/scheduler.py
+1
View File
@@ -3321,6 +3321,7 @@ class Scheduler(
time_stats=recv_req.time_stats, time_stats=recv_req.time_stats,
return_pooled_hidden_states=recv_req.return_pooled_hidden_states, return_pooled_hidden_states=recv_req.return_pooled_hidden_states,
multi_item_delimiter_indices=recv_req.multi_item_delimiter_indices, multi_item_delimiter_indices=recv_req.multi_item_delimiter_indices,
disable_radix_cache=self.disable_radix_cache,
) )
req.tokenizer = self.tokenizer req.tokenizer = self.tokenizer
self._maybe_namespace_elastic_radix_cache(req) self._maybe_namespace_elastic_radix_cache(req)