Skip to content

Known issues

  • Tracing reduces decode speed. Keep tracing off for speed.
  • More than two sessions at 262K context do not fit in 32 GB host RAM. Use no more than two sessions at this context size. Bubo refuses a request that needs new memory when free memory is below the floor. See Memory guard.
  • The disk prompt cache is an option. It is off by default. When on, it writes about 3 GB to the SSD for each long saved session. See Fast restarts with the disk cache.
  • The exact profile stops with an error when an explicit flag or environment value would change an exact-profile value (for example Error: profile=exact: explicit value re-enables GUFO_DENSE_TWIN_MAX_N). Remove that setting, or do not combine it with --profile exact.
  • Chat logprobs need thinking off, plain text (no tools or images) and no stop sequences. Otherwise the server returns 400. The /debug/prompt endpoint accepts plain text without tools only. See options.
  • With the prompt cache on (default profile), a restore from the cache is not bit-identical to a fresh prefill of the same prompt. The exact profile turns the cache off for this reason.
  • If you use --cache-disk, the first start of v0.1.2 deletes the disk-cache files of v0.1.1 in that folder (*.kvc and temporary files, top level only), and the first requests after that run cold. Use a separate folder for each server.
  • The option --cache-disk-staging-bytes is ignored on Qwen3.8 Flash Next, and the server logs a warning.
  • If the GPU stops responding, the server logs event=device_lost remedy=restart_process exit_status=75 and exits with status 75. A lost GPU context cannot recover inside the process. The Bubo app starts the server again once, after 5 seconds, with the same setup, and shows “The GPU stopped responding. Bubo restarted the server.” If the GPU stops again within 5 minutes, the app shows “The graphics device stopped responding.” and you select Start to start the server again. Without the app, nothing restarts the server by itself: start.cmd runs it once, so start the server again, or run it under a supervisor that restarts it on failure. bubo diagnose shows if the GPU memory reservation and the page file match the tested setup. See system setup. This error has not been seen on the tested setup.
  • Like many open-source Windows builds, Bubo v0.1.2 is not code-signed. SmartScreen can show a warning, and Smart App Control blocks unsigned programs. See Windows security prompts.