import math import torch from typing import Optional, Tuple from torch import nn from utils.nn.seq_utils import get_incremental_state, set_incremental_state, softmax, make_positions import torch.nn.functional as F # from flash_attn import flash_attn_qkvpacked_func, flash_attn_func DEFAULT_MAX_SOURCE_POSITIONS = 20000 DEFAULT_MAX_TARGET_POSITIONS = 20000 class RotaryEmbeddings(nn.Module): cos: torch.Tensor sin: torch.Tensor theta: torch.Tensor def __init__( self, width: int, *, seq_len: int = 4000, base: int = 10000, device: Optional[torch.device] = None, ): """Rotary embeddings (Su et al., 2021) layer. The rotary embedding will be precomputed for up to 'seq _len' positions. The embedding will be recomputed when a longer sequence is found in the input. :param width: Rotary embedding dimensionality, must be even. :param seq_len: Number of positons to initially precompute. :param base: The base used for Θ_i, determines the cycle length of the embeddings. :param device: Device on which the module is to be initialized. """ super().__init__() if width % 2: raise ValueError(f"Width of rotary embeddings must be even, was: {width}") # Ignore allocations on the meta device as we don't persist our buffer, # i.e., we don't expect the backing tensor to be replaced with pretrained weights. if device is not None and device.type == "meta": device = None # Θ_i = 10000^(-2(i-1)/d) theta = torch.pow( base, -torch.arange(0, width, 2, dtype=torch.float, device=device) / width ) self.register_buffer("theta", theta, persistent=False) self._create_rotary_embed(width=width, length=seq_len) def _create_rotary_embed(self, *, width: int, length: int): # mΘ position = torch.arange(length, device=self.theta.device).unsqueeze(1) m_theta = position * self.theta.unsqueeze(0) # We apply both sin and cos twice (see Eq 15, 34), but the ordering # is changed for compatibility with most common implementations. m_theta = torch.cat([m_theta, m_theta], dim=-1) re_cos = m_theta.cos().view([length, width]).half() re_sin = m_theta.sin().view([length, width]).half() self.register_buffer("cos", re_cos, persistent=False) self.register_buffer("sin", re_sin, persistent=False) def _rotate(self, input: torch.Tensor): """Rotate the input tensor by half of its innermost width. input (Tensor): array to rotate. RETURNS (Tensor): rotated array. Shapes: input - (..., width) output - (..., width) """ half_idx = input.shape[-1] // 2 input_1 = -input[..., half_idx:] input_2 = input[..., :half_idx] return torch.cat([input_1, input_2], dim=-1) def forward(self, input: torch.Tensor, *, positions: Optional[torch.Tensor] = None): """ Apply rotary embeddings to an array. :param input: Array to apply the rotary embeddings to. :param positions: positions of the inputs. If no positions are provided, they are assumed to be [0, seq_len). :return: Array with the rotary embeddings applied. Shapes: input - (batch_size, num_heads, seq_len, width_per_head) positions - (batch_size, seq_len) output - (batch_size, num_heads, seq_len, width_per_head) """ batch_size, _, seq_len, width = input.shape if positions is None: # Fastpath: positions from [0..seq_len), avoid indexing. if self.cos.size(-2) < seq_len: self._create_rotary_embed(width=width, length=seq_len) rot_cos = self.cos[:seq_len, :].view(1, 1, seq_len, width) rot_sin = self.sin[:seq_len, :].view(1, 1, seq_len, width) else: max_len = int(positions.max()) + 1 if self.cos.size(-2) < max_len: self._create_rotary_embed(width=width, length=max_len) # Flatten positions to index cos/sin arrays, then unflatten. # # Example shapes: # # positions_flat - (batch_size * seq_len) # self.cos - (max_len, width) # rot_cos - (batch_size, seq_len, width) positions_flat = positions.view(-1) rot_cos = self.cos[positions_flat].view(batch_size, 1, seq_len, width) rot_sin = self.sin[positions_flat].view(batch_size, 1, seq_len, width) # Eq 34 with ordering changed for compatibility. return rot_cos * input + rot_sin * self._rotate(input) class LayerNorm(nn.Module): """ LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False """ def __init__(self, ndim, bias=False): super().