perf: avoid output_hidden_states when only last_hidden_state is used

[ciaoyizhen]

What does this PR do?

Avoid unnecessary output_hidden_states=True when only the final layer hidden state is used
to compute reward logits.

Currently, the model enables output_hidden_states=True but only consumes
hidden_states[-1]. This forces the backbone to store all intermediate layer
activations, increasing memory usage and compute overhead.

This PR switches to output.last_hidden_state, which is functionally equivalent
to hidden_states[-1], while keeping the behavior unchanged.

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a GitHub issue? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

Anyone in the community is welcome to review this PR.

[ciaoyizhen]

from transformers import AutoModel, AutoTokenizer, AutoModelForSequenceClassification

reward_model_ckpt = "/data1/caoyizhen/caoyizhen/open-source-models/internlm2-1_8b-reward"

model = AutoModel.from_pretrained(reward_model_ckpt, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(reward_model_ckpt, trust_remote_code=True)
print(model.base_model_prefix)

a = "你好"
input_dict = tokenizer(a, return_tensors="pt")
output = model.model(
    **input_dict,
    return_dict=True,
    output_hidden_states=True,
    use_cache=False,  # otherwise mistral-based RM would error out
)
print(output.hidden_states[-1])


output = model.model(
    **input_dict,
    return_dict=True,
    output_hidden_states=False,
    use_cache=True,  # otherwise mistral-based RM would error out
)
print(output.last_hidden_state)

It is same.