Eval bug: Qwen3.6-35B-A3B-Q8_0.gguf multiple clients concurrently produce garbled output b9927 above（b9910 is ok）

[imweijh]

Name and Version

Pure CPU inference, with -np greater than 1, works normally in versions below b9910. In versions above b9927, multiple clients performing inference concurrently produce garbled output (b9927 may not necessarily be the earliest problematic version).

Operating systems

Linux

GGML backends

CPU

Hardware

CPU

Models

No response

Problem description & steps to reproduce

build/bin/llama-server -m ~/.cache/llama.cpp/unsloth_Qwen3.6-35B-A3B-Q8_0.gguf --mmproj ~/.cache/llama.cpp/unsloth_mmproj-BF16_Qwen3.6-35B-A3B-MTP.gguf --chat-template-kwargs {"preserve_thinking": true} --spec-type draft-mtp --spec-draft-n-max 3 --image-min-tokens 1024 --temp 0.6 --top_p 0.95 --top_k 20 --min_p 0.0 --presence-penalty 0.0 --repeat-penalty 1.0 --alias qwen3.6-35b-a3b

First Bad Commit

No response

Relevant log output

Logs

[nikwen]

b9927 may not necessarily be the earliest problematic version

If you can test what the earliest problematic version is, that will make it so much easier to fix the problem.

[kernelkaribou]

Noticed this issue as well, running server-intel release. Doing my own testing found that b10015 was the last good working version for my setup, moving to b10066 began producing the garbled text.

Utilizing Battlemage B70 GPU

Specific commands in docker compose for llama.cpp run.

command:
  - --model
  - /models/gemma-4/gemma-4-12b-it-UD-Q5_K_XL.gguf
  - --mmproj
  - /models/gemma-4/gemma-4-12B-it-mmproj-BF16.gguf
  - --model-draft
  - /models/gemma-4/gemma-4-12B-it-mtp.gguf
  - --alias
  - Gemma-4-12B
  - --spec-type
  - draft-mtp
  - --spec-draft-n-max
  - "2"
  - --ctx-size
  - "32768"
  - --batch-size
  - "2048"
  - --ubatch-size
  - "1024"
  - --threads
  - "6"
  - --n-gpu-layers
  - "99"
  - --flash-attn
  - "on"
  - --fit
  - "off"
  - --mlock
  - --parallel
  - "1"
  - --host
  - 0.0.0.0
  - --port
  - "8080"
  - --jinja