# coding=utf-8 # Copyright 2024 Kyutai and The HuggingFace Inc. team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """PyTorch Moshi model.""" import math from dataclasses import dataclass from typing import Any, Dict, List, Optional, Tuple, Union import torch import torch.nn as nn import torch.utils.checkpoint from torch.nn import CrossEntropyLoss from ...activations import ACT2FN from ...cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache from ...generation import ( GenerationConfig, GenerationMixin, ) from ...modeling_attn_mask_utils import AttentionMaskConverter from ...modeling_flash_attention_utils import flash_attn_supports_top_left_mask, is_flash_attn_available from ...modeling_outputs import ( BaseModelOutputWithPast, CausalLMOutputWithPast, ModelOutput, Seq2SeqLMOutput, ) from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update from ...modeling_utils import PreTrainedModel from ...pytorch_utils import ALL_LAYERNORM_LAYERS from ...utils import ( add_start_docstrings, add_start_docstrings_to_model_forward, is_torchdynamo_compiling, logging, replace_return_docstrings, ) from ...utils.deprecation import deprecate_kwarg from ..auto.modeling_auto import AutoModel from .configuration_moshi import MoshiConfig, MoshiDepthConfig if is_flash_attn_available(): from ...modeling_flash_attention_utils import _flash_attention_forward logger = logging.get_logger(__name__) _CONFIG_FOR_DOC = "MoshiConfig" _CHECKPOINT_FOR_DOC = "kmhf/hf-moshiko" @dataclass class MoshiConditionalGenerationGenerateOutput(ModelOutput): """ Outputs of [`MoshiForConditionalConditionalGeneration.generate`]. Args: audio_sequences (`torch.LongTensor` of shape `(batch_size*num_return_sequences, 1, sequence_length)`, *optional*): The generated audio waveforms. sequences (`torch.LongTensor` of shape `(batch_size*num_return_sequences, sequence_length)`): The generated text sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter if all batches finished early due to the `eos_token_id`. sequences_scores (`torch.FloatTensor` of shape `(batch_size*num_return_sequences)`, *optional*, returned when `output_scores=True`): Final beam scores of the generated `sequences`. scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True`): Beam transition scores for each vocabulary token at each generation step. Beam transition scores consisting of log probabilities of tokens conditioned on log softmax of previously generated tokens in this beam. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size*num_beams, config.vocab_size)`. logits (`tuple(torch.FloatTensor)` *optional*, returned when `output_logits=True`): Unprocessed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size, config.vocab_size)`. beam_indices (`torch.LongTensor`, *optional*, returned when `output_scores=True`): Beam indices of generated token id at each generation step. `torch.LongTensor` of shape `(batch_size*num_return_sequences, sequence_length)`. attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size*num_beams, num_heads, generated_length, sequence_length)`. hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size*num_beams*num_return_sequences, generated_length, hidden_size)`. past_key_values (`tuple(tuple(torch.FloatTensor)))`, *optional*, returned when `use_cache=True`): Returns the model cache, used to speed up decoding. Different models have a different cache format, check the model's documentation. Usually, a [`~cache_utils.Cache`] instance. audio_codes (`torch.LongTensor` of shape `(batch_size*num_return_sequences, num_codeooks, sequence_length)`, *optional*): The generated audio codes. Returned if `return_audio_codes=True`. Intermediate audio "tokens" which transforms to `audio_sequences` once passed through the audio decoder. """ audio_sequences: Optional[torch.Tensor] = None sequences: Optional[torch.LongTensor] = None sequences_scores: Optional[torch.FloatTensor] = None scores: Optional[Tuple[torch.FloatTensor]] = None logits: Optional[Tuple[torch.FloatTensor]] = None beam_indices: Optional[torch.LongTensor] = None attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None past_key_values: Optional[Tuple[Tuple[Tuple[torch.FloatTensor]]]] = None audio_codes: Optional[torch.LongTensor] = None @dataclass class MoshiCausalLMOutputWithPast(ModelOutput): """ `MoshiForCausalLM` outputs. Args: loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Language modeling loss (for next-token prediction). logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`): Sequence of hidden-states at the output of the last layer of the model. past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the optional initial embedding outputs. attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. """ loss: Optional[torch.FloatTensor] = None logits: Optional[torch.FloatTensor] = None last_hidden_state: Optional[torch.FloatTensor] = None past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None attentions: Optional[Tuple[torch.FloatTensor, ...]] = None @dataclass class MoshiConditionalGenerationOutputWithPast(ModelOutput): """ `MoshiForConditionalGeneration` outputs. Args: loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `text_labels` is provided): Text language modeling loss (for next-token prediction). logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the text language modeling head (scores for each vocabulary token before SoftMax). last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`): Sequence of hidden-states at the output of the last layer of the model. past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the optional initial embedding outputs. attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. depth_loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `audio_labels` is provided): Audio language modeling loss (for next-token prediction). audio_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the audio language modeling heads. depth_past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): Past key-values of the depth decoder. depth_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Hidden states of the depth decoder depth_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Depth decoder's Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. """ loss: Optional[torch.FloatTensor] = None logits: Optional[torch.FloatTensor] = None last_hidden_state: Optional[torch.FloatTensor] = None past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None attentions: Optional[Tuple[torch.FloatTensor, ...]] = None depth_loss: Optional[torch.FloatTensor] = None audio_logits: Optional[torch.FloatTensor] = None depth_past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None depth_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None depth_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None @dataclass class MoshiUnconditionalInput(ModelOutput): """ Args: input_ids (`torch.Tensor `of shape `(batch_size, sequence_length), *optional*): The sequence used as a text prompt for the generation. user_audio_codes (`torch.Tensor `of shape `(batch_size, num_codebooks, sequence_length), *optional*): The audio codes used as audio user prompt for the generation. Has priority over `user_input_values` and represents the audio "tokens" of `user_input_values` once passed through the audio encoder. moshi_audio_codes (`torch.Tensor `of shape `(batch_size, num_codebooks, sequence_length), *optional*): The audio codes used as audio Moshi prompt for the generation. Has priority over `moshi_input_values` and represents the audio "tokens" of `moshi_input_values` once passed through the audio encoder. attention_mask (`torch.LongTensor`) of shape `(batch_size, sequence_length)`, *optional*): Attention mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: 1 for tokens that are **not masked**, 0 for tokens that are **masked**. """ input_ids: Optional[torch.LongTensor] = None user_audio_codes: Optional[torch.Tensor] = None moshi_audio_codes: Optional[torch.Tensor] = None attention_mask: Optional[torch.LongTensor] = None # Copied from transformers.models.gemma.modeling_gemma.GemmaRMSNorm with Gemma->Moshi class MoshiRMSNorm(nn.Module): def __init__(self, dim: int, eps: float = 1e-6): super().__init__() self.eps = eps self.weight = nn.Parameter(torch.ones(dim)) # Ignore copy def _norm(self, x): return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) # Ignore copy def forward(self, x): output = self._norm(x.float()) output = output * self.weight.float() return output.type_as(x) def extra_repr(self): return f"{tuple(self.weight.shape)}, eps={self.eps}" ALL_LAYERNORM_LAYERS.append(MoshiRMSNorm) class MoshiFlexibleLinear(nn.Module): def __init__(self, input_size, output_size, num_layers): super().__init__() # Stack the weights for N layers into a single tensor (num_layers, output_size, input_size) self.weight = nn.Parameter(torch.randn(num_layers, output_size, input_size)) def forward(self, x, layer_idx=None): """ `MoshiFlexibleLinear` creates one linear layer per codebook. There's multiple ways to use it. In the default case, `sequence_length=num_layers`, so each element of the sequence will be matmul to the weights corresponding to its index on the sequence. For more advanced cases, one can specify which codebook's layer(s) to use with `layer_idx`. If `layer_idx` indicates a single integer, all of the element of the sequence will be matmul to this single codebook's layer. But if `layer_idx` is a tensor of shape `(seq_length,)`, it will matmul each i-th element of the input sequence to the corresponding layer `weight[i]`. Args: x (`torch.FloatTensor): input to the layer of shape `(batch, num_layers, embed_dim)` or of shape `(batch, seq_length, embed_dim)` layer_idx (`torch.Tensor`, *optional*): Can be used to specify which codebook's layers(s) to use. If it's a tensor of shape `(seq_length,)`, will matmul each element of the sequence to the corresponding weights. But if `layer_idx` is a tensor of shape `(seq_length,)`, it will matmul each i-th element of the input sequence to the corresponding layer `weight[i]`. """ # Use torch.gather to select the corresponding weights for each sample # (codebooks, output_size, hidden_size) selected_weights = torch.index_select(self.weight, 0, layer_idx) if layer_idx is not None else self.weight # (1, codebooks, hidden_size, output_size) selected_weights = selected_weights.transpose(1, 2)[None, :, :, :] # (batch_size, codebooks, 1, hidden_size) x (1, codebooks, hidden_size, output_size) # -> (batch_size, codebooks, 1, output_size) x = torch.matmul(x[:, :, None, :], selected_weights) # (batch_size, codebooks, output_size) return x.squeeze(2) class MoshiLinear(nn.Module): def __init__(self, input_dim, output_dim, num_codebooks, use_flexible_linear=False): super().