# Copyright (C) 2022-present Naver Corporation. All rights reserved. # Licensed under CC BY-NC-SA 4.0 (non-commercial use only). # croco: https://github.com/naver/croco # diffusers: https://github.com/huggingface/diffusers # -------------------------------------------------------- # Position embedding utils # -------------------------------------------------------- import numpy as np import torch def get_2d_sincos_pos_embed( embed_dim, grid_size, cls_token=False, extra_tokens=0, interpolation_scale=1.0, base_size=16 ): """ grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token) """ if isinstance(grid_size, int): grid_size = (grid_size, grid_size) grid_h = np.arange(grid_size[0], dtype=np.float32) / (grid_size[0] / base_size) / interpolation_scale grid_w = np.arange(grid_size[1], dtype=np.float32) / (grid_size[1] / base_size) / interpolation_scale grid = np.meshgrid(grid_w, grid_h) # here w goes first grid = np.stack(grid, axis=0) grid = grid.reshape([2, 1, grid_size[1], grid_size[0]]) pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid) if cls_token and extra_tokens > 0: pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0) return pos_embed def get_2d_sincos_pos_embed_from_grid(embed_dim, grid): if embed_dim % 2 != 0: raise ValueError("embed_dim must be divisible by 2") # use half of dimensions to encode grid_h emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2) emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2) emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D) return emb def get_1d_sincos_pos_embed( embed_dim, length, interpolation_scale=1.0, base_size=16 ): pos = torch.arange(0, length).unsqueeze(1) / interpolation_scale pos_embed = get_1d_sincos_pos_embed_from_grid(embed_dim, pos) return pos_embed def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): """ embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D) """ if embed_dim % 2 != 0: raise ValueError("embed_dim must be divisible by 2") omega = np.arange(embed_dim // 2, dtype=np.float64) omega /= embed_dim / 2.0 omega = 1.0 / 10000 ** omega # (D/2,) pos = pos.reshape(-1) # (M,) out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product emb_sin = np.sin(out) # (M, D/2) emb_cos = np.cos(out) # (M, D/2) emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D) return emb # ---------------------------------------------------------- # RoPE2D: RoPE implementation in 2D # ---------------------------------------------------------- try: from .curope import cuRoPE2D RoPE2D = cuRoPE2D except ImportError: print('Warning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version instead') class RoPE2D(torch.nn.Module): def __init__(self, freq=10000.0, F0=1.0, scaling_factor=1.0): super().__init__() self.base = freq self.F0 = F0 self.scaling_factor = scaling_factor self.cache = {} def get_cos_sin(self, D, seq_len, device, dtype): if (D, seq_len, device, dtype) not in self.cache: inv_freq = 1.0 / (self.base ** (torch.arange(0, D, 2).float().to(device) / D)) t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) freqs = torch.einsum("i,j->ij", t, inv_freq).to(dtype) freqs = torch.cat((freqs, freqs), dim=-1) cos = freqs.cos() # (Seq, Dim) sin = freqs.sin() self.cache[D, seq_len, device, dtype] = (cos, sin) return self.cache[D, seq_len, device, dtype] @staticmethod def rotate_half(x): x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2:] return torch.cat((-x2, x1), dim=-1) def apply_rope1d(self, tokens, pos1d, cos, sin): assert pos1d.ndim == 2 cos = torch.nn.functional.embedding(pos1d, cos)[:, None, :, :] sin = torch.nn.functional.embedding(pos1d, sin)[:, None, :, :] return (tokens * cos) + (self.rotate_half(tokens) * sin) def forward(self, tokens, positions): """ input: * tokens: batch_size x nheads x ntokens x dim * positions: batch_size x ntokens x 2 (y and x position of each token) output: * tokens after appplying RoPE2D (batch_size x nheads x ntokens x dim) """ assert tokens.size(3) % 2 == 0, "number of dimensions should be a multiple of two" D = tokens.size(3) // 2 assert positions.ndim == 3 and positions.shape[-1] == 2 # Batch, Seq, 2 cos, sin = self.get_cos_sin(D, int(positions.max()) + 1, tokens.device, tokens.dtype) # split features into two along the feature dimension, and apply rope1d on each half y, x = tokens.chunk(2, dim=-1) y = self.apply_rope1d(y, positions[:, :, 0], cos, sin) x = self.apply_rope1d(x, positions[:, :, 1], cos, sin) tokens = torch.cat((y, x), dim=-1) return tokens class LinearScalingRoPE2D(RoPE2D): """Code from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py#L148""" def forward(self, tokens, positions): # difference to the original RoPE: a scaling factor is aplied to the position ids dtype = positions.dtype positions = positions.float() / self.scaling_factor positions = positions.to(dtype) tokens = super().forward(tokens, positions) return tokens try: from .curope import cuRoPE1D RoPE1D = cuRoPE1D except ImportError: print('Warning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version instead') class RoPE1D(torch.nn.Module): def __init__(self, freq=10000.0, F0=1.0, scaling_factor=1.0): super().__init__() self.base = freq self.F0 = F0 self.scaling_factor = scaling_factor self.cache = {} def get_cos_sin(self, D, seq_len, device, dtype): if (D, seq_len, device, dtype) not in self.cache: inv_freq = 1.0 / (self.base ** (torch.arange(0, D, 2).float().to(device) / D)) t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) freqs = torch.einsum("i,j->ij", t, inv_freq).to(dtype) freqs = torch.cat((freqs, freqs), dim=-1) cos = freqs.cos() # (Seq, Dim) sin = freqs.sin() self.cache[D, seq_len, device, dtype] = (cos, sin) return self.cache[D, seq_len, device, dtype] @staticmethod def rotate_half(x): x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2:] return torch.cat((-x2, x1), dim=-1) def apply_rope1d(self, tokens, pos1d, cos, sin): assert pos1d.ndim == 2 cos = torch.nn.functional.embedding(pos1d, cos)[:, None, :, :] sin = torch.nn.functional.embedding(pos1d, sin)[:, None, :, :] return (tokens * cos) + (self.rotate_half(tokens) * sin) def forward(self, tokens, positions): """ input: * tokens: batch_size x nheads x ntokens x dim * positions: batch_size x ntokens (t position of each token) output: * tokens after appplying RoPE2D (batch_size x nheads x ntokens x dim) """ D = tokens.size(3) assert positions.ndim == 2 # Batch, Seq cos, sin = self.get_cos_sin(D, int(positions.max()) + 1, tokens.device, tokens.dtype) tokens = self.apply_rope1d(tokens, positions, cos, sin) return tokens class LinearScalingRoPE1D(RoPE1D): """Code from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py#L148""" def forward(self, tokens, positions): # difference to the original RoPE: a scaling factor is aplied to the position ids dtype = positions.dtype positions = positions.float() / self.scaling_factor positions = positions.to(dtype) tokens = super().forward(tokens, positions) return tokens class PositionGetter2D(object): """ return positions of patches """ def __init__(self): self.cache_positions = {} def __call__(self, b, h, w, device): if not (h,w) in self.cache_positions: x = torch.arange(w, device=device) y = torch.arange(h, device=device) self.cache_positions[h,w] = torch.cartesian_prod(y, x) # (h, w, 2) pos = self.cache_positions[h,w].view(1, h*w, 2).expand(b, -1, 2).clone() return pos class PositionGetter1D(object): """ return positions of patches """ def __init__(self): self.cache_positions = {} def __call__(self, b, l, device): if not (l) in self.cache_positions: x = torch.arange(l, device=device) self.cache_positions[l] = x # (l, ) pos = self.cache_positions[l].view(1, l).expand(b, -1).clone() return pos