__init__() self.weight = nn.Parameter(torch.ones(ndim)) self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None def forward(self, input): return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5) class CausalSelfAttention(nn.Module): def __init__(self, embed_dim, num_heads, dropout=0.): super().__init__() # Typically, bias = True in Linears and LayerNorms, like GPT-2. But we set bias = False: a bit better and faster (following https://github.com/karpathy/nanoGPT) assert embed_dim % num_heads == 0 self.embed_dim = embed_dim self.num_heads = num_heads self.dropout = dropout self.head_dim = embed_dim // num_heads assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads" self.scaling = self.head_dim ** -0.5 # key, query, value projections for all heads, but in a batch self.c_attn = nn.Linear(embed_dim, 3 * embed_dim, bias=False) # output projection self.out_proj = nn.Linear(embed_dim, embed_dim, bias=False) # rotary embeddings self.rotary_embeds = RotaryEmbeddings(width=embed_dim // num_heads) # flash attention make GPU go brrrrr but support is only in PyTorch >= 2.0 self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention') if not self.flash: print("WARNING: using slow attention. Flash Attention requires PyTorch >= 2.0") def forward( self, query, key, value, spk_pos_ids_flat=None, incremental_state=None, need_weights=True, static_kv=False, attn_mask=None, need_head_weights=False, enc_dec_attn_constraint_mask=None, ): """Input shape: Time x Batch x Channel Args: need_weights (bool, optional): return the attention weights, averaged over heads (default: False). attn_mask (ByteTensor, optional): typically used to implement causal attention, where the mask prevents the attention from looking forward in time (default: None). need_head_weights (bool, optional): return the attention weights for each head. Implies *need_weights*. Default: return the average attention weights over all heads. """ if need_head_weights: need_weights = True tgt_len, bsz, embed_dim = query.size() assert embed_dim == self.embed_dim assert list(query.size()) == [tgt_len, bsz, embed_dim] if incremental_state is not None: saved_state = self._get_input_buffer(incremental_state) else: saved_state = None # calculate query, key, values for all heads in batch and move head forward to be the batch dim q, k, v = self.c_attn(query).split(self.embed_dim, dim=2) q = q.contiguous().view(tgt_len, bsz * self.num_heads, self.head_dim).transpose(0, 1) k = k.contiguous().view(-1, bsz * self.num_heads, self.head_dim).transpose(0, 1) v = v.contiguous().view(-1, bsz * self.num_heads, self.head_dim).transpose(0, 1) # Apply rot embedding and store incremental_state q = self.rotary_embeds(q[None, :], positions=spk_pos_ids_flat)[0] if saved_state is not None: # saved states are stored with shape (bsz, num_heads, seq_len, head_dim) if 'prev_key' in saved_state: prev_key = saved_state['prev_key'].view(bsz * self.num_heads, -1, self.head_dim) if static_kv: k = prev_key else: k = torch.cat((prev_key, k), dim=1) if 'prev_value' in saved_state: prev_value = saved_state['prev_value'].view(bsz * self.num_heads, -1, self.head_dim) if static_kv: v = prev_value else: v = torch.cat((prev_value, v), dim=1) saved_state['prev_key'], saved_state['prev_value'] = k.view(bsz, self.num_heads, -1, self.head_dim), v.view( bsz, self.num_heads, -1, self.head_dim) self._set_input_buffer(incremental_state, saved_state) if