import math import copy import numpy as np import torch import torch.distributed as dist import torch.nn.functional as F def add_gumbel_noise(logits, temperature): ''' The Gumbel max is a method for sampling categorical distributions. According to arXiv:2409.02908, for MDM, low-precision Gumbel Max improves perplexity score but reduces generation quality. Thus, we use float64. ''' if math.isclose(temperature, 0.0): return logits logits = logits.to(torch.float64) noise = torch.rand_like(logits, dtype=torch.float64) gumbel_noise = (- torch.log(noise)) ** temperature return logits.exp() / gumbel_noise def get_num_transfer_tokens(mask_index, steps): ''' In the reverse process, the interval [0, 1] is uniformly discretized into steps intervals. Furthermore, because LLaDA employs a linear noise schedule (as defined in Eq. (8)), the expected number of tokens transitioned at each step should be consistent. This function is designed to precompute the number of tokens that need to be transitioned at each step. ''' mask_num = mask_index.sum(dim=1, keepdim=True) base = mask_num // steps remainder = mask_num % steps num_transfer_tokens = torch.zeros(mask_num.size(0), steps, device=mask_index.device, dtype=torch.int64) + base for i in range(mask_num.size(0)): num_transfer_tokens[i, :remainder[i]] += 1 return num_transfer_tokens def calculate_op_num(x, hidden_size=4096, mlp_hidden_size = 12288, vocab_size = 126464, num_hidden_layers=32, cache_length=0): cfg_factor = 1 qkv_ops = 4*x.shape[0]*hidden_size*hidden_size*x.shape[1]*2 attn_ops = x.shape[0]*(cache_length)*x.shape[1]*hidden_size*2 ffn_ops = 3*x.shape[0]*hidden_size*mlp_hidden_size*x.shape[1]*2 layer_ops = qkv_ops + attn_ops + ffn_ops op_num = cfg_factor * (num_hidden_layers*layer_ops + x.shape[0]*hidden_size*vocab_size*x.shape[1]*2) return op_num/1e12 class TokenArray: """ A token array to support read, update and expansion. We need to access the tokens that have been generated and write new tokens to the array. Some algorithms require to expand the token array. Parameters ---------- prompt : Torch.Tensor The array that contains the input prompt. gen_length : int The number of tokens to be generated. mask_id : int the mask id of the masked tokens device : Torch.Device The device where the token array is placed on. """ def __init__(self, prompt, gen_length, mask_id, eos_id, device): self.prompt = prompt.to(device) self.data = torch.full((prompt.shape[0], prompt.shape[1] + gen_length), mask_id, dtype=torch.long).to(device) self.data[:, :prompt.shape[1]] = prompt.clone() self.gen_length = gen_length self.eos_id = eos_id self.mask_id = mask_id @property def total_length(self): return self.prompt.shape[1] + self.gen_length @property def batch_size(self): return self.prompt.shape[0] @property def device(self): return self.data.device def expand(self, new_len): pass def get_generated_tokens(self): if self.batch_size == 1: return self.data[self.data != self.eos_id].unsqueeze(0) else: self.data[self.data == self.mask_id] = self.eos_id return self.data def select_seqs(self, idx): arr = copy.copy(self) arr.prompt = self.prompt[idx] arr.data = self.data[idx] return arr def __getitem__(self, idx): return self.data[idx] def __setitem__(self, idx, vals): self.data[idx] = vals class DistAlignedTokenArray: """ A token array to support read, update and expansion in the distributed setting. In this setting, each process still contains the full copy of the token array. The main difference from TokenArray is that this class makes sure that the length of the token array is rounded to the world size. Parameters ---------- prompt : Torch.Tensor The array that contains the input prompt. gen_length : int The number of tokens to be generated. mask_id : int the mask id of the masked tokens device : Torch.Device The device where the token array is placed on. rank : int The rank of the process world_size : int The number of processes. """ def __init__(self, prompt, gen_length, mask_id, eos_id, device, rank, world_size): total_length = prompt.shape[1] + gen_length if total_length % world_size != 0: total_length = (total_length // world_size + 1) * world_size self.data = torch.full((prompt.shape[0], total_length), mask_id, dtype=torch.long).to(device) self.data[:, :prompt.shape[1]] = prompt.clone() self.orig_gen_length = gen_length self.gen_length = total_length - prompt.shape[1] self.prompt = prompt self.eos_id = eos_id self.mask_id = mask_id @property def total_length(self): return self.prompt.shape[1] + self.gen_length @property def device(self): return self.data.device def get_generated_tokens(self): return self.data[self.data != self.eos_id].unsqueeze(0) def expand(self, new_len): pass def __getitem__(self, idx): return self.data[idx] def __setitem__(self, idx, vals): self.data[idx] = vals class BlockLoc: """ The location of the block in the token array. """ def __init__(self, start, end): self.start = start self.end = end class BlockIterator: """ Block iterator This performs block-wise iteration on the input token array for diffusion decoding. Parameters ---------- x : TokenArray The token array that contains decoded tokens and stores the new generated tokens block_length : int The length of the block start_block_align : bool Align the first decoding block to the block size. The first block may overlap with the prompt. """ def __init__(self, x, block_length, start_block_align=False): self.x = x self.iter = 0 self.block_length = block_length self.start_block_align = start_block_align if start_block_align: self.first_block_start = self._get_first_block_start() else: self.first_block_start = self.x.prompt.shape[1] def _get_first_block_start(self): prompt = self.x.prompt non_mask_number = (prompt != self.x.mask_id).sum(dim=-1).min().item() start = ((non_mask_number) // self.block_length) * self.block_length return start def __iter__(self): self.iter = 0 return self def __next__(self): current_block_start = self.first_block_start + self.iter * self.block_length if current_block_start >= self.x.total_length: raise StopIteration current_block_end = min(current_block_start + self.block_length, self.x.total_length) assert current_block_end <= self.x.total_length self.iter += 1 return BlockLoc(current_block_start, current_block_end), self.x[:, current_block_start:current_block_end] class BlockDiffusionIterator(): """ Block iterator This performs block-wise iteration on the input token array for diffusion decoding. Parameters ---------- x : TokenArray The token array that contains decoded tokens and stores the new generated tokens block_length : int The length of the block """ def __init__(self, x, block_length): self.x = x self.iter = 0 self.block_length = block_length self.first_block_start = self._get_first_block_start() def _get_first_block_start(self): prompt = self.x.prompt non_mask_number = (prompt != self.x.mask_id).sum(dim=-1).min().item() start = ((non_mask_number) // self.block_length) * self.block_length return start def __iter__(self): self.iter = 0 return self def __next__(self): current_block_start = self.first_block_start + self.iter * self.block_length if current_block_start >= self.x.total_length: raise StopIteration current_block_end = min(current_block_start + self.block_length, self.x.total_length) assert current_block_end <= self.x.total_length self.iter += 1 return BlockLoc(current_block_start, current_block_end), self.x[current_block_start:current_block_end] class BlockIteratorFactory: """ Iterator factory This generates iterators for DiffusionLLM to iterate over a sequence. Parameters ---------- start_block_align : bool Align the first decoding block to the block size. The first block may overlap with the prompt. use_block_diffusion: bool If this flag set to True, the block diffusion iteration will be used and start_block_algin will be ignored. Returns ------- BlockIterator : the block iterator. """ def __init__(self, start_block_align=False, use_block_diffusion=False): self._start_block_align = start_block_align self._use_bd = use_block_diffusion def create(self, x, block_length): if self._use_bd: return BlockDiffusionIterator(x, block_length) else: return BlockIterator(x, block_length, start_block_align=self._start_block_align) class KVCache: """ The KV-cache Parameters ---------- past_key_values : List[torch.Tensor] The keys and values of each transformer layer. """ def __init__(self, past_key_values, backend='vllm', length=2048, cache_align_size=128): if backend == 'vllm': assert len(past_key_values) % 2 == 0 self._data = past_key_values else: self.cache_align_size = cache_align_size assert len(past_key_values) % 2 == 0 self._raw_data = past_key_values self._consolidate_raw() n = -(-(self._raw_data.shape[4] + 64) // self.cache_align_size) next_pow2 = 1 << (n - 1).bit_length() if n > 1 else 1 self.length = next_pow2 * self.cache_align_size device = self._raw_data.device num_layer, _, batch_size, num_heads, seq_len, hidden_dim = self._raw_data.shape self._data = torch.zeros(num_layer, 2, batch_size, num_heads, self.length, hidden_dim, device=device, dtype=torch.bfloat16) self._data[:, :, :, :, :seq_len] = self._raw_data def consolidate(self): if isinstance(self._data, torch.Tensor): return num_layers = len(self._data) // 2 inner_shape = self._data[0].shape # The shape is [num_layers, 2, batch_size, num_heads, seq_len, hidden_dim] self._data = torch.stack(self._data, dim=0).reshape(num_layers, 2, *inner_shape) def _consolidate_raw(self): if isinstance(self._raw_data, torch.Tensor): return num_layers = len(self._raw_data) // 2 inner_shape = self._raw_data[0].shape # The shape is [num_layers, 2, batch_size, num_heads, seq_len, hidden_dim] self._raw_data = torch.stack(self._raw_data, dim=0).reshape(num_layers, 2, *inner_shape) @property def num_layers(self): assert