Skip to content

Gotcha: transformers < 4.58 doesn't recognize the gemma4 model type

Symptom

pydantic_core._pydantic_core.ValidationError: 1 validation error for ModelConfig
Value error, The checkpoint you are trying to load has model type
`gemma4` but Transformers does not recognize this architecture.

Root cause

Gemma 4 introduced a new model type gemma4 with architecture class Gemma4ForConditionalGeneration. This architecture is NOT in transformers==4.57.6 (the version bundled in both the Triton 26.03 and vllm-openai:latest containers at the time of our testing).

The auto_map field in the model's config.json is empty — meaning the model does NOT ship custom code on HuggingFace. It expects transformers itself to have the architecture class built in.

Fix

pip install --no-cache-dir 'transformers>=4.58'

For production: bake it into the Dockerfile: 

FROM vllm/vllm-openai:latest
RUN pip install --no-cache-dir 'transformers>=4.58'

Key detail

This is separate from the vLLM version issue. vLLM 0.19.0 has its own internal model implementation for Gemma 4, but it delegates config parsing to transformers' AutoConfig / PretrainedConfig path. If transformers can't recognize the model_type, vLLM fails during config validation before it ever gets to its own model loading code.

How to prevent

When a new model releases, check both: 1. Does vLLM support it? (check vllm/model_executor/models/) 2. Does the pinned transformers version in the container have the architecture class? (check transformers.models)

If #2 fails, you need a newer transformers. This is the "four independent release trains" problem from the infrastructure gap analysis.

Environment

  • Gemma 4 released 2026-04-02
  • transformers 4.57.6 — does NOT have gemma4
  • transformers 4.58+ (or 5.x dev) — HAS gemma4
  • Tested 2026-04-09 (one week after model release)