__init__() self.use_flexible_linear = use_flexible_linear if not use_flexible_linear: self.linear = nn.Linear(input_dim, output_dim, bias=False) else: self.linear = MoshiFlexibleLinear(input_dim, output_dim, num_layers=num_codebooks) def forward(self, x, layer_idx=None): if self.use_flexible_linear: return self.linear(x, layer_idx) else: return self.linear(x) # Copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding with Mistral->Moshi class MoshiRotaryEmbedding(nn.Module): def __init__(self, config: MoshiConfig, device=None): super().__init__() # BC: "rope_type" was originally "type" if hasattr(config, "rope_scaling") and config.rope_scaling is not None: self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type")) else: self.rope_type = "default" self.max_seq_len_cached = config.max_position_embeddings self.original_max_seq_len = config.max_position_embeddings self.config = config self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type] inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device) self.register_buffer("inv_freq", inv_freq, persistent=False) self.original_inv_freq = self.inv_freq @torch.no_grad() @dynamic_rope_update # power user: used with advanced RoPE types (e.g. dynamic rope) def forward(self, x, position_ids): inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device) position_ids_expanded = position_ids[:, None, :].float() device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu" with torch.autocast(device_type=device_type, enabled=False): # Force float32 freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) emb = torch.cat((freqs, freqs), dim=-1) cos = emb.cos() * self.attention_scaling sin = emb.sin() * self.attention_scaling return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) # Copied from transformers.models.llama.modeling_llama.rotate_half def rotate_half(x): """Rotates half the hidden dims of the input.""" x1 = x[..., : x.shape[-1] // 2] x2 = x[..., x.shape[-1] // 2 :] return torch.cat((-x2, x1), dim=-1) # Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): """Applies Rotary Position Embedding to the query and key tensors. Args: q (`torch.Tensor`): The query tensor. k (`torch.Tensor`): The key tensor. cos (`torch.Tensor`): The cosine part of the rotary embedding. sin (`torch.Tensor`): The sine part of the rotary embedding. position_ids (`torch.Tensor`, *optional*): Deprecated and unused. unsqueeze_dim (`int`, *optional*, defaults to 1): The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. Returns: `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. """ cos = cos.unsqueeze(unsqueeze_dim) sin = sin.unsqueeze(unsqueeze_dim) q_embed = (q * cos) + (rotate_half(q) * sin) k_embed = (k * cos) + (rotate_half(k) * sin) return q_embed, k_embed class MoshiGatingMLP(nn.Module): def __init__(self, config, use_flexible_linear=False): super().__init__() self.activation_fn = ACT2FN[config.hidden_act] ffn_dim = config.ffn_dim hidden_size = config.hidden_size num_layers = config.num_codebooks if use_flexible_linear else 1 if num_layers == 1: self.fc1 = nn.Linear(hidden_size, ffn_dim, bias=False) self.fc2 = nn.Linear(ffn_dim // 2, hidden_size, bias=False) else: self.fc1 = MoshiFlexibleLinear(hidden_size, ffn_dim, num_layers) self.fc2 = MoshiFlexibleLinear(ffn_dim // 2, hidden_size, num_layers) def forward(self, hidden_states: torch.Tensor, layer_idx: Optional[int] = None) -> torch.Tensor: hidden_states = self.fc1(hidden_states) if layer_idx is None else self.fc1(hidden_states, layer_idx) batch_size, sequence_length, _ = hidden_states.shape hidden_states = hidden_states.view(batch_size, sequence_length, 2, -1) hidden_states = self.activation_fn(hidden_states[..., 0, :]) * hidden_states[..., 1, :] hidden_states = self.fc2(hidden_states) if layer_idx is None else self.fc2(hidden_states, layer_idx) return hidden_states # Copied from transformers.models.llama.modeling_llama.repeat_kv def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: """ This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) """ batch, num_key_value_heads, slen, head_dim = hidden_states.shape if n_rep == 1: return hidden_states hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) class MoshiAttention(nn.Module): """Multi-headed attention from 'Attention Is All You Need' paper""" def __init__(self, config: MoshiConfig, layer_idx: Optional[int] = None, use_flexible_linear=False, use_rope=True): super().__init__() self.config = config self.layer_idx = layer_idx if layer_idx is None: logger.warning_once( f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will " "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` " "when creating this class." ) self.attention_dropout = config.attention_dropout self.hidden_size = config.hidden_size self.num_heads = config.num_attention_heads self.head_dim = config.head_dim self.num_key_value_heads = config.num_key_value_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.max_position_embeddings = config.max_position_embeddings self.is_causal = True self.scaling = 1 / math.sqrt(self.head_dim) if self.hidden_size % self.num_heads != 0: raise ValueError( f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}" f" and `num_heads`: {self.num_heads})." ) self.q_proj = MoshiLinear( self.hidden_size, self.num_heads * self.head_dim, config.num_codebooks, use_flexible_linear ) self.k_proj = MoshiLinear( self.hidden_size, self.num_key_value_heads * self.head_dim, config.num_codebooks, use_flexible_linear ) self.v_proj = MoshiLinear( self.hidden_size, self.num_key_value_heads * self.head_dim, config.num_codebooks, use_flexible_linear ) self.o_proj = MoshiLinear( self.num_heads * self.head_dim, self.hidden_size, config.num_codebooks, use_flexible_linear ) # rotary embeddings are not used in the depth decoder self.rotary_emb = None if use_rope: self.rope_theta = config.rope_theta self.rotary_emb = MoshiRotaryEmbedding(config) # copied from transformers.models.gemma.modeling_gemma.GemmaAttention.forward # no longer copied after attention refactors def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: bsz, q_len, _ = hidden_states.size() query_states = self.q_proj(hidden_states, cache_position) # Ignore copy key_states = self.k_proj(hidden_states, cache_position) # Ignore copy value_states = self.v_proj(hidden_states, cache_position) # Ignore copy query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) if self.rotary_emb is not None: # Ignore copy cos, sin = self.rotary_emb(value_states, position_ids) # Ignore copy query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) # Ignore copy if past_key_value is not None: # sin and cos are specific to RoPE models; cache_position needed for the static cache cache_kwargs = ( {"sin": sin, "cos": cos, "cache_position": cache_position} if self.rotary_emb is not None else {"cache_position": cache_position} ) # Ignore copy key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scaling if attention_mask is not None: # no matter the length, we just slice it causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] attn_weights = attn_weights + causal_mask # upcast attention to fp32 attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training) attn_output = torch.matmul(attn_weights, value_states) if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim): raise ValueError( f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is" f" {attn_output.size()}" ) attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.view(bsz, q_len, -1) attn_output = self.o_proj(attn_output, cache_position) # Ignore copy if not output_attentions: attn_weights = None return attn_output, attn_weights, past_key_value # NO LONGER EXIST Copied from transformers.models.gemma.modeling_gemma.GemmaFlashAttention2 with Gemma->Moshi # TODO cyril: modular class MoshiFlashAttention2(MoshiAttention): """ Moshi flash attention module. This module inherits from `MoshiAttention` as the weights of the module stays untouched. The only required change would be on the forward pass where it needs to correctly call the public API of flash attention and deal with padding tokens in case the input contains any of them. """ def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1. # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignment, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0. # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left). self._flash_attn_uses_top_left_mask = flash_attn_supports_top_left_mask() def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.LongTensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: if isinstance(past_key_value, StaticCache): raise ValueError( "`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` " "make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformers" ) output_attentions = False bsz, q_len, _ = hidden_states.size() query_states = self.q_proj(hidden_states, cache_position) # Ignore