Eval bug: context shift is disabled

[deific]

Name and Version

./llama.cpp/build/bin/llama-server  \
  -m /models/DeepSeek-R1-UD-IQ1_S-00001-of-00003.gguf \
  --cache-type-k q4_0 \
  --threads 64 \
  --temp 0.6 \
  --ctx-size 12288 \
  --parallel 3 \
  --n-gpu-layers 62 

Operating systems

Linux

GGML backends

BLAS

Hardware

AMD EPYC 9754 128-Core Processor
8 * RTX 4090D 24G

Models

DeepSeek-R1-UD-IQ1_S

Problem description & steps to reproduce

Error message after multiple rounds of conversation: context shift is disabled。

First Bad Commit

No response

Relevant log output

slot launch_slot_: id  1 | task 4640 | processing task
slot update_slots: id  1 | task 4640 | new prompt, n_ctx_slot = 4096, n_keep = 0, n_prompt_tokens = 3844
slot update_slots: id  1 | task 4640 | kv cache rm [3798, end)
slot update_slots: id  1 | task 4640 | prompt processing progress, n_past = 3844, n_tokens = 46, progress = 0.011967
slot update_slots: id  1 | task 4640 | prompt done, n_past = 3844, n_tokens = 46
slot      release: id  1 | task 4640 | stop processing: n_past = 4095, truncated = 0
srv    send_error: task id = 4640, error: context shift is disabled
srv  update_slots: no tokens to decode
srv  update_slots: all slots are idle
srv  cancel_tasks: cancel task, id_task = 4640
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200

[zshihang213]

I got the same error 'context shift is disabled' after a few rounds with --ctx-size 2048. Maybe the history is too long for the ctx-size?

[grimulkan]

#7343 is what is going on I think