from dataclasses import dataclass import logging import math import typing as tp import torch import torch.nn.functional as F from audiocraft.transformer import StreamingTransformer from dataclasses import dataclass from functools import partial from torch import nn from audiocraft.activations import get_activation_fn def sample_top_k(p, k=1, n_draw=None): """ p probabs 2048 ? num_draw : how many tokens to sample (for duplicate elongation) """ p = torch.softmax(p, dim=-1) # p/temp top_k_value, i250 = torch.topk(p, k, dim=-1) # probs: [1, 4, 2048] # print('\n_____TOPK________\n', top_k_value.shape, top_k_value[0, 0, :10], '\n___________END_TOPK____________\n') min_value_top_k = top_k_value[..., [-1]] # p *= (p >= min_value_top_k).float() p.div_(p.sum(dim=-1, keepdim=True)) # -- next_token = multinomial(probs, num_samples=num_draw) # RESHAPED into bs, 4, 250 p_ = p.reshape(-1, p.shape[-1]) out = torch.multinomial(p_, num_samples=n_draw, replacement=False) # [4, num_draw] return out.transpose(0, 1)[:, :, None] # [num_draw, 4, 1] # ============================================== From LM.py logger = logging.getLogger(__name__) TextCondition = tp.Optional[str] # a text condition can be a string or None (if doesn't exist) ConditionType = tp.Tuple[torch.Tensor, torch.Tensor] # condition, mask ConditionTensors = tp.Dict[str, ConditionType] CFGConditions = tp.Union[ConditionTensors, tp.Tuple[ConditionTensors, ConditionTensors]] def get_init_fn(method: str, input_dim: int, init_depth: tp.Optional[int] = None): """LM layer initialization. Inspired from xlformers: https://github.com/fairinternal/xlformers Args: method (str): Method name for init function. Valid options are: 'gaussian', 'uniform'. input_dim (int): Input dimension of the initialized module. init_depth (int, optional): Optional init depth value used to rescale the standard deviation if defined. """ # Compute std std = 1 / math.sqrt(input_dim) # Rescale with depth if init_depth is not None: std = std / math.sqrt(2 * init_depth) if method == 'gaussian': return partial( torch.nn.init.trunc_normal_, mean=0.0, std=std, a=-3 * std, b=3 * std ) elif method == 'uniform': bound = math.sqrt(3) * std # ensure the standard deviation is `std` return partial(torch.nn.init.uniform_, a=-bound, b=bound) else: raise ValueError("Unsupported layer initialization method") def init_layer(m: nn.Module, method: str, init_depth: tp.Optional[int] = None, zero_bias_init: bool = False): """Wrapper around ``get_init_fn`` for proper initialization of LM modules. Args: m (nn.Module): Module to initialize. method (str): Method name for the init function. init_depth (int, optional): Optional init depth value used to rescale the standard deviation if defined. zero_bias_init (bool): Whether to initialize the bias to 0 or not. """ if isinstance(m, nn.Linear): init_fn = get_init_fn(method, m.in_features, init_depth=init_depth) if m.weight.device.type == 'cpu' and m.weight.dtype == torch.float16: weight = m.weight.float() init_fn(weight) m.weight.data[:] = weight.half() else: init_fn(m.weight) if zero_bias_init and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.Embedding): init_fn = get_init_fn(method, m.embedding_dim, init_depth=None) if m.weight.device.type == 'cpu' and m.weight.dtype == torch.float16: weight = m.weight.float() init_fn(weight) m.weight.data[:] = weight.half() else: init_fn(m.weight) class ScaledEmbedding(nn.Embedding): """Boost learning rate for embeddings (with `scale`). """ def __init__(self, *args, lr=None, **kwargs): super().__init__(*args, **kwargs) self.lr = lr def make_optim_group(self): group = {"params": list(self.parameters())} if self.lr is not None: group["lr"] = self.lr return group @dataclass class LMOutput: # The logits are already re-aligned with the input codes # hence no extra shift is required, e.g. when computing CE logits: torch.Tensor # [B, K, T, card] mask: torch.Tensor # [B, K, T] class LMModel(nn.Module): """Transformer-based language model on multiple streams of codes. Args: pattern_provider (CodebooksPatternProvider): Pattern provider for codebook interleaving. condition_provider (MusicConditioningProvider): Conditioning provider from metadata. fuser (ConditionFuser): Fuser handling the fusing of conditions with language model input. n_q (int): Number of parallel streams to model. card (int): Cardinality, vocabulary size. dim (int): Dimension of the transformer encoder. num_heads (int): Number of heads for the