copy key_states = self.k_proj(hidden_states, cache_position) # Ignore copy value_states = self.v_proj(hidden_states, cache_position) # Ignore copy # Flash attention requires the input to have the shape # batch_size x seq_length x head_dim x hidden_dim # therefore we just need to keep the original shape query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) if self.rotary_emb is not None: # Ignore copy cos, sin = self.rotary_emb(value_states, position_ids) # Ignore copy query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) # Ignore copy if past_key_value is not None: # sin and cos are specific to RoPE models; cache_position needed for the static cache cache_kwargs = ( {"sin": sin, "cos": cos, "cache_position": cache_position} if self.rotary_emb is not None else {"cache_position": cache_position} ) # Ignore copy key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache # to be able to avoid many of these transpose/reshape/view. query_states = query_states.transpose(1, 2) key_states = key_states.transpose(1, 2) value_states = value_states.transpose(1, 2) dropout_rate = self.attention_dropout if self.training else 0.0 # In PEFT, usually we cast the layer norms in float32 for training stability reasons # therefore the input hidden states gets silently casted in float32. Hence, we need # cast them back in the correct dtype just to be sure everything works as expected. # This might slowdown training & inference so it is recommended to not cast the LayerNorms # in fp32. (MoshiRMSNorm handles it correctly) input_dtype = query_states.dtype if input_dtype == torch.float32: if torch.is_autocast_enabled(): target_dtype = torch.get_autocast_gpu_dtype() # Handle the case where the model is quantized elif hasattr(self.config, "_pre_quantization_dtype"): target_dtype = self.config._pre_quantization_dtype else: target_dtype = self.q_proj.weight.dtype logger.warning_once( f"The input hidden states seems to be silently casted in float32, this might be related to" f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in" f" {target_dtype}." ) query_states = query_states.to(target_dtype) key_states = key_states.to(target_dtype) value_states = value_states.to(target_dtype) attn_output = _flash_attention_forward( query_states, key_states, value_states, attention_mask, q_len, position_ids=position_ids, dropout=dropout_rate, sliding_window=getattr(self, "sliding_window", None), is_causal=self.is_causal, use_top_left_mask=self._flash_attn_uses_top_left_mask, ) attn_output = attn_output.reshape(bsz, q_len, -1).contiguous() attn_output = self.o_proj(attn_output, cache_position) # Ignore copy if not output_attentions: attn_weights = None return attn_output, attn_weights, past_key_value # NO LONGER EXIST Copied from transformers.models.gemma.modeling_gemma.GemmaSdpaAttention with Gemma->Moshi # TODO cyril: modular class MoshiSdpaAttention(MoshiAttention): """ Moshi attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from `MoshiAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to SDPA API. """ # Adapted from MoshiAttention.forward def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, **kwargs, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: if output_attentions: # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. logger.warning_once( "MoshiModel is using MoshiSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, " 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' ) return super().forward( hidden_states=hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, ) bsz, q_len, _ = hidden_states.size() query_states = self.q_proj(hidden_states, cache_position) # Ignore copy key_states = self.k_proj(hidden_states, cache_position) # Ignore copy value_states = self.v_proj(hidden_states, cache_position) # Ignore copy query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) if self.rotary_emb is not None: # Ignore copy cos, sin = self.rotary_emb(value_states, position_ids) # Ignore copy query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) # Ignore copy if past_key_value is not None: # sin and cos are specific to RoPE models; cache_position needed for the static cache cache_kwargs = ( {"sin": sin, "cos": cos, "cache_position": cache_position} if self.rotary_emb is not None else {"cache_position": cache_position} ) # Ignore copy key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) causal_mask = attention_mask if attention_mask is not None: causal_mask = causal_mask[:, :, :, : key_states.shape[-2]] # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, # Reference: https://github.com/pytorch/pytorch/issues/112577. if query_states.device.type == "cuda" and causal_mask is not None: query_states = query_states.contiguous() key_states = key_states.contiguous() value_states = value_states.contiguous() # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. is_causal = True if causal_mask is None and q_len > 1 else False attn_output = torch.nn.functional.scaled_dot_product_attention( query_states, key_states, value_states, attn_mask=causal_mask, dropout_p=self.attention_dropout if self.training else 0.0, is_causal=is_causal, ) attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.view(bsz, q_len, -1) attn_output = self.o_proj(attn_output, cache_position) # Ignore copy return attn_output, None, past_key_value MOSHI_ATTENTION_CLASSES = { "eager": MoshiAttention, "flash_attention_2": MoshiFlashAttention2, "sdpa": MoshiSdpaAttention, } class MoshiDecoderLayer(nn.Module): def __init__(self, config: MoshiConfig, layer_idx: int, use_flexible_linear: bool, use_rope=True): super().__init__() self.hidden_size = config.hidden_size self.use_flexible_linear = use_flexible_linear self.self_attn = MOSHI_ATTENTION_CLASSES[config._attn_implementation]( config=config, layer_idx=layer_idx, use_flexible_linear=use_flexible_linear, use_rope=use_rope ) self.mlp = MoshiGatingMLP(config, use_flexible_linear) self.input_layernorm = MoshiRMSNorm(self.hidden_size, eps=config.rms_norm_eps) self.post_attention_layernorm = MoshiRMSNorm(self.hidden_size, eps=config.rms_norm_eps) self.sliding_window = config.sliding_window self._attn_implementation = config._attn_implementation def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, cache_position: Optional[torch.LongTensor] = None, **kwargs, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`, *optional*): attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1, query_sequence_length, key_sequence_length)` if default attention is used. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence kwargs (`dict`, *optional*): Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code into the model """ residual = hidden_states hidden_states = self.input_layernorm(hidden_states) # Self Attention hidden_states, self_attn_weights, present_key_value = self.self_attn( hidden_states=hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, **kwargs, ) hidden_states = residual + hidden_states # Fully Connected residual = hidden_states hidden_states = self.post_attention_layernorm(hidden_states) hidden_states = ( self.mlp(hidden_states) if not self.use_flexible_linear else self.mlp(hidden_states, cache_position) ) hidden_states = residual + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) if use_cache: outputs += (present_key_value,) return outputs class MoshiPreTrainedModel(PreTrainedModel): """ An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. """ config_class = MoshiConfig base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = ["MoshiDecoderLayer", "MimiTransformerLayer"] _supports_flash_attn_2 = True _supports_sdpa = True _supports_cache_class = True main_input_name = "input_ids" def _init_weights(self, module): std = self.config.initializer_range if isinstance(module, (nn.Linear, nn.Conv1d)): module.weight.data.normal_(mean=0.0, std=std) if module.bias is not None: module.bias.data.zero_() elif isinstance(module, (nn.LayerNorm, nn.GroupNorm)): module.bias.data.zero_() module.weight.data.fill_(1.0) elif isinstance(module, nn.Conv1d): nn.init.kaiming_normal_(module.weight) if module.bias is not None: k = math.sqrt(module.groups / (module.in_channels * module.kernel_size[0])) nn.init.uniform_(module.bias, a=-k, b=k) elif isinstance(module, nn.Embedding): module.weight.data.normal_(mean=0.0, std=std) if module.padding_idx is not None: module.weight.data[module.padding_idx].zero_() MOSHI_START_DOCSTRING = r""" This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.) This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior. Parameters: config ([`MoshiConfig`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. """ MOSHI_INPUTS_DOCSTRING = r""" Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence text tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) user_input_values (`torch.Tensor `of shape `(batch_size, 1, audio_sequence_length), *optional*): The audio waveforms used as audio user prompt for the generation. user_audio_codes (`torch.Tensor `of shape `(batch_size, num_codebooks, sequence_length), *optional*): The audio codes used as audio user prompt for the generation. Has priority over `user_input_values` and represents the audio "tokens" of `user_input_values` once passed through the audio encoder. moshi_input_values (`torch.Tensor `of shape `(batch_size, 1, audio_sequence_length), *optional*): The audio waveforms used as audio Moshi prompt for the generation. moshi_audio_codes (`torch.Tensor `of shape `(batch_size, num_codebooks, sequence_length), *optional*): The audio codes used as audio Moshi prompt for the generation. Has priority over `moshi_input_values` and represents the audio "tokens" of `moshi_input_values` once passed through the audio encoder. