Rotary embeddings utilities

This page explains how the Rotary Embedding is computed and applied in Transformers and what types of RoPE are supported.

Overview

Rotary Position Embeddings are a technique used to inject positional information into attention mechanisms without relying on explicit position encodings.
Instead of adding position vectors to token embeddings, RoPE rotates query and key vectors in the complex plane according to their positions enabling relative positional awareness and better extrapolation to unseen sequence lengths.

The Transformers library provides a flexible and extensible implementation of various RoPE types defined in ~modeling_rope_utils.ROPE_VALIDATION_FUNCTIONS, including both the default and scaled variants:

Configuration in Model Configs

To enable and customize rotary embeddings, add a rope_parameters field to your model’s configuration file (config.json). This field controls the RoPE behavior across model layers. Note that each RoPE variant defines its own set of expected keys and missing keys will raise an error. See the example below which creates a llama config with default RoPE parameters:

from transformers import LlamaConfig

config = LlamaConfig()
config.rope_parameters = {
    "rope_type": "default", # type of RoPE to use
    # rope_theta is optional — omitting it uses the model’s default_theta (typically 10000.0)
}

# If we want to apply a scaled RoPE type, we need to pass extra parameters
config.rope_parameters = {
    "rope_type": "linear",
    "rope_theta": 10000.0,  # can be omitted to fall back to default_theta
    "factor": 8.0  # scale factor for context extension
}

Per-Layer-Type RoPE Configuration

Some models such as Gemma-3 use different layer types with different attention mechanisms, i.e. “full attention” in some blocks and “sliding-window attention” in others. Transformers supports specifying distinct RoPE parameters per layer type for these models. In this case, rope_parameters should be a nested dictionary, where top-level keys correspond to config.layer_types and values are per-type RoPE parameters. During model initialization, each decoder layer will automatically look up the matching RoPE configuration based on its declared layer type.

from transformers import Gemma3Config

config = Gemma3Config()
config.rope_parameters = {
    "full_attention": {
        "rope_type": "dynamic",
        "rope_theta": 1000000.0,
        "factor": 8.0,
        "original_max_position_embeddings": 8096,
    },
    "sliding_attention": {
        "rope_type": "default",
        "rope_theta": 10000.0,
    }
}

MRoPE

MRoPE is a type of rotation applied in multimodal models and defined by mrope_section. It is not a separate entry in the rope_type table. You can still apply rope scaling (linear, dynamic) with it.

mrope_section sizes contiguous frequency bands for the temporal, height, and width axes (those sizes sum to head_dim // 2). Frequencies are then repeated so the embedding spans the full head_dim. For multimodal inputs (usually vision), RoPE is applied in one shot (matmul or elementwise multiply of frequencies with positions). mrope_section only reorders those frequencies along (t, h, w) first, as in Qwen2-VL’s recomposition_frequencies. Prompt text and generated tokens keep normal 1D RoPE by using identical position ids on all three THW grids.

from transformers import Qwen2VLConfig

config = Qwen2VLConfig()
config.text_config.rope_parameters = {
    "rope_type": "default",
    "rope_theta": 1000000.0,
    "mrope_section": [16, 24, 24],  # temporal, height, width frequency band sizes
}

Qwen2-VL uses mrope_section = [16, 24, 24] (16 temporal, 24 height, 24 width when head_dim is 128). Multimodal models often use this layout in the text backbone (Qwen2-VL, GLM-4V). Some VLMs do not use MRoPE and stay on normal 2D text RoPE instead. Check mrope_section in the text config.

Axial RoPE

Separately, "axial" is a registered rope_type for vision models, but it is not listed in ROPE_INIT_FUNCTIONS. Frequency setup stays on the model. It usually applies the same frequencies (head_dim // 4 per spatial axis) for height and width positions and does not allow scaling on top. Examples include the Qwen2-VL vision model and other vision stacks such as Pixtral.

Utilities