ValueError: Cannot use apply_chat_template because this processor does not have a chat template.

[AlphaNext]

emmm, I get this error when using gemma-4-12B
transformers=5.10.1
ValueError: Cannot use apply_chat_template because this processor does not have a chat template.

look forward to your reply. Thanks very much.

The code is here

from transformers import AutoProcessor, AutoModelForMultimodalLM

MODEL_ID = "google/gemma-4-12B"

# Load model
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID,
    dtype="auto",
    device_map="auto"
)
# Prompt
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Write a short joke about saving RAM."},
]

# Process input
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=False
).to(model.device)
input_len = inputs["input_ids"].shape[-1]

# Generate output
outputs = model.generate(**inputs, max_new_tokens=1024)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)

# Parse output
processor.parse_response(response)

[arnav-144p]

@AlphaNext Hi, the issue is that apply_chat_template isn't delegated through AutoProcessor — it lives on the tokenizer.

Quick fix:

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True
).to(model.device)

To check if this is a transformers-side bug or a missing template in the model config, run:

print(processor.tokenizer.chat_template)

If that returns a template, the fix is making processor.apply_chat_template delegate to processor.tokenizer.apply_chat_template. Happy to open a PR for that if confirmed.

[AlphaNext]

@arnav-144p thanks your reply, print(processor.tokenizer.chat_template) get this error

    inputs = tokenizer.apply_chat_template(
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/miniconda3/lib/python3.12/site-packages/transformers/tokenization_utils_base.py", line 3087, in apply_chat_template
    chat_template = self.get_chat_template(chat_template, tools)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/miniconda3/lib/python3.12/site-packages/transformers/tokenization_utils_base.py", line 3269, in get_chat_template
    raise ValueError(
ValueError: Cannot use chat template functions because tokenizer.chat_template is not set and no template argument was passed! For information about writing templates and s
etting the tokenizer.chat_template attribute, please see the documentation at https://huggingface.co/docs/transformers/main/en/chat_templating

---

the code is

from transformers import AutoProcessor, AutoModelForMultimodalLM

#MODEL_ID = "google/gemma-4-12B-it"
#  local model path
MODEL_ID = "models/gemma-4-12B"  

# Load model
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID,
    dtype="auto",
    device_map="auto"
)

# Prompt
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Write a short joke about saving RAM."},
]

# added code
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True
).to(model.device)

print(processor.tokenizer.chat_template)

# Process input
inputs = processor.tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=False
).to(model.device)
input_len = inputs["input_ids"].shape[-1]

# Generate output
outputs = model.generate(**inputs, max_new_tokens=1024)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)

# Parse output
processor.parse_response(response)

[vasqu]

The base model does not have an instruction template, only the -it variation has it! Also see https://huggingface.co/google/gemma-4-12B/discussions/5

[AlphaNext]

This problem is solved by using google/gemma-4-12B-it, change google/gemma-4-12B to google/gemma-4-12B-it