Tokenizer mismatch when serving with vLLM
#6
by
one-man-won
- opened
I'm encountering a tokenizer issue when I'm serving the model with vLLM.
- ValueError: Tokenizer class TokenizersBackend does not exist or is not currently imported.
command: vllm serve "LiquidAI/LFM2.5-VL-1.6B" --trust-remote-code
Versions:
- tokenizers==0.22.2
- transformers==4.57.6
- vllm==0.15.1
Tried upgrading transforemers>5, still did not resolve this error.
Having the same issue with default vllm installtion with cu128.
Recommended setup in their vllm notebook had package version compatibility issues.
This is likely a dependency conflict — vLLM 0.15.1 pins or pulls in a specific transformers version that gets overwritten when you upgrade separately. Reinstalling in the right order should fix it:
uv pip install "vllm==0.15.1"
uv pip install "transformers==5.0.0"
Then activate your virtual env and run:
vllm serve "LiquidAI/LFM2.5-VL-1.6B" --trust-remote-code
The key is installing transformers after vLLM so it doesn't get downgraded by vLLM's dependency resolver. This worked for me with the same setup.