[Bug]: Text output degenerates into a repetitive subword loop until max_tokens #53

TUFF version

9.4.0

How did you install TUFF?

Release download or in-app update

Area

Chat and generation

Mac, chip and memory

Mac17,3, Apple M5, 16 GB

macOS version

26.7.1

Model

Gemma 4 26B-A4B

Steps to reproduce

1. Load a model
2. Write a prompt, even "hi".
3. After the first or second prompt, expect "Setting changed. Reload the model to continue."
4. Reload the model and submit another prompt, even "hi".
5. Prefill will start (takes minutes)
6. Wait for the output

What you expected

A normal text output

What happened instead

After the refill, I get as output a look of something like:
@m_f_s_s_m_f_s_s_m_f_s_s_m_f_s_s_m_f_s_s_m_f_s_s_m_f_s_s_m_f_s_s_m_f_s_s_m_f_s_s_s_m_f_s_s_s_s or
> _b_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s_s

It happens with Gemma 4 26B-A4B and Gemma 4 26B-A4B Heretic for sure. I don't remember if it happens also with Qwen 35B-A3B, as today I was able to reproduce this bug only on Gemma.

Diagnostic summary (optional)

TUFF 9.4.0
macOS: 26.7.1
Mac: Mac17,3
Chip: Apple M5
Unified memory: 17179869184 bytes
Model: qwen3.6-35b-a3b
Context: 16384 tokens
Sampling: temperature 0.2, top-k 64, top-p 0.95
Expert cache: 8 slots (lfu)
Prefill: off, chunk 128
Image residency: on-demand

Logs or crash report (optional)

Before submitting

• I removed credentials, private paths, personal prompts and chat content from this report.