incremental_state is not None: key_pos = torch.arange(k.shape[-2], device=q.device).unsqueeze(0) else: key_pos = spk_pos_ids_flat k = self.rotary_embeds(k[None, :], positions=key_pos)[0] src_len = k.size(1) # Start Attention if self.flash: # efficient attention using Flash Attention CUDA kernels attn = torch.nn.functional.scaled_dot_product_attention( q, k, v, attn_mask=attn_mask, dropout_p=0, is_causal=False) assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim] attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim) # Flash Attn 2 # from flash_attn import flash_attn_func # q, k, v = q.transpose(0, 1)[None, :], k.transpose(0, 1)[None, :], v.transpose(0, 1)[None, :] # attn = flash_attn_func(q, k, v, dropout_p=0.0, causal=False)[0].contiguous().view(tgt_len, bsz, embed_dim) attn = self.out_proj(attn) attn_logits = None else: attn_weights = torch.bmm(q, k.transpose(1, 2)) assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len] if attn_mask is not None: if len(attn_mask.shape) == 2: attn_mask = attn_mask.unsqueeze(0) elif len(attn_mask.shape) == 3: attn_mask = attn_mask[:, None].repeat([1, self.num_heads, 1, 1]).reshape( bsz * self.num_heads, tgt_len, src_len) attn_weights = attn_weights + attn_mask attn_logits = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) attn_weights_float = softmax(attn_weights, dim=-1) attn_weights = attn_weights_float.type_as(attn_weights) attn_probs = F.dropout(attn_weights_float.type_as(attn_weights), p=self.dropout, training=self.training) attn = torch.bmm(attn_probs, v) assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim] attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim) attn = self.out_proj(attn) if need_weights: attn_weights = attn_weights_float.view(bsz, self.num_heads, tgt_len, src_len).transpose(1, 0) if not need_head_weights: # average attention weights over heads attn_weights = attn_weights.mean(dim=0) else: attn_weights = None return attn, (attn_weights, attn_logits) def _get_input_buffer(self, incremental_state): return get_incremental_state( self, incremental_state, 'attn_state', ) or {} def _set_input_buffer(self, incremental_state, buffer): set_incremental_state( self, incremental_state, 'attn_state', buffer, ) def clear_buffer(self, incremental_state=None): if incremental_state is not None: saved_state = self._get_input_buffer(incremental_state) if 'prev_key' in saved_state: del saved_state['prev_key'] if 'prev_value' in saved_state: del saved_state['prev_value'] self._set_input_buffer(incremental_state, saved_state) class TransformerFFNLayer(nn.Module): def __init__(self, hidden_size, filter_size, padding="SAME", kernel_size=1, dropout=0., act='gelu'): super().__init__() self.kernel_size = kernel_size self.dropout = dropout self.act = act if padding == 'SAME': self.ffn_1 = nn.Conv1d(hidden_size, filter_size, kernel_size, padding=kernel_size // 2, bias=False) elif padding == 'LEFT': self.ffn_1 = nn.Sequential( nn.ConstantPad1d((kernel_size - 1, 0), 0.0), nn.Conv1d(hidden_size, filter_size, kernel_size, bias=False) ) self.ffn_2 = nn.Linear(filter_size, hidden_size, bias=False) def forward(self, x, incremental_state=None): # x: T x B x C if incremental_state is not None: T_inp = x.shape[0] saved_state = self._get_input_buffer(incremental_state) if 'prev_input' in saved_state: prev_input = saved_state['prev_input'] x = torch.cat((prev_input, x), dim=0) x = x[-self.kernel_size:] saved_state['prev_input'] = x self._set_input_buffer(incremental_state, saved_state) x = self.ffn_1(x.permute(1, 2, 0)).permute(2, 0, 1) x = x * self.kernel_size ** -0.5 if incremental_state is not None: x = x[-T_inp:] # if self.act == 'gelu': # x = F.gelu(x) # if self.act == 'relu': # x = F.relu(x) x = F.silu(x) x = F.dropout(x, self.dropout, training=self.training) x = self.ffn_2(x) return x def _get_input_buffer(self, incremental_state): return get_incremental_state( self, incremental_state, 'f', ) or {} def _set_input_buffer(self, incremental_state, buffer): set_incremental_state( self, incremental_state, 'f', buffer, ) def clear_buffer(self, incremental_state): if incremental_state is not None: saved_state = self._get_input_buffer(incremental_state) if 'prev_input' in saved_state: del saved_state['prev_input'] self._set_input_buffer(incremental_state, saved_state) class GPTBlock(nn.Module): def __init__(self, c, num_heads, dropout, attention_dropout=0.1, relu_dropout=0.1, kernel_size=9, ffn_hidden_size=1024, act='gelu', post_ln=False, norm_cls=LayerNorm): super().__init__() self.c = c self.dropout = dropout self.layer_norm1 = norm_cls(c) self.self_attn = CausalSelfAttention( c, num_heads, dropout=attention_dropout ) self.layer_norm2 = norm_cls(c) self.ffn = TransformerFFNLayer( c, ffn_hidden_size, padding='LEFT', kernel_size=kernel_size, dropout=relu_dropout, act=act) self.post_ln = post_ln def forward( self, x, encoder_out=None, encoder_padding_mask=None, incremental_state=None, self_attn_mask=None, attn_out=None, spk_pos_ids_flat=None, **kwargs, ): layer_norm_training = kwargs.get('layer_norm_training', None) if layer_norm_training is not None: self.layer_norm1.training = layer_norm_training self.layer_norm2.training = layer_norm_training residual = x if not self.post_ln: x = self.layer_norm1(x) x, _ = self.self_attn( query=x, key=x, value=x, incremental_state=incremental_state, attn_mask=self_attn_mask, spk_pos_ids_flat=spk_pos_ids_flat, need_weights=False ) x = F.dropout(x, self.dropout, training=self.training) x = residual + x if self.post_ln: x = self.layer_norm1(x) attn_logits = None residual = x if not self.post_ln: x = self.layer_norm2(x) x = self.ffn(x, incremental_state=incremental_state) x = F.dropout(x, self.dropout, training=self.training) x = residual + x if self.post_ln: x = self.layer_norm2(x) return x, attn_logits def clear_buffer(self, input, encoder_out=None, encoder_padding_mask=None, incremental_state=None): self.encoder_attn.clear_buffer(incremental_state) self.ffn.clear_buffer(incremental_state) def set_buffer(self, name, tensor, incremental_state): return set_incremental_state(self, incremental_state, name, tensor) class GPTLayer(nn.Module): def __init__(self, hidden_size, dropout, kernel_size=9, num_heads=8, ffn_hidden_size=1024, post_ln=False, lm_num_layers=10, norm_cls=LayerNorm): super().__init__() self.hidden_size = hidden_size self.dropout = dropout self.num_heads = num_heads self.op = GPTBlock( hidden_size, num_heads, dropout=dropout, attention_dropout=0.0, relu_dropout=dropout, kernel_size=kernel_size, ffn_hidden_size=ffn_hidden_size, post_ln=post_ln, norm_cls=norm_cls) # init all weights self.apply(self._init_weights) # apply special scaled init to the residual projections, per GPT-2 paper for pn, p in self.named_parameters(): if pn.endswith('ffn_2.weight') or pn.endswith('out_proj.weight'): torch.nn.init.normal_(p, mean=0.0, std=0.02 / math.sqrt(2 * lm_num_layers)) def _init_weights(self, module): if isinstance(module, nn.Linear): torch.nn.init.normal_(module.weight, mean=0.0, std=0.02) if module.bias is not None: torch.nn.init.zeros_(module.bias) elif isinstance(module, nn.Embedding): torch.nn.init.normal_(module.weight, mean=0.0, std=0.02) @torch.autocast(device_type='cuda') def forward(self, x, **kwargs): return self.op(x, **kwargs) def clear_buffer(self, *args): return self.op.clear_buffer(*args) def set_buffer(self, *args): return self.op.set_buffer(*args)