Fix decode OOM caused by retraction (#14939)

This commit is contained in:
Liangsheng Yin
2025-12-13 12:59:17 +08:00
committed by GitHub
parent 8698867479
commit 01e3b3f3a3
4 changed files with 10 additions and 4 deletions
+1 -1
View File
@@ -1965,7 +1965,7 @@ class Scheduler(
):
old_ratio = self.new_token_ratio
retracted_reqs, new_token_ratio, reqs_to_abort = batch.retract_decode(
self.server_args
self.server_args, self.decode_mem_cache_buf_multiplier
)
self.num_retracted_reqs = len(retracted_reqs)
self.new_token_ratio = new_token_ratio