isinstance(self._data, torch.Tensor) return self._data.shape[0] @property def seq_len(self): assert isinstance(self._data, torch.Tensor) return self._data.shape[4] def get_keys(self, layer_idx): """ Get the keys of a transformer layer. """ assert isinstance(self._data, torch.Tensor) return self._data[layer_idx][0] def get_values(self, layer_idx): """ Get the values of a transformer layer. """ assert isinstance(self._data, torch.Tensor) return self._data[layer_idx][1] def update(self, key_states, val_states, layer_idx, replace_position=None, backend='vllm'): """ Update the keys and values of a transformer layer. Parameters ---------- key_states : torch.Tensor The keys in a block of tokens. The shape is [batch_size, num_heads, seq_len, hidden_dim] val_states : torch.Tensor The values in a block of tokens. The shape is [batch_size, num_heads, seq_len, hidden_dim] layer_idx : int The index of the transformer layer replace_position : Tuple[int] The start and the end position where keys and values should be updated. Returns ------- torch.Tensor: the new keys for the entire sequence of the transformer layer. torch.Tensor: the new values for the entire sequence of the transformer layer. """ if backend == 'vllm': # This is dual cache. if replace_position is not None: keys = self.get_keys(layer_idx).slice_scatter(key_states, dim=2, start=replace_position[0], end=replace_position[1]) values = self.get_values(layer_idx).slice_scatter(val_states, dim=2, start=replace_position[0], end=replace_position[1]) else: # This is prefix cache. keys = torch.cat([self.get_keys(layer_idx), key_states], dim=2) values = torch.cat([self.get_values(layer_idx), val_states], dim=2) else: cache_length = self.get_keys(layer_idx).shape[2] block_length = key_states.shape[2] keys = self.get_keys(layer_idx).slice_scatter(key_states, dim=2, start=cache_length - block_length, end=cache_length) values = self.get_values(layer_idx).slice_scatter(val_states, dim=2, start=cache_length - block_length, end=cache_length) return keys, values class DiffusionKVCacheManager: """ KV-cache for diffusion LLM. The KV-cache caches the KV of the tokens before and after the block that is being decoded. Because diffusion LLM uses bidirectional attention, the KV-cache has to be updated frequently in the diffusion iterations. This class basically defines the KV-cache update policy in the diffusion iterations. This includes the locations where keys and values can be updated and the frequency of the keys and values can be updated. """ def __init__(self, cache_update_freq=None, cache_type='prefix', backend='vllm', max_length=2048): self.past_key_values = None self.block_start = None self.block_end = None self.cache_update_freq = cache_update_freq assert cache_type in ['prefix', 'dual'] self.cache_type = cache_type self.backend=backend self.max_length = max_length def require_update(self, iter_no, block_start, block_end): """ require to update the kv-cache. Parameters ---------- iter_no : int The diffusion iteration number block_start : int The start of the block that is being decoded. block_end : int The end of the block that is being decoded. """ if self.past_key_values is None: _require_update = True # If self.cache_update_freq is not specified, the KV-cache is updated when we enter a new block. if self.cache_update_freq is None: _require_update = self.block_start != block_start or self.block_end != block_end else: # Otherwise, we update the KV-cache when we enter a new block or the specified number of # diffusion iterations is reached. _require_update = iter_no % self.cache_update_freq == 0 \ or (self.block_start != block_start or self.block_end != block_end) # TODO(zhengda) change update logic to block idx self.block_start = block_start self.block_end = block_end return _require_update def update(self, past_key_values, range_start=None, range_end=None): """ update the KV-cache Parameters ---------- past_key_values : List[torch.Tensor] The key values in all transformer layers. range_start : int The start of the range that is being updated. range_end : int The end of the range that is being updated. """ if isinstance(past_key_values, KVCache): self.past_key_values = past_key_values else: self.past_key_values = KVCache(past_key_values, self.backend, self.max_length) # We should make sure the kv-cache in all layers are converted into a tensor. self.past_key_values.consolidate() def range_update(self, past_key_values, range_start=0, range_end=0, block_length=32): """ update the KV-cache Parameters ---------- past_key_values : List[torch.Tensor] The key values in all transformer layers. range_start : int The start of the range that is being updated. range_end : int The end of the range that is being updated. """ if isinstance(past_key_values, KVCache): # raise ValueError("past_key_values should be a list of tensors") self.past_key_values = past_key_values else: num_layers = len(past_key_values) // 2 inner_shape = past_key_values[0].shape compact_kvcache = torch.stack(past_key_values, dim=0).reshape(num_layers, 2, *inner_shape) if block_length>0: self.past_key_values = KVCache(torch.cat((compact_kvcache[:, :, :, :, range_start:range_end-block_length], compact_kvcache[:, :, :, :, -block_length:]), dim=4), self.backend, self.max_length) else: self.past_key_values = KVCache(compact_kvcache[:, :, :, :, range_start:range_end], self.backend, self.max_length) # We should make sure the kv-cache in all layers are converted into a tensor. self.past_key_values.consolidate() # print(self.past_key_values._data.shape) def get_key_values(self, block_start, block_end): """ Get the key-values given the block that is being decoded. Parameters ---------- block_start : int The start of the block that is being decoded. block_end : int The end of the block that is being decoded. Returns ------- List[List[torch.Tensor]] : the key-values required to decode the specified block. torch.Tensor : the tensor indicates the valid locations in the returned key-values. """ # The key-value cache cannot be empty. assert self.past_key_values is not None if self.cache_type == 'prefix': replace_position = (int(block_start), int(self.past_key_values.seq_len)) else: replace_position = (int(block_start), int(block_end)) return self.past_key_values, replace_position class BlockDiffusionPrefixCacheManager(DiffusionKVCacheManager): """ KVcache manager of block diffusion. """ def get_key_values(self, block_start, block_end): # use prefix cache for block diffusion. return self.past_key_values, (int(block_start), int(block_end)) def extend_cache(self, end): """ When move to new block, extend the kvcache length from previous block end to new block end location. Parameters ---------- length : int The extended length (equvelent to block length) """ if self.backend == 'vllm': cur_kv_length = self.past_key_values._data.shape[-2] extended_cache = F.pad(self.past_key_values._data, pad=(0, 0, 0, end-cur_kv_length), mode='constant', value=0) self.past_key_values._data = extended_cache else: cur_kv_length = self.past_key_values._data.shape[-2] n = -(-end // self.past_key_values.cache_align_size) # 等价于 ceil(target / cache_align_size)，纯整数 next_pow2 = 1 << (n - 1).bit_length() if n > 1 else 1 aligned_end = next_pow2 * self.past_key_values.cache_align_size if aligned_end <= cur_kv_length: return extended_cache = F.pad(self.past_key_values._data, pad=(0, 0, 0, aligned_end-cur_kv_length), mode='constant', value=0) self.past_key_values._data = extended_cache class KVCacheFactory: """ KV-cache factory. This class generates KV-cache for the diffusion LLM when it runs diffusion iterations. """ def __init__(self, cache_type, cache_update_freq=None, is_bd_model=False, backend='vllm', max_length=2048): self.cache_type = cache_type self.cache_update_freq = cache_update_freq self.is_bd_model = is_bd_model self.backend = backend self.max_length = max_length def create(self): if self.is_bd_model: return BlockDiffusionPrefixCacheManager(cache_update_freq=self.cache_update_freq, cache_type=self.cache_type, backend=self.backend, max_length=self.max_length) else: return DiffusionKVCacheManager(cache_update_freq=self.cache_update_freq, cache_type=self.cache_type) def gather_sequence_block(partial_data, partial_start, partial_end, block_start, block_end, rank, world_size): """ Gather the wanted block data from the partitioned data. Each process contains a partition specified by `partial_start` and `partial_end`. The wanted block is located between `block_start` and `block_end`. We want to gather the data within the block range from the partitioned data. """ if partial_start >= block_end or partial_end <= block_start: # there is no overlap, nothing is needed from partial_data arr = partial_data[:, 0:0] elif block_start >= partial_start and block_end <= partial_end: # the needed block is within partial_data. arr = partial_data[:, (block_start - partial_start):(block_end - partial_start)] elif block_start <= partial_start and block_end >= partial_end: # the needed partition is within the block. arr = partial_data elif partial_start >= block_start and partial_end >= block_end: # the needed block is overlapped in the front of partial_data arr = partial_data[:, 0:(block_end - partial_start)] else: # the needed block is overlapped at the end of partial_data arr = partial_data[:, (block_start - partial_start):(partial_end - partial_start)] arr = arr.contiguous() shape_list = [ torch.zeros(len(arr.shape), dtype=torch.int64, device=partial_data.device) for _ in range(world_size) ] dist.all_gather(shape_list, torch.tensor(arr.shape, dtype=torch.int64, device=partial_data.device)) part_list = [ torch.zeros(*tuple(shape.tolist()), dtype=partial_data.dtype, device=partial_data.device) for shape in shape_list ] dist.all_gather(part_list, arr) return torch.cat(part_list, dim=1)