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. If `past_key_values` is used, optionally only the last `inputs_embeds` have to be input (see `past_key_values`). This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. If `input_ids` and `inputs_embeds` are both unset, `inputs_embeds` takes the value of `inputs_embeds`. past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*): Pre-computed hidden-states (key and values in the self-attention blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. Two formats are allowed: - a [`~cache_utils.Cache`] instance; - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy cache format. The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the legacy cache format will be returned. text_labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for text language modeling. Note that the labels **are shifted** inside the model, i.e. you can set `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]` audio_labels (`torch.LongTensor` of shape `(batch_size, num_codebooks, sequence_length)`, *optional*): Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ..., config.audio_vocab_size]` use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. """ MOSHI_DECODER_INPUTS_DOCSTRING = r""" Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. If `past_key_values` is used, optionally only the last `input_ids` have to be input (see `past_key_values`). If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy. - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*): Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. Two formats are allowed: - a [`~cache_utils.Cache`] instance, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache); - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy cache format. The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the legacy cache format will be returned. If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, this tensor is not affected by padding. It is used to update the cache in the correct position and to infer the complete sequence length. """ class MoshiDepthDecoder(MoshiPreTrainedModel, GenerationMixin): """ Transformer depth decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MoshiTransformerLayer`] Args: config: MoshiConfig """ config_class = MoshiDepthConfig def __init__(self, config: MoshiDepthConfig): super().__init__(config) self.text_embed_tokens = nn.Embedding(config.vocab_size + 1, config.hidden_size) # the last codebook is never used as input self.embed_tokens = nn.ModuleList( [nn.Embedding(config.audio_vocab_size + 1, config.hidden_size) for _ in range(config.num_codebooks - 1)] ) self.input_projections = MoshiFlexibleLinear(config.input_size, config.hidden_size, config.num_codebooks) self.layers = nn.ModuleList( [ MoshiDecoderLayer(config, layer_idx, use_flexible_linear=True, use_rope=False) for layer_idx in range(config.num_hidden_layers) ] ) self.lm_heads = MoshiFlexibleLinear(config.hidden_size, config.audio_vocab_size, config.num_codebooks) self._attn_implementation = config._attn_implementation self.gradient_checkpointing = False self.config = config def forward( self, input_ids: Optional[torch.LongTensor] = None, last_hidden_state: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.BoolTensor] = None, past_key_values: Tuple[Tuple[torch.FloatTensor]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, position_ids: Optional[torch.LongTensor] = None, labels: Optional[torch.LongTensor] = None, cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, BaseModelOutputWithPast]: """ Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens. The first element of the sequence must the text token associated to the audio codebooks. The rest of the elements must be flatten audio codebooks. The `cache_position` argument can be used to indicate to which index is associated each token. last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`): Sequence of hidden-states at the output of the last layer of the main decoder. Used to contextualize `input_ids` attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. If `past_key_values` is used, optionally only the last `input_ids` have to be input (see `past_key_values`). If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy. - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*): Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. Two formats are allowed: - a [`~cache_utils.Cache`] instance; - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy cache format. The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the legacy cache format will be returned. If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert the inputs into associated vectors than the model's internal embedding lookup matrix. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. cache_position (`torch.Tensor`): Indices depicting the position of the input sequence tokens in the sequence. """ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict if self.gradient_checkpointing and self.training and use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`." ) use_cache = False if use_cache and past_key_values is None and not self.training: past_key_values = DynamicCache.from_legacy_cache(past_key_values) past_seen_tokens = 0 if past_key_values is None else past_key_values.get_seq_length() if cache_position is None: cache_position = torch.arange( past_seen_tokens, past_seen_tokens + input_ids.shape[1], device=input_ids.device ) if position_ids is None: position_ids = cache_position.unsqueeze(0) # If inputs_embeds is provided, it has the priority over input_ids, which won't be used if inputs_embeds is None: inputs_embeds = [] for position_idx in cache_position: position_idx = position_idx.item() if position_idx == 0: inputs_embeds.append(self.text_embed_tokens(input_ids[:, [position_idx]])) else: inputs_embeds.append( self.embed_tokens[(position_idx - 1)](input_ids[:, [position_idx - past_seen_tokens]]) ) inputs_embeds = torch.cat(inputs_embeds, dim=1) inputs_embeds += self.input_projections(last_hidden_state, cache_position) causal_mask = None if attention_mask is not None: causal_mask = self._update_causal_mask( attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions ) # decoder layers all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None next_decoder_cache = None hidden_states = inputs_embeds for decoder_layer in self.layers: if output_hidden_states: all_hidden_states += (hidden_states,) if self.gradient_checkpointing and self.training: layer_outputs = self._gradient_checkpointing_func( decoder_layer.__call__, hidden_states, causal_mask, position_ids, past_key_values, output_attentions, use_cache, cache_position, ) else: layer_outputs = decoder_layer( hidden_states, attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, ) hidden_states = layer_outputs[0] if use_cache: next_decoder_cache = layer_outputs[2 if output_attentions else 1] if output_attentions: all_self_attns += (layer_outputs[1],) # add hidden states from the last decoder layer if output_hidden_states: all_hidden_states += (hidden_states,) next_cache = next_decoder_cache if use_cache else None logits = self.lm_heads(hidden_states, cache_position) loss = None if labels is not None: # Upcast to float if we need to compute the loss to avoid potential precision issues logits = logits.float() loss_fct = CrossEntropyLoss() labels = labels.masked_fill(labels == self.config.audio_vocab_size, -100).reshape(-1) # Enable model parallelism labels = labels.to(logits.device) loss = loss_fct(logits.reshape(-1, self.config.audio_vocab_size), labels) if not return_dict: return tuple(v for v in [loss, logits, next_cache, all_hidden_states, all_self_attns] if v is not None) return CausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=next_cache, hidden_states=all_hidden_states, attentions=all_self_attns, ) # Copied from transformers.models.phi3.modeling_phi3.Phi3Model._update_causal_mask with Phi3->Moshi def _update_causal_mask( self, attention_mask: torch.Tensor, input_tensor: torch.Tensor, cache_position: torch.Tensor, past_key_values: Cache, output_attentions: bool = False, ): if self.config._attn_implementation == "flash_attention_2": if attention_mask is not None and past_key_values is not None: is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0] if is_padding_right: raise ValueError( "You are attempting to perform batched generation with padding_side='right'" " this may lead to unexpected behaviour for Flash Attention version of Moshi. Make sure to " " call `tokenizer.padding_side = 'left'` before tokenizing the input. " ) if attention_mask is not None and 0.0 in attention_mask: return attention_mask return None # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail # to infer the attention mask. past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 using_static_cache = isinstance(past_key_values, StaticCache) using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache) # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward if ( self.config._attn_implementation == "sdpa" and not (using_static_cache or using_sliding_window_cache) and not output_attentions ): if AttentionMaskConverter._ignore_causal_mask_sdpa( attention_mask, inputs_embeds=input_tensor, past_key_values_length=past_seen_tokens, sliding_window=self.config.sliding_window, is_training=self.training, ): return None dtype, device = input_tensor.dtype, input_tensor.device min_dtype = torch.finfo(dtype).min sequence_length = input_tensor.shape[1] # SlidingWindowCache or StaticCache if using_sliding_window_cache or using_static_cache: target_length = past_key_values.get_max_cache_shape() # DynamicCache or no cache else: target_length = ( attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else past_seen_tokens + sequence_length + 1 ) # In case the provided `attention` mask is 2D, we generate a causal mask here (4D). causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position( attention_mask, sequence_length=sequence_length, target_length=target_length, dtype=dtype, device=device, cache_position=cache_position, batch_size=input_tensor.shape[0], config=self.config, past_key_values=past_key_values, ) if ( self.config._attn_implementation == "sdpa" and attention_mask is not None and attention_mask.device.type in ["cuda", "xpu"] and not output_attentions ): # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. # Details: https://github.com/pytorch/pytorch/issues/110213 causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) return causal_mask @staticmethod # Copied from transformers.models.mistral.modeling_mistral.MistralModel._prepare_4d_causal_attention_mask_with_cache_position with Mistral->MoshiDepth def _prepare_4d_causal_attention_mask_with_cache_position( attention_mask: torch.Tensor, sequence_length: int, target_length: int, dtype: torch.dtype, device: torch.device, cache_position: torch.Tensor, batch_size: int, config: MoshiDepthConfig, past_key_values: Cache, ): """ Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing. Args: attention_mask (`torch.Tensor`): A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`. sequence_length (`int`): The sequence length being processed. target_length (`int`): The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet. dtype (`torch.dtype`): The dtype to use for the 4D attention mask. device (`torch.device`): The device to place the 4D attention mask on. cache_position (`torch.Tensor`): Indices depicting the position of the input sequence tokens in the sequence. batch_size (`torch.Tensor`): Batch size. config (`MoshiDepthConfig`): The model's configuration class past_key_values (`Cache`): The cache class that is being used currently to generate """ if attention_mask is not None and attention_mask.dim() == 4: # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing. causal_mask = attention_mask else: min_dtype = torch.finfo(dtype).min causal_mask = torch.full( (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device ) diagonal_attend_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) if config.sliding_window is not None: # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also # the check is needed to verify is current checkpoint was trained with sliding window or not if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length: sliding_attend_mask = torch.arange(target_length, device=device) <= ( cache_position.reshape(-1, 1) - config.sliding_window ) diagonal_attend_mask.bitwise_or_(sliding_attend_mask) causal_mask *= diagonal_attend_mask causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1) if attention_mask is not None: causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit if attention_mask.shape[-1] > target_length: attention_mask = attention_mask[:, :target_length] mask_length = attention_mask.shape[-1] padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to( causal_mask.device ) padding_mask = padding_mask == 0 causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( padding_mask, min_dtype ) return causal_mask @add_start_docstrings( "The bare Moshi Model outputting raw hidden-states without any specific head on top.", MOSHI_START_DOCSTRING, ) class MoshiModel(MoshiPreTrainedModel): """ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MoshiDecoderLayer`] Args: config: MoshiConfig """ def __init__(self, config: MoshiConfig): super().__init__(config) self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.embed_tokens = nn.Embedding(config.vocab_size + 1, config.hidden_size, self.padding_idx) self.layers = nn.ModuleList( [ MoshiDecoderLayer(config, layer_idx, use_flexible_linear=False) for layer_idx in range(config.num_hidden_layers) ] ) self.norm = MoshiRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.gradient_checkpointing = False # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self): return self.embed_tokens def set_input_embeddings(self, value): self.embed_tokens = value @add_start_docstrings_to_model_forward(MOSHI_DECODER_INPUTS_DOCSTRING) def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, BaseModelOutputWithPast]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict if self.gradient_checkpointing and self.training and use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`." ) use_cache = False if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) return_legacy_cache = False # noqa: F841 if ( use_cache and not isinstance(past_key_values, Cache) and not self.training ): # kept for BC (non `Cache` `past_key_values` inputs) return_legacy_cache = True # noqa: F841 past_key_values = DynamicCache.from_legacy_cache(past_key_values) if cache_position is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 cache_position = torch.arange( past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) if position_ids is None: position_ids = cache_position.unsqueeze(0) causal_mask = None if attention_mask is not None: causal_mask = self._update_causal_mask( attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions ) # embed positions hidden_states = inputs_embeds if ( use_cache and not isinstance(past_key_values, Cache) and not self.training ): # kept for BC (non `Cache` `past_key_values` inputs) return_legacy_cache = True past_key_values = DynamicCache.from_legacy_cache(past_key_values) logger.warning_once( "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/internal/generation_utils#transformers.Cache)" ) # decoder layers all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None next_decoder_cache = None for decoder_layer in self.layers: if output_hidden_states: all_hidden_states += (hidden_states,) if self.gradient_checkpointing and self.training: layer_outputs = self._gradient_checkpointing_func( decoder_layer.__call__, hidden_states, causal_mask, position_ids, past_key_values, output_attentions, use_cache, cache_position, ) else: layer_outputs = decoder_layer( hidden_states, attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, ) hidden_states = layer_outputs[0] if use_cache: next_decoder_cache = layer_outputs[2 if output_attentions else 1] if output_attentions: all_self_attns += (layer_outputs[1],) hidden_states = self.norm(hidden_states) # add hidden states from the last decoder layer if output_hidden_states: all_hidden_states += (hidden_states,) next_cache = next_decoder_cache if use_cache else None if return_legacy_cache: next_cache = next_cache.to_legacy_cache() if not return_dict: return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None) return BaseModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=next_cache, hidden_states=all_hidden_states, attentions=all_self_attns, ) # Copied from transformers.models.phi3.modeling_phi3.Phi3Model._update_causal_mask with Phi3->Moshi def _update_causal_mask( self, attention_mask: torch.Tensor, input_tensor: torch.Tensor, cache_position: torch.Tensor, past_key_values: Cache, output_attentions: bool = False, ): if self.config._attn_implementation == "flash_attention_2": if attention_mask is not None and past_key_values is not None: is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0] if is_padding_right: raise ValueError( "You are attempting to perform batched generation with padding_side='right'" " this may lead to unexpected behaviour for Flash Attention version of Moshi. Make sure to " " call `tokenizer.padding_side = 'left'` before tokenizing the input. " ) if attention_mask is not None and 0.0 in attention_mask: return attention_mask return None # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail # to infer the attention mask. past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 using_static_cache = isinstance(past_key_values, StaticCache) using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache) # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward if ( self.config._attn_implementation == "sdpa" and not (using_static_cache or using_sliding_window_cache) and not output_attentions ): if AttentionMaskConverter._ignore_causal_mask_sdpa( attention_mask, inputs_embeds=input_tensor, past_key_values_length=past_seen_tokens, sliding_window=self.config.sliding_window, is_training=self.training, ): return None dtype, device = input_tensor.dtype, input_tensor.device min_dtype = torch.finfo(dtype).min sequence_length = input_tensor.shape[1] # SlidingWindowCache or StaticCache if using_sliding_window_cache or using_static_cache: target_length = past_key_values.get_max_cache_shape() # DynamicCache or no cache else: target_length = ( attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else past_seen_tokens + sequence_length + 1 ) # In case the provided `attention` mask is 2D, we generate a causal mask here (4D). causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position( attention_mask, sequence_length=sequence_length, target_length=target_length, dtype=dtype, device=device, cache_position=cache_position, batch_size=input_tensor.shape[0], config=self.config, past_key_values=past_key_values, ) if ( self.config._attn_implementation == "sdpa" and attention_mask is not None and attention_mask.device.type in ["cuda", "xpu"] and not output_attentions ): # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. # Details: https://github.com/pytorch/pytorch/issues/110213 causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) return causal_mask @staticmethod # Copied from transformers.models.mistral.modeling_mistral.MistralModel._prepare_4d_causal_attention_mask_with_cache_position with Mistral->Moshi def _prepare_4d_causal_attention_mask_with_cache_position( attention_mask: torch.Tensor, sequence_length: int, target_length: int, dtype: torch.dtype, device: torch.device, cache_position: torch.Tensor, batch_size: int, config: MoshiConfig, past_key_values: Cache, ): """ Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing. Args: