Skip to content

description not much the actually shape #72

@wcyjerry

Description

@wcyjerry

in adpaters.py, the function run_transformer_block's description:
- ffn.w1.weight
Weight of the first linear transformation in the FFN.
Shape is (d_model, d_ff).
- ffn.w2.weight
Weight of the second linear transformation in the FFN.
Shape is (d_ff, d_model).
- ffn.w3.weight
Weight of the third linear transformation in the FFN.
Shape is (d_model, d_ff).

When I trying to implement this function as the shape in description, it doesn't match, I print it and find out the actually shape test_code input is
(d_ff, d_model). I think whatever your math is wtx or xw, but your input should match the description

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