Uncensored and NVMe offload Request

#1
by tnzwatermaloon - opened

This looks very promising for single-GPU deployment, especially on 64GB Ampere cards like the CMP 170HX / A100 64GB.

I have two feature requests that I think would make this model much more useful for local coding workloads:

An uncensored / abliterated / low-refusal variant

My main use case is coding and agentic coding, so a weight-level uncensored version of this pruned W4A16 checkpoint would be extremely interesting. Ideally this would preserve the same expert-pruning profile and coding-oriented routing behavior, instead of starting from a separately quantized uncensored model.

NVMe mmap for the PLE / n-gram table

It would be great to have an option similar to the recent Flash-Next NVMe-PLE implementations, where the PLE table is stored as a sidecar file on local NVMe and accessed via mmap / page cache instead of requiring ~100GB of resident system RAM.

The combination I’m especially interested in is:

Qwen3.8-Flash-Next + coding-oriented expert pruning + W4A16 + uncensored weights + NVMe-mmap PLE

That could potentially make this a very strong fit for a single 64GB SM80 GPU, while preserving much more coding quality than aggressive 3-bit quantization.

If you’re interested in supporting this, even publishing the pruning/export scripts or some guidance on how to apply the same expert mask to an uncensored checkpoint would be very helpful.

Thanks for releasing this — the coding-traffic-based expert pruning approach is particularly interesting.

Your feedback is invaluable; I understand that implementation shouldn't be too difficult, so I'll proceed to create a version incorporating your suggestions. This may involve model reweighting/optimization or simply adjusting the inference engine accordingly.

Your feedback is invaluable; I understand that implementation shouldn't be too difficult, so I'll proceed to create a version incorporating your suggestions. This may involve model reweighting/optimization or simply adjusting the inference engine accordingly.

This is incredible, I can't express how grateful I am. Thank you.

Sign up or log in to comment