fix(qwen2_5vl): replace in-place += with out-of-place + on expand view in decode path (#22634)

This commit is contained in:
Liu Zhenlong
2026-08-07 11:51:09 +08:00
committed by GitHub
parent f9e6888b5a
commit 453ea21dd3
@@ -1069,18 +1069,17 @@ class Qwen2_5_VLModel(nn.Module):
self.rope_deltas = rope_deltas self.rope_deltas = rope_deltas
else: else:
batch_size, seq_length, _ = inputs_embeds.shape batch_size, seq_length, _ = inputs_embeds.shape
position_ids = torch.arange(seq_length, device=inputs_embeds.device)
position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1)
if cache_position is not None: if cache_position is not None:
delta = (cache_position[0] + self.rope_deltas).to( delta = (cache_position[0] + self.rope_deltas).to(
inputs_embeds.device inputs_embeds.device
) )
else: else:
delta = torch.zeros( delta = torch.zeros(batch_size, 1, device=inputs_embeds.device)
(batch_size, seq_length), device=inputs_embeds.device position_ids = (
(torch.arange(seq_length, device=inputs_embeds.device) + delta)
.unsqueeze(0)
.expand(3, -1, -1)
) )
delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=1)
position_ids += delta.to(position_ids.device)
outputs = self.language_model( outputs = self.language_model(
input_ids=None, input_ids=None,