Convert cu_seqlens to CPU for npu_flash_attention_unpad operator (#15434)
This commit is contained in:
@@ -470,7 +470,7 @@ class VisionAscendAttention(nn.Module):
|
||||
Returns:
|
||||
[b * s, h, head_size]
|
||||
"""
|
||||
cu_seqlens = resolve_seqlens(cu_seqlens, bsz, seq_len, device=q.device)
|
||||
cu_seqlens = resolve_seqlens(cu_seqlens, bsz, seq_len, device="cpu")
|
||||
|
||||
seq_lens = cu_seqlens[1:] - cu_seqlens[:-1]
|
||||
if seq_lens.is_npu:
|
||||
|
||||
Reference in New Issue
Block a user