add files

coglm_strategy.py

@@ -0,0 +1,101 @@
+# -*- encoding: utf-8 -*-
+'''
+@File    :   coglm_strategy.py
+@Time    :   2021/10/08 22:22:42
+@Author  :   Ming Ding 
+@Contact :   dm18@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+import os
+import sys
+import math
+import random
+import torch
+import numpy as np
+import torch.nn.functional as F
+
+
+def top_k_logits(logits, top_k=0, top_p=0.0, filter_value=-65504):
+    # This function has been mostly taken from huggingface conversational ai code at
+    # https://medium.com/huggingface/how-to-build-a-state-of-the-art-conversational-ai-with-transfer-learning-2d818ac26313
+
+    if top_k > 0:
+        # Remove all tokens with a probability less than the last token of the top-k
+        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+        logits[indices_to_remove] = filter_value
+
+    if top_p > 0.0:
+        # convert to 1D
+        logits = logits.view(logits.size()[1]).contiguous()
+        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+
+        # Remove tokens with cumulative probability above the threshold
+        sorted_indices_to_remove = cumulative_probs > top_p
+        # Shift the indices to the right to keep also the first token above the threshold
+        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+        sorted_indices_to_remove[..., 0] = 0
+        indices_to_remove = sorted_indices[sorted_indices_to_remove]
+        logits[indices_to_remove] = filter_value
+        # going back to 2D
+        logits = logits.view(1, -1).contiguous()
+
+    return logits
+
+
+class CoglmStrategy:
+    def __init__(self, invalid_slices=[], temperature=1., top_k=200, eps=1e-4, top_p=0.0, end_tokens=None, temperature2=0.89):
+        self.invalid_slices = invalid_slices
+        self.temperature = temperature
+        self.temperature2 = temperature2
+        self.topk = top_k
+        self.top_p = top_p
+        self.eps = eps
+        if end_tokens is None:
+            end_tokens = []
+        self.end_tokens = end_tokens
+        self._is_done = False
+        self.outlier_count_down = torch.zeros(16) 
+        self.vis_list = [[]for i in range(16)]
+        self.cluster_labels = torch.tensor(np.load('cluster_label2.npy'), device='cuda', dtype=torch.long)
+        self.start_pos = -1
+        self.white_cluster = []
+        # self.fout = open('tmp.txt', 'w')
+
+    @property
+    def is_done(self) -> bool:
+        return self._is_done
+
+    def forward(self, logits, tokens, mems, temperature=None, temperature2=None):
+        if temperature is None:
+            temperature = self.temperature
+        if temperature2 is None:
+            temperature2 = self.temperature2
+        logits = logits / temperature
+        for invalid_slice in self.invalid_slices:
+            logits[..., invalid_slice] = -65504
+            
+        rprobs = F.softmax(logits.float(), dim=-1)
+        c = self.cluster_labels.expand(*rprobs.shape)
+        cprobs = torch.zeros(logits.shape[0], 500, device=logits.device).scatter_add_(1, c, rprobs)
+        # self.fout.write(str(tokens.shape[-1])+ ' ' + str(cprobs.topk(10)) + '\n')
+        # self.fout.flush()
+        best_scores, best_clusters = cprobs.topk(self.topk)
+        bz = logits.shape[0]
+        for i in range(bz):
+            selected_cluster = best_clusters[i][torch.multinomial(best_scores[i] / best_scores[i].sum(), num_samples=1)]
+            logits[i, self.cluster_labels != selected_cluster] = -65504
+            
+        # logits = top_k_logits(logits, self.topk, self.top_p)
+        probs = F.softmax(logits.float()/temperature2, dim=-1)  # float is essetial, due to a bug in Pytorch
+        pred = torch.multinomial(probs, num_samples=1)
+        
+        if pred.numel() == 1 and pred.item() in self.end_tokens:
+            self._is_done = True
+        tokens = torch.cat((tokens, pred.view(tokens.shape[0], 1)), dim=1)
+        return tokens, mems
+
+    def finalize(self, tokens, mems):
+        self._is_done = False
+        return tokens, mems

