Qwen3-32B demo failure

[djordje-tt]

Apart from CI failing I observed garbage output for demo script:
command: pytest models/demos/llama3_70b_galaxy/demo/text_qwen_demo.py -k "performance and batch-1"
branch: divanovic/qwen3_32b
output:
==REPEAT BATCH 0 ==USER 0 - PROMPT What is your favorite condiment? There are so many condiments to choose from, each bringing its uniq <long prompt not printed in full> what your favorite condiment is and why you love it. Does it remind you of a specific dish or meal? ==USER 0 - OUTPUT biological!"!"%&!!"*!$!!""(!"!!!#%($$"!"""�"<"$"!r!"A8"$"�4#�&"@$""l"�$$'$5*0J'E(/+$8$��*$$"$/&'$$$[&&$$!#$$*$/$$$$9$$$$&$$!$$$$$##*$$&$`

[mtairum]

For reference, there was a commit on main breaking the output of all our LLMs.

This has been reverted and I'm seeing good outputs on Qwen3-32B on T3K (f250fa3).
Running your command on this galaxy CI test (your branch): https://github.com/tenstorrent/tt-metal/actions/runs/22059284683

Keep in mind that I also removed the change to the trace size back to 90000000, instead of 99000000 (in tt-metal/models/tt_transformers/demo/trace_region_config.py)

[mtairum]

Ok. I ended up getting the trace size back up again.

Passing run here: https://github.com/tenstorrent/tt-metal/actions/runs/22060811777/job/63740987129

Gonna PR and merge commit.
EDIT: PR: #37933