Align rope init in lmdeploy

[RangiLyu]

Align all rope initialization with

lmdeploy/lmdeploy/pytorch/backends/default/rotary_embedding.py

Line 89 in 90245a3

inv_freq = 1.0 / (self.base**(torch.arange(0, self.dim, 2, dtype=torch.int64).float() / self.dim))

There is a slight difference in numerical precision between CPU and GPU, which leads to inconsistencies between training and inference in RL.

[Copilot]

Pull request overview

This PR standardizes RoPE inverse-frequency (inv_freq) initialization across several PyTorch model/backends to match the default rotary embedding implementation, aiming to reduce CPU vs GPU numerical differences that can cause RL training/inference inconsistencies.

Changes:

• Update Qwen3.5 default RoPE parameter computation to compute inv_freq in float on CPU first, then move to the target device.
• Update Qwen2-VL / Qwen2.5-VL vision rotary embeddings to follow the same arange(int64) -> float -> move device pattern.
• Update dlinfer rotary embedding initialization to follow the same pattern before moving to CUDA.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
lmdeploy/pytorch/models/qwen3_5.py	Adjusts default RoPE inv_freq initialization to align numerically with the default backend implementation.
lmdeploy/pytorch/models/qwen2_vl.py	Aligns vision rotary embedding inv_freq init pattern (CPU float compute, then .to(device)).
lmdeploy/pytorch/models/qwen2_5_vl.py	Same alignment for Qwen2.5-VL vision rotary embedding inv_freq.
lmdeploy/pytorch/backends/dlinfer/rotary_embedding.py	Aligns dlinfer base rotary embedding inv_freq init pattern (then moves to CUDA).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.