cogvideo_pipeline.py

@@ -0,0 +1,793 @@
+# -*- encoding: utf-8 -*-
+'''
+@File    :   cogvideo_pipeline.py
+@Time    :   2022/07/15 11:24:56
+@Author  :   Wenyi Hong 
+@Version :   1.0
+@Contact :   hwy22@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+
+import os
+import sys
+import torch
+import argparse
+import time
+from torchvision.utils import save_image
+import stat
+from icetk import icetk as tokenizer
+import logging, sys
+
+import torch.distributed as dist
+tokenizer.add_special_tokens(['<start_of_image>', '<start_of_english>', '<start_of_chinese>'])
+
+
+from SwissArmyTransformer import get_args
+from SwissArmyTransformer.data_utils import BinaryDataset, make_loaders
+from SwissArmyTransformer.generation.sampling_strategies import BaseStrategy
+from SwissArmyTransformer.generation.utils import timed_name, save_multiple_images, generate_continually
+from SwissArmyTransformer.resources import auto_create
+
+from models.cogvideo_cache_model import CogVideoCacheModel
+from coglm_strategy import CoglmStrategy
+
+
+def get_masks_and_position_ids_stage1(data, textlen, framelen):
+    # Extract batch size and sequence length.
+    tokens = data
+    seq_length = len(data[0])
+    # Attention mask (lower triangular).
+    attention_mask = torch.ones((1, textlen+framelen, textlen+framelen), device=data.device)
+    attention_mask[:, :textlen, textlen:] = 0
+    attention_mask[:, textlen:, textlen:].tril_()
+    attention_mask.unsqueeze_(1)
+    # Unaligned version
+    position_ids = torch.zeros(seq_length, dtype=torch.long,
+                                device=data.device)
+    torch.arange(textlen, out=position_ids[:textlen], 
+        dtype=torch.long, device=data.device)
+    torch.arange(512, 512+seq_length-textlen, out=position_ids[textlen:], 
+        dtype=torch.long, device=data.device)
+    position_ids = position_ids.unsqueeze(0)
+
+    return tokens, attention_mask, position_ids
+
+def get_masks_and_position_ids_stage2(data, textlen, framelen):
+    # Extract batch size and sequence length.
+    tokens = data
+    seq_length = len(data[0])
+
+    # Attention mask (lower triangular).
+    attention_mask = torch.ones((1, textlen+framelen, textlen+framelen), device=data.device)
+    attention_mask[:, :textlen, textlen:] = 0
+    attention_mask[:, textlen:, textlen:].tril_()
+    attention_mask.unsqueeze_(1)
+    
+    # Unaligned version
+    position_ids = torch.zeros(seq_length, dtype=torch.long,
+                                device=data.device)
+    torch.arange(textlen, out=position_ids[:textlen], 
+        dtype=torch.long, device=data.device)
+    frame_num = (seq_length-textlen)//framelen
+    assert frame_num == 5
+    torch.arange(512, 512+framelen, out=position_ids[textlen:textlen+framelen], 
+            dtype=torch.long, device=data.device)
+    torch.arange(512+framelen*2, 512+framelen*3, out=position_ids[textlen+framelen:textlen+framelen*2], 
+            dtype=torch.long, device=data.device)
+    torch.arange(512+framelen*(frame_num-1), 512+framelen*frame_num, out=position_ids[textlen+framelen*2:textlen+framelen*3], 
+            dtype=torch.long, device=data.device)
+    torch.arange(512+framelen*1, 512+framelen*2, out=position_ids[textlen+framelen*3:textlen+framelen*4], 
+            dtype=torch.long, device=data.device)
+    torch.arange(512+framelen*3, 512+framelen*4, out=position_ids[textlen+framelen*4:textlen+framelen*5], 
+            dtype=torch.long, device=data.device)
+
+    position_ids = position_ids.unsqueeze(0)
+
+    return tokens, attention_mask, position_ids
+
+def my_update_mems(hiddens, mems_buffers, mems_indexs, limited_spatial_channel_mem, text_len, frame_len):
+    if hiddens is None:
+        return None, mems_indexs
+    mem_num = len(hiddens)
+    ret_mem = []
+    with torch.no_grad():
+        for id in range(mem_num):
+            if hiddens[id][0] is None:
+                ret_mem.append(None)
+            else: 
+                if id == 0 and limited_spatial_channel_mem and mems_indexs[id]+hiddens[0][0].shape[1] >= text_len+frame_len:
+                    if mems_indexs[id] == 0:
+                        for layer, hidden in enumerate(hiddens[id]):
+                            mems_buffers[id][layer, :, :text_len] = hidden.expand(mems_buffers[id].shape[1], -1, -1)[:, :text_len]
+                    new_mem_len_part2 = (mems_indexs[id]+hiddens[0][0].shape[1]-text_len)%frame_len 
+                    if new_mem_len_part2 > 0:
+                        for layer, hidden in enumerate(hiddens[id]):
+                            mems_buffers[id][layer, :, text_len:text_len+new_mem_len_part2] = hidden.expand(mems_buffers[id].shape[1], -1, -1)[:, -new_mem_len_part2:]
+                    mems_indexs[id] = text_len+new_mem_len_part2
+                else:
+                    for layer, hidden in enumerate(hiddens[id]):
+                        mems_buffers[id][layer, :, mems_indexs[id]:mems_indexs[id]+hidden.shape[1]] = hidden.expand(mems_buffers[id].shape[1], -1, -1)
+                    mems_indexs[id] += hidden.shape[1]
+                ret_mem.append(mems_buffers[id][:, :, :mems_indexs[id]])
+    return ret_mem, mems_indexs
+
+
+def my_save_multiple_images(imgs, path, subdir, debug=True):
+    # imgs: list of tensor images
+    if debug:
+        imgs = torch.cat(imgs, dim=0)
+        print("\nSave to: ", path, flush=True)
+        save_image(imgs, path, normalize=True)
+    else:
+        print("\nSave to: ", path, flush=True)
+        single_frame_path = os.path.join(path, subdir)
+        os.makedirs(single_frame_path, exist_ok=True)
+        for i in range(len(imgs)):
+            save_image(imgs[i], os.path.join(single_frame_path, f'{str(i).rjust(4,"0")}.jpg'), normalize=True)
+            os.chmod(os.path.join(single_frame_path,f'{str(i).rjust(4,"0")}.jpg'), stat.S_IRWXO+stat.S_IRWXG+stat.S_IRWXU)
+        save_image(torch.cat(imgs, dim=0), os.path.join(single_frame_path,f'frame_concat.jpg'), normalize=True)
+        os.chmod(os.path.join(single_frame_path,f'frame_concat.jpg'), stat.S_IRWXO+stat.S_IRWXG+stat.S_IRWXU)
+        
+def calc_next_tokens_frame_begin_id(text_len, frame_len, total_len):
+    # The fisrt token's position id of the frame that the next token belongs to; 
+    if total_len < text_len:
+        return None
+    return (total_len-text_len)//frame_len * frame_len + text_len
+
+def my_filling_sequence(
+        model, 
+        args,
+        seq, 
+        batch_size,
+        get_masks_and_position_ids,
+        text_len,
+        frame_len,
+        strategy=BaseStrategy(),
+        strategy2=BaseStrategy(),
+        mems=None,
+        log_text_attention_weights=0, # default to 0: no artificial change
+        mode_stage1=True,
+        enforce_no_swin=False,
+        guider_seq=None,
+        guider_text_len=0,
+        guidance_alpha=1,
+        limited_spatial_channel_mem=False, # 空间通道的存储限制在本帧内
+        **kw_args
+        ):
+    '''
+        seq: [2, 3, 5, ..., -1(to be generated), -1, ...]
+        mems: [num_layers, batch_size, len_mems(index), mem_hidden_size]
+            cache, should be first mems.shape[1] parts of context_tokens.
+            mems are the first-level citizens here, but we don't assume what is memorized.
+            input mems are used when multi-phase generation.
+    '''
+    if guider_seq is not None:
+        logging.debug("Using Guidance In Inference")
+    if limited_spatial_channel_mem:
+        logging.debug("Limit spatial-channel's mem to current frame")
+    assert len(seq.shape) == 2
+
+    # building the initial tokens, attention_mask, and position_ids
+    actual_context_length = 0
+    
+    while seq[-1][actual_context_length] >= 0: # the last seq has least given tokens
+        actual_context_length += 1 # [0, context_length-1] are given
+    assert actual_context_length > 0
+    current_frame_num = (actual_context_length-text_len) // frame_len
+    assert current_frame_num >= 0
+    context_length = text_len + current_frame_num * frame_len
+    
+    tokens, attention_mask, position_ids = get_masks_and_position_ids(seq, text_len, frame_len)
+    tokens = tokens[..., :context_length]
+    input_tokens = tokens.clone()
+    
+    if guider_seq is not None:
+        guider_index_delta = text_len - guider_text_len
+        guider_tokens, guider_attention_mask, guider_position_ids = get_masks_and_position_ids(guider_seq, guider_text_len, frame_len)
+        guider_tokens = guider_tokens[..., :context_length-guider_index_delta]
+        guider_input_tokens = guider_tokens.clone()
+        
+    for fid in range(current_frame_num):
+        input_tokens[:, text_len+400*fid] = tokenizer['<start_of_image>']
+        if guider_seq is not None:
+            guider_input_tokens[:, guider_text_len+400*fid] = tokenizer['<start_of_image>']
+
+    attention_mask = attention_mask.type_as(next(model.parameters())) # if fp16
+    # initialize generation
+    counter = context_length - 1 # Last fixed index is ``counter'' 
+    index = 0 # Next forward starting index, also the length of cache.
+    mems_buffers_on_GPU = False
+    mems_indexs = [0, 0]
+    mems_len = [(400+74) if limited_spatial_channel_mem else 5*400+74, 5*400+74]
+    mems_buffers = [torch.zeros(args.num_layers, batch_size, mem_len, args.hidden_size*2, dtype=next(model.parameters()).dtype)
+                        for mem_len in mems_len]
+
+    
+    if guider_seq is not None: 
+        guider_attention_mask = guider_attention_mask.type_as(next(model.parameters())) # if fp16
+        guider_mems_buffers = [torch.zeros(args.num_layers, batch_size, mem_len, args.hidden_size*2, dtype=next(model.parameters()).dtype)
+                        for mem_len in mems_len]
+        guider_mems_indexs = [0, 0]
+        guider_mems = None
+    
+    torch.cuda.empty_cache()
+    # step-by-step generation
+    while counter < len(seq[0]) - 1:
+        # we have generated counter+1 tokens
+        # Now, we want to generate seq[counter + 1],
+        # token[:, index: counter+1] needs forwarding.
+        if index == 0:
+            group_size = 2 if (input_tokens.shape[0] == batch_size and not mode_stage1) else batch_size
+            
+            logits_all = None
+            for batch_idx in range(0, input_tokens.shape[0], group_size):
+                logits, *output_per_layers = model(
+                    input_tokens[batch_idx:batch_idx+group_size, index:],
+                    position_ids[..., index: counter+1],
+                    attention_mask, # TODO memlen
+                    mems=mems,
+                    text_len=text_len,
+                    frame_len=frame_len,
+                    counter=counter,
+                    log_text_attention_weights=log_text_attention_weights,
+                    enforce_no_swin=enforce_no_swin,
+                    **kw_args
+                )
+                logits_all = torch.cat((logits_all, logits), dim=0) if logits_all is not None else logits
+                mem_kv01 = [[o['mem_kv'][0] for o in output_per_layers], [o['mem_kv'][1] for o in output_per_layers]]
+                next_tokens_frame_begin_id = calc_next_tokens_frame_begin_id(text_len, frame_len, mem_kv01[0][0].shape[1])
+                for id, mem_kv in enumerate(mem_kv01):
+                    for layer, mem_kv_perlayer in enumerate(mem_kv):
+                        if limited_spatial_channel_mem and id == 0:
+                            mems_buffers[id][layer, batch_idx:batch_idx+group_size, :text_len] = mem_kv_perlayer.expand(min(group_size, input_tokens.shape[0]-batch_idx), -1, -1)[:, :text_len]
+                            mems_buffers[id][layer, batch_idx:batch_idx+group_size, text_len:text_len+mem_kv_perlayer.shape[1]-next_tokens_frame_begin_id] =\
+                                mem_kv_perlayer.expand(min(group_size, input_tokens.shape[0]-batch_idx), -1, -1)[:, next_tokens_frame_begin_id:]
+                        else:
+                            mems_buffers[id][layer, batch_idx:batch_idx+group_size, :mem_kv_perlayer.shape[1]] = mem_kv_perlayer.expand(min(group_size, input_tokens.shape[0]-batch_idx), -1, -1)
+                mems_indexs[0], mems_indexs[1] = mem_kv01[0][0].shape[1], mem_kv01[1][0].shape[1]
+                if limited_spatial_channel_mem:
+                    mems_indexs[0] -= (next_tokens_frame_begin_id - text_len)
+
+            mems = [mems_buffers[id][:, :, :mems_indexs[id]] for id in range(2)]
+            logits = logits_all
+            
+            # Guider 
+            if guider_seq is not None:
+                guider_logits_all = None
+                for batch_idx in range(0, guider_input_tokens.shape[0], group_size):
+                    guider_logits, *guider_output_per_layers = model(
+                        guider_input_tokens[batch_idx:batch_idx+group_size, max(index-guider_index_delta, 0):],
+                        guider_position_ids[..., max(index-guider_index_delta, 0): counter+1-guider_index_delta],
+                        guider_attention_mask,
+                        mems=guider_mems,
+                        text_len=guider_text_len,
+                        frame_len=frame_len,
+                        counter=counter-guider_index_delta,
+                        log_text_attention_weights=log_text_attention_weights,
+                        enforce_no_swin=enforce_no_swin,
+                        **kw_args
+                    )
+                    guider_logits_all = torch.cat((guider_logits_all, guider_logits), dim=0) if guider_logits_all is not None else guider_logits
+                    guider_mem_kv01 = [[o['mem_kv'][0] for o in guider_output_per_layers], [o['mem_kv'][1] for o in guider_output_per_layers]]
+                    for id, guider_mem_kv in enumerate(guider_mem_kv01):
+                        for layer, guider_mem_kv_perlayer in enumerate(guider_mem_kv):
+                            if limited_spatial_channel_mem and id == 0:
+                                guider_mems_buffers[id][layer, batch_idx:batch_idx+group_size, :guider_text_len] = guider_mem_kv_perlayer.expand(min(group_size, input_tokens.shape[0]-batch_idx), -1, -1)[:, :guider_text_len]
+                                guider_next_tokens_frame_begin_id = calc_next_tokens_frame_begin_id(guider_text_len, frame_len, guider_mem_kv_perlayer.shape[1])
+                                guider_mems_buffers[id][layer, batch_idx:batch_idx+group_size, guider_text_len:guider_text_len+guider_mem_kv_perlayer.shape[1]-guider_next_tokens_frame_begin_id] =\
+                                    guider_mem_kv_perlayer.expand(min(group_size, input_tokens.shape[0]-batch_idx), -1, -1)[:, guider_next_tokens_frame_begin_id:]
+                            else:
+                                guider_mems_buffers[id][layer, batch_idx:batch_idx+group_size, :guider_mem_kv_perlayer.shape[1]] = guider_mem_kv_perlayer.expand(min(group_size, input_tokens.shape[0]-batch_idx), -1, -1)
+                    guider_mems_indexs[0], guider_mems_indexs[1] = guider_mem_kv01[0][0].shape[1], guider_mem_kv01[1][0].shape[1]
+                    if limited_spatial_channel_mem:
+                        guider_mems_indexs[0] -= (guider_next_tokens_frame_begin_id-guider_text_len)
+                guider_mems = [guider_mems_buffers[id][:, :, :guider_mems_indexs[id]] for id in range(2)]
+                guider_logits = guider_logits_all
+        else:
+            if not mems_buffers_on_GPU:
+                if not mode_stage1:
+                    torch.cuda.empty_cache()
+                    for idx, mem in enumerate(mems):
+                        mems[idx] = mem.to(next(model.parameters()).device)
+                    if guider_seq is not None:
+                        for idx, mem in enumerate(guider_mems):
+                            guider_mems[idx] = mem.to(next(model.parameters()).device) 
+                else: 
+                    torch.cuda.empty_cache()
+                    for idx, mem_buffer in enumerate(mems_buffers):
+                        mems_buffers[idx] = mem_buffer.to(next(model.parameters()).device)
+                    mems = [mems_buffers[id][:, :, :mems_indexs[id]] for id in range(2)]
+                    if guider_seq is not None:
+                        for idx, guider_mem_buffer in enumerate(guider_mems_buffers):
+                            guider_mems_buffers[idx] = guider_mem_buffer.to(next(model.parameters()).device)
+                        guider_mems = [guider_mems_buffers[id][:, :, :guider_mems_indexs[id]] for id in range(2)]
+                    mems_buffers_on_GPU = True
+                    
+            logits, *output_per_layers = model(
+                input_tokens[:, index:],
+                position_ids[..., index: counter+1],
+                attention_mask, # TODO memlen
+                mems=mems,
+                text_len=text_len,
+                frame_len=frame_len,
+                counter=counter,
+                log_text_attention_weights=log_text_attention_weights,
+                enforce_no_swin=enforce_no_swin,
+                limited_spatial_channel_mem=limited_spatial_channel_mem,
+                **kw_args
+            )
+            mem_kv0, mem_kv1 = [o['mem_kv'][0] for o in output_per_layers], [o['mem_kv'][1] for o in output_per_layers]
+            
+            if guider_seq is not None:
+                guider_logits, *guider_output_per_layers = model(
+                    guider_input_tokens[:, max(index-guider_index_delta, 0):],
+                    guider_position_ids[..., max(index-guider_index_delta, 0): counter+1-guider_index_delta],
+                    guider_attention_mask,
+                    mems=guider_mems,
+                    text_len=guider_text_len,
+                    frame_len=frame_len,
+                    counter=counter-guider_index_delta,
+                    log_text_attention_weights=0,
+                    enforce_no_swin=enforce_no_swin,
+                    limited_spatial_channel_mem=limited_spatial_channel_mem,
+                    **kw_args
+                )
+                guider_mem_kv0, guider_mem_kv1 = [o['mem_kv'][0] for o in guider_output_per_layers], [o['mem_kv'][1] for o in guider_output_per_layers]
+            
+            if not mems_buffers_on_GPU:
+                torch.cuda.empty_cache()
+                for idx, mem_buffer in enumerate(mems_buffers):
+                    mems_buffers[idx] = mem_buffer.to(next(model.parameters()).device)
+                if guider_seq is not None:
+                    for idx, guider_mem_buffer in enumerate(guider_mems_buffers):
+                        guider_mems_buffers[idx] = guider_mem_buffer.to(next(model.parameters()).device)
+                mems_buffers_on_GPU = True
+
+            mems, mems_indexs = my_update_mems([mem_kv0, mem_kv1], mems_buffers, mems_indexs, limited_spatial_channel_mem, text_len, frame_len) 
+            if guider_seq is not None: 
+                guider_mems, guider_mems_indexs = my_update_mems([guider_mem_kv0, guider_mem_kv1], guider_mems_buffers, guider_mems_indexs, limited_spatial_channel_mem, guider_text_len, frame_len)
+
+       
+        counter += 1 
+        index = counter        
+
+        logits = logits[:, -1].expand(batch_size, -1) # [batch size, vocab size]
+        tokens = tokens.expand(batch_size, -1)
+        if guider_seq is not None:
+            guider_logits = guider_logits[:, -1].expand(batch_size, -1)
+            guider_tokens = guider_tokens.expand(batch_size, -1)
+        
+        if seq[-1][counter].item() < 0:
+            # sampling
+            guided_logits = guider_logits+(logits-guider_logits)*guidance_alpha if guider_seq is not None else logits
+            if mode_stage1 and counter < text_len + 400:
+                tokens, mems = strategy.forward(guided_logits, tokens, mems)
+            else:
+                tokens, mems = strategy2.forward(guided_logits, tokens, mems)
+            if guider_seq is not None:
+                guider_tokens = torch.cat((guider_tokens, tokens[:, -1:]), dim=1)
+                
+            if seq[0][counter].item() >= 0:
+                for si in range(seq.shape[0]):
+                    if seq[si][counter].item() >= 0:
+                        tokens[si, -1] = seq[si, counter]
+                        if guider_seq is not None:
+                            guider_tokens[si, -1] = guider_seq[si, counter-guider_index_delta]
+                    
+        else:
+            tokens = torch.cat((tokens, seq[:, counter:counter+1].clone().expand(tokens.shape[0], 1).to(device=tokens.device, dtype=tokens.dtype)), dim=1)
+            if guider_seq is not None: 
+                guider_tokens = torch.cat((guider_tokens, 
+                                           guider_seq[:, counter-guider_index_delta:counter+1-guider_index_delta]
+                                           .clone().expand(guider_tokens.shape[0], 1).to(device=guider_tokens.device, dtype=guider_tokens.dtype)), dim=1)
+        
+        input_tokens = tokens.clone()
+        if guider_seq is not None:
+            guider_input_tokens = guider_tokens.clone()
+        if (index-text_len-1)//400 < (input_tokens.shape[-1]-text_len-1)//400:
+            boi_idx = ((index-text_len-1)//400 +1)*400+text_len
+            while boi_idx < input_tokens.shape[-1]:
+                input_tokens[:, boi_idx] = tokenizer['<start_of_image>']
+                if guider_seq is not None:
+                    guider_input_tokens[:, boi_idx-guider_index_delta] = tokenizer['<start_of_image>']
+                boi_idx += 400
+        
+        if strategy.is_done:
+            break
+    return strategy.finalize(tokens, mems)
+
+class InferenceModel_Sequential(CogVideoCacheModel):
+    def __init__(self, args, transformer=None, parallel_output=True):
+        super().__init__(args, transformer=transformer, parallel_output=parallel_output, window_size=-1, cogvideo_stage=1)
+    # TODO: check it 
+    
+    def final_forward(self, logits, **kwargs):
+        logits_parallel = logits
+        logits_parallel = torch.nn.functional.linear(logits_parallel.float(), self.transformer.word_embeddings.weight[:20000].float())
+        return logits_parallel
+    
+class InferenceModel_Interpolate(CogVideoCacheModel):
+    def __init__(self, args, transformer=None, parallel_output=True):
+        super().__init__(args, transformer=transformer, parallel_output=parallel_output, window_size=10, cogvideo_stage=2)
+    # TODO: check it 
+
+    def final_forward(self, logits, **kwargs):
+        logits_parallel = logits
+        logits_parallel = torch.nn.functional.linear(logits_parallel.float(), self.transformer.word_embeddings.weight[:20000].float())
+        return logits_parallel
+
+def main(args):
+    assert int(args.stage_1) + int(args.stage_2) + int(args.both_stages) == 1
+    rank_id = args.device % args.parallel_size
+    generate_frame_num = args.generate_frame_num
+    
+    if args.stage_1 or args.both_stages:
+        model_stage1, args = InferenceModel_Sequential.from_pretrained(args, 'cogvideo-stage1')
+        model_stage1.eval()
+        if args.both_stages:
+            model_stage1 = model_stage1.cpu()
+
+    if args.stage_2 or args.both_stages:
+        model_stage2, args = InferenceModel_Interpolate.from_pretrained(args, 'cogvideo-stage2')
+        model_stage2.eval()
+        if args.both_stages:
+            model_stage2 = model_stage2.cpu()
+    
+    invalid_slices = [slice(tokenizer.num_image_tokens, None)]
+    strategy_cogview2 = CoglmStrategy(invalid_slices, 
+        temperature=1.0, top_k=16)
+    strategy_cogvideo = CoglmStrategy(invalid_slices, 
+        temperature=args.temperature, top_k=args.top_k,
+        temperature2=args.coglm_temperature2)
+    if not args.stage_1:
+        from sr_pipeline import DirectSuperResolution 
+        dsr_path = auto_create('cogview2-dsr', path=None) # path=os.getenv('SAT_HOME', '~/.sat_models')
+        dsr = DirectSuperResolution(args, dsr_path,
+                                    max_bz=12, onCUDA=False)
+    
+    def process_stage2(model, seq_text, duration, video_raw_text=None, video_guidance_text="视频", parent_given_tokens=None, conddir=None, outputdir=None, gpu_rank=0, gpu_parallel_size=1):
+        stage2_starttime = time.time()
+        use_guidance = args.use_guidance_stage2
+        if args.both_stages:
+            move_start_time = time.time()
+            logging.debug("moving stage-2 model to cuda")
+            model = model.cuda()
+            logging.debug("moving in stage-2 model takes time: {:.2f}".format(time.time()-move_start_time))
+            
+        try:
+            if parent_given_tokens is None:
+                assert conddir is not None
+                parent_given_tokens = torch.load(os.path.join(conddir, 'frame_tokens.pt'), map_location='cpu')
+            sample_num_allgpu = parent_given_tokens.shape[0]
+            sample_num = sample_num_allgpu // gpu_parallel_size
+            assert sample_num * gpu_parallel_size == sample_num_allgpu
+            parent_given_tokens = parent_given_tokens[gpu_rank*sample_num:(gpu_rank+1)*sample_num]
+        except:
+            logging.critical("No frame_tokens found in interpolation, skip")
+            return False
+        
+        # CogVideo Stage2 Generation
+        while duration >= 0.5: # TODO: You can change the boundary to change the frame rate
+            parent_given_tokens_num = parent_given_tokens.shape[1]
+            generate_batchsize_persample = (parent_given_tokens_num-1)//2
+            generate_batchsize_total = generate_batchsize_persample * sample_num
+            total_frames = generate_frame_num
+            frame_len = 400
+            enc_text = tokenizer.encode(seq_text)
+            enc_duration = tokenizer.encode(str(float(duration))+"秒")
+            seq = enc_duration + [tokenizer['<n>']] + enc_text + [tokenizer['<start_of_image>']] + [-1]*400*generate_frame_num
+            text_len = len(seq) - frame_len*generate_frame_num - 1
+            
+            logging.info("[Stage2: Generating Frames, Frame Rate {:d}]\nraw text: {:s}".format(int(4/duration), tokenizer.decode(enc_text)))
+            
+            # generation
+            seq = torch.cuda.LongTensor(seq, device=args.device).unsqueeze(0).repeat(generate_batchsize_total, 1)
+            for sample_i in range(sample_num):
+                for i in range(generate_batchsize_persample):
+                    seq[sample_i*generate_batchsize_persample+i][text_len+1:text_len+1+400] = parent_given_tokens[sample_i][2*i]
+                    seq[sample_i*generate_batchsize_persample+i][text_len+1+400:text_len+1+800] = parent_given_tokens[sample_i][2*i+1]
+                    seq[sample_i*generate_batchsize_persample+i][text_len+1+800:text_len+1+1200] = parent_given_tokens[sample_i][2*i+2]
+                
+            if use_guidance:
+                guider_seq = enc_duration + [tokenizer['<n>']] + tokenizer.encode(video_guidance_text) + [tokenizer['<start_of_image>']] + [-1]*400*generate_frame_num                 
+                guider_text_len = len(guider_seq) - frame_len*generate_frame_num - 1
+                guider_seq = torch.cuda.LongTensor(guider_seq, device=args.device).unsqueeze(0).repeat(generate_batchsize_total, 1)
+                for sample_i in range(sample_num):
+                    for i in range(generate_batchsize_persample):
+                        guider_seq[sample_i*generate_batchsize_persample+i][text_len+1:text_len+1+400] = parent_given_tokens[sample_i][2*i]
+                        guider_seq[sample_i*generate_batchsize_persample+i][text_len+1+400:text_len+1+800] = parent_given_tokens[sample_i][2*i+1]
+                        guider_seq[sample_i*generate_batchsize_persample+i][text_len+1+800:text_len+1+1200] = parent_given_tokens[sample_i][2*i+2]
+                video_log_text_attention_weights = 0
+            else:
+                guider_seq=None
+                guider_text_len=0
+                video_log_text_attention_weights = 1.4
+
+            mbz = args.max_inference_batch_size
+
+            assert generate_batchsize_total < mbz or generate_batchsize_total % mbz == 0
+            output_list = []
+            start_time = time.time()
+            for tim in range(max(generate_batchsize_total // mbz, 1)):
+                input_seq = seq[:min(generate_batchsize_total, mbz)].clone() if tim == 0 else seq[mbz*tim:mbz*(tim+1)].clone()
+                guider_seq2 = (guider_seq[:min(generate_batchsize_total, mbz)].clone() if tim == 0 else guider_seq[mbz*tim:mbz*(tim+1)].clone()) if guider_seq is not None else None
+                output_list.append(
+                    my_filling_sequence(model, args, input_seq,
+                        batch_size=min(generate_batchsize_total, mbz),
+                        get_masks_and_position_ids=get_masks_and_position_ids_stage2,
+                        text_len=text_len, frame_len=frame_len,
+                        strategy=strategy_cogview2,
+                        strategy2=strategy_cogvideo,
+                        log_text_attention_weights=video_log_text_attention_weights,
+                        mode_stage1=False,
+                        guider_seq=guider_seq2,
+                        guider_text_len=guider_text_len,
+                        guidance_alpha=args.guidance_alpha,
+                        limited_spatial_channel_mem=True,
+                        )[0]
+                    )
+            logging.info("Duration {:.2f}, Taken time {:.2f}\n".format(duration, time.time() - start_time))
+
+            output_tokens = torch.cat(output_list, dim=0)
+            output_tokens = output_tokens[:, text_len+1:text_len+1+(total_frames)*400].reshape(sample_num, -1, 400*total_frames)
+            output_tokens_merge = torch.cat((output_tokens[:, :, :1*400], 
+                                            output_tokens[:, :, 400*3:4*400],
+                                            output_tokens[:, :, 400*1:2*400],
+                                            output_tokens[:, :, 400*4:(total_frames)*400]), dim=2).reshape(sample_num, -1, 400)     
+
+            output_tokens_merge = torch.cat((output_tokens_merge, output_tokens[:, -1:, 400*2:3*400]), dim=1)
+            duration /= 2
+            parent_given_tokens = output_tokens_merge
+            
+        if args.both_stages:
+            move_start_time = time.time()
+            logging.debug("moving stage 2 model to cpu")
+            model = model.cpu()
+            torch.cuda.empty_cache()
+            logging.debug("moving out model2 takes time: {:.2f}".format(time.time()-move_start_time))
+            
+        logging.info("CogVideo Stage2 completed. Taken time {:.2f}\n".format(time.time() - stage2_starttime))
+
+        # decoding
+        # imgs = [torch.nn.functional.interpolate(tokenizer.decode(image_ids=seq.tolist()), size=(480, 480)) for seq in output_tokens_merge]
+        # os.makedirs(output_dir_full_path, exist_ok=True)
+        # my_save_multiple_images(imgs, output_dir_full_path,subdir="frames", debug=False)
+        # torch.save(output_tokens_merge.cpu(), os.path.join(output_dir_full_path, 'frame_token.pt'))
+        # os.system(f"gifmaker -i '{output_dir_full_path}'/frames/0*.jpg -o '{output_dir_full_path}/{str(float(duration))}_concat.gif' -d 0.2")
+        
+        # direct super-resolution by CogView2
+        logging.info("[Direct super-resolution]")
+        dsr_starttime = time.time()
+        enc_text = tokenizer.encode(seq_text)
+        frame_num_per_sample = parent_given_tokens.shape[1]
+        parent_given_tokens_2d = parent_given_tokens.reshape(-1, 400)
+        text_seq = torch.cuda.LongTensor(enc_text, device=args.device).unsqueeze(0).repeat(parent_given_tokens_2d.shape[0], 1)
+        sred_tokens = dsr(text_seq, parent_given_tokens_2d)
+        decoded_sr_videos = []
+        
+        for sample_i in range(sample_num):
+            decoded_sr_imgs = []
+            for frame_i in range(frame_num_per_sample):
+                decoded_sr_img = tokenizer.decode(image_ids=sred_tokens[frame_i+sample_i*frame_num_per_sample][-3600:])
+                decoded_sr_imgs.append(torch.nn.functional.interpolate(decoded_sr_img, size=(480, 480)))
+            decoded_sr_videos.append(decoded_sr_imgs)
+
+        for sample_i in range(sample_num):
+            my_save_multiple_images(decoded_sr_videos[sample_i], outputdir,subdir=f"frames/{sample_i+sample_num*gpu_rank}", debug=False)
+            os.system(f"gifmaker -i '{outputdir}'/frames/'{sample_i+sample_num*gpu_rank}'/0*.jpg -o '{outputdir}/{sample_i+sample_num*gpu_rank}.gif' -d 0.125")
+        
+        logging.info("Direct super-resolution completed. Taken time {:.2f}\n".format(time.time() - dsr_starttime))
+
+        return True
+   
+    
+    def process_stage1(model, seq_text, duration, video_raw_text=None, video_guidance_text="视频", image_text_suffix="", outputdir=None, batch_size=1):
+        process_start_time = time.time()
+        use_guide = args.use_guidance_stage1
+        if args.both_stages:
+            move_start_time = time.time()
+            logging.debug("moving stage 1 model to cuda")
+            model = model.cuda()
+            logging.debug("moving in model1 takes time: {:.2f}".format(time.time()-move_start_time))
+            
+        if video_raw_text is None: 
+            video_raw_text = seq_text
+        mbz = args.stage1_max_inference_batch_size if args.stage1_max_inference_batch_size > 0 else args.max_inference_batch_size
+        assert batch_size < mbz or batch_size % mbz == 0
+        frame_len = 400
+        
+        # generate the first frame:
+        enc_text = tokenizer.encode(seq_text+image_text_suffix)
+        seq_1st = enc_text + [tokenizer['<start_of_image>']] + [-1]*400 # IV!!  # test local!!! # test randboi!!!
+        logging.info("[Generating First Frame with CogView2]Raw text: {:s}".format(tokenizer.decode(enc_text)))
+        text_len_1st = len(seq_1st) - frame_len*1 - 1
+
+        seq_1st = torch.cuda.LongTensor(seq_1st, device=args.device).unsqueeze(0)
+        output_list_1st = []
+        for tim in range(max(batch_size // mbz, 1)):
+            start_time = time.time()
+            output_list_1st.append(
+                my_filling_sequence(model, args,seq_1st.clone(),
+                    batch_size=min(batch_size, mbz),
+                    get_masks_and_position_ids=get_masks_and_position_ids_stage1,
+                    text_len=text_len_1st, 
+                    frame_len=frame_len,
+                    strategy=strategy_cogview2,
+                    strategy2=strategy_cogvideo,
+                    log_text_attention_weights=1.4,
+                    enforce_no_swin=True,
+                    mode_stage1=True,
+                    )[0]
+                )
+            logging.info("[First Frame]Taken time {:.2f}\n".format(time.time() - start_time))
+        output_tokens_1st = torch.cat(output_list_1st, dim=0)
+        given_tokens = output_tokens_1st[:, text_len_1st+1:text_len_1st+401].unsqueeze(1) # given_tokens.shape: [bs, frame_num, 400]
+
+        # generate subsequent frames:
+        total_frames = generate_frame_num
+        enc_duration = tokenizer.encode(str(float(duration))+"秒")
+        if use_guide:
+            video_raw_text = video_raw_text + " 视频"
+        enc_text_video = tokenizer.encode(video_raw_text)
+        seq = enc_duration + [tokenizer['<n>']] + enc_text_video + [tokenizer['<start_of_image>']] + [-1]*400*generate_frame_num
+        guider_seq = enc_duration + [tokenizer['<n>']] + tokenizer.encode(video_guidance_text) + [tokenizer['<start_of_image>']] + [-1]*400*generate_frame_num
+        logging.info("[Stage1: Generating Subsequent Frames, Frame Rate {:.1f}]\nraw text: {:s}".format(4/duration, tokenizer.decode(enc_text_video)))
+
+        text_len = len(seq) - frame_len*generate_frame_num - 1
+        guider_text_len = len(guider_seq) - frame_len*generate_frame_num - 1
+        seq = torch.cuda.LongTensor(seq, device=args.device).unsqueeze(0).repeat(batch_size, 1)
+        guider_seq = torch.cuda.LongTensor(guider_seq, device=args.device).unsqueeze(0).repeat(batch_size, 1)
+
+        for given_frame_id in range(given_tokens.shape[1]):
+            seq[:, text_len+1+given_frame_id*400: text_len+1+(given_frame_id+1)*400] = given_tokens[:, given_frame_id]
+            guider_seq[:, guider_text_len+1+given_frame_id*400:guider_text_len+1+(given_frame_id+1)*400] = given_tokens[:, given_frame_id]
+        output_list = []
+        
+        if use_guide:
+            video_log_text_attention_weights = 0
+        else: 
+            guider_seq = None
+            video_log_text_attention_weights = 1.4
+            
+        for tim in range(max(batch_size // mbz, 1)):
+            start_time = time.time()
+            input_seq = seq[:min(batch_size, mbz)].clone() if tim == 0 else seq[mbz*tim:mbz*(tim+1)].clone()
+            guider_seq2 = (guider_seq[:min(batch_size, mbz)].clone() if tim == 0 else guider_seq[mbz*tim:mbz*(tim+1)].clone()) if guider_seq is not None else None
+            output_list.append(
+                my_filling_sequence(model, args,input_seq,
+                    batch_size=min(batch_size, mbz),
+                    get_masks_and_position_ids=get_masks_and_position_ids_stage1,
+                    text_len=text_len, frame_len=frame_len,
+                    strategy=strategy_cogview2,
+                    strategy2=strategy_cogvideo,
+                    log_text_attention_weights=video_log_text_attention_weights,
+                    guider_seq=guider_seq2,
+                    guider_text_len=guider_text_len,
+                    guidance_alpha=args.guidance_alpha,
+                    limited_spatial_channel_mem=True,
+                    mode_stage1=True,
+                    )[0]
+                )
+
+        output_tokens = torch.cat(output_list, dim=0)[:, 1+text_len:]
+        
+        if args.both_stages:
+            move_start_time = time.time()
+            logging.debug("moving stage 1 model to cpu")
+            model = model.cpu()
+            torch.cuda.empty_cache()
+            logging.debug("moving in model1 takes time: {:.2f}".format(time.time()-move_start_time))
+        
+        # decoding
+        imgs, sred_imgs, txts = [], [], []
+        for seq in output_tokens:
+            decoded_imgs = [torch.nn.functional.interpolate(tokenizer.decode(image_ids=seq.tolist()[i*400: (i+1)*400]), size=(480, 480)) for i in range(total_frames)]
+            imgs.append(decoded_imgs) # only the last image (target)
+
+        assert len(imgs) == batch_size
+        save_tokens = output_tokens[:, :+total_frames*400].reshape(-1, total_frames, 400).cpu()
+        if outputdir is not None:
+            for clip_i in range(len(imgs)):
+                # os.makedirs(output_dir_full_paths[clip_i], exist_ok=True)
+                my_save_multiple_images(imgs[clip_i], outputdir, subdir=f"frames/{clip_i}", debug=False)
+                os.system(f"gifmaker -i '{outputdir}'/frames/'{clip_i}'/0*.jpg -o '{outputdir}/{clip_i}.gif' -d 0.25")
+            torch.save(save_tokens, os.path.join(outputdir, 'frame_tokens.pt'))
+        
+        logging.info("CogVideo Stage1 completed. Taken time {:.2f}\n".format(time.time() - process_start_time))
+        
+        return save_tokens
+               
+    # ======================================================================================================
+    
+    if args.stage_1 or args.both_stages:
+        if args.input_source != "interactive":
+            with open(args.input_source, 'r') as fin:
+                promptlist = fin.readlines()
+            promptlist = [p.strip() for p in promptlist]
+        else:
+            promptlist = None 
+            
+        now_qi = -1
+        while True:
+            now_qi += 1
+            
+            if promptlist is not None: # with input-source
+                if args.multi_gpu:
+                    if now_qi % dist.get_world_size() != dist.get_rank():
+                        continue
+                    rk = dist.get_rank()
+                else: 
+                    rk = 0
+                raw_text = promptlist[now_qi]
+                raw_text = raw_text.strip()
+                print(f'Working on Line No. {now_qi} on {rk}... [{raw_text}]')
+            else: # interactive
+                raw_text = input("\nPlease Input Query (stop to exit) >>> ") 
+                raw_text = raw_text.strip()
+                if not raw_text:
+                    print('Query should not be empty!')
+                    continue
+                if raw_text == "stop":
+                    return 
+                
+            try:
+                path = os.path.join(args.output_path, f"{now_qi}_{raw_text}")
+                parent_given_tokens = process_stage1(model_stage1, raw_text, duration=4.0, video_raw_text=raw_text, video_guidance_text="视频",
+                                                     image_text_suffix=" 高清摄影",
+                                                     outputdir=path if args.stage_1 else None, batch_size=args.batch_size)
+                if args.both_stages:
+                    process_stage2(model_stage2, raw_text, duration=2.0, video_raw_text=raw_text+" 视频", 
+                            video_guidance_text="视频", parent_given_tokens=parent_given_tokens, 
+                            outputdir=path,
+                            gpu_rank=0, gpu_parallel_size=1) # TODO: 修改
+            except (ValueError, FileNotFoundError) as e:
+                print(e)
+                continue 
+            
+    elif args.stage_2: 
+        sample_dirs = os.listdir(args.output_path)
+        for sample in sample_dirs:    
+            raw_text = sample.split('_')[-1]
+            path = os.path.join(args.output_path, sample, 'Interp')
+            parent_given_tokens = torch.load(os.path.join(args.output_path, sample, "frame_tokens.pt"))
+            
+            process_stage2(raw_text, duration=2.0, video_raw_text=raw_text+" 视频", 
+                            video_guidance_text="视频", parent_given_tokens=parent_given_tokens, 
+                            outputdir=path,
+                            gpu_rank=0, gpu_parallel_size=1) # TODO: 修改
+            
+    else: 
+        assert False
+
+
+if __name__ == "__main__":    
+    logging.basicConfig(stream=sys.stderr, level=logging.DEBUG)
+    
+    py_parser = argparse.ArgumentParser(add_help=False)
+    py_parser.add_argument('--generate-frame-num', type=int, default=5)
+    py_parser.add_argument('--coglm-temperature2', type=float, default=0.89)
+    # py_parser.add_argument("--interp-duration", type=float, default=-1) # -1是顺序生成，0是超分，0.5/1/2是插帧
+    # py_parser.add_argument("--total-duration", type=float, default=4.0) # 整个的时间
+    py_parser.add_argument('--use-guidance-stage1', action='store_true')
+    py_parser.add_argument('--use-guidance-stage2', action='store_true')
+    py_parser.add_argument('--guidance-alpha', type=float, default=3.0)
+    py_parser.add_argument('--stage-1', action='store_true') # stage 1: sequential generation
+    py_parser.add_argument('--stage-2', action='store_true') # stage 2: interp + dsr
+    py_parser.add_argument('--both-stages', action='store_true') # stage 1&2: sequential generation; interp + dsr
+    py_parser.add_argument('--parallel-size', type=int, default=1)
+    py_parser.add_argument('--stage1-max-inference-batch-size', type=int, default=-1) # -1: use max-inference-batch-size
+    py_parser.add_argument('--multi-gpu', action='store_true')
+
+    CogVideoCacheModel.add_model_specific_args(py_parser)
+
+    known, args_list = py_parser.parse_known_args()
+    args = get_args(args_list)
+    args = argparse.Namespace(**vars(args), **vars(known))
+    args.layout = [int(x) for x in args.layout.split(',')]
+    args.do_train = False
+
+    torch.cuda.set_device(args.device)
+
+    with torch.no_grad():
+        main(args)

