Exllamav2 tokenizer kwargs are not used

[cpfiffer]

Several of the kwargs in the docstring for the exllamav2 inference engine do not seem to be in use,

outlines/outlines/models/exllamav2.py

Lines 244 to 290 in 36f1bf2

	def exl2(
	model_path: str,
	draft_model_path: Optional[str] = None,
	max_seq_len: Optional[int] = None,
	cache_q4: bool = False,
	paged: bool = True,
	max_chunk_size: Optional[int] = None,
	) -> ExLlamaV2Model:
	"""
	Load an ExLlamaV2 model.
	Parameters
	----------
	model_path (str)
	Path to the model directory.
	device (str)
	Device to load the model on. Pass in 'cuda' for GPU or 'cpu' for CPU
	max_seq_len (Optional[int], optional)
	Maximum sequence length. Defaults to None.
	scale_pos_emb (Optional[float], optional)
	Scale factor for positional embeddings. Defaults to None.
	scale_alpha_value (Optional[float], optional)
	Scale alpha value. Defaults to None.
	no_flash_attn (Optional[bool], optional)
	Disable flash attention. Defaults to None.
	num_experts_per_token (Optional[int], optional)
	Number of experts per token. Defaults to None.
	cache_q4 (bool, optional)
	Use Q4 cache. Defaults to False.
	tokenizer_kwargs (dict, optional)
	Additional keyword arguments for the tokenizer. Defaults to {}.
	gpu_split (str)
	\"auto\", or VRAM allocation per GPU in GB. Auto will use exllama's autosplit feature
	low_mem (bool, optional)
	Enable VRAM optimizations, potentially trading off speed
	verbose (bool, optional)
	Enable if you want debugging statements
	Returns
	-------
	An `ExLlamaV2Model` instance.
	Raises
	------
	`ImportError` if the `exllamav2` library is not installed.
	"""

The following kwargs do not seem to be used, but are mentioned in the doc string:

• tokenizer_kwargs
• scale_pos_emb
• scale_alpha_value
• no_flash_attention
• num_experts_per_token
• gpu_split
• low_mem
• verbose

Used but not documented:

• draft_model_path
• paged
• max_chunk_size