I’ｍ staying at 0.3.23, because....

[SwHaraday]

HI there,

Thank you for your great work !
I've been playing with your llama-cpp-python to use 'Qwen3VL-8B-Instruct-Q8_0.gguf'.
It works fine, however newer ones than 0.3.23 give me following warnings.

Exception ignored on calling ctypes callback function: <function CustomSampler.__init__.<locals>._cb_apply at 0x000002003EA4BA60>
Traceback (most recent call last):
  File "C:\ProgramData\Anaconda3\envs\vlmgguf-py311\Lib\site-packages\llama_cpp\_internals.py", line 1459, in _cb_apply
    self.apply_func(cur_p.contents)
  File "C:\ProgramData\Anaconda3\envs\vlmgguf-py311\Lib\site-packages\llama_cpp\llama.py", line 1317, in adapter
    new_scores = logits_processor(self._input_ids, current_scores)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\ProgramData\Anaconda3\envs\vlmgguf-py311\Lib\site-packages\transformers\generation\logits_process.py", line 93, in __call__
    scores = processor(input_ids, scores)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "c:\llm\vlm\vlmrag_ui_multi_gguf.py", line 202, in __call__
    seen = set(input_ids[b].tolist())
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: 'int' object is not iterable

I don't care as far as my system works. Just for information.
Project is here, https://github.com/SwHaraday/local-vlm-rag

Thanks again for your continuing efforts !

[JamePeng]

The sampler has been completely refactored and now natively supports the present_penalty parameter. You can check the parameter types and comments in the LlamaSamplingParams class for more information.