models/cogvideo_cache_model.py

@@ -0,0 +1,695 @@
+# -*- encoding: utf-8 -*-
+'''
+@File    :   cogvideo_cache_model.py
+@Time    :   2022/07/15 11:22:19
+@Author  :   Wenyi Hong 
+@Version :   1.0
+@Contact :   hwy22@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+
+from multiprocessing import context
+from tkinter import E
+import torch
+from SwissArmyTransformer.model.base_model import BaseModel, BaseMixin
+
+from SwissArmyTransformer.mpu.utils import split_tensor_along_last_dim
+from SwissArmyTransformer.model.transformer import unscaled_init_method
+from SwissArmyTransformer.mpu import ColumnParallelLinear, RowParallelLinear
+import torch.nn.functional as F
+from deepspeed.runtime.activation_checkpointing.checkpointing import get_cuda_rng_tracker
+import math
+
+
+class PositionEmbeddingMixin(BaseMixin):
+    def __init__(self, additional_sequence_length, hidden_size,
+                 init_method_std=0.02, reinit_slice=slice(512, 912), 
+                 ):
+        super(PositionEmbeddingMixin, self).__init__()
+        self.reinit_slice = reinit_slice
+        self.position_embeddings = torch.nn.Embedding(additional_sequence_length, hidden_size)
+        torch.nn.init.normal_(self.position_embeddings.weight, mean=0.0, std=init_method_std)
+
+    def reinit(self, parent_model=None):
+        old_weights = self.transformer.position_embeddings.weight.data[self.reinit_slice]
+        old_len, hidden_size = old_weights.shape
+        assert hidden_size == self.position_embeddings.weight.shape[-1]
+        self.position_embeddings.weight.data.view(-1, old_len, hidden_size).copy_(old_weights)
+        
+        
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, framenum, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, frame_num, window_size, window_size, C)
+    """
+    B, framenum, H, W, C = x.shape
+    x = x.view(B, framenum, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 2, 4, 1, 3, 5, 6).contiguous().view(-1, framenum, window_size, window_size, C)
+    return windows
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, frame_num, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, frame_num, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    framenum = windows.shape[1]
+    x = windows.view(B, H // window_size, W // window_size, framenum, window_size, window_size, -1)
+    x = x.permute(0, 3, 1, 4, 2, 5, 6).contiguous().view(B, framenum, H, W, -1)
+    return x
+
+class WindowAttentionMixin(BaseMixin):
+    def __init__(self, num_layers,
+                hidden_size, 
+                frame_resolution,
+                window_size,
+                shift_size,
+                n_head,
+                frame_num, 
+                init_method=unscaled_init_method(0.02),
+                output_layer_init_method=unscaled_init_method(0.02),
+                time_dim_attend_length=0
+        ):
+        super(WindowAttentionMixin, self).__init__()
+        self.num_layers = num_layers # replace attention in the LAST n layers
+        self.query_key_value = torch.nn.ModuleList(
+            [ColumnParallelLinear(hidden_size, 3*hidden_size,stride=3,
+                gather_output=False,init_method=init_method)
+                for layer_id in range(num_layers)
+            ])
+        self.dense = torch.nn.ModuleList(
+            [RowParallelLinear(
+                hidden_size,
+                hidden_size,
+                input_is_parallel=True,
+                init_method=output_layer_init_method,
+                bias=True,
+                module=self,
+                name="dense")
+                for layer_id in range(num_layers)
+            ])
+
+        self.n_head = n_head
+        self.window_size = window_size
+        self.frame_resolution = frame_resolution
+        self.frame_len = frame_resolution * frame_resolution
+        self.time_dim_attend_length = time_dim_attend_length
+        assert frame_resolution % window_size == 0
+        assert 0 < shift_size < window_size
+        nW = (self.frame_resolution // self.window_size) ** 2
+        ws_squre = self.window_size * self.window_size
+        
+        # odd non-shift, even shift
+        img_mask = torch.zeros((1, 1, frame_resolution, frame_resolution, 1))
+        h_slices = (slice(0, -shift_size),
+                    slice(-shift_size, None))
+        w_slices = (slice(0, -shift_size),
+                    slice(-shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, :, h, w, :] = cnt
+                cnt += 1
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, 1, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        sub_attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) #[nW, self.window_size * self.window_size, self.window_size * self.window_size]
+        sub_attn_mask = sub_attn_mask.masked_fill(sub_attn_mask != 0, float(0.0)).masked_fill(sub_attn_mask == 0, float(1.00))
+        attn_mask = sub_attn_mask.repeat(1, frame_num, frame_num)
+        attn_mask = attn_mask.tril()
+        
+        causal_mask = torch.ones(ws_squre*frame_num, ws_squre*frame_num)
+        causal_mask = causal_mask.tril()
+
+        self.shift_sizes = [0, shift_size]
+        self.attn_mask = attn_mask
+        self.causal_mask = causal_mask
+        self.mask_initialized = False
+        
+        self.attn_distribution = torch.nn.ParameterList([
+            torch.nn.Parameter(torch.zeros(hidden_size))
+            for _ in range(num_layers)
+        ])
+        
+    def reinit(self, *pre_mixins):
+        start_layer = len(self.transformer.layers) - self.num_layers
+        assert start_layer >= 0
+        for layer_id in range(self.num_layers):
+            old_attention = self.transformer.layers[start_layer + layer_id].attention
+            self.query_key_value[layer_id].weight.data.copy_(old_attention.query_key_value.weight.data)
+            self.query_key_value[layer_id].bias.data.copy_(old_attention.query_key_value.bias.data)
+            
+    def attention_extra_NAR_inference(self, frame_hidden_state, layer_id, attn_dropout=None, memkv_text=None, stage=1):
+        # frame_hidden_state [batchsize, frame_num*frame_size, n_head*hiddensize_perhead]
+        if not self.mask_initialized:
+            self.attn_mask = self.attn_mask.to(device=frame_hidden_state.device, dtype=frame_hidden_state.dtype)
+            self.causal_mask = self.causal_mask.to(device=frame_hidden_state.device, dtype=frame_hidden_state.dtype)
+            self.mask_initialized = True
+        b0, s1, h0 = frame_hidden_state.shape 
+        h = h0 // self.n_head
+        frame_len = self.frame_resolution * self.frame_resolution
+        frame_num = s1 // frame_len
+        if stage == 2:
+            assert frame_num == 3
+        assert frame_num*frame_len == s1
+        wind_square = self.window_size * self.window_size
+        nW = frame_len // wind_square
+        bswin = b0 * nW
+        
+        if memkv_text is not None:
+            s0 = memkv_text.shape[-2]
+            k_text = memkv_text[..., :h0].expand(b0, -1, -1).reshape(b0, s0, self.n_head, h).permute(0, 2, 1, 3)
+            v_text = memkv_text[..., h0:].expand(b0, -1, -1).reshape(b0, s0, self.n_head, h).permute(0, 2, 1, 3)
+        
+        # shift
+        frame_hidden_state = frame_hidden_state.reshape(b0, frame_num, self.frame_resolution, self.frame_resolution, h0)
+        if self.shift_sizes[layer_id%2] > 0:
+            frame_hidden_state = torch.roll(frame_hidden_state, shifts=(-self.shift_sizes[layer_id%2], -self.shift_sizes[layer_id%2]), dims=(2,3))
+        # window partition    
+        frame_hidden_state = window_partition(frame_hidden_state, self.window_size).reshape(bswin, frame_num*wind_square, h0)
+        qkv = self.query_key_value[layer_id](frame_hidden_state).reshape(bswin, frame_num*wind_square, 3, self.n_head, h)\
+                .permute(2, 0, 3, 1, 4) #[3, bswin, n_head, frame_num*wind_size*wind_size, h]
+        q, k, v = qkv[0], qkv[1], qkv[2] 
+        attn = torch.matmul(q / math.sqrt(h), k.transpose(-1, -2))
+        
+        if stage == 1:
+            if self.shift_sizes[layer_id%2] > 0:
+                attn = torch.mul(attn.view(bswin // nW, nW, self.n_head, frame_num*wind_square, frame_num*wind_square),
+                                self.attn_mask[:,:frame_num*wind_square, :frame_num*wind_square].unsqueeze(1).unsqueeze(0))\
+                    - 10000.0 * (1.0 - self.attn_mask[:,:frame_num*wind_square, :frame_num*wind_square].unsqueeze(1).unsqueeze(0))
+                attn = attn.view(bswin, self.n_head, frame_num*wind_square, frame_num*wind_square)
+            else:
+                attn = torch.mul(attn, self.causal_mask[:frame_num*wind_square, :frame_num*wind_square].unsqueeze(0).unsqueeze(0))\
+                    - 10000.0 * (1.0 - self.causal_mask[:frame_num*wind_square, :frame_num*wind_square].unsqueeze(0).unsqueeze(0))
+        
+        if memkv_text is None:
+            attn = F.softmax(attn, dim=-1)
+            if attn_dropout is not None:
+                with get_cuda_rng_tracker().fork():
+                    attn = attn_dropout(attn)
+            context_swin = torch.matmul(attn, v).permute(0, 2, 1, 3).reshape(bswin, frame_num, self.window_size, self.window_size, h0)
+        else:
+            attn_frame2text = torch.matmul(q.reshape(b0, -1, self.n_head, frame_num*wind_square, h) / math.sqrt(h), k_text.unsqueeze(1).transpose(-1, -2))
+            attn_frame2text = attn_frame2text.reshape(bswin, self.n_head, frame_num*wind_square, s0)
+            attn = torch.cat((attn, attn_frame2text), dim=-1)
+            attn = F.softmax(attn, dim=-1)
+            
+            if attn_dropout is not None:
+                with get_cuda_rng_tracker().fork():
+                    attn = attn_dropout(attn)
+                    
+            context_swin = (torch.matmul(attn[..., :-s0], v) + 
+                            torch.matmul(attn[..., -s0:].reshape(b0, -1, self.n_head,frame_num*wind_square, s0), v_text.unsqueeze(1))\
+                                .reshape(bswin, self.n_head, frame_num*wind_square, h))\
+                .permute(0, 2, 1, 3).reshape(bswin, frame_num, self.window_size, self.window_size, h0)
+                
+        context_swin = window_reverse(context_swin, self.window_size, self.frame_resolution, self.frame_resolution)
+        
+        # reverse cycle shift
+        if self.shift_sizes[layer_id%2] > 0:
+            context_swin = torch.roll(context_swin, shifts=(self.shift_sizes[layer_id%2], self.shift_sizes[layer_id%2]), dims=(2,3))
+        ret_context = context_swin.reshape(b0, s1, h0)
+    
+        # for mem
+        memk = k.permute(0, 2, 1, 3).reshape(bswin, frame_num, self.window_size, self.window_size, h0)
+        memv = v.permute(0, 2, 1, 3).reshape(bswin, frame_num, self.window_size, self.window_size, h0)
+        memk = window_reverse(memk, self.window_size, self.frame_resolution, self.frame_resolution)
+        memv = window_reverse(memv, self.window_size, self.frame_resolution, self.frame_resolution)
+        if self.shift_sizes[layer_id%2] > 0:
+            memk = torch.roll(memk, shifts=(self.shift_sizes[layer_id%2], self.shift_sizes[layer_id%2]), dims=(2,3))
+            memv = torch.roll(memv, shifts=(self.shift_sizes[layer_id%2], self.shift_sizes[layer_id%2]), dims=(2,3))
+        memk, memv = memk.reshape(b0, s1, h0), memv.reshape(b0, s1, h0)
+                
+        ret_mem = torch.cat((memk, memv), dim=-1)
+        return ret_context, ret_mem
+        
+    def attention_extra_AR_inference(self, frame_hidden_state, memkv, pos, layer_id, log_text_attention_weights=0, attn_dropout=None, memkv_text=None, stage=1):
+        # frame_hidden_state [batchsize, 1, n_head*hiddensize_perhead]
+        # memkv [batchsize, pos, hidden_size*2] (include frames only)
+        # if memkv_text is not None: will attend to text
+        # pos: token's pos
+        b0, sin, h0 = frame_hidden_state.shape 
+        h = h0 // self.n_head
+        assert sin == 1
+        this_qkv = self.query_key_value[layer_id](frame_hidden_state)
+        thisq, thisk, thisv = this_qkv[..., :h0], this_qkv[..., h0:2*h0], this_qkv[..., 2*h0:]
+        s1 = memkv.shape[1] if memkv is not None else 0
+        frame_len = self.frame_resolution * self.frame_resolution
+        frame_num_before = s1 // frame_len
+        
+        
+        if memkv is not None:
+            pos_inframe = pos - frame_num_before * frame_len
+            
+            xpos = pos_inframe // self.frame_resolution # pos = xpos*self.frame_resolution + ypos
+            ypos = pos_inframe % self.frame_resolution
+            # [start, end)
+            if self.shift_sizes[layer_id%2] > 0:
+                xstart = ((xpos+self.shift_sizes[layer_id%2]) // self.window_size) * self.window_size - self.shift_sizes[layer_id%2]
+                ystart = ((ypos+self.shift_sizes[layer_id%2]) // self.window_size) * self.window_size - self.shift_sizes[layer_id%2]
+                xend = xstart + self.window_size
+                yend = ystart + self.window_size
+                xstart, ystart = max(0, xstart), max(0, ystart)
+                xend, yend = min(xend, self.frame_resolution), min(yend, self.frame_resolution)
+            else:
+                xstart = (xpos // self.window_size) * self.window_size
+                ystart = (ypos // self.window_size) * self.window_size
+                xend, yend = xstart + self.window_size, ystart+self.window_size
+            
+            # select index 
+            selected_index = list()
+            if frame_num_before > 0:
+                # frames before
+                frame_attended_start = max(0, frame_num_before-self.time_dim_attend_length+1) if self.time_dim_attend_length > 0 else 0
+                for x in range(xstart, xend):
+                    for y in range(ystart, yend):
+                        selected_index.append(x*self.frame_resolution+y+frame_len*frame_attended_start)
+                cnt_per_frame = len(selected_index)
+                for _ in range((frame_num_before-frame_attended_start-1)*cnt_per_frame):
+                    selected_index.append(selected_index[-cnt_per_frame]+frame_len)
+                    
+            # the last frame
+            for x in range(xstart, xend):
+                for y in range(ystart, yend):
+                    tmppos = x*self.frame_resolution+y + frame_num_before * frame_len
+                    if tmppos < pos:
+                        selected_index.append(tmppos)
+                    else:
+                        break
+            cnt_all = len(selected_index)+1
+            selected_index = torch.tensor(selected_index, device=memkv.device)
+            used_memkv = torch.index_select(memkv, 1, selected_index)
+            used_k, used_v = used_memkv[..., :h0], used_memkv[..., h0:]
+            used_k = torch.cat((used_k.expand(thisk.shape[0], -1, -1), thisk), dim=-2)
+            used_v = torch.cat((used_v.expand(thisv.shape[0], -1, -1), thisv), dim=-2)
+            if memkv_text is not None:
+                cnt_all += memkv_text.shape[-2]
+                used_k = torch.cat((memkv_text[..., :h0].expand(thisk.shape[0], -1, -1), used_k), dim=-2)
+                used_v = torch.cat((memkv_text[..., h0:].expand(thisv.shape[0], -1, -1), used_v), dim=-2)
+            used_k = used_k.reshape(b0, cnt_all, self.n_head, h).permute(0, 2, 1, 3)
+            used_v = used_v.reshape(b0, cnt_all, self.n_head, h).permute(0, 2, 1, 3)
+        else: 
+            used_k = thisk
+            used_v = thisv
+            
+            if memkv_text is not None:
+                used_k = torch.cat((memkv_text[..., :h0].expand(thisk.shape[0], -1, -1), used_k), dim=-2)
+                used_v = torch.cat((memkv_text[..., h0:].expand(thisv.shape[0], -1, -1), used_v), dim=-2)
+                used_k = used_k.reshape(b0, 1+memkv_text.shape[-2], self.n_head, h).permute(0, 2, 1, 3)
+                used_v = used_v.reshape(b0, 1+memkv_text.shape[-2], self.n_head, h).permute(0, 2, 1, 3)
+            else:
+                used_k = used_k.reshape(b0, 1, self.n_head, h).permute(0, 2, 1, 3)
+                used_v = used_v.reshape(b0, 1, self.n_head, h).permute(0, 2, 1, 3)
+        
+        thisq = thisq.reshape(b0, 1, self.n_head, h).permute(0, 2, 1, 3) # [b0, n_head, 1, h]
+        attn = torch.matmul(thisq / math.sqrt(h), used_k.transpose(-1, -2))
+        if memkv_text is not None:
+            attn[..., :memkv_text.shape[-2]] += log_text_attention_weights
+        attn = F.softmax(attn, dim=-1)
+        context_swin = torch.matmul(attn, used_v).permute(0, 2, 1, 3).reshape(b0, 1, h0)
+
+        return context_swin, this_qkv[..., h0:]
+        
+class FullAttentionMixin(BaseMixin):
+    def __init__(self, num_layers,
+                hidden_size, 
+                frame_resolution,
+                n_head,
+                frame_num, 
+                init_method=unscaled_init_method(0.02),
+                output_layer_init_method=unscaled_init_method(0.02),
+                **kwargs,
+        ):
+        super(FullAttentionMixin, self).__init__()
+        self.num_layers = num_layers # replace attention in the LAST n layers
+        self.query_key_value = torch.nn.ModuleList(
+            [ColumnParallelLinear(hidden_size, 3*hidden_size,stride=3,
+                gather_output=False,init_method=init_method)
+                for layer_id in range(num_layers)
+            ])
+        self.dense = torch.nn.ModuleList(
+            [RowParallelLinear(
+                hidden_size,
+                hidden_size,
+                input_is_parallel=True,
+                init_method=output_layer_init_method,
+                bias=True,
+                module=self,
+                name="dense")
+                for layer_id in range(num_layers)
+            ])
+
+        self.n_head = n_head
+        self.frame_resolution = frame_resolution
+        self.frame_len = frame_resolution * frame_resolution
+        
+        self.attn_distribution = torch.nn.ParameterList([
+            torch.nn.Parameter(torch.zeros(hidden_size))
+            for _ in range(num_layers)
+        ])
+        
+    def reinit(self, *pre_mixins):
+        start_layer = len(self.transformer.layers) - self.num_layers
+        assert start_layer >= 0
+        for layer_id in range(self.num_layers):
+            old_attention = self.transformer.layers[start_layer + layer_id].attention
+            self.query_key_value[layer_id].weight.data.copy_(old_attention.query_key_value.weight.data)
+            self.query_key_value[layer_id].bias.data.copy_(old_attention.query_key_value.bias.data)
+            
+            
+    def attention_extra_NAR_inference(self, frame_hidden_state, layer_id, attn_dropout=None, memkv_text=None, stage=1):
+        # frame_hidden_state [batchsize, frame_num*frame_size, n_head*hiddensize_perhead]
+        assert stage == 1
+        
+        b0, s1, h0 = frame_hidden_state.shape 
+        h = h0 // self.n_head
+        frame_len = self.frame_resolution * self.frame_resolution
+        frame_num = s1 // frame_len
+        assert frame_num*frame_len == s1
+
+        if memkv_text is not None:
+            s0 = memkv_text.shape[-2]
+            k_text = memkv_text[..., :h0].expand(b0, -1, -1).reshape(b0, s0, self.n_head, h).permute(0, 2, 1, 3)
+            v_text = memkv_text[..., h0:].expand(b0, -1, -1).reshape(b0, s0, self.n_head, h).permute(0, 2, 1, 3)
+        qkv = self.query_key_value[layer_id](frame_hidden_state).reshape(b0, s1, 3, self.n_head, h)\
+                .permute(2, 0, 3, 1, 4) #[3, b0, n_head, s1, h]
+        q, k, v = qkv[0], qkv[1], qkv[2] 
+        attn = torch.matmul(q / math.sqrt(h), k.transpose(-1, -2))
+        attn = attn - 10000.0 * (1.0-torch.ones(b0, self.n_head, s1, s1, device=attn.device, dtype=attn.dtype).tril())
+                
+        if memkv_text is None:
+            attn = F.softmax(attn, dim=-1)
+            if attn_dropout is not None:
+                with get_cuda_rng_tracker().fork():
+                    attn = attn_dropout(attn)
+            context_swin = torch.matmul(attn, v).permute(0, 2, 1, 3).reshape(b0, s1, h0)
+        else:
+            attn_frame2text = torch.matmul(q / math.sqrt(h), k_text.transpose(-1, -2)) #[b0, s1, s0]
+            attn = torch.cat((attn, attn_frame2text), dim=-1)
+            attn = F.softmax(attn, dim=-1)
+            if attn_dropout is not None:
+                with get_cuda_rng_tracker().fork():
+                    attn = attn_dropout(attn)
+            context_swin = (torch.matmul(attn[..., :-s0], v) + torch.matmul(attn[..., -s0:], v_text))\
+                .permute(0, 2, 1, 3).reshape(b0, s1, h0)
+                    
+        # for mem
+        memk = k.permute(0, 2, 1, 3).reshape(b0, s1, h0)
+        memv = v.permute(0, 2, 1, 3).reshape(b0, s1, h0)
+        ret_mem = torch.cat((memk, memv), dim=-1)
+        
+        return context_swin, ret_mem
+        
+    def attention_extra_AR_inference(self, frame_hidden_state, memkv, pos, layer_id, log_text_attention_weights=0, attn_dropout=None, memkv_text=None, stage=1):
+        # pos: current token's pos
+        b0, sin, h0 = frame_hidden_state.shape 
+        h = h0 // self.n_head
+        assert sin == 1
+        assert stage == 1
+        
+        this_qkv = self.query_key_value[layer_id](frame_hidden_state)
+        thisq, thisk, thisv = this_qkv[..., :h0], this_qkv[..., h0:2*h0], this_qkv[..., 2*h0:]
+        
+        if memkv is not None:
+            used_k, used_v = memkv[..., :h0], memkv[..., h0:]
+            used_k = torch.cat((used_k.expand(thisk.shape[0], -1, -1), thisk), dim=-2)
+            used_v = torch.cat((used_v.expand(thisv.shape[0], -1, -1), thisv), dim=-2)
+        else: 
+            used_k, used_v = thisk, thisv
+            
+        if memkv_text is not None:
+            used_k = torch.cat((memkv_text[..., :h0].expand(thisk.shape[0], -1, -1), used_k), dim=-2)
+            used_v = torch.cat((memkv_text[..., h0:].expand(thisv.shape[0], -1, -1), used_v), dim=-2)
+        
+        used_k = used_k.reshape(b0, -1, self.n_head, h).permute(0, 2, 1, 3)
+        used_v = used_v.reshape(b0, -1, self.n_head, h).permute(0, 2, 1, 3)
+        thisq = thisq.reshape(b0, 1, self.n_head, h).permute(0, 2, 1, 3) # [b0, n_head, 1, h]
+        attn = torch.matmul(thisq / math.sqrt(h), used_k.transpose(-1, -2))
+        if memkv_text is not None:
+            attn[..., :memkv_text.shape[-2]] += log_text_attention_weights
+        attn = F.softmax(attn, dim=-1)
+        
+        context_swin = torch.matmul(attn, used_v).permute(0, 2, 1, 3).reshape(b0, 1, h0)
+
+        return context_swin, this_qkv[..., h0:]
+        
+
+def attention_localframe_and_text_NAR(q0, k0, v0, attention_mask, 
+                             n_head, text_len, frame_len, frame_num, 
+                             attention_dropout=None, log_text_attention_weights=0, stage=1, **kwargs):
+    b, s0, h0 = q0.shape
+    s1 = s0 - text_len
+    h = h0 // n_head
+    assert q0.shape[1] == v0.shape[1] == k0.shape[1] == text_len+frame_len*frame_num
+    # attention_mask.shape [4, b or 1, 1, text_len+frame_len, text_len+frame_len]
+    if stage == 2:
+        assert frame_num == 3
+
+    q0 = q0.reshape(b, s0, n_head, h).permute(0, 2, 1, 3)
+    v0 = v0.reshape(b, s0, n_head, h).permute(0, 2, 1, 3)
+    k0 = k0.reshape(b, s0, n_head, h).permute(0, 2, 1, 3)
+    k0T = k0.transpose(-1, -2)
+
+    score_any2text = torch.matmul(q0 / math.sqrt(q0.shape[-1]), k0T[..., :text_len])
+    score_any2text += log_text_attention_weights
+    score_any2text_part1 = torch.mul(score_any2text[..., :text_len, :], attention_mask[..., :text_len, :text_len]) \
+        - 10000.0 * (1.0 - attention_mask[..., :text_len, :text_len])
+    # context for text
+    attention_probs_text = F.softmax(score_any2text_part1, dim=-1)
+    if attention_dropout is not None:
+        with get_cuda_rng_tracker().fork():
+            attention_probs_text = attention_dropout(attention_probs_text)
+    context_text2text = torch.matmul(attention_probs_text, v0[..., :text_len, :])
+    context_text2text = context_text2text.transpose(1, 2).reshape(b, text_len, h0)
+    
+    if frame_num > 0:
+        score_any2text_part2 = score_any2text[..., text_len:, :]
+        
+        # score: frame local
+        q0_frame = q0[:, :, text_len:].reshape(b, n_head, frame_num, frame_len, h)
+        v0_frame = v0[:, :, text_len:].reshape(b, n_head, frame_num, frame_len, h)
+        k0T_frame = k0[:, :, text_len:].reshape(b, n_head, frame_num, frame_len, h).transpose(-1, -2)
+        score_frame_local0 = torch.matmul(q0_frame / math.sqrt(q0_frame.shape[-1]), k0T_frame)
+        if stage == 1:
+            score_frame_local0 = torch.mul(score_frame_local0, attention_mask[..., text_len:, text_len:].unsqueeze(1)) \
+                - 10000.0 * (1.0 - attention_mask[..., text_len:, text_len:].unsqueeze(1))
+        
+        # context for frame
+        score_frame_all = torch.cat((score_any2text_part2, 
+                                    score_frame_local0.view(b, n_head, s1, frame_len)), dim=-1)
+        attention_probs_frame = F.softmax(score_frame_all, dim=-1)
+        if attention_dropout is not None:
+            with get_cuda_rng_tracker().fork():
+                attention_probs_frame = attention_dropout(attention_probs_frame)
+        context_frame2text = torch.matmul(attention_probs_frame[..., :text_len], v0[..., :text_len, :]) # [b, n_head, s1, h]
+        context_frame_local0 = torch.matmul(attention_probs_frame[..., text_len:text_len+frame_len].\
+            view(b, n_head, frame_num, frame_len, frame_len), v0_frame).view(b, n_head, s1, h)
+        
+        context_frame = (context_frame2text + context_frame_local0).transpose(1, 2).reshape(b, s1, h0)
+    else: 
+        context_frame = None
+
+    return context_text2text, context_frame
+    
+def attention_localframe_and_text_AR(q0, k0, v0, n_head, text_len, frame_len, frame_num, 
+                                        attention_dropout=None, log_text_attention_weights=0, layer_id=None, limited_spatial_channel_mem=False, stage=1, **kwargs):
+    # limited_spatial_channel_mem=True means: mems in spatial channel is consisted of {mem_text, mem_current_frame}
+    b, s0, h0 = k0.shape
+    frame_num_before = (s0-text_len-1) // frame_len # frame_num == frame_num_before or frame_num == frame_num_before+1
+    h = h0 // n_head
+    assert q0.shape[1] == 1
+    assert v0.shape[1] == k0.shape[1]
+
+    q0 = q0.reshape(b, 1, n_head, h).permute(0, 2, 1, 3)
+    v0 = v0.reshape(b, s0, n_head, h).permute(0, 2, 1, 3)
+    k0T = k0.reshape(b, s0, n_head, h).permute(0, 2, 3, 1)
+    
+    if limited_spatial_channel_mem:
+        assert frame_num_before == 0
+        assert stage == 1 # not implemented for stage-2 yet
+        score = torch.matmul(q0 / math.sqrt(q0.shape[-1]), k0T)
+        score[..., :text_len] += log_text_attention_weights
+        attention_probs_frame = F.softmax(score, dim=-1)
+        context_frame = torch.matmul(attention_probs_frame, v0).transpose(1, 2).reshape(b, 1, h0)
+        
+    else:
+        score_token2text = torch.matmul(q0 / math.sqrt(q0.shape[-1]), k0T[..., :text_len])
+        score_token2text += log_text_attention_weights
+        score_frame_local0 = torch.matmul(q0 / math.sqrt(q0.shape[-1]), k0T[..., text_len+frame_num_before*frame_len:])
+        score_frame_all = torch.cat((score_token2text, 
+                                    score_frame_local0), dim=-1)
+        attention_probs_frame = F.softmax(score_frame_all, dim=-1)
+        
+        context_token2text = torch.matmul(attention_probs_frame[..., :text_len], v0[..., :text_len, :]) # [b, n_head, s1, h]
+        context_frame_local0 = torch.matmul(attention_probs_frame[..., text_len:], \
+            v0[:, :, text_len+frame_num_before*frame_len:, :])
+        context_frame = (context_token2text + context_frame_local0).transpose(1, 2).reshape(b, 1, h0)
+    
+    return context_frame
+
+    
+class CogVideoCacheModel(BaseModel):
+    def __init__(self, args, transformer=None, parallel_output=True, window_size=None, cogvideo_stage=None):
+        super().__init__(args, transformer=transformer, parallel_output=parallel_output)
+        self.layout = args.layout # [64, 64+1024, 64+6*1024]
+        self.stage = cogvideo_stage if cogvideo_stage is not None else args.cogvideo_stage # 1 or 2
+        self.n_head = args.num_attention_heads
+        self.window_size = window_size if window_size is not None else args.window_size
+        
+        frame_resolution = int(math.sqrt(self.layout[1]-self.layout[0]))
+        self.add_mixin('extra_position_embedding', PositionEmbeddingMixin(
+            args.additional_seqlen, args.hidden_size
+        ))
+        
+        if self.stage == 1:
+            self.add_mixin('attention_plus', FullAttentionMixin(
+                num_layers=args.num_layers,
+                hidden_size=args.hidden_size,
+                frame_resolution=frame_resolution,
+                n_head=args.num_attention_heads,
+                frame_num=(args.layout[2]-args.layout[0])//(args.layout[1]-args.layout[0]),
+            ))
+        else: 
+            self.add_mixin('attention_plus', WindowAttentionMixin(
+                num_layers=args.num_layers,
+                hidden_size=args.hidden_size,
+                frame_resolution=frame_resolution,
+                window_size=self.window_size,
+                shift_size=self.window_size//2,
+                n_head=args.num_attention_heads,
+                frame_num=(args.layout[2]-args.layout[0])//(args.layout[1]-args.layout[0]),
+            ))
+
+        
+    @classmethod
+    def add_model_specific_args(cls, parser):
+        group = parser.add_argument_group('VideoSwinLocalModel', 'video swin local model configurations')
+        group.add_argument("--layout", type=str, default='64, 464, 2064') 
+        group.add_argument("--window-size", type=int, default=10) # 优先级在直接参数赋值之后
+        group.add_argument("--additional-seqlen", type=int, default=2000)
+        group.add_argument("--cogvideo-stage", type=int, default=1, choices=[1,2])  # 优先级在直接参数赋值之后
+        return parser
+    
+    def disable_untrainable_params(self):
+        pass
+    
+    def position_embedding_forward(self, position_ids, **kw_args):
+        if position_ids.shape[-1] > 1:
+            if self.stage == 1:
+                if position_ids[0,-1] >= (512+400):
+                    frame_num = position_ids.shape[-1] // 400
+                    position_embeddings = torch.cat(
+                    (
+                        self.transformer.position_embeddings(position_ids[..., :-400*(frame_num-1)]),
+                        self.get_mixin('extra_position_embedding').position_embeddings(position_ids[..., -400*(frame_num-1):]-(512+400))
+                    ),
+                    dim=-2
+                )
+                else:
+                    position_embeddings = self.transformer.position_embeddings(position_ids)
+            else:
+                # given 3, interpolate 2
+                position_embeddings = torch.cat(
+                    (
+                        self.transformer.position_embeddings(position_ids[..., :-800]),
+                        self.get_mixin('extra_position_embedding').position_embeddings(position_ids[..., -800:]-(512+400))
+                    ),
+                    dim=-2
+                )
+        else:
+            if position_ids[0, 0] >= (512+400):
+                position_embeddings = self.get_mixin('extra_position_embedding').position_embeddings(position_ids-(512+400))
+            else:
+                position_embeddings = self.transformer.position_embeddings(position_ids)
+        return position_embeddings
+        
+    def attention_forward(self, hidden_states, mask, layer_id, mems=None, log_text_attention_weights=0, text_len=0, frame_len=0, counter=0, enforce_no_swin=False, limited_spatial_channel_mem=False, **kw_args):
+        attn_module = self.transformer.layers[layer_id].attention
+        hidden_size = hidden_states.shape[-1]
+        
+        # base model qkv
+        if mems is None:
+            mixed_raw_layer = attn_module.query_key_value(hidden_states)
+            q0, k0, v0 = split_tensor_along_last_dim(mixed_raw_layer, 3)
+            assert (q0.shape[1]-text_len) % frame_len == 0
+            memkv0 = torch.cat((k0, v0), dim=-1)
+            context_text, context_frame_local_text = attention_localframe_and_text_NAR(
+                    q0, k0, v0,
+                    mask,
+                    n_head=attn_module.num_attention_heads_per_partition,
+                    text_len=text_len,
+                    frame_len=frame_len,
+                    frame_num=(q0.shape[1]-text_len)//frame_len,
+                    log_text_attention_weights=log_text_attention_weights,
+                    stage=self.stage
+                )
+
+            # change: self.swin_attend_to_text默认为True:
+            memkv1_text = self.get_mixin('attention_plus').query_key_value[layer_id](hidden_states[..., :text_len, :])[..., hidden_size:]
+            output_text = attn_module.dense(context_text)
+                
+            if (q0.shape[1]-text_len)//frame_len > 0:
+                assert (q0.shape[1]-text_len) % frame_len == 0
+                context_frame_swin, memkv1_frame = self.get_mixin('attention_plus').attention_extra_NAR_inference(
+                    hidden_states[:,text_len:], layer_id, memkv_text=memkv1_text, stage=self.stage)
+                if not enforce_no_swin:
+                    attn_distrib = torch.sigmoid(self.get_mixin('attention_plus').attn_distribution[layer_id])
+                    attn_distrib = attn_distrib.unsqueeze(0).unsqueeze(0)
+                    output_frame = torch.mul(attn_module.dense(context_frame_local_text), attn_distrib)\
+                        +torch.mul(self.get_mixin('attention_plus').dense[layer_id](context_frame_swin), 1-attn_distrib)
+                else:
+                    output_frame = attn_module.dense(context_frame_local_text[..., :frame_len, :])
+                output = torch.cat((output_text, output_frame), dim=-2)
+                memkv1 = torch.cat((memkv1_text, memkv1_frame), dim=-2) if memkv1_text is not None else memkv1_frame
+            else: 
+                output = output_text
+                memkv1 = memkv1_text
+            kw_args['output_this_layer']['mem_kv'] = (memkv0, memkv1)
+            
+            
+        else:
+            mixed_raw_layer = attn_module.query_key_value(hidden_states)
+            q0, k0, v0 = split_tensor_along_last_dim(mixed_raw_layer, 3)
+            new_memkv0 = torch.cat((k0, v0), dim=-1)
+            old_k0, old_v0 = mems[0][layer_id][..., :hidden_size], mems[0][layer_id][..., hidden_size:] 
+            
+            context_frame_local_text = attention_localframe_and_text_AR(
+                    q0, 
+                    torch.cat((old_k0.expand(k0.shape[0], -1, -1), k0), dim=-2),
+                    torch.cat((old_v0.expand(v0.shape[0], -1, -1), v0), dim=-2),
+                    n_head=attn_module.num_attention_heads_per_partition,
+                    text_len=text_len,
+                    frame_len=frame_len,
+                    frame_num=None,
+                    log_text_attention_weights=log_text_attention_weights,
+                    layer_id=layer_id,
+                    limited_spatial_channel_mem=limited_spatial_channel_mem,
+                )
+            
+            old_memkv1 = mems[1][layer_id] if mems[1] is not None else None
+            
+            context_frame_swin, new_memkv1 = self.get_mixin('attention_plus').attention_extra_AR_inference(hidden_states, 
+                                                                                                            old_memkv1[..., text_len:, :] if old_memkv1.shape[-2]>text_len else None,
+                                                                                                            counter-text_len,
+                                                                                                            layer_id, 
+                                                                                                            memkv_text=old_memkv1[..., :text_len, :],
+                                                                                                            log_text_attention_weights=log_text_attention_weights)
+            if not enforce_no_swin:
+                attn_distrib = torch.sigmoid(self.get_mixin('attention_plus').attn_distribution[layer_id])
+                attn_distrib = attn_distrib.unsqueeze(0).unsqueeze(0)
+                output = torch.mul(attn_module.dense(context_frame_local_text), attn_distrib)\
+                    +torch.mul(self.get_mixin('attention_plus').dense[layer_id](context_frame_swin), 1-attn_distrib)
+            else: 
+                output = attn_module.dense(context_frame_local_text)
+                
+            kw_args['output_this_layer']['mem_kv'] = (new_memkv0, new_memkv1)
+            
+        return output

models/cogvideo_model.py

@@ -0,0 +1,543 @@
+# -*- encoding: utf-8 -*-
+'''
+@File    :   cogvideo_model.py
+@Time    :   2022/07/11 16:12:05
+@Author  :   Wenyi Hong 
+@Version :   1.0
+@Contact :   hwy22@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+
+import torch
+from SwissArmyTransformer.model.base_model import BaseModel, BaseMixin
+
+from SwissArmyTransformer.mpu.utils import split_tensor_along_last_dim
+from SwissArmyTransformer.model.transformer import unscaled_init_method
+from SwissArmyTransformer.mpu import ColumnParallelLinear, RowParallelLinear
+import torch.nn.functional as F
+from deepspeed.runtime.activation_checkpointing.checkpointing import get_cuda_rng_tracker
+import math
+
+class PositionEmbeddingMixin(BaseMixin):
+    def __init__(self, additional_sequence_length, hidden_size,
+                 init_method_std=0.02, reinit_slice=slice(512, 912), 
+                 ):
+        super(PositionEmbeddingMixin, self).__init__()
+        self.reinit_slice = reinit_slice
+        self.position_embeddings = torch.nn.Embedding(additional_sequence_length, hidden_size)
+        torch.nn.init.normal_(self.position_embeddings.weight, mean=0.0, std=init_method_std)
+
+    def reinit(self, parent_model=None):
+        old_weights = self.transformer.position_embeddings.weight.data[self.reinit_slice]
+        old_len, hidden_size = old_weights.shape
+        assert hidden_size == self.position_embeddings.weight.shape[-1]
+        self.position_embeddings.weight.data.view(-1, old_len, hidden_size).copy_(old_weights)
+        
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, framenum, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, frame_num, window_size, window_size, C)
+    """
+    B, framenum, H, W, C = x.shape
+    x = x.view(B, framenum, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 2, 4, 1, 3, 5, 6).contiguous().view(-1, framenum, window_size, window_size, C)
+    return windows
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, frame_num, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, frame_num, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    framenum = windows.shape[1]
+    x = windows.view(B, H // window_size, W // window_size, framenum, window_size, window_size, -1)
+    x = x.permute(0, 3, 1, 4, 2, 5, 6).contiguous().view(B, framenum, H, W, -1)
+    return x
+
+class WindowAttentionMixin(BaseMixin):
+    def __init__(self, num_layers,
+                hidden_size, 
+                frame_resolution,
+                window_size,
+                shift_size,
+                n_head,
+                frame_num, 
+                init_method=unscaled_init_method(0.02),
+                output_layer_init_method=unscaled_init_method(0.02),
+        ):
+        super(WindowAttentionMixin, self).__init__()
+        self.num_layers = num_layers # replace attention in the LAST n layers
+        self.query_key_value = torch.nn.ModuleList(
+            [ColumnParallelLinear(hidden_size, 3*hidden_size,stride=3,
+                gather_output=False,init_method=init_method)
+                for layer_id in range(num_layers)
+            ])
+        self.dense = torch.nn.ModuleList(
+            [RowParallelLinear(
+                hidden_size,
+                hidden_size,
+                input_is_parallel=True,
+                init_method=output_layer_init_method,
+                bias=True,
+                module=self,
+                name="dense",
+                ) 
+                for layer_id in range(num_layers)
+            ])
+
+        self.n_head = n_head
+        self.window_size = window_size
+        self.frame_resolution = frame_resolution
+        self.frame_len = frame_resolution * frame_resolution
+        assert frame_resolution % window_size == 0
+        assert 0 < shift_size < window_size
+        nW = (self.frame_resolution // self.window_size) ** 2
+        ws_squre = self.window_size * self.window_size
+        
+        # odd non-shift, even shift
+        img_mask = torch.zeros((1, 1, frame_resolution, frame_resolution, 1))
+        h_slices = (slice(0, -shift_size),
+                    slice(-shift_size, None))
+        w_slices = (slice(0, -shift_size),
+                    slice(-shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, :, h, w, :] = cnt
+                cnt += 1
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, 1, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        sub_attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) #[nW, self.window_size * self.window_size, self.window_size * self.window_size]
+        sub_attn_mask = sub_attn_mask.masked_fill(sub_attn_mask != 0, float(0.0)).masked_fill(sub_attn_mask == 0, float(1.00))
+        attn_mask = sub_attn_mask.repeat(1, frame_num, frame_num)
+        
+        self.attn_mask_sequential = attn_mask.clone().tril()
+        self.causal_mask_sequential = torch.ones(1, ws_squre*frame_num, ws_squre*frame_num).tril()
+        
+        self.causal_mask_interp = torch.ones(1, ws_squre*frame_num, ws_squre*frame_num)
+        self.attn_mask_interp = attn_mask.clone()
+
+        # bi-dir 
+        for bi_idx in range(0, frame_num, 2):
+            for uni_idx in range(1, frame_num, 2):
+                self.attn_mask_interp[:, bi_idx*ws_squre:(bi_idx+1)*ws_squre, uni_idx*ws_squre:(uni_idx+1)*ws_squre] = 0
+                self.causal_mask_interp[:, bi_idx*ws_squre:(bi_idx+1)*ws_squre, uni_idx*ws_squre:(uni_idx+1)*ws_squre] = 0
+        # uni-dir
+        for uni_idx in range(1, frame_num, 2):
+            self.attn_mask_interp[:, ws_squre*uni_idx:ws_squre*(uni_idx+1), ws_squre*uni_idx:ws_squre*(uni_idx+1)].tril_()
+            self.causal_mask_interp[:, ws_squre*uni_idx:ws_squre*(uni_idx+1), ws_squre*uni_idx:ws_squre*(uni_idx+1)].tril_()
+            for uni_idx2 in range(uni_idx+2, frame_num, 2):
+                self.attn_mask_interp[:, ws_squre*uni_idx:ws_squre*(uni_idx+1), ws_squre*uni_idx2:ws_squre*(uni_idx2+1)] = 0
+                self.causal_mask_interp[:, ws_squre*uni_idx:ws_squre*(uni_idx+1), ws_squre*uni_idx2:ws_squre*(uni_idx2+1)] = 0
+
+        # expand dim
+        self.attn_mask_sequential = self.attn_mask_sequential[None, None, :, None]
+        self.attn_mask_interp = self.attn_mask_interp[None, None, :, None]
+        self.causal_mask_sequential = self.causal_mask_sequential[None, None, :, None]
+        self.causal_mask_interp = self.causal_mask_interp[None, None, :, None]
+                
+        self.shift_sizes = [0, shift_size]
+        # self.register_buffer("attn_mask", attn_mask)
+        # self.register_buffer("causal_mask", causal_mask)
+        self.mask_initialized = False
+        
+        self.attn_distribution = torch.nn.ParameterList([
+            torch.nn.Parameter(torch.zeros(hidden_size))
+            for _ in range(num_layers)
+        ])
+        
+    def reinit(self, *pre_mixins):
+        start_layer = len(self.transformer.layers) - self.num_layers
+        assert start_layer >= 0
+        for layer_id in range(self.num_layers):
+            old_attention = self.transformer.layers[start_layer + layer_id].attention
+            self.query_key_value[layer_id].weight.data.copy_(old_attention.query_key_value.weight.data)
+            self.query_key_value[layer_id].bias.data.copy_(old_attention.query_key_value.bias.data)
+            
+    def attention_extra(self, frame_hidden_state, layer_id, attn_dropout, text_hidden_state=None, 
+                       text_attn_mask=None, mode_sequential=True):
+        # pb relax 
+        swin_pb_relax = True
+        alpha = 16
+        
+        # frame_hidden_state [batchsize, frame_num*frame_size, n_head*hiddensize_perhead]
+        if not self.mask_initialized:
+            self.attn_mask_sequential = self.attn_mask_sequential.to(device=frame_hidden_state.device, dtype=frame_hidden_state.dtype)
+            self.causal_mask_sequential = self.causal_mask_sequential.to(device=frame_hidden_state.device, dtype=frame_hidden_state.dtype)
+            self.attn_mask_interp = self.attn_mask_interp.to(device=frame_hidden_state.device, dtype=frame_hidden_state.dtype)
+            self.causal_mask_interp = self.causal_mask_interp.to(device=frame_hidden_state.device, dtype=frame_hidden_state.dtype)
+            self.mask_initialized = True
+        b0, s1, h0 = frame_hidden_state.shape 
+        h = h0 // self.n_head
+        frame_len = self.frame_resolution * self.frame_resolution
+        frame_num = s1 // frame_len
+        assert frame_num*frame_len == s1
+        wind_square = self.window_size * self.window_size
+        nW = frame_len // wind_square
+        bswin = b0 * nW
+        
+        causal_mask = self.causal_mask_sequential if mode_sequential else self.causal_mask_interp
+        attn_mask = self.attn_mask_sequential if mode_sequential else self.attn_mask_interp
+        if text_hidden_state is not None:
+            s0 = text_hidden_state.shape[1]
+            qkv_text = self.query_key_value[layer_id](text_hidden_state).reshape(b0, s0, 3, self.n_head, h).permute(2, 0, 3, 1, 4) #[3, b0, n_head, s0, h]
+            q_text, k_text, v_text = qkv_text[0], qkv_text[1], qkv_text[2]
+            
+        # shift
+        frame_hidden_state = frame_hidden_state.reshape(b0, frame_num, self.frame_resolution, self.frame_resolution, h0)
+        if self.shift_sizes[layer_id%2] > 0:
+            frame_hidden_state = torch.roll(frame_hidden_state, shifts=(-self.shift_sizes[layer_id%2], -self.shift_sizes[layer_id%2]), dims=(2,3))
+        # window partition    
+        frame_hidden_state = window_partition(frame_hidden_state, self.window_size).reshape(bswin, frame_num*wind_square, h0)
+        qkv = self.query_key_value[layer_id](frame_hidden_state).reshape(bswin, frame_num*wind_square, 3, self.n_head, h)\
+                .permute(2, 0, 3, 1, 4) #[3, bswin, n_head, frame_num*wind_size*wind_size, h]
+        q, k, v = qkv[0], qkv[1], qkv[2]
+        
+        # pb-relax
+        if swin_pb_relax:
+            attn = torch.matmul(q / (math.sqrt(h)*alpha), k.transpose(-1, -2))
+        else: 
+            attn = torch.matmul(q / math.sqrt(h), k.transpose(-1, -2))
+
+        if self.shift_sizes[layer_id%2] > 0:
+            # attn = attn.view(bswin // nW, nW, self.n_head, frame_num*wind_square, frame_num*wind_square) + self.attn_mask.unsqueeze(1).unsqueeze(0)
+            attn = torch.mul(attn.view(bswin // nW, nW, self.n_head, frame_num*wind_square, frame_num*wind_square), attn_mask)\
+                 - 10000.0 * (1.0 - attn_mask)
+            attn = attn.view(bswin, self.n_head, frame_num*wind_square, frame_num*wind_square)
+        else:
+            attn = torch.mul(attn.view(bswin // nW, nW, self.n_head, frame_num*wind_square, frame_num*wind_square), causal_mask)\
+                 - 10000.0 * (1.0 - causal_mask)
+            attn = attn.view(bswin, self.n_head, frame_num*wind_square, frame_num*wind_square)
+        if swin_pb_relax:
+            swin_pb_relax_const = torch.max(attn.reshape(bswin, self.n_head, -1), dim=-1, keepdim=True)[0].detach().unsqueeze(-1)
+            attn = (attn - swin_pb_relax_const)*alpha
+            
+        if text_hidden_state is None:
+            attn = F.softmax(attn, dim=-1)
+            if attn_dropout is not None:
+                with get_cuda_rng_tracker().fork():
+                    attn = attn_dropout(attn)
+            context_swin = torch.matmul(attn, v).permute(0, 2, 1, 3).reshape(bswin, frame_num, self.window_size, self.window_size, h0)
+        else:
+            assert text_attn_mask is not None
+            text_attn_mask = text_attn_mask.unsqueeze(2).unsqueeze(2)
+            # pb-relax
+            if swin_pb_relax:
+                attn_frame2text = torch.matmul(q.reshape(b0, -1, self.n_head, frame_num*wind_square, h) / (math.sqrt(h)*alpha), k_text.unsqueeze(1).transpose(-1, -2))
+                attn_frame2text = (attn_frame2text-swin_pb_relax_const.reshape(b0, -1, self.n_head, 1, 1))*alpha
+            else:
+                attn_frame2text = torch.matmul(q.reshape(b0, -1, self.n_head, frame_num*wind_square, h) / math.sqrt(h), k_text.unsqueeze(1).transpose(-1, -2))
+
+            attn_frame2text = torch.mul(text_attn_mask, attn_frame2text) - 10000.0 * (1.0 - text_attn_mask)
+            attn_frame2text = attn_frame2text.reshape(bswin, self.n_head, frame_num*wind_square, s0)
+            attn = torch.cat((attn, attn_frame2text), dim=-1)
+            attn = F.softmax(attn, dim=-1)
+            
+            if attn_dropout is not None:
+                with get_cuda_rng_tracker().fork():
+                    attn = attn_dropout(attn)
+                    
+            context_swin = (torch.matmul(attn[..., :-s0], v) + 
+                            torch.matmul(attn[..., -s0:].reshape(b0, -1, self.n_head,frame_num*wind_square, s0), v_text.unsqueeze(1))\
+                                .reshape(bswin, self.n_head, frame_num*wind_square, h))\
+                .permute(0, 2, 1, 3).reshape(bswin, frame_num, self.window_size, self.window_size, h0)
+                
+        context_swin = window_reverse(context_swin, self.window_size, self.frame_resolution, self.frame_resolution)
+        # reverse cycle shift
+        if self.shift_sizes[layer_id%2] > 0:
+            context_swin = torch.roll(context_swin, shifts=(self.shift_sizes[layer_id%2], self.shift_sizes[layer_id%2]), dims=(2,3))
+        context_swin = context_swin.reshape(b0, s1, h0)
+
+        return context_swin
+    
+
+class FullAttentionMixin(BaseMixin):
+    def __init__(self, num_layers,
+                hidden_size, 
+                frame_resolution,
+                n_head,
+                frame_num, 
+                init_method=unscaled_init_method(0.02),
+                output_layer_init_method=unscaled_init_method(0.02),
+        ):
+        super(FullAttentionMixin, self).__init__()
+        self.num_layers = num_layers # replace attention in the LAST n layers
+        self.query_key_value = torch.nn.ModuleList(
+            [ColumnParallelLinear(hidden_size, 3*hidden_size,stride=3,
+                gather_output=False,init_method=init_method) 
+                for layer_id in range(num_layers)
+            ])
+        self.dense = torch.nn.ModuleList(
+            [RowParallelLinear(
+                hidden_size,
+                hidden_size,
+                input_is_parallel=True,
+                init_method=output_layer_init_method,
+                bias=True,
+                module=self,
+                name="dense",)
+                for layer_id in range(num_layers)
+            ])
+
+        self.n_head = n_head
+        self.frame_resolution = frame_resolution
+        self.frame_len = frame_resolution * frame_resolution
+        self.causal_mask = torch.ones(1, 1, self.frame_len*frame_num, self.frame_len*frame_num).tril()
+        
+        self.mask_initialized = False
+        
+        self.attn_distribution = torch.nn.ParameterList([
+            torch.nn.Parameter(torch.zeros(hidden_size))
+            for _ in range(num_layers)
+        ])
+        
+    def reinit(self, *pre_mixins):
+        start_layer = len(self.transformer.layers) - self.num_layers
+        assert start_layer >= 0
+        for layer_id in range(self.num_layers):
+            base_attention = self.transformer.layers[start_layer + layer_id].attention
+            self.query_key_value[layer_id].weight.data.copy_(base_attention.query_key_value.weight.data)
+            self.query_key_value[layer_id].bias.data.copy_(base_attention.query_key_value.bias.data)
+    
+    def attention_extra(self, frame_hidden_state, layer_id, attn_dropout, text_hidden_state=None, 
+                       text_attn_mask=None, mode_sequential=False):
+        # pb relax
+        # frame_hidden_state [batchsize, frame_num*frame_size, n_head*hiddensize_perhead]
+        assert mode_sequential == True # only 
+        swin_pb_relax = True
+        alpha = 16
+        
+        if not self.mask_initialized:
+            self.causal_mask = self.causal_mask.to(device=frame_hidden_state.device, dtype=frame_hidden_state.dtype)
+            self.mask_initialized = True
+        b0, s1, h0 = frame_hidden_state.shape 
+        h = h0 // self.n_head
+        frame_len = self.frame_resolution * self.frame_resolution
+        frame_num = s1 // frame_len
+        assert frame_num*frame_len == s1
+            
+        qkv = self.query_key_value[layer_id](frame_hidden_state).reshape(b0, s1, 3, self.n_head, h)\
+                .permute(2, 0, 3, 1, 4) #[3, b0, n_head, s1, h]
+        q, k, v = qkv[0], qkv[1], qkv[2]
+        
+        # frames-to-frames 
+        if swin_pb_relax:
+            attn = torch.matmul(q / (math.sqrt(h)*alpha), k.transpose(-1, -2))
+        else: 
+            attn = torch.matmul(q / math.sqrt(h), k.transpose(-1, -2))
+        attn = torch.mul(attn, self.causal_mask) - 10000.0 * (1.0 - self.causal_mask)
+        if swin_pb_relax:
+            swin_pb_relax_const = torch.max(attn.reshape(b0, self.n_head, -1), dim=-1, keepdim=True)[0].detach().unsqueeze(-1)
+            attn = (attn - swin_pb_relax_const)*alpha
+
+        if text_hidden_state is None:
+            attn = F.softmax(attn, dim=-1)
+            if attn_dropout is not None:
+                with get_cuda_rng_tracker().fork():
+                    attn = attn_dropout(attn)
+            context_swin = torch.matmul(attn, v).permute(0, 2, 1, 3).reshape(b0, s1, h0)
+        else:
+            # frame-to-text
+            assert text_attn_mask is not None
+            s0 = text_hidden_state.shape[1]
+            qkv_text = self.query_key_value[layer_id](text_hidden_state).reshape(b0, s0, 3, self.n_head, h).permute(2, 0, 3, 1, 4) #[3, b0, n_head, s0, h]
+            q_text, k_text, v_text = qkv_text[0], qkv_text[1], qkv_text[2]
+            text_attn_mask = text_attn_mask.unsqueeze(2)
+            if swin_pb_relax:
+                attn_frame2text = torch.matmul(q.reshape(b0, self.n_head, s1, h) / (math.sqrt(h)*alpha), k_text.transpose(-1, -2))
+                attn_frame2text = (attn_frame2text-swin_pb_relax_const.reshape(b0, self.n_head, 1, 1))*alpha
+            else:
+                attn_frame2text = torch.matmul(q.reshape(b0, self.n_head, s1, h) / math.sqrt(h), k_text.transpose(-1, -2))
+            attn_frame2text = torch.mul(text_attn_mask, attn_frame2text) - 10000.0 * (1.0 - text_attn_mask)
+            attn_frame2text = attn_frame2text.reshape(b0, self.n_head, s1, s0)
+            
+            attn = torch.cat((attn, attn_frame2text), dim=-1)
+            attn = F.softmax(attn, dim=-1)
+            
+            if attn_dropout is not None:
+                with get_cuda_rng_tracker().fork():
+                    attn = attn_dropout(attn)
+                    
+            context_frame = (torch.matmul(attn[..., :-s0], v) + 
+                            torch.matmul(attn[..., -s0:].reshape(b0, self.n_head,s1, s0), v_text))\
+                .permute(0, 2, 1, 3).reshape(b0, s1, h0)
+                
+        return context_frame
+        
+
+def attention_localframe_and_text(q0, k0, v0, attention_mask_totxt, attention_mask_local, 
+                             n_head, text_len, frame_len, frame_num, attention_dropout=None, layer_id=0, **kwargs):
+    b, s0, h0 = q0.shape
+    s1 = s0 - text_len
+    h = h0 // n_head
+    assert q0.shape[1] == v0.shape[1] == k0.shape[1] == text_len+frame_len*frame_num
+    # attention_mask_totxt [b, 1, 1, text_len]
+    # attention_mask_local [1, 1, frame_num, frame_len, frame_len]
+    # attention_mask: [1, 1, text_len+frame_len, text_len+frame_len]
+
+    q0 = q0.reshape(b, s0, n_head, h).permute(0, 2, 1, 3)
+    v0 = v0.reshape(b, s0, n_head, h).permute(0, 2, 1, 3)
+    k0 = k0.reshape(b, s0, n_head, h).permute(0, 2, 1, 3)
+    k0T = k0.transpose(-1, -2)
+
+    # score: any2text
+    score_any2text = torch.matmul(q0 / math.sqrt(q0.shape[-1]), k0T[..., :text_len])
+    score_any2text_part1 = torch.mul(score_any2text[..., :text_len, :], attention_mask_totxt) \
+        - 10000.0 * (1.0 - attention_mask_totxt)
+    score_any2text_part2 = torch.mul(score_any2text[..., text_len:, :], attention_mask_totxt) - \
+                                     10000.0 * (1.0 - attention_mask_totxt)
+    
+    # score: frame local
+    q0_frame = q0[:, :, text_len:].reshape(b, n_head, frame_num, frame_len, h)
+    v0_frame = v0[:, :, text_len:].reshape(b, n_head, frame_num, frame_len, h)
+    k0T_frame = k0[:, :, text_len:].reshape(b, n_head, frame_num, frame_len, h).transpose(-1, -2)
+    score_frame_local0 = torch.matmul(q0_frame / math.sqrt(q0_frame.shape[-1]), k0T_frame)
+    score_frame_local0 = torch.mul(score_frame_local0, attention_mask_local) \
+        - 10000.0 * (1.0 - attention_mask_local)
+    
+    # context for frame
+    score_frame_all = torch.cat((score_any2text_part2, 
+                                 score_frame_local0.view(b, n_head, s1, frame_len)), dim=-1)
+    attention_probs_frame = F.softmax(score_frame_all, dim=-1)
+    
+    if attention_dropout is not None:
+        with get_cuda_rng_tracker().fork():
+            attention_probs_frame = attention_dropout(attention_probs_frame)
+            
+    context_frame2text = torch.matmul(attention_probs_frame[..., :text_len], v0[..., :text_len, :]) # [b, n_head, s1, h]
+    context_frame_local0 = torch.matmul(attention_probs_frame[..., text_len:text_len+frame_len].\
+        view(b, n_head, frame_num, frame_len, frame_len), v0_frame).view(b, n_head, s1, h)
+    context_frame = (context_frame2text + context_frame_local0).transpose(1, 2).reshape(b, s1, h0)
+
+    # context for text
+    attention_probs_text = F.softmax(score_any2text_part1, dim=-1)
+    if attention_dropout is not None:
+        with get_cuda_rng_tracker().fork():
+            attention_probs_text = attention_dropout(attention_probs_text)
+    context_text2text = torch.matmul(attention_probs_text, v0[..., :text_len, :])
+    context_text2text = context_text2text.transpose(1, 2).reshape(b, text_len, h0)
+    
+    return context_text2text, context_frame
+    
+    
+class CogVideoModel(BaseModel):
+    def __init__(self, args, transformer=None, parallel_output=True):
+        super().__init__(args, transformer=transformer, parallel_output=parallel_output)
+        self.stage = args.cogvideo_stage # 1 or 2
+        self.mode_sequential = True if self.stage==1 else False
+        self.layout = args.layout # [64, 64+400, 64+5*400]
+        self.n_head = args.num_attention_heads
+        frame_resolution = int(math.sqrt(self.layout[1]-self.layout[0]))
+        frame_num = (args.layout[2]-args.layout[0])//(args.layout[1]-args.layout[0])
+        frame_len = self.layout[1]-self.layout[0]
+        
+        self.add_mixin('extra_position_embedding', PositionEmbeddingMixin(
+            args.additional_seqlen, args.hidden_size
+        ))
+        
+        if args.window_size == -1:
+            # full attention
+            assert self.stage == 1
+            self.add_mixin('attention_plus', FullAttentionMixin(
+                num_layers=args.num_layers,
+                hidden_size=args.hidden_size,
+                frame_resolution=frame_resolution,
+                n_head=args.num_attention_heads,
+                frame_num=frame_num,
+            ))
+        else:
+            self.add_mixin('attention_plus', WindowAttentionMixin(
+                num_layers=args.num_layers,
+                hidden_size=args.hidden_size,
+                frame_resolution=frame_resolution,
+                window_size=args.window_size,
+                shift_size=args.window_size//2,
+                n_head=args.num_attention_heads,
+                frame_num=frame_num,
+            ))
+        # attention_mask_local
+        self.attention_mask_local_sequential = torch.ones(1, 1, frame_num, frame_len, frame_len).tril().unsqueeze(0)
+        self.attention_mask_local_interp = torch.ones(1, 1, frame_num, frame_len, frame_len)
+
+        for idx in range(1, frame_num, 2):
+            self.attention_mask_local_interp[:, :, idx:idx+1].tril_()
+        self.attention_mask_local_interp = self.attention_mask_local_interp.unsqueeze(0)
+        self.mask_initialized = False
+        
+    @classmethod
+    def add_model_specific_args(cls, parser):
+        group = parser.add_argument_group('CogVideoModel', 'CogVideo model configurations')
+        group.add_argument("--layout", type=str, default='64, 464, 2064', help='text_len, textlen+frame_len, textlen+frame_len*frame_num')
+        group.add_argument("--window-size", type=int, default=10, help="swin attention's window size in temperal channel, -1 represents full attention")
+        group.add_argument("--additional-seqlen", type=int, default=2000)
+        group.add_argument("--cogvideo-stage", type=int, default=1, choices=[1,2])
+        return parser
+    
+    def disable_untrainable_params(self):
+        self.transformer.requires_grad_(False)
+        
+    def position_embedding_forward(self, position_ids, **kw_args):
+        position = position_ids[..., :(64+400)]
+        position_plus = position_ids[..., (64+400):]
+        position_embeddings = torch.cat(
+            (
+                self.transformer.position_embeddings(position),
+                self.get_mixin('extra_position_embedding').position_embeddings(position_plus-(512+400))
+            ),
+            dim=-2
+        )
+        return position_embeddings
+        
+    def attention_forward(self, hidden_states, mask, layer_id, **kw_args):
+        # mask.shape=[bs, 1, 1, 64]
+        if not self.mask_initialized:
+            self.attention_mask_local_sequential = self.attention_mask_local_sequential.to(device=hidden_states.device, dtype=hidden_states.dtype)
+            self.attention_mask_local_interp = self.attention_mask_local_interp.to(device=hidden_states.device, dtype=hidden_states.dtype)
+            self.mask_initialized = True
+        
+        attn_module = self.transformer.layers[layer_id].attention
+        hidden_size = hidden_states.shape[-1]
+        bs = hidden_states.shape[0]
+        
+        # base model qkv
+        mixed_raw_layer = attn_module.query_key_value(hidden_states)
+        q0, k0, v0 = split_tensor_along_last_dim(mixed_raw_layer, 3)
+        dropout_fn = self.transformer.layers[layer_id].attention.attention_dropout if self.training else None
+
+        attention_mask_local = self.attention_mask_local_sequential if self.mode_sequential else self.attention_mask_local_interp
+        context_text, context_frame_local_text = attention_localframe_and_text(
+                q0, k0, v0,
+                attention_mask_totxt=mask,
+                attention_mask_local=attention_mask_local,
+                n_head=attn_module.num_attention_heads_per_partition,
+                text_len=self.layout[0],
+                frame_len=self.layout[1]-self.layout[0],
+                frame_num=(self.layout[2]-self.layout[0])//(self.layout[1]-self.layout[0]),
+                attention_dropout=dropout_fn, 
+                layer_id=layer_id,
+            )
+
+        context_frame_swin = self.get_mixin('attention_plus').attention_extra(
+            hidden_states[:, self.layout[0]:], layer_id, dropout_fn, 
+            text_hidden_state=hidden_states[:, :self.layout[0]], 
+            text_attn_mask=mask[..., 0, :],
+            mode_sequential=self.mode_sequential)
+            
+        attn_distrib = torch.sigmoid(self.get_mixin('attention_plus').attn_distribution[layer_id])
+        attn_distrib = attn_distrib.unsqueeze(0).unsqueeze(0)
+        
+        output_text = attn_module.dense(context_text)
+        output_frame = torch.mul(attn_module.dense(context_frame_local_text), attn_distrib)\
+            +torch.mul(self.get_mixin('attention_plus').dense[layer_id](context_frame_swin), 1-attn_distrib)
+        output = torch.cat((output_text, output_frame), dim=-2)
+
+        return output

