Skip to content

Reproduction tooling for Qwen3.8-27B GSQ-RCO mixed-precision GGUF release #9

Description

@SoppyLeon

Hello IST-DASLab team,

Thank you for publishing GSQ, RCO, and the
Qwen3.8-27B-GSQ-RCO-GGUF
release.

We are preparing a reproducible experiment applying the same method to the
BF16 weights of JonathanColetti/Qwen3.8-27B-Uncensored. Our intent is to
generate fresh candidates and run a new RCO search on the modified weights
rather than reuse the stock-Qwen allocation.

We audited the following public revisions:

  • GSQ: 03fc16484c369e3127225615d5e03e8d3a6043e3
  • RCO: 9a1e09c07d468109cbe60a1b87d5036034a79d10
  • llama.cpp: 0cae43063cf15170e91a2ff4d034da0ecef4a1b2

We also inspected the published allocation reports and release artifacts. They
allow us to recover the final selected tensor/qtype maps, but not the
candidate-generation path or enough information to reproduce the optimization
on different weights.

The current public repositories appear to be missing several release-specific
components required for exact reproduction:

  1. BF16 → GSQ candidate generation for the native GGUF/IQuant formats used in
    the release;
  2. the exact per-tensor/per-candidate serialized byte-cost model supplied to
    RCO;
  3. the RCO command/configuration used for the 2.75 and 3.00 bpw targets;
  4. mixed-per-tensor GGUF assembly/export from the selected RCO allocation; and
  5. any llama.cpp fork, patch, or auxiliary code used for those steps.

Could you please point us to the repository, branch, scripts, or commit
containing those components, or publish them if possible?

Additional reproduction details would also be useful if available, especially
calibration/importance-matrix generation, GSQ optimizer settings, handling of
embeddings/lm_head/GatedDeltaNet/MTP tensors, and the exact stock BF16
revision used for the published release.

If some of the release pipeline is not currently public, even confirmation of
that—and a description of the expected candidate and allocation schemas—would
help us avoid claiming an invalid reproduction.

Thanks for your time and for releasing the models and allocation reports.

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

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions