Use helper function instead of looping through attribute names

[hmellor]

Since we have the helper function, it'd be better to use it so that the behaviour is consistent.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.

[gemini-code-assist]

Code Review

This pull request aims to refactor attribute access to use a helper function for consistency. While the change in get_inputs_embeds_size is correct, the change in get_total_num_kv_heads introduces a critical bug. The refactoring alters the logic for finding the number of KV heads. The original code correctly skipped attributes that were present but None, while the new implementation with getattr_iter will return None if it's the value of the first attribute found. This can lead to a TypeError downstream. I've left a specific comment with a suggested fix.

[gemini-code-assist]

This refactoring introduces a critical bug. The original implementation iterated through attributes and returned the value of the first attribute that was not None. The new implementation using getattr_iter will return the value of the first attribute that exists, even if that value is None.

For example, if self.hf_text_config has an attribute n_head_kv=None and num_kv_heads=8, the old code would correctly return 8. The new code would return None, which would cause a TypeError when this function's return value is used in comparisons.

Please revert to the original logic, which was correct.

        for attr in attributes:
            num_kv_heads = getattr(self.hf_text_config, attr, None)
            if num_kv_heads is not None:
                return num_kv_heads

        # For non-grouped-query attention models, the number of KV heads is
        # equal to the number of attention heads.
        return self.hf_text_config.num_attention_heads

[ProExpertProg]

Does self.hf_text_config.num_attention_heads always exist? If not, existing code might not fail but current code certainly will. Could we pass a lambda default instead?

Suggested change

	default = self.hf_text_config.num_attention_heads
	default = lambda: self.hf_text_config.num_attention_heads