pretrain_cogvideo.py

@@ -0,0 +1,184 @@
+# -*- encoding: utf-8 -*-
+'''
+@File    :   pretrain_cogvideo.py
+@Time    :   2021/10/06 00:58:32
+@Author  :   Wenyi Hong
+@Contact :   hwy22@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+import os
+import sys
+import math
+import random
+import torch
+import argparse
+import numpy as np
+from icetk import icetk as tokenizer
+tokenizer.add_special_tokens(['<start_of_image>', '<start_of_english>', '<start_of_chinese>'])
+
+from models.cogvideo_model import CogVideoModel
+from SwissArmyTransformer import mpu, get_args
+from SwissArmyTransformer.training.deepspeed_training import training_main
+from SwissArmyTransformer.data_utils import BinaryDataset
+
+def get_masks_and_position_ids_video(data, attention_mask_totxt=None, args=None):
+    # Extract batch size and sequence length.
+    batch_size, seq_length = data.size()
+    assert attention_mask_totxt is not None
+    layout = args.layout
+    assert seq_length == layout[-1]
+    n_pads = layout[0] - attention_mask_totxt.sum(dim=-1).long()
+    frame_len = layout[1]-layout[0]
+    position_ids = torch.zeros(batch_size, layout[2], dtype=torch.long,
+                                device=data.device)
+    for i in range(batch_size):
+        torch.arange(layout[0] - n_pads[i], out=position_ids[i, n_pads[i]:layout[0]], 
+            dtype=torch.long, device=data.device)
+        torch.arange(512, 512+layout[2]-layout[0], 
+            out=position_ids[i, layout[0]:], dtype=torch.long, device=data.device)
+    return position_ids
+
+
+def get_batch(data_iterator, args, timers):
+    # Items and their type.
+    keys = ['text', 'loss_mask', 'attention_mask_totxt']
+    datatype = torch.int64
+
+    # Broadcast data.
+    timers('data loader').start()
+    if data_iterator is not None:
+        data = next(data_iterator)
+    else:
+        data = None
+    timers('data loader').stop()
+
+    data_b = mpu.broadcast_data(keys, data, datatype)
+    # Unpack.
+    tokens_ = data_b['text'].long()
+    loss_mask = data_b['loss_mask'].float()
+    attention_mask_totxt = data_b['attention_mask_totxt'].float()
+
+    labels = tokens_[:, 1:].clone().contiguous()
+    loss_mask = loss_mask[:, 1:].contiguous()
+    tokens = tokens_[:, :-1].clone().contiguous()
+    
+    for idx in range(args.layout[0], args.layout[2], 400):
+        tokens[:, idx] = tokenizer['<start_of_image>']
+    # Get the masks and postition ids.
+    position_ids = get_masks_and_position_ids_video(
+        tokens,
+        attention_mask_totxt=attention_mask_totxt,
+        args=args
+        )
+    attention_mask_totxt = attention_mask_totxt.unsqueeze(1).unsqueeze(1)
+    # Convert
+    if args.fp16:
+        attention_mask_totxt = attention_mask_totxt.half()
+    return tokens, labels, loss_mask, attention_mask_totxt, position_ids
+
+
+def forward_step(data_iterator, model, args, timers):
+    """Forward step."""
+
+    # Get the batch.
+    timers('batch generator').start()
+    tokens, labels, loss_mask, attention_mask_totxt, position_ids = get_batch(
+        data_iterator, args, timers)
+    timers('batch generator').stop()
+
+    # Forward model.
+    logits, *mems = model(tokens, position_ids, attention_mask_totxt)
+    # ======= hyper params =======#
+    perframe_len = 400
+    text_len=64
+    frame_num = 5
+    logits_img_tokens = logits[:, text_len:, :tokenizer.num_image_tokens].float().contiguous()
+    losses = mpu.vocab_parallel_cross_entropy(logits_img_tokens, labels[:, text_len:])
+    # scaling loss mask
+    loss_mask = loss_mask[:, text_len:].reshape(-1)  
+
+    losses_1d = losses.reshape(-1) * loss_mask
+    loss = torch.sum(losses_1d) / loss_mask.sum()
+    # =====================   Log partial losses   ======================== #
+    log_loss_dict = {}
+    bs = losses.shape[0]
+
+    if args.cogvideo_stage == 1:
+        for i in range(frame_num):
+            log_loss_dict[f'AR_f{i}_loss'] = losses[:, i*perframe_len:(i+1)*perframe_len].contiguous().reshape(-1).detach().sum() / max((perframe_len*bs), 1) 
+    else:
+        for i in range(1, frame_num-1):
+            log_loss_dict[f'ITP_f{i}_loss'] = losses[:, i*perframe_len:(i+1)*perframe_len].contiguous().reshape(-1).detach().sum() / max((perframe_len*bs), 1) 
+ 
+    # ===================== END OF BLOCK ======================= #
+    return loss, log_loss_dict
+    
+
+def create_dataset_function(path, args):
+    dataset_layout = [64, 464, 2064]
+    input_layout = [64, 464, 2064]
+    # frame_num = 6
+    # frame_interval = 2 # DEBUG!!!
+    def process_fn(row):
+        row = row.astype(np.int64)
+        text = row[:dataset_layout[0]]
+        frames = row[dataset_layout[0]:]
+        
+        if text[0] == tokenizer['<pad>']:
+            text = text[1:] # due to our way of data processing
+        if args.cogvideo_stage == 1:
+            text, loss_mask, frames = make_text_video_generation(text, frames)
+        else: 
+            text, loss_mask, frames = mask_video_frame_interpolation(text, frames)
+            
+        n_pad = input_layout[0] - len(text)
+        parts = [
+            np.array([tokenizer['<pad>']] * n_pad, dtype=np.int64),
+            text,
+            np.array([tokenizer['<start_of_image>']], dtype=np.int64),
+            frames,
+        ]
+        ret = np.concatenate(parts, axis=0)
+        
+        attention_mask_totxt = np.array([0] * n_pad + [1] * (input_layout[0]-n_pad))
+        return {'text': ret, 
+            'loss_mask':  loss_mask,
+            'attention_mask_totxt': attention_mask_totxt,
+            }
+    return BinaryDataset(path, process_fn, length_per_sample=dataset_layout[-1])
+
+def make_text_video_generation(text, frames):
+    input_layout = [64, 464, 2064]
+    text = text[text!= tokenizer['<pad>']][:input_layout[0]] # dataset format: 1.0秒<n>{text}<pad><pad> ... 
+    loss_mask = np.array([0] * (input_layout[1]+1) + [1] * (input_layout[2] - input_layout[1])) # 按照input的，之后loss_mask会左移一位
+    return text, loss_mask, frames
+
+def mask_video_frame_interpolation(text, frames):
+    input_layout = [64, 464, 2064]
+    frame_len = input_layout[1]-input_layout[0]
+    # text format: <pad> 1.0秒 <n> {text} <pad> <pad>
+    text = text[text!= tokenizer['<pad>']][:input_layout[0]]
+    loss_mask = np.array([0] * (input_layout[1]+1) 
+                        + [1] * (input_layout[1]-input_layout[0])
+                        + [0] * (input_layout[1]-input_layout[0])
+                        + [1] * (input_layout[1]-input_layout[0])
+                        + [0] * (input_layout[1]-input_layout[0]) )# 按照input的，之后loss_mask会左移一位
+        
+    return text, loss_mask, frames
+    
+
+
+if __name__ == '__main__':    
+    py_parser = argparse.ArgumentParser(add_help=False)
+    py_parser.add_argument('--txt-loss-scale', type=float, default=1)    
+    CogVideoModel.add_model_specific_args(py_parser)
+    
+    known, args_list = py_parser.parse_known_args()
+    
+    args = get_args(args_list)
+    args = argparse.Namespace(**vars(args), **vars(known))
+    
+    args.layout = [int(x) for x in args.layout.split(',')]
+    
+    training_main(args, model_cls=CogVideoModel, forward_step_function=forward_step, create_dataset_function=create_dataset_function)

requirements.txt

@@ -0,0 +1,4 @@
+SwissArmyTransformer>=0.2.9
+icetk
+gifmaker
+torchvision

scripts/ds_brain_pretrain_cogvideo_stage1.sh

@@ -0,0 +1,108 @@
+#! /bin/bash
+
+# Change for multinode config
+
+NUM_WORKERS=1
+NUM_GPUS_PER_WORKER=8
+MP_SIZE=1
+
+script_path=$(realpath $0)
+script_dir=$(dirname $script_path)
+main_dir=$(dirname $script_dir)
+
+OPTIONS_NCCL="NCCL_DEBUG=warning NCCL_IB_DISABLE=0 NCCL_NET_GDR_LEVEL=2"
+HOST_FILE_PATH="hostfile"
+# HOST_FILE_PATH="hostfile_single"
+
+video_data_test="" # TODO
+CHECKPOINT_PATH="" # TODO: CogView2 ckpt
+
+config_json="$script_dir/ds_config_zero.json"
+gpt_options=" \
+       --experiment-name pretrain-cogvideo-stage1 \
+       --tokenizer-type fake \
+       --vocab-size 150010 \
+       --model-parallel-size ${MP_SIZE} \
+       --mode finetune \
+       --num-workers 0 \
+       --num-layers 48 \
+       --hidden-size 3072 \
+       --num-attention-heads 48 \
+       --layout 64,464,2064 \
+       --window-size -1 \
+       --cogvideo-stage 1 \
+       --additional-seqlen 2000 \
+       --train-iters 500000 \
+       --resume-dataloader \
+       --train-data ${video_data_test}  \
+       --train-data-weights 1 \
+       --split 949,50,1 \
+       --distributed-backend nccl \
+       --lr-decay-style cosine \
+       --warmup .001 \
+       --checkpoint-activations \
+       --max-sequence-length 1024 \
+       --fp16 \
+       --save-interval 2000 \
+       --eval-interval 500 \
+       --eval-iters 15 \
+       --log-interval 50 \
+       --save $main_dir/checkpoints \
+       --sandwich-ln \
+       --load $CHECKPOINT_PATH \
+"
+       # --load $CHECKPOINT_PATH \
+       #  \       --sandwich-ln
+
+
+gpt_options="${gpt_options}
+       --deepspeed \
+       --deepspeed_config ${config_json} \
+"
+
+#!/bin/bash
+
+# Distribute Example
+#export NCCL_SOCKET_IFNAME=eth0
+export NCCL_IB_DISABLE=0
+export NCCL_NET_GDR_LEVEL=2
+#export NCCL_IB_CUDA_SUPPORT=1
+#export NCCL_IB_GID_INDEX=3
+#export NCCL_IB_HCA=$(pushd /sys/class/infiniband/ > /dev/null; for i in mlx5_*; do cat $i/ports/1/gid_attrs/types/* 2>/dev/null | grep v >/dev/null && echo $i ; done; popd > /dev/null)
+export NCCL_DEBUG=info
+export OMP_NUM_THREADS=4
+
+if [ $RLAUNCH_REPLICA == "0" ]; then
+	ifconfig eth0 | grep inet | grep -v inet6 | awk '{print $2}' > master_ip
+fi
+
+function finish {
+	rm -rf master_ip
+}
+
+trap finish EXIT INT TERM
+
+while [ ! -f master_ip ]; do
+	echo "wait master_ip..."
+	ls > /dev/null && sleep 1;
+done
+
+export MASTER_ADDR=$(cat master_ip)
+echo "master_ip: $MASTER_ADDR"
+
+MP_SIZE=1
+task_set=$2
+source $1
+DATESTR=$(date +"%m-%d-%H-%M")
+
+mkdir logs
+run_cmd="sudo /opt/conda/bin/python -m torch.distributed.launch --nproc_per_node=8 \
+       	--nnodes=$RLAUNCH_REPLICA_TOTAL --node_rank=$RLAUNCH_REPLICA \
+	--master_addr=$MASTER_ADDR --master_port=12355 pretrain_cogvideo.py $@ ${gpt_options}  2>&1 | tee logs/log-${DATESTR}-${RLAUNCH_REPLICA}.txt"
+              
+
+# run_cmd="${OPTIONS_NCCL} deepspeed  --num_nodes ${NUM_WORKERS} --num_gpus ${NUM_GPUS_PER_WORKER} --hostfile ${HOST_FILE_PATH} pretrain_video_swin_cond_glm_interp.py $@ ${gpt_options}"
+echo ${run_cmd}
+eval ${run_cmd}
+
+set +x

scripts/ds_brain_pretrain_cogvideo_stage2.sh

@@ -0,0 +1,108 @@
+#! /bin/bash
+
+# Change for multinode config
+
+NUM_WORKERS=1
+NUM_GPUS_PER_WORKER=8
+MP_SIZE=1
+
+script_path=$(realpath $0)
+script_dir=$(dirname $script_path)
+main_dir=$(dirname $script_dir)
+
+OPTIONS_NCCL="NCCL_DEBUG=warning NCCL_IB_DISABLE=0 NCCL_NET_GDR_LEVEL=2"
+HOST_FILE_PATH="hostfile"
+# HOST_FILE_PATH="hostfile_single"
+
+video_data_test="" # TODO
+CHECKPOINT_PATH="" # TODO: CogView2 ckpt
+
+config_json="$script_dir/ds_config_zero.json"
+gpt_options=" \
+       --experiment-name pretrain-cogvideo-stage2 \
+       --tokenizer-type fake \
+       --vocab-size 150010 \
+       --model-parallel-size ${MP_SIZE} \
+       --mode finetune \
+       --num-workers 0 \
+       --num-layers 48 \
+       --hidden-size 3072 \
+       --num-attention-heads 48 \
+       --layout 64,464,2064 \
+       --window-size 10 \
+       --cogvideo-stage 2 \
+       --additional-seqlen 2000 \
+       --train-iters 500000 \
+       --resume-dataloader \
+       --train-data ${video_data_test}  \
+       --train-data-weights 1 \
+       --split 949,50,1 \
+       --distributed-backend nccl \
+       --lr-decay-style cosine \
+       --warmup .001 \
+       --checkpoint-activations \
+       --max-sequence-length 1024 \
+       --fp16 \
+       --save-interval 2000 \
+       --eval-interval 500 \
+       --eval-iters 15 \
+       --log-interval 50 \
+       --save $main_dir/checkpoints \
+       --sandwich-ln \
+       --load $CHECKPOINT_PATH \
+"
+       # --load $CHECKPOINT_PATH \
+       #  \       --sandwich-ln
+
+
+gpt_options="${gpt_options}
+       --deepspeed \
+       --deepspeed_config ${config_json} \
+"
+
+#!/bin/bash
+
+# Distribute Example
+#export NCCL_SOCKET_IFNAME=eth0
+export NCCL_IB_DISABLE=0
+export NCCL_NET_GDR_LEVEL=2
+#export NCCL_IB_CUDA_SUPPORT=1
+#export NCCL_IB_GID_INDEX=3
+#export NCCL_IB_HCA=$(pushd /sys/class/infiniband/ > /dev/null; for i in mlx5_*; do cat $i/ports/1/gid_attrs/types/* 2>/dev/null | grep v >/dev/null && echo $i ; done; popd > /dev/null)
+export NCCL_DEBUG=info
+export OMP_NUM_THREADS=4
+
+if [ $RLAUNCH_REPLICA == "0" ]; then
+	ifconfig eth0 | grep inet | grep -v inet6 | awk '{print $2}' > master_ip
+fi
+
+function finish {
+	rm -rf master_ip
+}
+
+trap finish EXIT INT TERM
+
+while [ ! -f master_ip ]; do
+	echo "wait master_ip..."
+	ls > /dev/null && sleep 1;
+done
+
+export MASTER_ADDR=$(cat master_ip)
+echo "master_ip: $MASTER_ADDR"
+
+MP_SIZE=1
+task_set=$2
+source $1
+DATESTR=$(date +"%m-%d-%H-%M")
+
+mkdir logs
+run_cmd="sudo /opt/conda/bin/python -m torch.distributed.launch --nproc_per_node=8 \
+       	--nnodes=$RLAUNCH_REPLICA_TOTAL --node_rank=$RLAUNCH_REPLICA \
+	--master_addr=$MASTER_ADDR --master_port=12355 pretrain_cogvideo.py $@ ${gpt_options}  2>&1 | tee logs/log-${DATESTR}-${RLAUNCH_REPLICA}.txt"
+              
+
+# run_cmd="${OPTIONS_NCCL} deepspeed  --num_nodes ${NUM_WORKERS} --num_gpus ${NUM_GPUS_PER_WORKER} --hostfile ${HOST_FILE_PATH} pretrain_video_swin_cond_glm_interp.py $@ ${gpt_options}"
+echo ${run_cmd}
+eval ${run_cmd}
+
+set +x

scripts/ds_config_zero.json

@@ -0,0 +1,42 @@
+{
+    "train_micro_batch_size_per_gpu": 4,
+    "gradient_accumulation_steps": 1,
+    "steps_per_print": 1,
+    "gradient_clipping": 0.1,
+    "zero_optimization": {
+      "stage": 2,
+      "cpu_offload": true,
+      "contiguous_gradients": false,
+      "overlap_comm": true,
+      "reduce_scatter": false,
+      "reduce_bucket_size": 100000000,
+      "allgather_bucket_size": 1000000000,
+      "load_from_fp32_weights": false
+    },
+    "zero_allow_untested_optimizer": true,
+    "fp16": {
+      "enabled": true,
+      "loss_scale": 0,
+      "loss_scale_window": 400,
+      "hysteresis": 2,
+      "min_loss_scale": 1
+    },
+    "optimizer": {
+      "type": "Adam",
+      "params": {
+        "lr": 0.0002,
+        "betas": [
+          0.9,
+          0.95
+        ],
+        "eps": 1e-8,
+        "weight_decay": 1e-4
+      }
+    },
+    "activation_checkpointing": {
+      "partition_activations": false,
+      "contiguous_memory_optimization": false
+    },
+    "wall_clock_breakdown": false
+  }
+  

scripts/inference_cogvideo_pipeline.sh

@@ -0,0 +1,38 @@
+#!/bin/bash
+
+NLAYERS=48
+NHIDDEN=3072
+NATT=48
+MAXSEQLEN=1024
+MASTER_PORT=$(shuf -n 1 -i 10000-65535)
+MPSIZE=1
+
+#SAMPLING ARGS
+TEMP=1.05
+TOPK=12
+
+script_path=$(realpath $0)
+script_dir=$(dirname $script_path)
+
+MASTER_PORT=${MASTER_PORT} SAT_HOME=/sharefs/cogview-new python cogvideo_pipeline.py \
+        --input-source interactive \
+        --output-path ./output \
+        --parallel-size 1 \
+        --both-stages \
+        --use-guidance-stage1 \
+        --guidance-alpha 3.0 \
+        --generate-frame-num 5 \
+        --tokenizer-type fake \
+        --mode inference \
+        --distributed-backend nccl \
+        --fp16 \
+        --model-parallel-size $MPSIZE \
+        --temperature $TEMP \
+        --coglm-temperature2 0.89 \
+        --top_k $TOPK \
+        --sandwich-ln \
+        --seed 1234 \
+        --num-workers 0 \
+        --batch-size 4 \
+        --max-inference-batch-size 8 \
+        $@

sr_pipeline/__init__.py

@@ -0,0 +1,17 @@
+# -*- encoding: utf-8 -*-
+'''
+@File    :   __init__.py
+@Time    :   2022/03/02 13:57:09
+@Author  :   Ming Ding 
+@Contact :   dm18@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+import os
+import sys
+import math
+import random
+
+from .direct_sr import DirectSuperResolution
+from .iterative_sr import IterativeSuperResolution
+from .sr_group import SRGroup

sr_pipeline/direct_sr.py

@@ -0,0 +1,117 @@
+# -*- encoding: utf-8 -*-
+'''
+@File    :   direct_sr.py
+@Time    :   2022/03/02 13:58:11
+@Author  :   Ming Ding 
+@Contact :   dm18@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+import os
+import sys
+import math
+import random
+import torch
+
+# -*- encoding: utf-8 -*-
+'''
+@File    :   inference_cogview2.py
+@Time    :   2021/10/10 16:31:34
+@Author  :   Ming Ding 
+@Contact :   dm18@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+import os
+import sys
+import math
+import random
+from PIL import ImageEnhance, Image
+
+import torch
+import argparse
+from torchvision import transforms
+
+from SwissArmyTransformer import get_args
+from SwissArmyTransformer.training.model_io import load_checkpoint
+from .dsr_sampling import filling_sequence_dsr, IterativeEntfilterStrategy
+from SwissArmyTransformer.generation.utils import timed_name, save_multiple_images, generate_continually
+
+from .dsr_model import DsrModel
+
+from icetk import icetk as tokenizer
+
+class DirectSuperResolution:
+    def __init__(self, args, path, max_bz=4, topk=6, onCUDA=False):
+        args.load = path
+        args.kernel_size = 5
+        args.kernel_size2 = 5
+        args.new_sequence_length = 4624
+        args.layout = [96,496,4096]
+        
+        model = DsrModel(args)
+        if args.fp16:
+            model = model.half()
+        
+        load_checkpoint(model, args) # on cpu
+        model.eval()
+        self.model = model
+        self.onCUDA = onCUDA
+        if onCUDA:
+            self.model = self.model.cuda()
+        
+        invalid_slices = [slice(tokenizer.num_image_tokens, None)]
+    
+        self.strategy = IterativeEntfilterStrategy(invalid_slices,
+            temperature=1.0, topk=topk) # temperature not used # Temperature Freezed Here!!
+        self.max_bz = max_bz
+        
+    def __call__(self, text_tokens, image_tokens, enhance=False):
+        if len(text_tokens.shape) == 1:
+            text_tokens.unsqueeze_(0)
+        if len(image_tokens.shape) == 1:
+            image_tokens.unsqueeze_(0)
+        # =====================   Debug   ======================== #
+        # new_image_tokens = []
+        # for small_img in image_tokens:
+        #     decoded = tokenizer.decode(image_ids=small_img)
+        #     decoded = torch.nn.functional.interpolate(decoded, size=(480, 480)).squeeze(0)
+        #     ndarr = decoded.mul(255).add_(0.5).clamp_(0, 255).permute(1, 2, 0).to('cpu', torch.uint8).numpy()
+        #     image_pil_raw = ImageEnhance.Sharpness(Image.fromarray(ndarr))
+        #     small_img2 = tokenizer.encode(image_pil=image_pil_raw.enhance(1.5), image_size=480).view(-1)
+        #     new_image_tokens.append(small_img2)
+        # image_tokens = torch.stack(new_image_tokens)
+        # return image_tokens
+        # ===================== END OF BLOCK ======================= #
+        if enhance:
+            new_image_tokens = []
+            for small_img in image_tokens:
+                decoded = tokenizer.decode(image_ids=small_img).squeeze(0)
+                ndarr = decoded.mul(255).add_(0.5).clamp_(0, 255).permute(1, 2, 0).to('cpu', torch.uint8).numpy()
+                image_pil_raw = ImageEnhance.Sharpness(Image.fromarray(ndarr))
+                small_img2 = tokenizer.encode(image_pil=image_pil_raw.enhance(1.), image_size=160).view(-1)
+                new_image_tokens.append(small_img2)
+            image_tokens = torch.stack(new_image_tokens)
+                
+        seq = torch.cat((text_tokens,image_tokens), dim=1)
+        seq1 = torch.tensor([tokenizer['<start_of_image>']]*3601, device=image_tokens.device).unsqueeze(0).expand(text_tokens.shape[0], -1)
+        if not self.onCUDA:
+            print('Converting Dsr model...')
+            model = self.model.cuda()
+        else:
+            model = self.model
+        print('Direct super-resolution...')
+        output_list = []
+        for tim in range(max((text_tokens.shape[0]+self.max_bz-1) // self.max_bz, 1)): 
+            output1 = filling_sequence_dsr(model,
+                seq[tim*self.max_bz:(tim+1)*self.max_bz], 
+                seq1[tim*self.max_bz:(tim+1)*self.max_bz], 
+                warmup_steps=1, block_hw=(1, 0),
+                strategy=self.strategy
+                )
+            output_list.extend(output1[1:])
+        if not self.onCUDA:
+            print('Moving back Dsr to cpu...')
+            model = model.cpu()
+            torch.cuda.empty_cache()
+        return torch.cat(output_list, dim=0)

