Eval bug: Trapped into garbled "/" loop with a specific conversation #26215

Name and Version

$ ./llama-cli --version
version: 10040 (c7d8722)
built with GNU 11.4.0 for Linux x86_64

Operating systems

Linux

GGML backends

HIP

Hardware

AMD RYZEN AI MAX+ 395

Models

Any qwen35 arch models that are loaded with mmproj.

Problem description & steps to reproduce

Use any qwen35 arch models (e.g., qwen3.6 35B A3B) with mmproj loaded; Edit the user prompts "B" without changing it, commit, and then the model starts reasoning. Stop the reasoning, and then click the regenerated icon one or more times, and the reasoning will collapse into an endless "///" until the model is unloaded.

payload.zip

Using the trigonometric identity $\sin^2 x = \frac{1 - \cos(2x)}{2}$, we get:
$ \int_{0}^{\pi} e^x \sin^2 x , dx = \int_{0}^{\pi} e^x \frac{1 - \cos(2x)}{2} dx $
$ = \frac{1}{2} \int_{0}^{\pi} e^x dx - \frac{1}{2} \int_{0}^{\pi} e^x \cos(2x) dx $

• The first part: $\frac{1}{2} [e^x]_0^\pi = \frac{1}{2}(e^\pi - 1)$
• The second part (using the integral formula $\int e^{ax}\cos(bx)dx$):
  $ \int_{0}^{\pi} e^x \cos(2x) dx = \left[ \frac{e^x}{5}(\cos(2x) + 2\sin(2x)) \right]_0^\pi $

This is a snippet of dialogue extracted from an unexpected conversation anomaly. I modified the AI’s response to this very smallest amount necessary to reproduce the BUG.

The initial “A” and that image (could be replaced with others) in the first prompt are also essential. The system prompt should be empty, and no tools should be enabled in the webui.

First Bad Commit

b10040.
Confirmed that b10038 works normally.

