Skip to content

Wrong side chain geometry in multimer predictions #365

Description

@nmishra93

Hi ESM Team,

I am trying to fold several antibodies and have noticed weird side chain geometries (with extremely squished rings for aromatic residues); this only happens in multimer input (chain A + chain B). When I fold either chain A or chain B alone, all side chains look correct.

Related to but distinct from #315 (diffusion transformer misbehaving when flash_attn is missing, fixed by requiring flash_attn) — this repro has flash_attn installed and available; the bug is in how chain boundaries are (not) represented in the atom-attention cu_seqlens, not in flash-attn availability.

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