Hi,
I am trying to serve gemma-4 with nightly VLLM docker and upated transformers.
The erro I get is
AttributeError: ‘NoneType’ object has no attribute ‘dtype’
[rank0]:[W402 18:10:34.327798055 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources.
(EngineCore pid=82) ERROR 04-02 18:10:35 [core.py:1108] EngineCore failed to..
Als it says before:
TransformersMultiModalForCausalLM has no vLLM implementation, falling back to Transformers implementation. Some features may not be supported and performance may not be optimal.
The error AttributeError: 'NoneType' object has no attribute 'dtype' when serving Gemma-4 with nightly vLLM and updated transformers is not explicitly documented in the provided data sources. However, the warning about “TransformersMultiModalForCausalLM has no vLLM implementation, falling back to Transformers implementation” indicates vLLM is using the Transformers backend, which may lack full feature support and optimal performance for some models, especially new or multimodal ones. This fallback can lead to compatibility issues, including attribute errors like the one you encountered.
For newly released or complex models like Gemma-4, such errors often stem from version mismatches or incomplete support in either vLLM or Transformers. The recommended steps are: (1) ensure both vLLM and Transformers are installed from their latest main branches, not just the latest release, and (2) monitor the vLLM supported models documentation and GitHub issues for updates on official support and bug fixes. If the issue persists after updating, it likely means Gemma-4 is not yet fully supported in vLLM, especially for multimodal or non-standard architectures.
Would you like more detail on troubleshooting or version management for vLLM and Transformers?