import torch from torch import nn import torch.nn.functional as F from typing import Any, Dict, List, Optional, Tuple, Union def get_rope_index( config, input_ids: Optional[torch.LongTensor] = None, image_grid_thw: Optional[torch.LongTensor] = None, video_grid_thw: Optional[torch.LongTensor] = None, second_per_grid_ts: Optional[torch.Tensor] = None, attention_mask: Optional[torch.Tensor] = None, ) -> Tuple[torch.Tensor, torch.Tensor]: """ Calculate the 3D rope index based on image and video's temporal, height and width in LLM. Explanation: Each embedding sequence contains vision embedding and text embedding or just contains text embedding. For pure text embedding sequence, the rotary position embedding has no difference with modern LLMs. Examples: input_ids: [T T T T T], here T is for text. temporal position_ids: [0, 1, 2, 3, 4] height position_ids: [0, 1, 2, 3, 4] width position_ids: [0, 1, 2, 3, 4] For vision and text embedding sequence, we calculate 3D rotary position embedding for vision part and 1D rotary position embeddin for text part. Examples: Temporal (Time): 3 patches, representing different segments of the video in time. Height: 2 patches, dividing each frame vertically. Width: 2 patches, dividing each frame horizontally. We also have some important parameters: fps (Frames Per Second): The video's frame rate, set to 1. This means one frame is processed each second. tokens_per_second: This is a crucial parameter. It dictates how many "time-steps" or "temporal tokens" are conceptually packed into a one-second interval of the video. In this case, we have 25 tokens per second. So each second of the video will be represented with 25 separate time points. It essentially defines the temporal granularity. temporal_patch_size: The number of frames that compose one temporal patch. Here, it's 2 frames. interval: The step size for the temporal position IDs, calculated as tokens_per_second * temporal_patch_size / fps. In this case, 25 * 2 / 1 = 50. This means that each temporal patch will be have a difference of 50 in the temporal position IDs. input_ids: [V V V V V V V V V V V V T T T T T], here V is for vision. vision temporal position_ids: [0, 0, 0, 0, 50, 50, 50, 50, 100, 100, 100, 100] vision height position_ids: [0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1] vision width position_ids: [0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1] text temporal position_ids: [101, 102, 103, 104, 105] text height position_ids: [101, 102, 103, 104, 105] text width position_ids: [101, 102, 103, 104, 105] Here we calculate the text start position_ids as the max vision position_ids plus 1. Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*): The temporal, height and width of feature shape of each image in LLM. video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*): The temporal, height and width of feature shape of each video in LLM. second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*): The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs. attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. Returns: position_ids (`torch.LongTensor` of shape `(3, batch_size, sequence_length)`) mrope_position_deltas (`torch.Tensor` of shape `(batch_size)`) """ spatial_merge_size = config.vision_config.spatial_merge_size image_token_id = config.image_token_id video_token_id = config.video_token_id vision_start_token_id = config.vision_start_token_id mrope_position_deltas = [] if input_ids is not None and (image_grid_thw is not None or video_grid_thw is not None): total_input_ids = input_ids if attention_mask is None: attention_mask = torch.ones_like(total_input_ids) position_ids = torch.ones( 3, input_ids.shape[0], input_ids.shape[1], dtype=input_ids.dtype, device=input_ids.device, ) image_index, video_index = 0, 0 attention_mask = attention_mask.to(total_input_ids.device) for i, input_ids in enumerate(total_input_ids): input_ids = input_ids[attention_mask[i] == 1] image_nums, video_nums = 0, 0 vision_start_indices = torch.argwhere(input_ids == vision_start_token_id).squeeze(1) vision_tokens = input_ids[vision_start_indices + 1] image_nums = (vision_tokens == image_token_id).sum() video_nums = (vision_tokens == video_token_id).sum() input_tokens = input_ids.tolist() llm_pos_ids_list: list = [] st = 0 remain_images, remain_videos = image_nums, video_nums for _ in range(image_nums + video_nums): if image_token_id in input_tokens and remain_images > 0: ed_image = input_tokens.index(image_token_id, st) else: ed_image = len(input_tokens) + 1 if video_token_id in input_tokens and remain_videos > 0: ed_video = input_tokens.index(video_token_id, st) else: ed_video = len(input_tokens) + 1 if ed_image < ed_video: t, h, w = ( image_grid_thw[image_index][0], image_grid_thw[image_index][1], image_grid_thw[image_index][2], ) second_per_grid_t = 0 image_index += 1 remain_images -= 1 ed = ed_image else: t, h, w = ( video_grid_thw[video_index][0], video_grid_thw[video_index][1], video_grid_thw[video_index][2], ) if second_per_grid_ts is not None: second_per_grid_t = second_per_grid_ts[video_index] else: second_per_grid_t = 1.0 video_index += 1 remain_videos -= 1 ed = ed_video llm_grid_t, llm_grid_h, llm_grid_w = ( t.item(), h.item() // spatial_merge_size, w.item() // spatial_merge_size, ) text_len = ed - st