[Bugfix] RoBERTa position_id accumulation in CUDA graph padding region (#37873)

Signed-off-by: dass90 <3053034939@qq.com>
This commit is contained in:
yanghui1-arch
2026-03-23 14:59:21 +00:00
committed by GitHub
parent 45bd5c8e75
commit 7151ae6528
+2
View File
@@ -3084,6 +3084,8 @@ class GPUModelRunner(
positions = self.xdrope_positions.gpu[:, :num_input_tokens]
else:
positions = self.positions.gpu[:num_input_tokens]
if num_input_tokens > num_scheduled_tokens:
self.positions.gpu[num_scheduled_tokens:num_input_tokens].zero_()
if is_first_rank:
intermediate_tensors = None