sr_pipeline/dsr_model.py

@@ -0,0 +1,225 @@
+# -*- encoding: utf-8 -*-
+'''
+@File    :   cuda2d_model.py
+@Time    :   2021/10/02 01:36:32
+@Author  :   Ming Ding 
+@Contact :   dm18@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+import os
+import sys
+import math
+import random
+import torch
+import torch.nn.functional as F
+
+
+from SwissArmyTransformer.model.base_model import BaseModel, BaseMixin
+
+from SwissArmyTransformer.model.transformer import split_tensor_along_last_dim, unscaled_init_method
+from SwissArmyTransformer.mpu.utils import sqrt
+from deepspeed.runtime.activation_checkpointing.checkpointing import get_cuda_rng_tracker
+from SwissArmyTransformer.mpu import ColumnParallelLinear, RowParallelLinear
+
+class PositionEmbeddingMixin(BaseMixin):
+    def __init__(self, additional_sequence_length, hidden_size,
+                 init_method_std=0.02, reinit_slice=slice(512, 512+400)
+                 ):
+        super(PositionEmbeddingMixin, self).__init__()
+        self.reinit_slice = reinit_slice
+        self.position_embeddings = torch.nn.Embedding(additional_sequence_length, hidden_size)
+        torch.nn.init.normal_(self.position_embeddings.weight, mean=0.0, std=init_method_std)
+
+    def reinit(self, parent_model=None):
+        old_weights = self.transformer.position_embeddings.weight.data[self.reinit_slice]
+        old_len, hidden_size = old_weights.shape
+        assert hidden_size == self.position_embeddings.weight.shape[-1]
+        old_edge, new_edge = sqrt(old_len), sqrt(self.position_embeddings.weight.shape[-2])
+        assert new_edge % old_edge == 0
+        self.position_embeddings.weight.data.view(new_edge // old_edge, old_edge, new_edge // old_edge, old_edge, hidden_size).copy_(old_weights.view(1, old_edge, 1, old_edge, hidden_size))
+        # self.position_embeddings.weight.data.view(-1, old_len, hidden_size).copy_(old_weights)
+
+
+class AttentionMixin(BaseMixin):
+    def __init__(self, num_layers,
+                 hidden_size,
+                 init_method=unscaled_init_method(0.02),
+                 output_layer_init_method=unscaled_init_method(0.02)
+                 ):
+        super(AttentionMixin, self).__init__()
+        self.num_layers = num_layers  # replace attention in the LAST n layers
+        self.query_key_value = torch.nn.ModuleList(
+            [ColumnParallelLinear(hidden_size, 3 * hidden_size, stride=3,
+                                  gather_output=False, init_method=init_method)
+             for layer_id in range(num_layers)
+             ])
+        self.dense = torch.nn.ModuleList(
+            [RowParallelLinear(hidden_size,
+                               hidden_size,
+                               input_is_parallel=True,
+                               init_method=output_layer_init_method)
+             for layer_id in range(num_layers)
+             ])
+
+    def reinit(self, parent_model=None):
+        start_layer = len(self.transformer.layers) - self.num_layers
+        assert start_layer >= 0
+        for layer_id in range(self.num_layers):
+            old_attention = self.transformer.layers[start_layer + layer_id].attention
+            self.query_key_value[layer_id].weight.data.copy_(old_attention.query_key_value.weight.data)
+            self.query_key_value[layer_id].bias.data.copy_(old_attention.query_key_value.bias.data)
+            self.dense[layer_id].weight.data.copy_(old_attention.dense.weight.data)
+            self.dense[layer_id].bias.data.copy_(old_attention.dense.bias.data)
+
+class DsrModel(BaseModel):
+    def __init__(self, args, transformer=None):
+        super().__init__(args, transformer=transformer)
+        self.original_sequence_length = args.max_sequence_length
+        additional_seqlen = args.new_sequence_length - args.max_sequence_length
+        self.add_mixin('extra_position_embedding', PositionEmbeddingMixin(
+            additional_seqlen, args.hidden_size
+        ))
+        self.add_mixin('attention_plus', AttentionMixin(
+            num_layers=args.num_layers,
+            hidden_size=args.hidden_size
+        ))
+        self.layout = args.layout
+        # [PAD]... [ROI1] text ... [BOI1] {layout[0]} 1024 {layout[1]} [EOI1] 4095 {layout[2]}
+        self.kernel_size = args.kernel_size
+        self.kernel_size2 = args.kernel_size2
+        self.log_attention_weights = None
+    
+    def position_embedding_forward(self, position_ids, **kw_args):
+        position = position_ids[..., :self.layout[1]]
+        position_plus = position_ids[..., self.layout[1]:] - self.original_sequence_length
+        position_embeddings = torch.cat(
+                (
+                    self.transformer.position_embeddings(position),
+                    self.get_mixin('extra_position_embedding').position_embeddings(position_plus)
+                ),
+                dim=-2
+            )
+        return position_embeddings
+        
+    def attention_forward(self, hidden_states, mask, 
+                        layer_id=None, log_attention_weights=None, **kw_args):
+        attn_module = self.transformer.layers[layer_id].attention
+        # attention_plus on all layers
+        query_key_value_plus = self.get_mixin('attention_plus').query_key_value[layer_id] 
+        dense_plus = self.get_mixin('attention_plus').dense[layer_id]
+        # split two parts
+        hidden_states_plus = hidden_states[:, self.layout[1]:]
+        hidden_states = hidden_states[:, :self.layout[1]]
+        # base model qkv
+        mixed_raw_layer = attn_module.query_key_value(hidden_states)
+        q0, k0, v0 = split_tensor_along_last_dim(mixed_raw_layer, 3)
+        # cuda2d model qkv
+        mixed_raw_layer = query_key_value_plus(hidden_states_plus)
+        q1, k1, v1 = split_tensor_along_last_dim(mixed_raw_layer, 3)
+        
+        dropout_fn = attn_module.attention_dropout if self.training else None
+
+        # cuda2d attention
+        context_layer0, context_layer1 = sparse_attention_2d_light(
+                q0, k0, v0,
+                q1, k1, v1,
+                mask,
+                n_head=attn_module.num_attention_heads_per_partition,
+                text_len=self.layout[0],
+                kernel_size=self.kernel_size,
+                kernel_size2=self.kernel_size2,
+                attention_dropout=dropout_fn,
+                log_attention_weights=log_attention_weights,
+                add_scalar=(kw_args['add_scalar'] if 'add_scalar' in kw_args else 0)
+            )
+
+        output_0 = attn_module.dense(context_layer0)
+        output_1 = dense_plus(context_layer1)
+        output = torch.cat((output_0, output_1), dim=1)
+        
+        return output
+
+    def final_forward(self, logits, **kwargs):
+        logits_parallel = logits
+        logits_parallel = torch.nn.functional.linear(logits_parallel.float(), self.transformer.word_embeddings.weight[:20000].float())
+        # logits_parallel = torch.nn.functional.linear(logits_parallel, self.transformer.word_embeddings.weight[:20000])
+        return logits_parallel
+    
+    def disable_untrainable_params(self):
+        self.transformer.requires_grad_(False)
+    
+    @classmethod
+    def add_model_specific_args(cls, parser):
+        group = parser.add_argument_group('Cuda2dModel', 'cuda2d model configurations')
+        group.add_argument("--kernel-size", type=int, default=5)
+        group.add_argument("--kernel-size2", type=int, default=5)
+        group.add_argument("--layout", type=str, default='96,496,4096')
+        group.add_argument("--new-sequence-length", type=int, default=4096)
+        return parser
+
+def sparse_attention_2d_light(q0, k0, v0, q1, k1, v1, attention_mask, n_head, text_len, kernel_size=9, kernel_size2=7, attention_dropout=None, log_attention_weights = None, add_scalar=0, **kwargs):
+    '''
+    q0, k0, v0: [batch_size, 1088, hidden_size]
+    q1, k1, v1: [batch_size, 4096, h2]
+    n_head: int
+    attention_mask: [batch_size, 1088, 1088]
+    '''
+    from SwissArmyTransformer.ops.local_attention_function import f_similar, f_weighting
+
+    b, s0, h0 = q0.shape
+    b, s1, h1 = q1.shape
+    h, l0, l1 = h0 // n_head, sqrt(s0-text_len), sqrt(s1)
+
+    q0 = q0.reshape(b, s0, n_head, h).permute(0, 2, 1, 3)
+    v0 = v0.reshape(b, s0, n_head, h).permute(0, 2, 1, 3)
+    k0T = k0.reshape(b, s0, n_head, h).permute(0, 2, 3, 1)
+    
+    # standard attention for level 0
+    attention_scores = torch.matmul(q0 / math.sqrt(q0.shape[-1]), k0T)
+    
+    if log_attention_weights is not None:
+        attention_scores += log_attention_weights
+    attention_scores = torch.mul(attention_scores, attention_mask) - \
+                    10000.0 * (1.0 - attention_mask)
+    
+    attention_probs0 = F.softmax(attention_scores, dim=-1)
+    
+    # local attention for level 1
+    q1 = (q1.view(b, s1, n_head, h1 // n_head).permute(0, 2, 3, 1) / math.sqrt(h1//n_head)).contiguous().view(b*n_head, h1//n_head, l1, l1)
+    k1 = k1.view(b, s1, n_head, h1 // n_head).permute(0, 2, 3, 1).contiguous().view(b*n_head, h1//n_head, l1, l1)
+    v1 = v1.view(b, s1, n_head, h1 // n_head).permute(0, 2, 3, 1).contiguous().view(b*n_head, h1//n_head, l1, l1)
+    # scores_1_to_1 = f_similar(q1, k1, kernel_size*2-1, kernel_size, True)   
+    scores_1_to_1 = f_similar(q1, k1, kernel_size*2-1, kernel_size, False)    
+
+    # cross attention
+    k0T = k0T[..., -l0**2:].reshape(b*n_head, h, l0, l0).contiguous()
+    scores_1_to_0 = f_similar(q1, k0T, kernel_size2, kernel_size2, False) # [b*n_head, l1, l1, field]
+    scores_1 = torch.cat(
+        (
+            scores_1_to_0.view(b*n_head, -1, scores_1_to_0.shape[3]) + add_scalar,
+            scores_1_to_1.view(b*n_head, -1, scores_1_to_1.shape[3])
+        ),
+        dim=-1)
+    attention_probs1 = F.softmax(scores_1, dim=-1)
+    
+    if attention_dropout is not None:
+        # with get_cuda_rng_tracker().fork():
+            attention_probs0 = attention_dropout(attention_probs0)
+            attention_probs1 = attention_dropout(attention_probs1)
+        
+    # weighting for level 0
+    context0 = torch.matmul(attention_probs0, v0) # [b, n_head, s0, h]
+    # weighting for level 1
+    probs_1_to_1 = attention_probs1[:, :, -scores_1_to_1.shape[3]:].view_as(scores_1_to_1)
+    # context1_to_1 = f_weighting(v1, probs_1_to_1.contiguous(), kernel_size*2-1, kernel_size, True)
+    context1_to_1 = f_weighting(v1, probs_1_to_1.contiguous(), kernel_size*2-1, kernel_size, False)
+
+    context1 = context1_to_1.view(b, n_head * h, l1**2)
+    # weighting for cross attention
+    probs_1_to_0 = attention_probs1[:, :, :scores_1_to_0.shape[3]].view_as(scores_1_to_0)
+    v0_part = v0[:, :, -l0**2:].transpose(-1, -2).contiguous().view(b*n_head, h, l0, l0)
+    context1_to_0 = f_weighting(v0_part, probs_1_to_0.contiguous(), kernel_size2, kernel_size2, False)
+    context1_to_0 = context1_to_0.view(b, n_head * h, l1**2)
+    context1 = context1 + context1_to_0
+    return context0.transpose(1, 2).reshape(b, s0, h0), context1.transpose(-1, -2)

sr_pipeline/dsr_sampling.py

@@ -0,0 +1,159 @@
+# -*- encoding: utf-8 -*-
+'''
+@File    :   cuda2d_sampling.py
+@Time    :   2021/10/09 00:46:04
+@Author  :   Ming Ding 
+@Contact :   dm18@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+import os
+import sys
+import math
+import random
+from cv2 import reduce
+import torch
+
+import torch
+import torch.nn.functional as F
+import numpy as  np
+
+def top_k_logits_(logits, top_k=0, filter_value=-float('Inf')):
+    indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+    logits[indices_to_remove] = filter_value     
+    return logits
+
+class IterativeEntfilterStrategy:
+    def __init__(self, invalid_slices=[], temperature=1., topk=6):
+        self.invalid_slices = invalid_slices
+        self.temperature = temperature
+        self.topk = topk        
+        self.cluster_labels = torch.tensor(np.load('cluster_label2.npy'), device='cuda', dtype=torch.long)
+
+
+    def forward(self, logits_, tokens, temperature=None, entfilter=None, filter_topk=5, temperature2=None):
+        # In interative strategy, logits are of shape [batch_size, seq_length, hidden_size]
+        if temperature is None:
+            temperature = self.temperature 
+            
+        logits = logits_.float() / temperature
+        for invalid_slice in self.invalid_slices:
+            logits[..., invalid_slice] = -float('Inf')
+        logits = logits.view(-1, logits.shape[-1])
+        
+        rprobs = F.softmax(logits.float(), dim=-1)
+        c = self.cluster_labels.expand(*rprobs.shape)
+        cprobs = torch.zeros(logits.shape[0], 500, device=logits.device).scatter_add_(1, c, rprobs)
+    
+        best_scores, best_clusters = cprobs.topk(self.topk)
+        bz = logits.shape[0]
+        best_scores = best_scores / best_scores.sum(dim=-1, keepdim=True)
+        sampled_ids = torch.multinomial(best_scores, num_samples=1)
+        selected_clusters = torch.gather(best_clusters, dim=1, index=sampled_ids)
+        selected_mask = (self.cluster_labels.unsqueeze(0).expand(bz, -1) != selected_clusters) # cluster_labels [1, 20000] \in [0,500)
+        logits[selected_mask] = -65504
+        # for i in range(bz):
+        #     selected_cluster = best_clusters[i][torch.multinomial(best_scores[i] / best_scores[i].sum(), num_samples=1)]
+        #     logits[i, self.cluster_labels != selected_cluster] = -65504
+            
+        # logits = top_k_logits(logits, self.topk, self.top_p)
+        probs = F.softmax(logits.float()/0.6, dim=-1)  # float is essetial, due to a bug in Pytorch
+        pred = torch.multinomial(probs, num_samples=1).view(*logits_.shape[:2])
+        
+        assert tokens.shape[1] == pred.shape[1] + 1
+        tokens = torch.cat((tokens[:, :1], pred), dim=1)
+        return tokens
+
+def filling_sequence_dsr(
+        model, 
+        seq0,
+        seq1, 
+        warmup_steps=3,
+        block_hw=(4, 4),
+        strategy=IterativeEntfilterStrategy(topk=10),
+        ):
+    '''
+        seq: [PAD]... [ROI1] text ... [BOI1] {layout[0]} 1024 {layout[1]} [EOI1]
+            4095 {layout[2]} final_token.
+        Attention:
+        The sampling temperature are changing, temporally we hard code them here.
+        The temperature in the strategy is not used.
+    '''
+    assert hasattr(model, 'layout')
+    layout = model.layout
+    assert len(seq0.shape) == 2 and len(seq1.shape) == 2 \
+        and seq0.shape[0] == seq1.shape[0]
+    assert len(layout) == 3
+    assert seq1.shape[1] == layout[-1] - layout[-2] + 1
+    assert (seq1 >= 0).all() and (seq0 >= 0).all()
+    device = seq0.device
+    # concat and pad sequences
+    batch_size = seq0.shape[0]
+    n_pad = layout[1] - seq0.shape[1]
+    assert n_pad > 0, "You should truncate long input before filling."
+    seq = torch.cat((
+        torch.tensor([0]*n_pad, device=device, dtype=seq0.dtype)
+            .unsqueeze(0).expand(batch_size, n_pad),
+        seq0, seq1), dim=1) # [b, layout[-1]+1]
+    assert seq.shape[1] == layout[-1] + 1
+
+    # build initial tokens, attention_mask, and position_ids
+    tokens = seq.clone()
+    attention_mask = torch.ones(layout[1], layout[1]).to(device)
+    attention_mask[:layout[0], layout[0]:] = 0
+    attention_mask[n_pad:, :n_pad] = 0
+    attention_mask = attention_mask.type_as(next(model.parameters())) # if fp16
+    position_ids = torch.cat((
+        torch.zeros(n_pad, dtype=torch.long),
+        torch.arange(0, layout[0] - n_pad), 
+        torch.arange(513, 513 + layout[1] - layout[0]),
+        torch.arange(1024, 1024+layout[2]-layout[1]))).to(device)
+    log_attention_weights = torch.zeros(layout[1], layout[1], 
+            device=device).type_as(next(model.parameters()))
+    log_attention_weights[layout[0]:, n_pad:layout[0]] = 0.
+
+    # prepare for interation
+    unfixed = (tokens < 0) # just init an all-False tensor
+    unfixed[:, -layout[-1] + layout[-2]:] = True
+    
+    ll, rr = block_hw
+    edge_len = int(math.sqrt(layout[-1] - layout[-2]) + 1e-4)
+    num_steps = warmup_steps + ll - 1 + rr
+    # interative refining
+    
+    # unfixed[..., -(layout[-1] - layout[-2]):].view(
+    #     batch_size, edge_len//ll, ll, edge_len//rr, rr)[:, :, :, :, -1] = False
+    
+    
+    ret = []
+    ret.append(tokens[:, layout[-2]+1:].clone())
+    for step_cnt in range(1, num_steps+1):
+        if step_cnt <= warmup_steps:
+            logits, *_dump = model(tokens[:,:-1], position_ids, attention_mask, log_attention_weights=log_attention_weights)
+            real_temp = 1.
+            new_tokens = strategy.forward(logits, tokens, real_temp)
+            tokens[unfixed] = new_tokens[unfixed]
+        else:
+            logits, *_dump = model(tokens[:,:-1], position_ids, attention_mask, log_attention_weights=log_attention_weights)
+            real_temp = 1.
+            new_tokens = strategy.forward(
+                logits, tokens, real_temp,
+                entfilter=1.3,
+                filter_topk=5,
+                temperature2=0.6
+            )
+            # tokens[unfixed] = new_tokens[unfixed]
+            # fixed tokens (update unfixed)
+            unfixed2 = (tokens > 10000000)
+            for x in range(min(ll, step_cnt - warmup_steps)):
+                y = step_cnt - warmup_steps - x - 1
+                if y < rr:
+                    unfixed[..., -(layout[-1] - layout[-2]):].view(
+                        batch_size, edge_len//ll, ll, edge_len//rr, rr)[:, :, x, :, y] = False
+                    unfixed2[..., -(layout[-1] - layout[-2]):].view(
+                        batch_size, edge_len//ll, ll, edge_len//rr, rr)[:, :, x, :, y] = True
+            tokens[unfixed2] = new_tokens[unfixed2]
+                
+        ret.append(tokens[:, layout[-2]+1:].clone())
+
+    return ret

sr_pipeline/iterative_sr.py

@@ -0,0 +1,118 @@
+# -*- encoding: utf-8 -*-
+'''
+@File    :   iterative_sr.py
+@Time    :   2022/03/02 15:57:45
+@Author  :   Ming Ding 
+@Contact :   dm18@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+import os
+import sys
+import math
+import random
+
+# here put the import lib
+import os
+import sys
+import math
+import random
+from PIL import ImageEnhance, Image
+
+import torch
+import argparse
+from torchvision import transforms
+
+from SwissArmyTransformer.training.model_io import load_checkpoint
+from SwissArmyTransformer import get_args
+from .itersr_sampling import filling_sequence_itersr, IterativeEntfilterStrategy
+from SwissArmyTransformer.generation.utils import timed_name, save_multiple_images, generate_continually
+
+from .itersr_model import ItersrModel
+
+from icetk import icetk as tokenizer
+
+class IterativeSuperResolution:
+    def __init__(self, args, path, max_bz=4, shared_transformer=None):
+        args.load = path
+        args.kernel_size = 5
+        args.kernel_size2 = 5
+        args.new_sequence_length = 4624
+        args.layout = [16,3616]
+        
+        model = ItersrModel(args, transformer=shared_transformer)
+        if args.fp16:
+            model = model.half()
+        
+        load_checkpoint(model, args) # on cpu
+        model.eval()
+        self.model = model.cuda()
+
+        # save cpu weights
+        self.saved_weights = dict((k,v.cpu()) 
+            for k, v in model.named_parameters()
+            if 'transformer' in k
+        )
+        
+        invalid_slices = [slice(tokenizer.num_image_tokens, None)]
+    
+        self.strategy = IterativeEntfilterStrategy(invalid_slices,
+            temperature=args.temp_all_itersr, topk=args.topk_itersr)
+        self.max_bz = max_bz
+
+    def _restore_transformer_from_cpu(self, non_blocking=False):
+        for k, v in self.model.named_parameters():
+            if k in self.saved_weights:
+                v.copy_(self.saved_weights[k])
+        
+    def __call__(self, text_tokens, image_tokens, enhance=False, input_mask=None):
+        if len(text_tokens.shape) == 1:
+            text_tokens.unsqueeze_(0)
+        text_tokens = text_tokens.clone()[..., :16]
+        if len(image_tokens.shape) == 1:
+            image_tokens.unsqueeze_(0)
+        if enhance:
+            new_image_tokens = []
+            for big_img in image_tokens:
+                decoded = tokenizer.decode(image_ids=big_img).squeeze(0)
+                ndarr = decoded.mul(255).add_(0.5).clamp_(0, 255).permute(1, 2, 0).to('cpu', torch.uint8).numpy()
+                image_pil_raw = ImageEnhance.Sharpness(Image.fromarray(ndarr))
+                big_img2 = tokenizer.encode(image_pil=image_pil_raw.enhance(1.5), image_size=480).view(-1)
+                new_image_tokens.append(big_img2)
+            image_tokens = torch.stack(new_image_tokens)
+        print('Converting Itersr model...')
+        self._restore_transformer_from_cpu()
+        model = self.model
+        print('iterative super-resolution...')
+        output_list = []
+        for tim in range(max(text_tokens.shape[0] // self.max_bz, 1)): 
+                big_img = image_tokens[tim*self.max_bz:(tim+1)*self.max_bz]
+                text_seq = text_tokens[tim*self.max_bz:(tim+1)*self.max_bz]
+                mask_raw = torch.tensor(
+                    [
+                        -1, 0, 1, 2, 3, 4,
+                        0, -1, 2, -1, -2, 5,
+                        1, -2, 3, 4, 5, 6,
+                        2, 3, 4, 5, -1, 1,
+                        3, -1, -2, 0, -1, 2,
+                        4, 5, 6, 1, 3, -2
+                    ]
+                ).view(1, 6, 1, 6).expand(10, 6, 10, 6).reshape(-1).contiguous()
+
+                topks = [60, 40, 40, 40, 20, 20, 10]
+
+                for mask_ratio in range(1, 7):  
+                    self.strategy.topk = topks[mask_ratio]
+                    mask = (mask_raw.to(big_img.device) >= mask_ratio)
+                    if input_mask is not None:
+                        mask = mask & input_mask
+                    big_img.masked_fill_(mask, tokenizer['<start_of_image>'])
+                    seq1 = big_img
+                    output1 = filling_sequence_itersr(model, text_seq, seq1, 
+                        warmup_steps=1, block_hw=(1, 0),
+                        strategy=self.strategy
+                        )
+                    big_img = output1
+                    print(f'Iter {mask_ratio} times.')
+                output_list.append(output1.clone())
+        return torch.cat(output_list, dim=0)

sr_pipeline/itersr_model.py

@@ -0,0 +1,232 @@
+# -*- encoding: utf-8 -*-
+'''
+@File    :   itersr_model.py
+@Time    :   2021/10/02 01:36:32
+@Author  :   Ming Ding 
+@Contact :   dm18@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+import os
+import sys
+import math
+import random
+import torch
+import torch.nn.functional as F
+
+
+from SwissArmyTransformer.model.base_model import BaseModel, BaseMixin
+
+from SwissArmyTransformer.mpu.utils import sqrt
+from deepspeed.runtime.activation_checkpointing.checkpointing import get_cuda_rng_tracker
+from SwissArmyTransformer.mpu import ColumnParallelLinear, RowParallelLinear
+from SwissArmyTransformer.model.transformer import unscaled_init_method, split_tensor_along_last_dim
+
+class PositionEmbeddingMixin(BaseMixin):
+    def __init__(self, additional_sequence_length, hidden_size,
+                 init_method_std=0.02, reinit_slice=slice(512, 512+400)
+                 ):
+        super(PositionEmbeddingMixin, self).__init__()
+        self.reinit_slice = reinit_slice
+        self.position_embeddings = torch.nn.Embedding(additional_sequence_length, hidden_size)
+        torch.nn.init.normal_(self.position_embeddings.weight, mean=0.0, std=init_method_std)
+
+    def reinit(self, parent_model=None):
+        old_weights = self.transformer.position_embeddings.weight.data[self.reinit_slice]
+        old_len, hidden_size = old_weights.shape
+        assert hidden_size == self.position_embeddings.weight.shape[-1]
+        old_edge, new_edge = sqrt(old_len), sqrt(self.position_embeddings.weight.shape[-2])
+        assert new_edge % old_edge == 0
+        self.position_embeddings.weight.data.view(new_edge // old_edge, old_edge, new_edge // old_edge, old_edge, hidden_size).copy_(old_weights.view(1, old_edge, 1, old_edge, hidden_size))
+
+class ItersrModel(BaseModel):
+    def __init__(self, args, transformer=None):
+        super().__init__(args, transformer=transformer)
+        self.original_sequence_length = args.max_sequence_length
+        additional_seqlen = args.new_sequence_length - args.max_sequence_length
+        self.add_mixin('extra_position_embedding', PositionEmbeddingMixin(
+            additional_seqlen, args.hidden_size
+        ))
+        # self.add_mixin('attention_plus', AttentionMixin(
+        #     num_layers=args.num_layers,
+        #     hidden_size=args.hidden_size
+        # ))
+        self.layout = args.layout
+        # [PAD]... [ROI1] text ... [BOI1] {layout[0]} 1024 {layout[1]} [EOI1] 4095 {layout[2]}
+        self.kernel_size = args.kernel_size
+        self.kernel_size2 = args.kernel_size2
+        self.log_attention_weights = None
+    
+    def position_embedding_forward(self, position_ids, **kw_args):
+        position = position_ids[..., :self.layout[0]]
+        position_plus = position_ids[..., self.layout[0]:] - self.original_sequence_length
+        position_embeddings = torch.cat(
+                (
+                    self.transformer.position_embeddings(position),
+                    self.get_mixin('extra_position_embedding').position_embeddings(position_plus)
+                ),
+                dim=-2
+            )
+        return position_embeddings
+        
+    def attention_forward(self, hidden_states, mask, 
+                        layer_id=None, log_attention_weights=None, **kw_args):
+        attn_module = self.transformer.layers[layer_id].attention
+        # base model qkv
+        mixed_raw_layer = attn_module.query_key_value(hidden_states)
+        q0, k0, v0 = split_tensor_along_last_dim(mixed_raw_layer[:, :self.layout[0]], 3)
+        # cuda2d model qkv
+        q1, k1, v1 = split_tensor_along_last_dim(mixed_raw_layer[:, self.layout[0]:], 3)
+        
+        dropout_fn = attn_module.attention_dropout if self.training else None
+
+        # cuda2d attention
+        context_layer = sparse_attention_2d_text(
+                q0, k0, v0,
+                q1, k1, v1,
+                mask,
+                n_head=attn_module.num_attention_heads_per_partition,
+                text_len=self.layout[0],
+                kernel_size=self.kernel_size,
+                attention_dropout=dropout_fn,
+                log_attention_weights=log_attention_weights,
+            )
+
+        output = attn_module.dense(context_layer)
+        
+        return output
+    
+    def final_forward(self, logits, **kwargs):
+        logits_parallel = logits
+        logits_parallel = torch.nn.functional.linear(logits_parallel, self.transformer.word_embeddings.weight[:20000]).float()
+        # logits_parallel = torch.nn.functional.linear(logits_parallel, self.transformer.word_embeddings.weight[:20000])
+        return logits_parallel
+    
+    # def disable_untrainable_params(self):
+    #     self.transformer.requires_grad_(False)
+    
+    @classmethod
+    def add_model_specific_args(cls, parser):
+        group = parser.add_argument_group('Cuda2dModel', 'cuda2d model configurations')
+        group.add_argument("--kernel-size", type=int, default=5)
+        group.add_argument("--kernel-size2", type=int, default=5)
+        group.add_argument("--layout", type=str, default='16,3616')
+        group.add_argument("--new-sequence-length", type=int, default=4096)
+        return parser
+
+def sparse_attention_2d_text(q0, k0, v0, q1, k1, v1, attention_mask, n_head, text_len, kernel_size=9,  attention_dropout=None, log_attention_weights = None, **kwargs):
+    '''
+    q0, k0, v0: [batch_size, 16, hidden_size]
+    q1, k1, v1: [batch_size, 3600, hidden_size]
+    n_head: int
+    attention_mask: [batch_size, 16]
+    '''
+    from SwissArmyTransformer.ops.local_attention_function import f_similar, f_weighting
+    b, s0, h0 = q0.shape
+    b, s1, h1 = q1.shape
+    h, l1 = h0 // n_head, sqrt(s1)
+    assert attention_mask.shape[-1] == s0, f"Mask Shape: {attention_mask.shape}"
+
+    q0 = q0.reshape(b, s0, n_head, h).permute(0, 2, 1, 3)
+    v0 = v0.reshape(b, s0, n_head, h).permute(0, 2, 1, 3)
+    k0T = k0.reshape(b, s0, n_head, h).permute(0, 2, 3, 1)
+    
+    # standard attention for level 0
+    attention_scores = torch.matmul(q0 / math.sqrt(q0.shape[-1]), k0T)
+    
+    attention_scores = torch.mul(attention_scores, attention_mask) - \
+                    10000.0 * (1.0 - attention_mask)
+    
+    attention_probs0 = F.softmax(attention_scores, dim=-1)
+    
+    # local attention for level 1
+    q1 = (q1.view(b, s1, n_head, h1 // n_head).permute(0, 2, 3, 1) / math.sqrt(h1//n_head)).contiguous().view(b*n_head, h1//n_head, l1, l1)
+    k1 = k1.view(b, s1, n_head, h1 // n_head).permute(0, 2, 3, 1).contiguous().view(b*n_head, h1//n_head, l1, l1)
+    v1 = v1.view(b, s1, n_head, h1 // n_head).permute(0, 2, 3, 1).contiguous().view(b*n_head, h1//n_head, l1, l1)
+    scores_1_to_1 = f_similar(q1, k1, kernel_size*2-1, kernel_size, False)    
+
+    # cross attention
+    scores_1_to_0 = torch.matmul(q1.view(b, n_head, h, s1).transpose(-1, -2), k0T)
+    if log_attention_weights is not None:
+        scores_1_to_0 += log_attention_weights
+    scores_1_to_0 = torch.mul(scores_1_to_0, attention_mask) - \
+                    10000.0 * (1.0 - attention_mask)
+    scores_1 = torch.cat(
+        (
+            scores_1_to_0.view(b*n_head, s1, s0),
+            scores_1_to_1.view(b*n_head, -1, scores_1_to_1.shape[3])
+        ),
+        dim=-1)
+    attention_probs1 = F.softmax(scores_1, dim=-1)
+    
+    if attention_dropout is not None:
+        with get_cuda_rng_tracker().fork():
+            attention_probs1 = attention_dropout(attention_probs1)
+        
+    # weighting for level 0
+    context0 = torch.matmul(attention_probs0, v0) # [b, n_head, s0, h]
+    # weighting for level 1
+    probs_1_to_1 = attention_probs1[:, :, -scores_1_to_1.shape[3]:].view_as(scores_1_to_1)
+    context1_to_1 = f_weighting(v1, probs_1_to_1.contiguous(), kernel_size*2-1, kernel_size, False)
+
+    context1 = context1_to_1.view(b, n_head, h, l1**2)
+    # weighting for cross attention
+    probs_1_to_0 = attention_probs1[:, :, :scores_1_to_0.shape[3]].view(b, n_head, -1, scores_1_to_0.shape[3])
+    
+    context1_to_0 = torch.matmul(probs_1_to_0, v0)
+    context1 = context1.transpose(-1, -2) + context1_to_0
+    
+    output = torch.cat((context0, context1), dim=2).transpose(1, 2).reshape(b, s0+s1, h0)
+
+    return output
+
+def sparse_attention_2d_notext(q0, k0, v0, q1, k1, v1, attention_mask, n_head, text_len, kernel_size=9,  attention_dropout=None, log_attention_weights = None, **kwargs):
+    '''
+    q0, k0, v0: [batch_size, 16, hidden_size]
+    q1, k1, v1: [batch_size, 3600, hidden_size]
+    n_head: int
+    attention_mask: [batch_size, 16]
+    '''
+    from SwissArmyTransformer.mpu.local_attention_function import f_similar, f_weighting
+    b, s0, h0 = q0.shape
+    b, s1, h1 = q1.shape
+    h, l1 = h0 // n_head, sqrt(s1)
+    assert len(attention_mask.shape) == 4 and attention_mask.shape[-1] == s0, f"Mask Shape: {attention_mask.shape}"
+
+    q0 = q0.reshape(b, s0, n_head, h).permute(0, 2, 1, 3)
+    v0 = v0.reshape(b, s0, n_head, h).permute(0, 2, 1, 3)
+    k0T = k0.reshape(b, s0, n_head, h).permute(0, 2, 3, 1)
+    
+    # standard attention for level 0
+    attention_scores = torch.matmul(q0 / math.sqrt(q0.shape[-1]), k0T)
+    
+    attention_scores = torch.mul(attention_scores, attention_mask) - \
+                    10000.0 * (1.0 - attention_mask)
+    
+    attention_probs0 = F.softmax(attention_scores, dim=-1)
+    
+    # local attention for level 1
+    q1 = (q1.view(b, s1, n_head, h1 // n_head).permute(0, 2, 3, 1) / math.sqrt(h1//n_head)).contiguous().view(b*n_head, h1//n_head, l1, l1)
+    k1 = k1.view(b, s1, n_head, h1 // n_head).permute(0, 2, 3, 1).contiguous().view(b*n_head, h1//n_head, l1, l1)
+    v1 = v1.view(b, s1, n_head, h1 // n_head).permute(0, 2, 3, 1).contiguous().view(b*n_head, h1//n_head, l1, l1)
+    scores_1_to_1 = f_similar(q1, k1, kernel_size*2-1, kernel_size, False)    
+
+    attention_probs1 = F.softmax(scores_1_to_1, dim=-1)
+    
+    if attention_dropout is not None:
+        with get_cuda_rng_tracker().fork():
+            attention_probs1 = attention_dropout(attention_probs1)
+        
+    # weighting for level 0
+    context0 = torch.matmul(attention_probs0, v0) # [b, n_head, s0, h]
+    # weighting for level 1
+    probs_1_to_1 = attention_probs1
+    context1_to_1 = f_weighting(v1, probs_1_to_1.contiguous(), kernel_size*2-1, kernel_size, False)
+
+    context1 = context1_to_1.view(b, n_head, h, l1**2)
+    # weighting for cross attention    
+    context1 = context1.transpose(-1, -2)
+    
+    output = torch.cat((context0, context1), dim=2).transpose(1, 2).reshape(b, s0+s1, h0)
+
+    return output

sr_pipeline/itersr_sampling.py

@@ -0,0 +1,168 @@
+# -*- encoding: utf-8 -*-
+'''
+@File    :   itersr_sampling.py
+@Time    :   2022/03/03 14:24:28
+@Author  :   Ming Ding 
+@Contact :   dm18@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+import os
+import sys
+import math
+import random
+import numpy as np
+
+import torch
+import torch.nn.functional as F
+from icetk import icetk as tokenizer
+
+def top_k_logits_(logits, top_k=0, filter_value=-float('Inf')):
+    indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+    logits[indices_to_remove] = filter_value     
+    return logits
+
+# class IterativeEntfilterStrategy:
+#     def __init__(self, invalid_slices=[], temperature=1., topk=10):
+#         self.invalid_slices = invalid_slices
+#         self.temperature = temperature
+#         self.topk = topk        
+#         self.cluster_labels = torch.tensor(np.load('cluster_label.npy'), device='cuda', dtype=torch.long)
+
+
+#     def forward(self, logits_, tokens, temperature=None, entfilter=None, filter_topk=5, temperature2=None):
+#         # In interative strategy, logits are of shape [batch_size, seq_length, hidden_size]
+#         if temperature is None:
+#             temperature = self.temperature 
+            
+#         logits = logits_.float() / temperature
+#         for invalid_slice in self.invalid_slices:
+#             logits[..., invalid_slice] = -float('Inf')
+#         logits = logits.view(-1, logits.shape[-1])
+        
+#         rprobs = F.softmax(logits.float(), dim=-1)
+#         c = self.cluster_labels.expand(*rprobs.shape)
+#         cprobs = torch.zeros(logits.shape[0], 500, device=logits.device).scatter_add_(1, c, rprobs)
+    
+#         best_scores, best_clusters = cprobs.topk(self.topk)
+#         bz = logits.shape[0]
+#         best_scores = best_scores / best_scores.sum(dim=-1, keepdim=True)
+#         sampled_ids = torch.multinomial(best_scores, num_samples=1)
+#         selected_clusters = torch.gather(best_clusters, dim=1, index=sampled_ids)
+#         selected_mask = (self.cluster_labels.unsqueeze(0).expand(bz, -1) != selected_clusters) # cluster_labels [1, 20000] \in [0,500)
+#         logits[selected_mask] = -65504
+#         # for i in range(bz):
+#         #     selected_cluster = best_clusters[i][torch.multinomial(best_scores[i] / best_scores[i].sum(), num_samples=1)]
+#         #     logits[i, self.cluster_labels != selected_cluster] = -65504
+            
+#         # logits = top_k_logits(logits, self.topk, self.top_p)
+#         probs = F.softmax(logits.float(), dim=-1)  # float is essetial, due to a bug in Pytorch
+#         pred = torch.multinomial(probs, num_samples=1).view(*logits_.shape[:2])
+        
+#         assert tokens.shape[1] == pred.shape[1]
+#         tokens = pred
+#         return tokens
+
+class IterativeEntfilterStrategy:
+    def __init__(self, invalid_slices=[], temperature=1., topk=10):
+        self.invalid_slices = invalid_slices
+        self.temperature = temperature
+        self.topk = topk
+
+    def forward(self, logits, tokens, temperature=None, entfilter=None, filter_topk=5, temperature2=None):
+        # In interative strategy, logits are of shape [batch_size, seq_length, hidden_size]
+        if temperature is None:
+            temperature = self.temperature 
+        # check entropy filter
+        # if entfilter is not None:
+        #     assert temperature2 is not None
+        #     topraw = (torch.topk(logits, filter_topk, dim=-1)[0]).softmax(dim=-1)
+        #     ent = -(topraw * topraw.log()).sum(dim=-1) # [batch_size, seq_length]
+        #     temperature = torch.tensor([[[temperature - temperature2]]], device=logits.device).expand(*logits.shape[:2], 1) * (ent > entfilter).unsqueeze(-1) + temperature2
+            
+        logits = logits.float() / temperature
+        for invalid_slice in self.invalid_slices:
+            logits[..., invalid_slice] = -float('Inf')
+        
+        # debiased topk
+        # probs = F.softmax(logits, dim=-1)
+        # tk_value, tk_idx = torch.topk(probs, self.topk, dim=-1)
+        # pred = torch.multinomial(probs.view(-1, logits.shape[-1]), num_samples=1).view(*logits.shape[:2], 1)
+        # edge_idx = tk_idx[:, :, -1:]
+        # edge_value = tk_value[:, :, -1:]
+        # edge_mask = probs.gather(dim=-1, index=pred) < edge_value
+        # pred[edge_mask] = edge_idx[edge_mask] # replace outliers as the "filter_topk"-th token
+        # pred.squeeze_(-1) # [batch_size, seq_length]
+        
+        top_k_logits_(logits, self.topk)
+        probs = F.softmax(logits, dim=-1)
+        pred = torch.multinomial(probs.view(-1, logits.shape[-1]), num_samples=1).view(*logits.shape[:2], 1)
+        pred.squeeze_(-1)
+        
+        assert tokens.shape[1] == pred.shape[1]
+        tokens = pred
+        return tokens
+
+def filling_sequence_itersr(
+        model, 
+        seq0,
+        seq1, 
+        warmup_steps=3,
+        block_hw=(4, 4),
+        strategy=IterativeEntfilterStrategy(topk=10),
+        ):
+    '''
+        seq: [PAD]... [ROI1] text ... [BOI1] {layout[0]} 1024 {layout[1]} [EOI1]
+            4095 {layout[2]} final_token.
+        Attention:
+        The sampling temperature are changing, temporally we hard code them here.
+        The temperature in the strategy is not used.
+    '''
+    assert hasattr(model, 'layout')
+    layout = model.layout
+    
+    device = seq0.device
+    # concat and pad sequences
+    batch_size = seq0.shape[0]
+    n_pad = layout[0] - seq0.shape[1]
+    assert n_pad >= 0, "You should truncate long input before filling."
+    seq = torch.cat((
+        torch.tensor([0]*n_pad, device=device, dtype=seq0.dtype)
+            .unsqueeze(0).expand(batch_size, n_pad),
+        seq0, seq1), dim=1) # [b, layout[-1]+1]
+    assert seq.shape[1] == layout[-1]
+
+    # build initial tokens, attention_mask, and position_ids
+    tokens = seq.clone()
+    attention_mask = torch.ones(layout[0]).to(device)
+    attention_mask[:n_pad] = 0
+    attention_mask = attention_mask.unsqueeze(0).type_as(next(model.parameters())) # if fp16
+    position_ids = torch.cat((
+        torch.zeros(n_pad, dtype=torch.long),
+        torch.arange(0, layout[0] - n_pad), 
+        torch.arange(1024, 1024+layout[1]-layout[0]))).to(device)
+    log_attention_weights = torch.zeros(layout[0], device=device).type_as(next(model.parameters()))
+    log_attention_weights[n_pad:layout[0]] = 0.
+    log_attention_weights = log_attention_weights.unsqueeze(0)
+
+    # prepare for interation
+    unfixed = (tokens == tokenizer['<start_of_image>']) 
+    ll, rr = block_hw
+    edge_len = int(math.sqrt(layout[-1] - layout[-2]) + 1e-4)
+    num_steps = 1
+    # interative refining
+    
+    # unfixed[..., -(layout[-1] - layout[-2]):].view(
+    #     batch_size, edge_len//ll, ll, edge_len//rr, rr)[:, :, :, :, -1] = False
+    
+    
+    ret = []
+    # ret.append(tokens[:, layout[-2]:-1].clone())
+    for step_cnt in range(1, num_steps+1):
+        logits, *_dump = model(tokens, position_ids, attention_mask, log_attention_weights=log_attention_weights)
+        real_temp = 1.
+        new_tokens = strategy.forward(logits, tokens, real_temp)
+        tokens[unfixed] = new_tokens[unfixed]
+                
+        ret.append(tokens[:, layout[-2]:].clone())
+    return torch.cat(ret, dim=0)

sr_pipeline/sr_group.py

@@ -0,0 +1,49 @@
+# -*- encoding: utf-8 -*-
+'''
+@File    :   sr_group.py
+@Time    :   2022/04/02 01:17:21
+@Author  :   Ming Ding 
+@Contact :   dm18@mails.tsinghua.edu.cn
+'''
+
+# here put the import lib
+import os
+import sys
+import math
+import random
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from SwissArmyTransformer.resources import auto_create
+from .direct_sr import DirectSuperResolution
+from .iterative_sr import IterativeSuperResolution
+
+class SRGroup:
+    def __init__(self, args, home_path=None,):
+        dsr_path = auto_create('cogview2-dsr', path=home_path)
+        itersr_path = auto_create('cogview2-itersr', path=home_path)
+        dsr = DirectSuperResolution(args, dsr_path)
+        itersr = IterativeSuperResolution(args, itersr_path, shared_transformer=dsr.model.transformer)
+        self.dsr = dsr
+        self.itersr = itersr
+
+    def sr_base(self, img_tokens, txt_tokens):
+        assert img_tokens.shape[-1] == 400 and len(img_tokens.shape) == 2
+        batch_size = img_tokens.shape[0]
+        txt_len = txt_tokens.shape[-1]
+        if len(txt_tokens.shape) == 1:
+            txt_tokens = txt_tokens.unsqueeze(0).expand(batch_size, txt_len)
+        sred_tokens = self.dsr(txt_tokens, img_tokens)
+        iter_tokens = self.itersr(txt_tokens, sred_tokens[:, -3600:].clone())
+        return iter_tokens[-batch_size:]
+    
+    # def sr_patch(self, img_tokens, txt_tokens):
+    #     assert img_tokens.shape[-1] == 3600 and len(img_tokens.shape) == 2
+    #     batch_size = img_tokens.shape[0] * 9
+    #     txt_len = txt_tokens.shape[-1]
+    #     if len(txt_tokens.shape) == 1:
+    #         txt_tokens = txt_tokens.unsqueeze(0).expand(batch_size, txt_len)
+    #     img_tokens = img_tokens.view(img_tokens.shape[0], 3, 20, 3, 20).permute(0, 1, 3, 2, 4).reshape(batch_size, 400)
+    #     iter_tokens = self.sr_base(img_tokens, txt_tokens)
+    #     return iter_tokens