[Bug]: When using hosted_vLLM, the generic token counter is used even if custom_tokenizer is configured

[m4oc]

What happened?

When using hosted_vLLM, the system uses the generic token counter even when a custom_tokenizer is configured in the settings.

model_info={'id': 'ID', 'db_model': False, 'custom_tokenizer': {'identifier': 'deepseek-ai/deepseek-coder-6.7b-instruct', 'revision': 'main'}, 'input_cost_per_token': 5e-08, 'output_cost_per_token': 2.5e-07}, timeout=6000, max_retries=0)

LiteLLM:DEBUG: utils.py:301 - Token Counter - using generic token counter, for model=deepseek-ai/deepseek-coder-6.7b-instruct

NotImplementedError: num_tokens_from_messages() is not implemented for model gpt-3.5-turbo. See https://github.com/openai/openai-python/blob/main/chatml.md for information on how messages are converted to tokens.

Relevant log output

Are you a ML Ops Team?

No

What LiteLLM version are you on ?

1.61.0

Twitter / LinkedIn details

No response

[m4oc]

Digging deeper, I noticed that /utils/token_counter is not being called, so custom_tokenizer always remains None.
From the documentation, it's not clear how to enable it