transformer encoder. hidden_scale (int): Scale for hidden feed forward dimension of the transformer encoder. norm (str): Normalization method. norm_first (bool): Use pre-norm instead of post-norm. emb_lr (float, optional): Embedding-specific learning rate. bias_proj (bool): Use bias for output projections. weight_init (str, optional): Method for weight initialization. depthwise_init (str, optional): Method for depthwise weight initialization. zero_bias_init (bool): If true and bias in Linears, initialize bias to zeros. cfg_dropout (float): Classifier-free guidance dropout. cfg_coef (float): Classifier-free guidance coefficient. attribute_dropout (dict): Attribute dropout probabilities. two_step_cfg (bool): Whether to run classifier free-guidance with 2 distinct steps. **kwargs: Additional parameters for the transformer encoder. """ def __init__(self, pattern_provider, condition_provider, fuser, n_q: int = 8, card: int = 1024, dim: int = 128, num_heads: int = 8, hidden_scale: int = 4, norm: str = 'layer_norm', norm_first: bool = False, emb_lr: tp.Optional[float] = None, bias_proj: bool = True, weight_init: tp.Optional[str] = None, depthwise_init: tp.Optional[str] = None, zero_bias_init: bool = False, cfg_dropout: float = 0, cfg_coef: float = 1.0, attribute_dropout: tp.Dict[str, tp.Dict[str, float]] = {}, two_step_cfg: bool = False, **kwargs): super().__init__() self.cfg_coef = cfg_coef self.n_draw = 1 self.condition_provider = condition_provider self.fuser = fuser self.card = card # 2048 ? embed_dim = self.card + 1 self.n_q = n_q self.dim = dim self.pattern_provider = pattern_provider self.two_step_cfg = two_step_cfg self.emb = nn.ModuleList([ScaledEmbedding(embed_dim, dim, lr=emb_lr) for _ in range(n_q)]) if 'activation' in kwargs: kwargs['activation'] = get_activation_fn(kwargs['activation']) # ======================================================================== # { # 'dtype': torch.float16, 'device': 'cuda', # 'num_layers': 48, 'dropout': 0.0, 'activation': 'gelu', # 'bias_ff': False, 'bias_attn': False, # 'past_context': None, 'causal': True, # 'custom': False, 'memory_efficient': True, # 'attention_as_float32': False, 'positional_embedding': 'sin', 'xpos': False, # 'checkpointing': 'none', 'cross_attention': True, 'qk_layer_norm': False, # 'qk_layer_norm_cross': False, 'attention_dropout': None, 'kv_repeat': 1 # } # ========================================================================== kwargs.pop('layer_scale') # nn.Indentity() self.transformer = StreamingTransformer( d_model=dim, num_heads=num_heads, dim_feedforward=int(hidden_scale * dim), norm=norm, norm_first=norm_first, **kwargs) self.out_norm: tp.Optional[nn.Module] = None if norm_first: self.out_norm = nn.LayerNorm(dim, eps=1e-5) self.linears = nn.ModuleList([nn.Linear(dim, self.card, bias=bias_proj) for _ in range(n_q)]) self._init_weights(weight_init, depthwise_init, zero_bias_init) self._fsdp: tp.Optional[nn.Module] self.__dict__['_fsdp'] = None def _init_weights(self, weight_init: tp.Optional[str], depthwise_init: tp.Optional[str], zero_bias_init: bool): """Initialization of the transformer module weights. Args: weight_init (str, optional): Weight initialization strategy. See ``get_init_fn`` for valid options. depthwise_init (str, optional): Depthwise initialization strategy. The following options are valid: 'current' where the depth corresponds to the current layer index or 'global' where the total number of layer is used as depth. If not set, no depthwise initialization strategy is used. zero_bias_init (bool): Whether to initialize bias to zero or not. """ assert depthwise_init is None or depthwise_init in ['current', 'global'] assert depthwise_init is None or weight_init is not None, \ "If 'depthwise_init' is defined, a 'weight_init' method should be provided." assert not zero_bias_init or weight_init is not None, \ "If 'zero_bias_init', a 'weight_init' method should be provided" if weight_init is None: return for emb_layer in self.emb: init_layer(emb_layer, method=weight_init, init_depth=None, zero_bias_init=zero_bias_init) for layer_idx, tr_layer in enumerate(self.transformer.layers): depth = None if depthwise_init == 'current': depth = layer_idx + 1 elif depthwise_init == 'global': depth = len(self.transformer.layers) init_fn = partial(init_layer, method=weight_init, init_depth=depth, zero_bias_init=zero_bias_init) tr_layer.apply(init_fn) for linear in self.linears: init_layer(linear, method=weight_init, init_depth=None, zero_bias_init=zero_bias_init) @property def special_token_id(self) -> int: return self.card @property def num_codebooks(self) -> int: return self.n_q def forward(self, sequence, condition_tensors=None, token_count=None): B, K, S = sequence.shape # linears are n_q input_ = sum([self.emb[k](sequence[:, k]) for k in range(K)]) # input_, cross_attention_input = self.fuser(input_, condition_tensors) cross_attention_input = condition_tensors['description'][0] # print(f'{input_.shape=}') out = self.transformer(input_, cross_attention_src=cross_attention_input, token_count=token_count) if self.out_norm: out = self.out_norm(out) # K = 2 because of llm producing 2 tokens? # so only 2 x sel.flinear() of 4 are used ? # WHy torch.stack is in dim=1 logits = torch.stack([self.linears[k](out) for k in range(K)], dim=1) # [B, K, S, card] # print(f'{input_.shape=} {out.shape=} {cross_attention_input.shape=} {logits.shape=} FUSER LLM') # remove the prefix from the model outputs # if len(self.fuser.fuse2cond['prepend']) > 0: # logits = logits[:, :, -S:] # print('==========================================PRESFIX') return logits # [B, K, S, card] # GENERATE class revert_codebook_patterns() @torch.no_grad() def generate(self, prompt = None, conditions = [], num_samples = 1, # N next token max_gen_len=256): print(f'{prompt=} {conditions=}') first_param = next(iter(self.parameters())) device = first_param.device tokenized = self.condition_provider.tokenize(conditions) # print('TOKENIZ', tokenized) # 'description' # TOKENIZ {'description': {'input_ids': tensor([[3887, 16, 2815, 1], # [3887, 16, 2815, 1]], device='cuda:0'), 'attention_mask': tensor([[1, 1, 1, 1], # [1, 1, 1, 1]], device='cuda:0')}} cfg_conditions = self.condition_provider(tokenized) if prompt is None: assert num_samples > 0 prompt = torch.zeros((num_samples, self.num_codebooks, 0), dtype=torch.long, device=device) print('\n\n\n\n DEFAULT PROMPT ZERO \n\n-') B, K, T = prompt.shape start_offset = T pattern = self.pattern_provider.get_pattern(max_gen_len) # duplicate sequence # this token is used as default value for codes that are not generated yet ? unknown_token = -1 gen_codes = torch.full((B, K, max_gen_len), unknown_token, dtype=torch.long, device=device) gen_codes[..., :start_offset] = prompt # place 0 _gen_sequence, _, mask = pattern.build_pattern_sequence(gen_codes, self.special_token_id) # -- # print(mask.shape, mask.sum(), 'MSK LM') # torch.Size([4, 39]) tensor(140, device='cuda:0') MSK LM ? Fully 1 normal no special token # --\ # list - Elongation for take-5 next tokens - n_draw 5 tokens at each time-step # append them at end of sequence duplicate_draw = [ _gen_sequence[:, :, 0:1].repeat(self.n_draw, 1, 1) ] for offset in range(1, _gen_sequence.shape[2]): logits = self.forward(_gen_sequence[:, :, offset-1:offset], # bs/n_draw, 4, 1 condition_tensors=cfg_conditions, token_count=offset) # print(f'BEF {logits.shape=} BEF utils.SampleTop5') # AGREES 4 BEF logits.shape=torch.Size([1, 4, 1, 2048]) BEF utils.SampleTop5 next_token = sample_top_k(logits, n_draw=self.n_draw) # [1,4,2048] logits _gen_sequence[:, :, offset] = next_token[0, :, 0] # next_token=[1,4,6] gen_seq=[1, 4, 39] duplicate_draw.append(next_token) gen_sequence = torch.cat(duplicate_draw, 2) # RESHAPE -> N_DRAW -> TIME # revert codes as "batch" # In decoder - flatten # _, tokd, len_seq = gen_sequence.shape # gen_sequence = gen_sequence.transpose(0, 1).reshape(tokd, self.n_draw * len_seq)[None, :, :] print(f' <=> BEFORE CODES {gen_sequence.shape=} {_gen_sequence.shape=}\n') # ARRIVES here also if special # revert_pattern_logits ~ NOT CALLED EXPLICIT out_codes, _, _ = pattern.revert_pattern_sequence(gen_sequence, special_token=unknown_token) # set(out_codes.unique().tolist()) - set(gen_sequence.unique().tolist()) # set() # UNIQUE are the SAME ---------------?> is it rearrange # ARE SOME PARTS IGNORED OR RE-ARRANGED # print(f'{unknown_token=} {gen_sequence.shape=} {out_codes.shape=}') # -> unknown tokn = -1 or 2048 # unknown_token=-1 print(f' <=> CODES {out_codes.shape=} {out_codes.min()} {out_codes.max()}\n') # ARRIVES here also if special # unknown_token=-1 gen_sequence.shape=torch.Size([1, 4, 39]) out_codes.shape=torch.Size([1, 4, 35]) # <=> CODES out_codes.shape=torch.Size([1, 4, 35]) 30 2024 # Clean Transformer MHA k_history v_history for lay in self.transformer.layers: lay.self_attn.k_history = None lay.self_attn.v_history = None return out_codes #