Rotary Embedding Implementation with Cache

Rotary Embedding Implementation with Cache | Generated by AI

Home 2026.07

Here’s the full file at /Users/lzwjava/projects/nano-vllm/nanovllm/layers/rotary_embedding.py:

from functools import lru_cache
import torch
from torch import nn


def apply_rotary_emb(
    x: torch.Tensor,
    cos: torch.Tensor,
    sin: torch.Tensor,
) -> torch.Tensor:
    x1, x2 = torch.chunk(x.float(), 2, dim=-1)
    y1 = x1 * cos - x2 * sin
    y2 = x2 * cos + x1 * sin
    return torch.cat((y1, y2), dim=-1).to(x.dtype)


class RotaryEmbedding(nn.Module):

    def __init__(
        self,
        head_size: int,
        rotary_dim: int,
        max_position_embeddings: int,
        base: float,
    ) -> None:
        super().__init__()
        self.head_size = head_size
        assert rotary_dim == head_size
        inv_freq = 1.0 / (base**(torch.arange(0, rotary_dim, 2, dtype=torch.float) / rotary_dim))
        t = torch.arange(max_position_embeddings, dtype=torch.float)
        freqs = torch.einsum("i,j -> ij", t, inv_freq)
        cos = freqs.cos()
        sin = freqs.sin()
        cache = torch.cat((cos, sin), dim=-1).unsqueeze_(1)
        self.register_buffer("cos_sin_cache", cache, persistent=False)

    def forward(
        self,
        positions: torch.Tensor,
        query: torch.Tensor,
        key: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        cos_sin = self.cos_sin_cache[positions]
        cos, sin = cos_sin.chunk(2, dim=-1)
        query = apply_rotary_emb(query, cos, sin)
        key = apply_rotary_emb(key, cos, sin)
        return query, key


@lru_cache(1)
def get_rope(
    head_size: int,
    rotary_dim: int,
    max_position: int,
    base: float,
):
    rotary_emb = RotaryEmbedding(head_size, rotary_dim, max_position, base)
    return rotary_emb

Note: In the actual file, there is no trailing comma after persistent=False) — it’s clean as shown above.

If you were asking about the trailing comma in your pasted snippet, that would be a harmless syntax error (a stray , after )). But the file on disk is correct.

How this RoPE implementation works

__init__ — Precomputes the full cos/sin table up to max_position_embeddings:
- inv_freq = 1 / (base^(2i/dim)) for i = 0, 2, 4, ... (standard RoPE frequency bands)
- freqs = outer product of position indices t × inv_freq
- cos, sin = element-wise trig of freqs
- Stored as a buffer cos_sin_cache of shape [max_pos, 1, rotary_dim*2] (the unsqueeze_(1) adds a head dimension)
forward — Looks up precomputed cos/sin for given positions, splits into cos/sin halves, and applies the rotation to both query and key.
apply_rotary_emb — Splits x into two halves (x1, x2) and applies:
```
y1 = x1·cos - x2·sin
y2 = x2·cos + x1·sin
```
This is the standard 2D rotation per pair of dimensions.
get_rope — A singleton factory cached with lru_cache so you only create one RotaryEmbedding per config.

Back Donate