fix(pcg,mm): fix zeroing of input_embeds when replay PCG (#22229)

This commit is contained in:
Yujun Dong
2026-04-07 20:33:17 +08:00
committed by GitHub
parent 98f38b14df
commit 233f3e31bf
@@ -609,7 +609,7 @@ class PiecewiseCudaGraphRunner:
buffers.input_ids[num_tokens:static_num_tokens].zero_()
buffers.positions[num_tokens:static_num_tokens].zero_()
if self.is_multimodal:
buffers.input_embeds[:, num_tokens:static_num_tokens].zero_()
buffers.input_embeds[num_tokens:static_num_tokens].zero_()
if forward_batch.mrope_positions is not None:
buffers.mrope_positions[:, num_tokens:static_num_tokens].zero_()