Skip to content

feat(GFX1100-TG200): T21 keep-quant for V-head row-permuted GDN projections #2803

Description

@ghazni101

Row: `GFX1100-TG200`

T21 adds keep-quant support for V-head row-permuted GDN projections in the Qwen3.5 GGUF weight loader. The row-permuted projection reorders the output weight matrix so that the GDN in_proj can consume the K-quant V-head tiles without a dequantize-requantize round trip.

Depends on #2782 (F1), #2790 (T4a), #2792 (lever-C), #2796 (T27), and #2800 (T24).

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions