Make default RoPE path explicit in RotaryEmbedding classes

[DevPatel-11]

🧩 Pull Request: Make default RoPE path explicit in Llama RotaryEmbedding initialization

📜 Summary

This PR addresses issue #39753, ensuring that the default Rotary Positional Embedding (RoPE) initialization path is made explicit in the Llama model.

Previously, inv_freq was implicitly assigned, which went against the library’s philosophy of explicit initialization for reproducibility and clarity.

This change:

• Makes the "default" RoPE path explicit by selecting ROPE_INIT_FUNCTIONS["default"] when self.rope_type == "default".
• Uses ROPE_INIT_FUNCTIONS[self.rope_type] for non-default cases.
• Keeps the initialization logic (inv_freq, attention_scaling) delegated to the RoPE function—no manual computation is done in the constructor.

The modification is minimal, isolated, and maintains backward compatibility.

---

🧠 Motivation

Following the discussion in #39753, the goal is to make RoPE initialization behavior transparent and avoid implicit defaults.
This aligns with 🤗 Transformers’ philosophy of clarity and explicit model configuration.

---

🛠️ Changes Made

In transformers/src/transformers/models/llama/modeling_llama.py and more 20 models:

self.config = config
if self.rope_type == "default":
    self.rope_init_fn = ROPE_INIT_FUNCTIONS["default"]
else:
    self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)

• Delegates all logic to the existing RoPE functions.
• Reduces potential maintenance risk by avoiding any new computation paths.
• Keeps inv_freq and attention_scaling initialization consistent with the rest of the library.

---

✅ Checklist

• The PR title summarizes the contribution.
• Linked to issue #39753.
• Tested locally — no change in model outputs or test failures.
• Backward compatible (no breaking changes).

---

🧪 Testing

• Verified initialization behavior for both "default" and non-default RoPE types.
• Confirmed no change in model outputs.
• Ran relevant test subset:

  pytest

---

💬 Additional Notes

This is a small but meaningful cleanup that improves code readability and aligns initialization logic with the library’s explicit philosophy.
No other files or RoPE logic were touched.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: apertus, arcee, aria, bamba, bitnet, blt, cohere, cohere2, csm, dbrx, deepseek_v3, dia, diffllama, doge, dots1, emu3

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: apertus, arcee, aria, bamba, bitnet, blt, cohere, cohere2, csm, dbrx, deepseek_v3, dia, diffllama, doge, dots1, emu3

[zucchini-nlp]

Thanks for the PR @DevPatel-11. I am currently reworking RoPE for all models in #39847 which also includes moving the default freq computation to model files. I'd prefer to not merge a separate PR for rope freq in the meanwhile as it would results in huge merge conflicts

[DevPatel-11]

Ok @zucchini-nlp.
Would love to help you for other issues too!