st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0 llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx) range_tensor = torch.arange(llm_grid_t).view(-1, 1) expanded_range = range_tensor.expand(-1, llm_grid_h * llm_grid_w) time_tensor = expanded_range * second_per_grid_t * config.vision_config.tokens_per_second time_tensor_long = time_tensor.long() t_index = time_tensor_long.flatten() h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten() w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten() llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + text_len + st_idx) st = ed + llm_grid_t * llm_grid_h * llm_grid_w if st < len(input_tokens): st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0 text_len = len(input_tokens) - st llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx) llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1) position_ids[..., i, attention_mask[i] == 1] = llm_positions.to(position_ids.device) mrope_position_deltas.append(llm_positions.max() + 1 - len(total_input_ids[i])) mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device).unsqueeze(1) return position_ids, mrope_position_deltas else: if attention_mask is not None: position_ids = attention_mask.long().cumsum(-1) - 1 position_ids.masked_fill_(attention_mask == 0, 1) position_ids = position_ids.unsqueeze(0).expand(3, -1, -1).to(attention_mask.device) max_position_ids = position_ids.max(0, keepdim=False)[0].max(-1, keepdim=True)[0] mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1] else: position_ids = ( torch.arange(input_ids.shape[1], device=input_ids.device) .view(1, 1, -1) .expand(3, input_ids.shape[0], -1) ) mrope_position_deltas = torch.zeros( [input_ids.shape[0], 1], device=input_ids.device, dtype=input_ids.dtype, ) return position_ids, mrope_position_deltas def get_window_index(grid_thw, window_size=112,spatial_merge_size=2,patch_size=14): spatial_merge_unit = spatial_merge_size * spatial_merge_size window_index: list = [] cu_window_seqlens: list = [0] window_index_id = 0 vit_merger_window_size = window_size // spatial_merge_size // patch_size for grid_t, grid_h, grid_w in grid_thw: llm_grid_h, llm_grid_w = ( grid_h // spatial_merge_size, grid_w // spatial_merge_size, ) index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(grid_t, llm_grid_h, llm_grid_w) pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size index_padded = F.pad(index, (0, pad_w, 0, pad_h), "constant", -100) index_padded = index_padded.reshape( grid_t, num_windows_h, vit_merger_window_size, num_windows_w, vit_merger_window_size, ) index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape( grid_t, num_windows_h * num_windows_w, vit_merger_window_size, vit_merger_window_size, ) seqlens = (index_padded != -100).sum([2, 3]).reshape(-1) index_padded = index_padded.reshape(-1) index_new = index_padded[index_padded != -100] window_index.append(index_new + window_index_id) cu_seqlens_tmp = seqlens.cumsum(0) * spatial_merge_unit + cu_window_seqlens[-1] cu_window_seqlens.extend(cu_seqlens_tmp.tolist()) window_index_id += (grid_t * llm_grid_h * llm_grid_w).item() window_index = torch.cat(window_index, dim=0) return window_index, cu_window_seqlens class Qwen2_5_VisionRotaryEmbedding(nn.Module): def __init__(self, dim: int, theta: float = 10000.0) -> None: super().__init__() inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim)) self.register_buffer("inv_freq", inv_freq, persistent=False) def forward(self, seqlen: int) -> torch.Tensor: seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype) freqs = torch.outer(seq, self.inv_freq) return freqs def rot_pos_emb( grid_thw, spatial_merge_size=2, hidden_size=2048, num_heads=16): head_dim = hidden_size // num_heads rotary_pos_emb = Qwen2_5_VisionRotaryEmbedding(head_dim // 2) pos_ids = [] for t, h, w in grid_thw: hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w) hpos_ids = hpos_ids.reshape( h // spatial_merge_size, spatial_merge_size, w // spatial_merge_size, spatial_merge_size, ) hpos_ids = hpos_ids.permute(0, 2, 1, 3) hpos_ids = hpos_ids.flatten() wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1) wpos_ids = wpos_ids.reshape( h // spatial_merge_size, spatial_merge_size, w // spatial_merge_size, spatial_merge_size, ) wpos_ids = wpos_ids.permute(0, 2, 1, 3) wpos_ids = wpos_ids.flatten() print("hpos_ids",hpos_ids.shape) pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1)) pos_ids = torch.cat(pos_ids, dim=0) max_grid_size = grid_thw[:, 1:].max() # return max_grid_size, pos_ids rotary_pos_emb_full = rotary_pos_emb(max_grid_size) rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1) return rotary_pos_emb def rot_pos_id(grid_thw, spatial_merge_size=2): pos_ids = [] for t, h, w in grid_thw: hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w) hpos_ids = hpos_ids.reshape( h // spatial_merge_size, spatial_merge_size, w // spatial_merge_size, spatial_merge_size, ) hpos_ids = hpos_ids.permute(0, 2, 1, 3) hpos_ids = hpos_ids.flatten() wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1) wpos_ids = wpos_ids.reshape( h // spatial_merge_size, spatial_merge_size, w // spatial_merge_size, spatial_merge_size, ) wpos_ids = wpos_ids.permute(0, 2, 1, 3) wpos_ids = wpos_ids.flatten() pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1)) pos_ids = torch.cat(pos_ids, dim=0) return pos_ids