google

/

gemma-3-27b-it

like 489

Follow

Google 9.26k

Here's some more detail

Description of Issue:

I am encountering an AttributeError: 'Gemma3Config' object has no attribute 'vocab_size' when attempting to load the google/gemma-3-27b-it model using the transformers library from source. This error occurs because the config.json file for this model is missing the vocab_size attribute, which the transformers library expects.

Steps to Reproduce:

1. Install the latest transformers library from source: %pip install git+https://github.com/huggingface/transformers.git

2. Attempt to load the model using the following code:

   from transformers import Gemma3ForCausalLM, Gemma3Config
   from huggingface_hub import hf_hub_download
   import json
   
   model_id = "google/gemma-3-27b-it"
   huggingface_token = "YOUR_HUGGINGFACE_TOKEN" #replace with your token
   config = Gemma3Config.from_pretrained(model_id, token = huggingface_token)
   model = Gemma3ForCausalLM.from_pretrained(model_id, config=config, token = huggingface_token)
   

3. The following error occurs:

   AttributeError: 'Gemma3Config' object has no attribute 'vocab_size'
   File .../transformers/configuration_utils.py:214, in PretrainedConfig.__getattribute__(self, key)
       212 if key != "attribute_map" and key in super().__getattribute__("attribute_map"):
       213     key = super().__getattribute__("attribute_map")[key]
   --> 214 return super().__getattribute__(key)
   

config.json Output:

{
  "architectures": ["Gemma3ForConditionalGeneration"],
  "boi_token_index": 255999,
  "eoi_token_index": 256000,
  "eos_token_id": [1, 106],
  "image_token_index": 262144,
  "initializer_range": 0.02,
  "mm_tokens_per_image": 256,
  "model_type": "gemma3",
  "text_config": {
    "head_dim": 128,
    "hidden_size": 5376,
    "intermediate_size": 21504,
    "model_type": "gemma3_text",
    "num_attention_heads": 32,
    "num_hidden_layers": 62,
    "num_key_value_heads": 16,
    "query_pre_attn_scalar": 168,
    "rope_scaling": {
      "factor": 8.0,
      "rope_type": "linear"
    },
    "sliding_window": 1024
  },
  "torch_dtype": "bfloat16",
  "transformers_version": "4.50.0.dev0",
  "vision_config": {
    "hidden_size": 1152,
    "image_size": 896,
    "intermediate_size": 4304,
    "model_type": "siglip_vision_model",
    "num_attention_heads": 16,
    "num_hidden_layers": 27,
    "patch_size": 14,
    "vision_use_head": false
  }
}

@Renu11 any thoughts on this one?

im getting the same error when trying to run off terminal, I am on linux

vllm serve "google/gemma-3-27b-it"

gives me a wall of text and this error: INFO 03-12 11:46:46 cuda.py:229] Using Flash Attention backend.
INFO 03-12 11:46:47 model_runner.py:1110] Starting to load model google/gemma-3-27b-it...
WARNING 03-12 11:46:47 utils.py:78] Gemma3ForConditionalGeneration has no vLLM implementation, falling back to Transformers implementation. Some features may not be supported and performance may not be optimal.
INFO 03-12 11:46:47 transformers.py:129] Using Transformers backend.
ERROR 03-12 11:46:47 engine.py:400] 'Gemma3Config' object has no attribute 'vocab_size'

The following works for me, though some other problems come into being:

config = AutoConfig.from_pretrained(base_model_name)
for key, value in vars(config.text_config).items():
setattr(config, key, value)

base_model = AutoModelForCausalLM.from_pretrained(
base_model_name,
config=config,
quantization_config=quantization_config,
use_flash_attention_2=True,
torch_dtype=torch.bfloat16,
device_map = "cpu")

This should work:

device_map = (
        {"": f"xpu:{Accelerator().local_process_index}"}
        if is_xpu_available()
        else {"": Accelerator().local_process_index}
    )

base_model = Gemma3ForConditionalGeneration.from_pretrained(
        base_model_name,
        quantization_config=quantization_config,
        use_flash_attention_2=True,
        torch_dtype=torch.bfloat16,
        device_map = device_map
    )