Skip to content

Can fp8_paged_mqa_logits support non-contiguous layout? #221

@MARD1NO

Description

@MARD1NO

I notice that there exist an assertion

DG_HOST_ASSERT(fused_kv_cache.stride(1) == head_dim_with_sf);

When we sliced paged kvcache by layer_idx, the sliced paged kvcache may not be contiguous.

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