Relevant log output

0.02.762.150 W srv operator(): (CORS) skip non-localhost origin:
0.02.762.317 I srv ensure_model: model name=TMP-Qwen3.6-35B-A3B-Original is not loaded, loading...
0.02.762.387 I srv load: spawning server instance with name=TMP-Qwen3.6-35B-A3B-Original on port 55251
0.02.762.410 I srv load: spawning server instance with args:
0.02.762.412 I srv load: /mnt/data/ai/llama-orig-test/llama-b10040/llama-server
0.02.762.413 I srv load: --cache-prompt
0.02.762.413 I srv load: --host
0.02.762.413 I srv load: 127.0.0.1
0.02.762.413 I srv load: --image-max-tokens
0.02.762.414 I srv load: 2048
0.02.762.414 I srv load: --log-file
0.02.762.416 I srv load: debug.txt
0.02.762.417 I srv load: --mlock
0.02.762.417 I srv load: --no-mmap
0.02.762.417 I srv load: --port
0.02.762.417 I srv load: 55251
0.02.762.418 I srv load: --slot-save-path
0.02.762.418 I srv load: ./models/_cache
0.02.762.418 I srv load: --tools
0.02.762.418 I srv load: get_datetime
0.02.762.418 I srv load: --alias
0.02.762.419 I srv load: TMP-Qwen3.6-35B-A3B-Original
0.02.762.419 I srv load: --ctx-size
0.02.762.419 I srv load: 0
0.02.762.419 I srv load: --cache-ram
0.02.762.419 I srv load: 512
0.02.762.420 I srv load: --cache-type-k
0.02.762.420 I srv load: bf16
0.02.762.420 I srv load: --cache-type-v
0.02.762.420 I srv load: bf16
0.02.762.420 I srv load: --device
0.02.762.421 I srv load: ROCm0
0.02.762.421 I srv load: --flash-attn
0.02.762.421 I srv load: on
0.02.762.421 I srv load: --fit
0.02.762.421 I srv load: off
0.02.762.422 I srv load: --kv-unified
0.02.762.422 I srv load: --lookup-cache-dynamic
0.02.762.422 I srv load: ./models/cache/prompt.cache
0.02.762.422 I srv load: --model
0.02.762.422 I srv load: ./models/Qwen3.6-35B-A3B-Orig-APEX-I-Quality.gguf
0.02.762.423 I srv load: --mmproj
0.02.762.423 I srv load: ./models/Qwen3.6-35B-A3B.mmproj-Q8_0.gguf
0.02.762.423 I srv load: --n-gpu-layers
0.02.762.423 I srv load: all
0.02.762.424 I srv load: --parallel
0.02.762.424 I srv load: 4
0.02.762.424 I srv load: --reasoning
0.02.762.424 I srv load: off
0.02.762.725 I srv ensure_model: waiting until model name=TMP-Qwen3.6-35B-A3B-Original is fully loaded...
[55251] 0.00.064.639 W server tools are enabled, using localhost as default CORS origin (change via --cors-origins)
[55251] 0.00.064.735 I cmn common_param: common_params_print_info: verbosity = 3 (adjust with the -lv N CLI arg)
[55251] 0.00.065.135 W srv llama_server: -----------------
[55251] 0.00.065.136 W srv llama_server: the following feature(s) are enabled:
[55251] 0.00.065.137 W srv llama_server: built-in tools (experimental)
[55251] 0.00.065.137 W srv llama_server: do not expose the server to untrusted environments
[55251] 0.00.065.137 W srv llama_server: -----------------
[55251] cmd_child_to_router:state:{"state":"loading","payload":{"stages":["text_model","mmproj_model"],"current":"text_model","value":0.0}}
[55251] 0.00.066.398 I srv load_model: loading model './models/Qwen3.6-35B-A3B-Orig-APEX-I-Quality.gguf'
[55251] cmd_child_to_router:state:{"state":"loading","payload":{"stages":["text_model","mmproj_model"],"current":"text_model","value":0.0}}
[55251] cmd_child_to_router:state:{"state":"loading","payload":{"stages":["text_model","mmproj_model"],"current":"text_model","value":0.047545719891786575}}
[55251] cmd_child_to_router:state:{"state":"loading","payload":{"stages":["text_model","mmproj_model"],"current":"text_model","value":1.0}}
[55251] cmd_child_to_router:state:{"state":"loading","payload":{"stage":"mmproj_model"}}
[55251] 0.03.571.260 W load_hparams: Qwen-VL models require at minimum 1024 image tokens to function correctly on grounding tasks
[55251] 0.03.571.262 W load_hparams: if you encounter problems with accuracy, try adding --image-min-tokens 1024
[55251] 0.03.571.262 W load_hparams: more info: #16842
[55251]
[55251] cmd_child_to_router:state:{"state":"loading","payload":{"stages":["text_model","mmproj_model"],"current":"mmproj_model","value":0.0}}
[55251] cmd_child_to_router:state:{"state":"loading","payload":{"stages":["text_model","mmproj_model"],"current":"mmproj_model","value":1.0}}
[55251] 0.03.699.877 I srv load_model: loaded multimodal model, './models/Qwen3.6-35B-A3B.mmproj-Q8_0.gguf'
[55251] 0.03.786.376 I srv load_model: initializing, n_slots = 4, n_ctx_slot = 262144, kv_unified = 'true'
[55251] 0.03.797.201 I srv init: chat template supports preserving reasoning, consider enabling it via --reasoning-preserve
[55251] 0.03.797.244 I srv llama_server: model loaded
[55251] 0.03.797.249 I srv llama_server: listening on http://127.0.0.1:55251
[55251] 0.03.797.326 I srv operator(): child server monitoring thread started, waiting for EOF on stdin...
[55251] cmd_child_to_router:state:{"state":"ready","payload":{"id":"TMP-Qwen3.6-35B-A3B-Original","aliases":["TMP-Qwen3.6-35B-A3B-Original"],"tags":[],"object":"model","created":1785210456,"owned_by":"llamacpp","meta":{"vocab_type":2,"n_vocab":248320,"n_ctx":262144,"n_ctx_train":262144,"n_embd":2048,"n_params":34660610688,"size":22808412672,"ftype":"Q6_K"}}}
0.06.563.839 I srv proxy_reques: proxying request to model TMP-Qwen3.6-35B-A3B-Original on port 55251
[55251] 0.03.798.213 W srv operator(): (CORS) skip non-localhost origin:
[55251] 0.03.812.027 I slot get_availabl: id 3 | task -1 | selected slot by LRU, t_last = -1
[55251] 0.03.812.098 I slot launch_slot: id 3 | task 0 | processing task, is_child = 0
[55251] 0.03.905.599 W find_slot: non-consecutive token position 11 after 10 for sequence 3 with 9 new tokens
[55251] 0.03.905.621 W find_slot: non-consecutive token position 11 after 10 for sequence 3 with 9 new tokens
[55251] 0.03.917.017 W find_slot: non-consecutive token position 271 after 11 for sequence 3 with 258 new tokens
[55251] 0.03.917.121 W find_slot: non-consecutive token position 271 after 11 for sequence 3 with 258 new tokens
[55251] 0.07.329.463 I slot print_timing: id 3 | task 0 | n_decoded = 166, tg = 55.29 t/s, tg_3s = 55.29 t/s
0.10.260.168 E srv operator(): http client error: Connection handling canceled
[55251] 0.07.512.106 W srv stop: cancel task, id_task = 0
[55251] 0.07.530.311 I slot release: id 3 | task 0 | stop processing: n_tokens = 467, truncated = 0
^C[55251] 0.09.306.194 I srv operator(): operator(): cleaning up before exit...
0.12.073.468 I srv operator(): operator(): cleaning up before exit...
0.12.073.537 I srv unload_all: stopping model instance name=TMP-Qwen3.6-35B-A3B-Original
0.12.073.553 I srv operator(): stopping model instance name=TMP-Qwen3.6-35B-A3B-Original
[55251] 0.09.307.282 I srv operator(): exit command received, exiting...
0.12.298.037 I srv operator(): instance name=TMP-Qwen3.6-35B-A3B-Original exited with status 0