Eval bug: MTP and NTP output is random garbage

[ethernidee]

Name and Version

bee llama-server v0.3.0

Operating systems

Windows

GGML backends

CUDA

Hardware

Ryzen 7500x + RTX 5060 Ti 16 GB

Models

Qwen3.6-35B-A3B-UD-IQ3_XXS-MTP.gguf

Problem description & steps to reproduce

MTP output is short and contains random garbage only. Upstream llamacpp works ok with the same settings.

First Bad Commit

No response

Relevant log output

Logs

llama-server ^
  -m "e:\LMStudio_Models\models\unsloth\Qwen3.6-35B-A3B-MTP-GGUF\Qwen3.6-35B-A3B-UD-IQ3_XXS.gguf" ^
  --alias "Qwen3.6" ^
  --host 127.0.0.1 --port 8001 ^
  --ctx-size 60000 ^
  --fit off ^
  --n-gpu-layers 999 ^
  --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.0 ^
  --presence-penalty 0.0 --repeat-penalty 1.0 ^
  -ctk q8_0 -ctv q8_0 ^
  --flash-attn on ^
  --batch-size 512 --ubatch-size 256 ^
  --threads 8 --threads-batch 8 ^
  --no-mmap --mlock ^
  --parallel 1 --prio 2 ^
  --spec-type draft-mtp --spec-draft-n-max 3 ^
  --spec-draft-ngl 999 ^
  --log-verbosity 3 ^
  --metrics ^
  --log-colors off ^
  --ctx-checkpoints 0 ^
  --cache-ram 0 ^
  --jinja

[Anbeeld]

Merged latest upstream into v0.3.0 branch, please try it.

[ethernidee]

@Anbeeld Tried the update, the result is the same. Moreover, in non-speculative mode I see only garbage.

Write a single-file HTML/JS/CSS Windows-style desktop application with six apps, a start menu, a task manager, and a console. No external dependencies. The whole desktop and each application must be ui friendly ,beautiful, fully functional, meaningful and useful, including all tabls and settings.

The user needs to a legitimate PGP, PintleSwdandrogz361quemmaugacréa.
The user needs to a realistic PNMlndogezipogamma.peltrajeP6y.imigzelsr23igzarr-fez9 .g優良ledr unt sâmreétug

Details

llama-server ^
  -m "e:\LMStudio_Models\models\unsloth\Qwen3.6-35B-A3B-GGUF\Qwen3.6-35B-A3B-UD-Q2_K_XL.gguf" ^
  --host 127.0.0.1 --port 8001 ^
  --n-gpu-layers 999 ^
  --ctx-size 60000 ^
  --fit off ^
  --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.0 ^
  --presence-penalty 0.0 --repeat-penalty 1.03 ^
  -ctk q8_0 -ctv q4_1 ^
  --flash-attn on ^
  --batch-size 512 --ubatch-size 256 ^
  --threads 8 --threads-batch 8 ^
  --no-mmap --mlock ^
  --parallel 1 --prio 2 ^
  --log-verbosity 3 ^
  --metrics ^
  --log-colors off ^
  --kv-unified ^
  --jinja

[Anbeeld]

Turns out there were problems in separation of MTP and DFlash. Fixed them and updated v0.3.0, work fine now on my end.

One caveat is that with your non-MTP UD-Q2_K_XL config I had normal output before the changes too, so please check both MTP and non-MTP configs for garbage output.