Skip to content

light_attention精度有误 #132

@pdx1989

Description

@pdx1989

在triton_ascend下,调整prefill参数:
(1, 64, 5, 128, 128, torch.float32, 8), # 64
(1, 64, 72, 128, 128, torch.float32, 8), # 64
(1, 64, 72, 128, 128, torch.float32, 16), # 128
调整decode参数:
(8, 64, 1, 128, 128, torch.float32, 64),
(16, 64, 1, 128, 128, torch.float32, 64), # 128

运行dlBLAS仓库下:python tests/kernels/test_lightning_attn.py ,执行通过
在目前版本的dlcompiler上精度报错

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions