[Model][QwenVL] Simplify cos/sin rotary embedding indexing

[lgeiger]

Purpose

This is a small followup from #28798 which simplifies the indexing logic. /cc @gcanlin @Isotr0py

For Qwen3VL #28798 slightly changed behaviour where get_cos_sin() now already returns a GPU tensor. This introduces synchronous CPU to GPU copies of pos_ids when using it to index. cc1f0c5 Fixes it by moving the indices onto the GPU in a non-blocking way.

Before:

After:

Test Plan

VLLM_WORKER_MULTIPROC_METHOD=spawn lm_eval --model vllm-vlm --model_args "pretrained=Qwen/Qwen3-VL-30B-A3B-Instruct-FP8,max_model_len=10000" --tasks chartqa --batch_size auto --apply_chat_template

Test Result

Before:

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
chartqa	0	none	0	anywhere_accuracy	↑	0.8752	±	0.0066
		none	0	exact_match	↑	0.6448	±	0.0096
		none	0	relaxed_accuracy	↑	0.8656	±	0.0068

After:

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
chartqa	0	none	0	anywhere_accuracy	↑	0.8760	±	0.0066
		none	0	exact_match	↑	0.6388	±	0.0096
		none	0	relaxed_accuracy	↑	0.8656	±	0.0068

[gemini-code-assist]

Code Review

This pull request simplifies the rotary embedding indexing logic across several models, which improves code readability and maintainability. It also addresses a performance issue in qwen3_vl.py by moving pos_ids to the GPU asynchronously, preventing synchronous CPU-to-GPU copies.

However, the same performance optimization is missing in other models modified in this PR (glm4_1v.py, qwen2_5_vl.py, qwen2_vl.py, and qwen3_omni_moe_thinker.py), where pos_ids is still created on the CPU, leading to synchronous copies during indexing. I've added specific comments to apply the same fix to these models for consistency and performance improvement.