[VLM][feat] Support encoder DP for Qwen2.5-VL (#13126)

Co-authored-by: Shangming Cai <csmthu@gmail.com>
Co-authored-by: liusy58 <xiehang.lsy@alibaba-inc.com>
Co-authored-by: Yuan Luo <yuan.luo@hotmail.com>
This commit is contained in:
Nicholas
2025-11-18 16:13:18 +08:00
committed by GitHub
co-authored by Shangming Cai liusy58 Yuan Luo
parent 33905005ee
commit ac81db66c2
6 changed files with 595 additions and 6 deletions
+3 -4
View File
@@ -486,13 +486,12 @@ class VisionAttention(nn.Module):
customized_position_embedding_applier: Callable[
[torch.Tensor, torch.Tensor, Any, Any], Tuple[torch.Tensor, torch.Tensor]
] = None,
use_data_parallel: bool = False,
**kwargs,
):
super().__init__()
attn_tp_rank = get_attention_tp_rank()
attn_tp_size = get_attention_tp_size()
self.tp_size = attn_tp_size
self.tp_rank = attn_tp_rank
self.tp_size = 1 if use_data_parallel else get_attention_tp_size()
self.tp_rank = 0 if use_data_parallel else get_attention_tp_rank()
self.dropout = dropout
self.head_size = embed_dim // num_heads
self.hidden_size_per_attention_head = dist_utils.divide(