[DLLM] Add initial cuda graph support (#14203)

This commit is contained in:
Tiwei Bie
2025-12-08 14:12:35 +08:00
committed by GitHub
parent 661e9775d0
commit 36361adcbf
6 changed files with 93 additions and 8 deletions
+3 -1
View File
@@ -1318,7 +1318,9 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
), f"Expected {len(self.out_cache_loc)}, got {self.extend_num_tokens}"
def prepare_for_extend(self):
self.forward_mode = ForwardMode.EXTEND
self.forward_mode = (
ForwardMode.DLLM_EXTEND if self.is_dllm() else ForwardMode.EXTEND
)
# Init tensors
reqs = self.reqs