Eval bug: Repeatable Crashes when an image and a large prompt is pasted or loaded as file.

[KevinFleischer]

Name and Version

version: 9775 (be4a6a6)
built with Clang 20.1.8 for Windows x86_64

Operating systems

Windows

GGML backends

Vulkan

Hardware

GPU: AMD 7900xtx
CPU: AMD 7800x3d (not used)

Models

.\llama-server.exe -hf unsloth/gemma-4-26B-A4B-it-qat-GGUF:UD-Q4_K_XL --jinja --chat-template-file C:\llamaCpp\templates\gemma4_26B_chat_template.jinja --reasoning-format auto -ngl 999 --ctx-size 262144 -np 2 --cache-type-k q8_0 --cache-type-v q8_0 --cache-ram 4096 --ctx-checkpoints 8 --no-context-shift --ubatch-size 4096 --temp 1.0 --top-p 0.95 --top-k 64 --repeat-penalty 1.0 --port 8080 --host 127.0.0.1

Problem description & steps to reproduce

I've attached two files, a Screenshot and the extraction prompt.

IMPORTANT: If I load the prompt bit by bit into the text field, the execution is successfull. If I copy the whole prompt in one go and paste it, it is shown as a "text file" in the prompt area. This will fail.
If you load both files as attachments it will also fail.

First Bad Commit

don know.

Relevant log output

Logs

�[34m0.30.615.200�[0m �[32mI �[0msrv  update_slots: all slots are idle
�[34m1.11.240.526�[0m �[32mI �[0msrv    operator(): Chat format: peg-gemma4
�[34m1.11.240.643�[0m �[32mI �[0mslot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
�[34m1.11.240.643�[0m �[32mI �[0msrv  get_availabl: updating prompt cache
�[34m1.11.240.648�[0m �[32mI �[0msrv          load:  - looking for better prompt, base f_keep = -1.000, sim = 0.000
�[34m1.11.240.650�[0m �[32mI �[0msrv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 4096.000 MiB, 262144 tokens, 4294967296 est)
�[34m1.11.240.651�[0m �[32mI �[0msrv  get_availabl: prompt cache update took 0.01 ms
�[34m1.11.240.673�[0m �[32mI �[0mreasoning-budget: activated, budget=2147483647 tokens
�[34m1.11.240.674�[0m �[32mI �[0mreasoning-budget: deactivated (natural end)
�[34m1.11.240.684�[0m �[32mI �[0mslot launch_slot_: id  0 | task 1895 | processing task, is_child = 0
�[34m1.11.240.685�[0m �[32mI �[0mslot process_sing: id  1 | task -1 | saving idle slot to prompt cache
�[34m1.11.241.050�[0m �[35mW srv   prompt_save:  - saving prompt with length 3184, total state size = 139.336 MiB (draft: 0.000 MiB)
�[0m�[34m1.11.273.204�[0m �[32mI �[0msrv        update:  - cache state: 1 prompts, 139.440 MiB (limits: 4096.000 MiB, 262144 tokens, 262144 est)
�[34m1.11.273.207�[0m �[32mI �[0msrv        update:    - prompt 000001FD587DEE80:    3184 tokens, checkpoints:  1,   139.440 MiB
�[34m1.11.291.580�[0m �[32mI �[0mslot create_check: id  0 | task 1895 | created context checkpoint 1 of 8 (pos_min = 0, pos_max = 0, n_tokens = 1, size = 0.104 MiB)
�[34m1.11.300.555�[0m �[32mI �[0mslot process_mtmd: id  0 | task 1895 | encoding mtmd batch from idx = 860, n_chunks = 1

PS C:\llamaCpp>

[KevinFleischer]

extraction_prompt.md

[chejh-amd]

Hi @KevinFleischer, The log stops right at the first image encode during prompt processing, so the vision/mtmd path is a more likely trigger than prompt length on its own.

You can try:

• the big prompt with no image attached: still crashes?
• the image with a one-line prompt: crashes on its own?

[KevinFleischer]

As I already wrote: If I send the exact same prompt in over the text field it works (by pasting the prompt in 2 pieces behind each other).

Whenever I use a prompt that is delivered via the text field, it works. If I paste a longer prompt and the textfields "converts it automatically into a file" it will not work.