Skip to content

escha-w2-dense: loader queries blk.N.ssm_out.weight, Q8E file ships escha-native triplets #3

Description

@mettbrot0815

Branch: escha-w2-dense @ 2940b80. File: Escha-Qwen3.8-27B-W2-Q8E.gguf (2054 tensors, card-verified).

Load fails immediately:

llama_model_load: error loading model: check_tensor_dims: tensor 'blk.0.ssm_out.weight' not found

The file has no .weight tensor for ssm_out (or any matmul). Block 0 ships 32 tensors; the projections use Escha-native triplets:

  • blk.0.ssm_out.escha_code / .escha_rin / .escha_rout / .bias
  • same pattern for attn_qkv, attn_gate, ffn_gate/up/down

So the dense-branch SSM mapping still queries the classic weight name while the file (like the F16 build, 2052/2054 byte-identical per the model card) stores escha triplets. Attention/ffn escha paths load; the SSM block does not get that far.

Repro: stock run flags from the model card (-ngl 99 -fa on --jinja, q8_0 KV) on RTX 3060 12 GB, CUDA build. Ask: map the ssm_* escha triplets in the dense path, or confirm which file revision this branch loads.

Side note for 12 GB users: the card's own table puts 64k ctx at 12.31 GiB total — over a 12 GB card. 32k (11.25 GiB) is the realistic fit line there.

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