attention_mask (`torch.Tensor`): A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`. sequence_length (`int`): The sequence length being processed. target_length (`int`): The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet. dtype (`torch.dtype`): The dtype to use for the 4D attention mask. device (`torch.device`): The device to place the 4D attention mask on. cache_position (`torch.Tensor`): Indices depicting the position of the input sequence tokens in the sequence. batch_size (`torch.Tensor`): Batch size. config (`MoshiConfig`): The model's configuration class past_key_values (`Cache`): The cache class that is being used currently to generate """ if attention_mask is not None and attention_mask.dim() == 4: # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing. causal_mask = attention_mask else: min_dtype = torch.finfo(dtype).min causal_mask = torch.full( (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device ) diagonal_attend_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) if config.sliding_window is not None: # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also # the check is needed to verify is current checkpoint was trained with sliding window or not if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length: sliding_attend_mask = torch.arange(target_length, device=device) <= ( cache_position.reshape(-1, 1) - config.sliding_window ) diagonal_attend_mask.bitwise_or_(sliding_attend_mask) causal_mask *= diagonal_attend_mask causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1) if attention_mask is not None: causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit if attention_mask.shape[-1] > target_length: attention_mask = attention_mask[:, :target_length] mask_length = attention_mask.shape[-1] padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to( causal_mask.device ) padding_mask = padding_mask == 0 causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( padding_mask, min_dtype ) return causal_mask @add_start_docstrings( "The Moshi decoder model with a text language modelling head on top. Only usable for text.", MOSHI_START_DOCSTRING, ) class MoshiForCausalLM(MoshiPreTrainedModel, GenerationMixin): _tied_weights_keys = ["model.embed_tokens.weight", "lm_head.weight"] # Copied from transformers.models.gemma.modeling_gemma.GemmaForCausalLM.__init__ with Gemma->Moshi def __init__(self, config): super().__init__(config) self.model = MoshiModel(config) self.vocab_size = config.vocab_size self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self): return self.model.embed_tokens def set_input_embeddings(self, value): self.model.embed_tokens = value def get_output_embeddings(self): return self.lm_head def set_output_embeddings(self, new_embeddings): self.lm_head = new_embeddings def set_decoder(self, decoder): self.model = decoder def get_decoder(self): return self.model @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep") @add_start_docstrings_to_model_forward(MOSHI_DECODER_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=MoshiCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, labels: Optional[torch.LongTensor] = None, logits_to_keep: Union[int, torch.Tensor] = 0, **kwargs, ) -> Union[Tuple, MoshiCausalLMOutputWithPast]: r""" labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. logits_to_keep (`int` or `torch.Tensor`, *optional*): If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that token can save memory, which becomes pretty significant for long sequences or large vocabulary size. If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension. This is useful when using packed tensor format (single dimension for batch and sequence length). Returns: Example: ```python >>> from transformers import AutoTokenizer, MoshiForCausalLM >>> model = MoshiForCausalLM.from_pretrained("kmhf/hf-moshiko") >>> tokenizer = AutoTokenizer.from_pretrained("kmhf/hf-moshiko") >>> prompt = "What is your favorite condiment?" >>> inputs = tokenizer(prompt, return_tensors="pt") >>> # Generate >>> generate_ids = model.generate(inputs.input_ids, max_length=30) >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] "What is your favorite condiment?" ```""" output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) outputs = self.model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, cache_position=cache_position, ) hidden_states = outputs[0] if labels is None and not is_torchdynamo_compiling(): logger.warning_once( "Starting from v4.46, the `logits` model output will have the same type as the model (except at train time, where it will always be FP32)" ) # Only compute necessary logits, and do not upcast them to float if we are not computing the loss slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep logits = self.lm_head(hidden_states[:, slice_indices, :]) loss = None if labels is not None: # Upcast to float if we need to compute the loss to avoid potential precision issues logits = logits.float() # Shift so that tokens < n predict n shift_logits = logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() # Flatten the tokens shift_logits = shift_logits.view(-1, self.config.vocab_size) shift_labels = shift_labels.view(-1) # Enable model parallelism shift_labels = shift_labels.to(shift_logits.device) loss = self.loss_function( shift_logits, shift_labels, vocab_size=self.config.vocab_size, **kwargs, ) if not return_dict: output = ( logits, hidden_states, ) + outputs[1:] return (loss,) + output if loss is not None else output return MoshiCausalLMOutputWithPast( loss=loss, logits=logits, last_hidden_state=hidden_states, # Ignore copy past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) @add_start_docstrings( "The original Moshi model with an audio encoder, a Moshi depth decoder and a Moshi decoder, for speech-to-speech.", MOSHI_START_DOCSTRING, ) class MoshiForConditionalGeneration(MoshiPreTrainedModel, GenerationMixin): _tied_weights_keys = ["decoder.model.embed_tokens.weight", "decoder.lm_head.weight"] config_class = MoshiConfig main_input_name = "input_ids" supports_gradient_checkpointing = True _supports_flash_attn_2 = True _supports_sdpa = True def __init__(self, config: MoshiConfig): super().__init__(config) # We have 2 * num_codebooks audio embedding layers because we have the user input channel and the model output channel. self.embed_tokens = nn.ModuleList( [nn.Embedding(config.audio_vocab_size + 1, config.hidden_size) for _ in range(2 * config.num_codebooks)] ) self.audio_encoder = AutoModel.from_config(config.audio_encoder_config) self.decoder = MoshiForCausalLM(config) self.depth_decoder = MoshiDepthDecoder(config.depth_decoder_config) self.num_codebooks = config.num_codebooks self.post_init() def get_audio_encoder(self): return self.audio_encoder def get_depth_decoder(self): return self.depth_decoder def get_decoder(self): return self.decoder @add_start_docstrings_to_model_forward(MOSHI_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC) def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.BoolTensor] = None, user_input_values: Optional[torch.FloatTensor] = None, user_audio_codes: Optional[torch.Tensor] = None, moshi_input_values: Optional[torch.FloatTensor] = None, moshi_audio_codes: Optional[torch.Tensor] = None, past_key_values: Tuple[Tuple[torch.FloatTensor]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, text_labels: Optional[torch.LongTensor] = None, audio_labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, **kwargs, ) -> Union[Tuple, Seq2SeqLMOutput]: r""" Returns: Examples: ```python >>> from transformers import MoshiForConditionalGeneration >>> import torch >>> model = MoshiForConditionalGeneration.from_pretrained("kmhf/hf-moshiko") >>> inputs = moshi.get_unconditional_inputs() >>> logits = model(**inputs, ).logits >>> logits.shape # (bsz, seq_len, text_vocab_size) torch.Size([1, 1, 32000]) ```""" return_dict = return_dict if return_dict is not None else self.config.use_return_dict kwargs_audio_encoder = { argument[len("audio_encoder_")]: value for argument, value in kwargs.items() if argument.startswith("audio_encoder_") } kwargs_decoder = { argument[len("decoder_") :]: value for argument, value in kwargs.items() if argument.startswith("decoder_") } kwargs_depth_decoder = { argument[len("depth_decoder_") :]: value for argument, value in kwargs.items() if argument.startswith("depth_decoder_") } # If inputs_embeds is provided, it has the priority over input_ids and audio_codes, which won't be used if inputs_embeds is None: if user_input_values is not None and user_audio_codes is None: user_audio_codes = self.audio_encoder.encode( user_input_values, num_quantizers=self.num_codebooks, **kwargs_audio_encoder )[0] if moshi_input_values is not None and moshi_audio_codes is None: moshi_audio_codes = self.audio_encoder.encode( moshi_input_values, num_quantizers=self.num_codebooks, **kwargs_audio_encoder )[0] audio_codes = torch.cat([moshi_audio_codes, user_audio_codes], dim=1) if input_ids is None and audio_codes is None: raise ValueError( "You must provide at least one of `input_ids`, `inputs_embeds`, `input_values` and `audio_codes`." ) if input_ids is not None: inputs_embeds = self.decoder.model.embed_tokens(input_ids) if audio_codes is not None: audio_inputs_embeds = sum( [self.embed_tokens[codebook](audio_codes[:, codebook]) for codebook in range(audio_codes.shape[1])] ) inputs_embeds = ( audio_inputs_embeds if inputs_embeds is None else audio_inputs_embeds + inputs_embeds.to(audio_inputs_embeds.device) ) # Decode decoder_outputs = self.decoder( attention_mask=attention_mask, inputs_embeds=inputs_embeds, output_attentions=output_attentions, output_hidden_states=output_hidden_states, use_cache=use_cache, past_key_values=past_key_values, return_dict=True, labels=text_labels, **kwargs_decoder, ) decoder_last_hidden_state = decoder_outputs.last_hidden_state depth_decoder_outputs = None final_loss = decoder_outputs.loss if text_labels is not None and audio_labels is not None: # To use depth decoder forward here, we actually need oracle input ids since we're supposed to pass the true input ids audio_labels = self.build_delay_pattern_mask( audio_labels, bos_token_id=self.config.audio_vocab_size, pad_token_id=self.config.audio_vocab_size, max_length=audio_labels.shape[-1] + 1, )[0] # (batch_size, sequence_length) -> (batch_size * sequence_length, 1) text_labels = text_labels.view(-1, 1) # (batch_size, num_codebooks, sequence_length) -> (batch_size * sequence_length, num_codebooks) audio_labels = audio_labels.transpose(1, 2).reshape(-1, audio_labels.shape[1]) depth_input_ids = torch.cat([text_labels, audio_labels], dim=1) # keep the last codebook out of input_ids depth_input_ids = depth_input_ids[:, :-1] # (batch_size, sequence_length, dim) -> (batch_size * sequence_length, 1, dim) decoder_last_hidden_state = decoder_last_hidden_state.view(-1, 1, decoder_last_hidden_state.shape[-1]) depth_decoder_outputs = self.depth_decoder( last_hidden_state=decoder_last_hidden_state, input_ids=depth_input_ids, attention_mask=attention_mask, labels=audio_labels, **kwargs_depth_decoder, ) final_loss += depth_decoder_outputs.loss if not return_dict: outputs = decoder_outputs.to_tuple() if depth_decoder_outputs is not None: outputs += depth_decoder_outputs.to_tuple() return outputs return MoshiConditionalGenerationOutputWithPast( loss=decoder_outputs.loss, logits=decoder_outputs.logits, last_hidden_state=decoder_last_hidden_state, past_key_values=decoder_outputs.past_key_values, hidden_states=decoder_outputs.hidden_states, attentions=decoder_outputs.attentions, depth_loss=None if depth_decoder_outputs is None else depth_decoder_outputs.loss, audio_logits=None if depth_decoder_outputs is None else depth_decoder_outputs.logits, depth_past_key_values=None if decoder_outputs is None else decoder_outputs.past_key_values, depth_hidden_states=None if decoder_outputs is None else decoder_outputs.hidden_states, depth_attentions=None if decoder_outputs is None else decoder_outputs.attentions, ) def _prepare_attention_mask_for_generation( self, input_ids: torch.LongTensor, generation_config: GenerationConfig, kwargs: Dict[str, Any], ) -> torch.LongTensor: pad_token_id = generation_config.pad_token_id eos_token_id = generation_config.eos_token_id default_attention_mask = torch.ones(input_ids.shape, dtype=torch.long, device=input_ids.device) if pad_token_id is None: return default_attention_mask is_pad_token_in_inputs = (pad_token_id is not None) and torch.isin(input_ids, pad_token_id).any() is_pad_token_not_equal_to_eos_token_id = (eos_token_id is None) or ~torch.isin( eos_token_id, pad_token_id ).any() can_infer_attention_mask = is_pad_token_in_inputs * is_pad_token_not_equal_to_eos_token_id attention_mask_from_padding = input_ids.ne(pad_token_id).long() attention_mask = ( attention_mask_from_padding * can_infer_attention_mask + default_attention_mask * ~can_infer_attention_mask ) return attention_mask def _prepare_inputs_embeds_for_generation( self, input_ids: Optional[torch.LongTensor] = None, user_input_values: Optional[torch.FloatTensor] = None, user_audio_codes: Optional[torch.Tensor] = None, moshi_input_values: Optional[torch.FloatTensor] = None, moshi_audio_codes: Optional[torch.Tensor] = None, inputs_embeds: Optional[torch.FloatTensor] = None, attention_mask: Optional[torch.Tensor] = None, generation_config: Optional[GenerationConfig] = None, apply_delay_pattern_mask: bool = False, concat_unconditional_inputs: bool = False, ): user_delay_pattern_mask = None moshi_delay_pattern_mask = None if ( inputs_embeds is None and input_ids is None and user_input_values is None and user_audio_codes is None and moshi_input_values is None and moshi_audio_codes is None ): raise ValueError( "You must provide at least one of `input_ids`, `user_input_values`, `moshi_input_values`, `user_audio_codes`, `moshi_audio_codes` or `inputs_embeds`." ) # in case inputs_embeds is passed, we might still need to create delay pattern masks if inputs_embeds is None or apply_delay_pattern_mask: if user_input_values is not None and user_audio_codes is None: user_audio_codes = self.audio_encoder.encode(user_input_values, num_quantizers=self.num_codebooks)[0] if moshi_input_values is not None and moshi_audio_codes is None: moshi_audio_codes = self.audio_encoder.encode(moshi_input_values, num_quantizers=self.num_codebooks)[0] if inputs_embeds is None and concat_unconditional_inputs: unconditional_inputs = self.get_unconditional_inputs(num_samples=user_audio_codes.shape[0]) moshi_audio_codes = torch.cat([unconditional_inputs.moshi_audio_codes, moshi_audio_codes], dim=2) user_audio_codes = torch.cat([unconditional_inputs.user_audio_codes, user_audio_codes], dim=2) input_ids = torch.cat([unconditional_inputs.input_ids, input_ids], dim=1) if attention_mask is not None: attention_mask = torch.cat([unconditional_inputs.attention_mask, attention_mask], dim=1) if inputs_embeds is None or apply_delay_pattern_mask: if apply_delay_pattern_mask and user_audio_codes is not None: user_audio_codes, user_delay_pattern_mask = self.build_delay_pattern_mask( user_audio_codes, bos_token_id=self.config.audio_vocab_size, pad_token_id=self.config.audio_vocab_size, max_length=generation_config.max_length, ) if apply_delay_pattern_mask and moshi_audio_codes is not None: moshi_audio_codes, moshi_delay_pattern_mask = self.build_delay_pattern_mask( moshi_audio_codes, bos_token_id=self.config.audio_vocab_size, pad_token_id=self.config.audio_vocab_size, max_length=generation_config.max_length, ) # If inputs_embeds is provided, it has the priority over input_ids and audio_codes, which won't be used if inputs_embeds is None: audio_inputs_embeds = None if user_audio_codes is not None and moshi_audio_codes is not None: audio_codes = torch.cat([moshi_audio_codes, user_audio_codes], dim=1) audio_inputs_embeds = sum( [self.embed_tokens[codebook](audio_codes[:, codebook]) for codebook in range(audio_codes.shape[1])] ) elif moshi_audio_codes is not None: audio_codes = moshi_audio_codes audio_inputs_embeds = sum( [self.embed_tokens[codebook](audio_codes[:, codebook]) for codebook in range(audio_codes.shape[1])] ) elif user_audio_codes is not None: audio_codes = user_audio_codes audio_inputs_embeds = sum( [ self.embed_tokens[codebook](audio_codes[:, codebook + self.num_codebooks]) for codebook in range(audio_codes.shape[1]) ] ) if input_ids is not None: inputs_embeds = self.decoder.model.embed_tokens(input_ids) if audio_inputs_embeds is not None: inputs_embeds = ( audio_inputs_embeds if inputs_embeds is None else audio_inputs_embeds + inputs_embeds.to(audio_inputs_embeds.device) ) return ( inputs_embeds, input_ids, user_audio_codes, moshi_audio_codes, user_delay_pattern_mask, moshi_delay_pattern_mask, attention_mask, ) @torch.no_grad() def generate( self, input_ids: Optional[torch.LongTensor] = None, user_input_values: Optional[torch.FloatTensor] = None, user_audio_codes: Optional[torch.Tensor] = None, moshi_input_values: Optional[torch.FloatTensor] = None, moshi_audio_codes: Optional[torch.Tensor] = None, inputs_embeds: Optional[torch.FloatTensor] = None, return_audio_waveforms: Optional[bool] = True, return_audio_codes: Optional[bool] = None, concat_unconditional_inputs: Optional[bool] = True, **kwargs, ) -> torch.LongTensor: """ Generates sequences of text token ids and audio tokens ids. Parameters: input_ids (`torch.Tensor `of shape `(batch_size, sequence_length), *optional*): The sequence used as a text prompt for the generation. user_input_values (`torch.Tensor `of shape `(batch_size, 1, audio_sequence_length), *optional*): The audio waveforms used as audio user prompt for the generation. user_audio_codes (`torch.Tensor `of shape `(batch_size, num_codebooks, sequence_length), *optional*): The audio codes used as audio user prompt for the generation. Has priority over `user_input_values` and represents the audio "tokens" of `user_input_values` once passed through the audio encoder. moshi_input_values (`torch.Tensor `of shape `(batch_size, 1, audio_sequence_length), *optional*): The audio waveforms used as audio Moshi prompt for the generation. moshi_audio_codes (`torch.Tensor `of shape `(batch_size, num_codebooks, sequence_length), *optional*): The audio codes used as audio Moshi prompt for the generation. Has priority over `moshi_input_values` and represents the audio "tokens" of `moshi_input_values` once passed through the audio encoder. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` and the audio inputs you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert the inputs into associated vectors than the model's internal embedding lookup matrix. return_audio_waveforms (`bool`, *optional*, defaults to `True`): If `False`, won't generate the audio waveforms. return_audio_codes (`bool`, *optional*): If `True`, will also returns the generated audio codes, i.e the intermediate audio "tokens" which transforms to `audio_sequences` once passed through the audio decoder. concat_unconditional_inputs (`bool`, *optional*, defaults to `True`): If `False`, won't concatenate initial audio and text tokens. kwargs (`Dict[str, Any]`, *optional*): Remaining dictionary of keyword arguments that are passed to the `generate` method. Refers to the original [`generate` docstrings](https://huggingface.co/docs/transformers/main/en/main_classes/text_generation#transformers.GenerationMixin.generate) for more information on how to use them. Note that keywords with a *depth_* prefix will be input for the `generate` method of the depth decoder. Otherwise, the latter will use its default generation config. Return: [`MoshiConditionalGenerationGenerateOutput`] """ # multiple generate -> need to create/update device map if hasattr(self, "hf_device_map") and not hasattr(self.depth_decoder, "hf_device_map"): self.depth_decoder.hf_device_map = {} if "" in self.hf_device_map: self.depth_decoder.hf_device_map = self.hf_device_map else: main_device = [d for d in self.hf_device_map.values() if d not in ["cpu", "disk"]][0] self.depth_decoder.hf_device_map = { key[len("depth_decoder") :]: main_device if value in ["cpu", "disk"] else value for key, value in self.hf_device_map.items() if key.startswith("depth_decoder") } # need to remove depth_decoder from the top device_map so that we assign correctly the device for each layer idx in the cache self.hf_device_map = { key: value for key, value in self.hf_device_map.items() if not key.startswith("depth_decoder") } # retrieve depth decoder kwargs depth_decoder_kwargs_keys = {argument for argument in kwargs if argument.startswith("depth_decoder_")} kwargs_depth_decoder = { argument[len("depth_decoder_") :]: kwargs.pop(argument) for argument in depth_decoder_kwargs_keys } # needs to prepare generation config, even though it'll be done again in `generate` generation_config, kwargs = self._prepare_generation_config(kwargs.pop("generation_config", None), **kwargs) input_ids, user_audio_codes, moshi_audio_codes, concat_unconditional_inputs = ( self._check_and_maybe_initalize_inputs( input_ids=input_ids, user_input_values=user_input_values, user_audio_codes=user_audio_codes, moshi_input_values=moshi_input_values, moshi_audio_codes=moshi_audio_codes, inputs_embeds=inputs_embeds, concat_unconditional_inputs=concat_unconditional_inputs, ) ) inputs = inputs_embeds if input_ids is None else input_ids input_ids_length = inputs.shape[-1] + 1 if concat_unconditional_inputs else inputs.shape[-1] has_default_max_length = kwargs.get("max_length") is None and generation_config.max_length is not None has_default_min_length = kwargs.get("min_length") is None and generation_config.min_length is not None generation_config = self._prepare_generated_length( generation_config=generation_config, has_default_max_length=has_default_max_length, has_default_min_length=has_default_min_length, model_input_name="inputs_embeds" if input_ids is None else "input_ids", inputs_tensor=inputs, input_ids_length=input_ids_length, ) # retrieve depth decoder generation config if it exists if hasattr(generation_config, "depth_decoder_config"): depth_decoder_generation_config = generation_config.depth_decoder_config else: # we need to control the number of tokens generated by the depth decoder depth_decoder_generation_config = { "min_length": self.num_codebooks + 1, "max_length": self.num_codebooks + 1, "cache_implementation": "sliding_window", } # update kwargs_depth_decoder: kwargs_depth_decoder have priority over depth_decoder_generation_config depth_decoder_generation_config.update(kwargs_depth_decoder) kwargs_depth_decoder = depth_decoder_generation_config attention_mask = kwargs.pop("attention_mask", None) if attention_mask is None: attention_mask = self._prepare_attention_mask_for_generation( input_ids=input_ids, generation_config=generation_config, kwargs=kwargs, ) ( inputs_embeds, input_ids, user_audio_codes, moshi_audio_codes, user_delay_pattern_mask, moshi_delay_pattern_mask, attention_mask, ) = self._prepare_inputs_embeds_for_generation( input_ids=input_ids, user_input_values=user_input_values, user_audio_codes=user_audio_codes, moshi_input_values=moshi_input_values, moshi_audio_codes=moshi_audio_codes, inputs_embeds=inputs_embeds, attention_mask=attention_mask, generation_config=generation_config, apply_delay_pattern_mask=True, concat_unconditional_inputs=concat_unconditional_inputs, ) # create blank user inputs - moshi needs a constant stream of user inputs blank_input_values = torch.zeros( (inputs_embeds.shape[0], 1, int(self.config.sampling_rate / self.config.audio_encoder_config.frame_rate)), dtype=self.dtype, device=self.device, ) blank_user_audio_codes = self.audio_encoder.encode(blank_input_values, num_quantizers=self.num_codebooks)[0] # set delay pattern mask for the rest of the generation kwargs["user_delay_pattern_mask"] = ( user_delay_pattern_mask if user_delay_pattern_mask is not None else kwargs.get("user_delay_pattern_mask") ) kwargs["moshi_delay_pattern_mask"] = ( moshi_delay_pattern_mask if moshi_delay_pattern_mask is not None else kwargs.get("moshi_delay_pattern_mask") ) self.generated_audio_codes = torch.repeat_interleave( moshi_audio_codes, max(generation_config.num_beams, generation_config.num_return_sequences), dim=0 ) return_dict_in_generate = generation_config.num_beams > 1 or generation_config.return_dict_in_generate output_scores = generation_config.num_beams > 1 or generation_config.output_scores outputs = super().generate( inputs_embeds=inputs_embeds, input_ids=input_ids, generation_config=generation_config, blank_user_audio_codes=blank_user_audio_codes, kwargs_depth_decoder=kwargs_depth_decoder, return_dict_in_generate=return_dict_in_generate, output_scores=output_scores, attention_mask=attention_mask, **kwargs, ) if not return_audio_waveforms and not return_audio_codes: if return_dict_in_generate and not generation_config.return_dict_in_generate: return outputs.sequences return outputs # check if outputs is a dict or tokens if not return_dict_in_generate: output_text_ids = outputs else: output_text_ids = outputs.sequences if generation_config.num_return_sequences > 1: moshi_delay_pattern_mask = torch.repeat_interleave( moshi_delay_pattern_mask, generation_config.num_return_sequences, dim=0 ) if generation_config.num_beams > 1: # we need to reorganize self.last_hidden_states and generated audio codes according to the beam_indices # Beam indices are of shape `input_length + number_generated_tokens` but actually starts # indexing indices at index 0 instead of index `input_length-1`. # We thus discard the last `input_length` indices that are never used. beam_indices = outputs.beam_indices[:, : -moshi_audio_codes.shape[-1]] generated_audio_codes = self.generated_audio_codes[:, :, moshi_audio_codes.shape[-1] :] # we've generated audio tokens `number_generated_tokens-1` times, so we use the corresponding beam indices to # retrieve the right audio tokens expanded_beam_indices = beam_indices[:, :-1].unsqueeze(1).expand(-1, self.num_codebooks, -1) generated_audio_codes = torch.gather(generated_audio_codes, dim=0, index=expanded_beam_indices) # now, rebuild generated audio codes, this time with the right beam tracking moshi_audio_codes = torch.repeat_interleave( moshi_audio_codes, generation_config.num_return_sequences, dim=0 ) self.generated_audio_codes = torch.cat((moshi_audio_codes, generated_audio_codes), dim=2) # use the last beam indice to retrieve the right self.last_hidden_state self.last_hidden_state = torch.index_select(self.last_hidden_state, dim=0, index=beam_indices[:, -1]) # we need to make a last generation with the latest generated tokens last_hidden_state = self.last_hidden_state.view(-1, 1, self.last_hidden_state.shape[-1]) last_generated_audio_codes = self.depth_decoder.generate( last_hidden_state=last_hidden_state, input_ids=output_text_ids[:, -1:].view(-1, 1), **kwargs_depth_decoder, ) last_generated_audio_codes = last_generated_audio_codes[:, 1:].unsqueeze(2) self.generated_audio_codes = torch.cat([self.generated_audio_codes, last_generated_audio_codes], dim=2) # apply the pattern mask to the final audio ids output_audio_codes = self.apply_delay_pattern_mask(self.generated_audio_codes, moshi_delay_pattern_mask) # revert the pattern delay mask by filtering the pad token id and bos token ids mask = moshi_delay_pattern_mask != self.config.audio_vocab_size output_audio_codes = output_audio_codes[mask].reshape(mask.shape[0], self.num_codebooks, -1) output_values = None if return_audio_waveforms: output_values = self.audio_encoder.decode( output_audio_codes, ).audio_values output_audio_codes = output_audio_codes if return_audio_codes else None if generation_config.return_dict_in_generate: return MoshiConditionalGenerationGenerateOutput( audio_sequences=output_values, audio_codes=output_audio_codes, **outputs ) return MoshiConditionalGenerationGenerateOutput( audio_sequences=output_values, sequences=output_text_ids, audio_codes=output_audio_codes ) def prepare_inputs_for_generation( self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, cache_position=None, position_ids=None, use_cache=True, logits_to_keep=None, user_delay_pattern_mask=None, moshi_delay_pattern_mask=None, kwargs_depth_decoder=None, blank_user_audio_codes: Optional[torch.FloatTensor] = None, **kwargs, ): # Overwritten -- Moshi has custom post-processing on the prepared inputs. # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens # Exception 1: when passing input_embeds, input_ids may be missing entries # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here # Exception 3: with synced GPUs cache_position may go out of bounds, but we only want dummy token in that case. # (we can't check exception 3 while compiling) if past_key_values is not None: if ( inputs_embeds is not None # Exception 1 or cache_position[-1] >= input_ids.shape[1] # Exception 3 ): input_ids = input_ids[:, -cache_position.shape[0] :] elif input_ids.shape[1] != cache_position.shape[0]: # Default case (the "else", a no op, is Exception 2) input_ids = input_ids[:, cache_position] # if `inputs_embeds` are passed, we only want to use them in the 1st generation step if inputs_embeds is not None and cache_position[0] == 0: model_inputs = {"inputs_embeds": inputs_embeds, "input_ids": None} else: model_inputs = {"input_ids": input_ids, "inputs_embeds": None} if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2: if model_inputs["inputs_embeds"] is not None: batch_size, sequence_length, _ = inputs_embeds.shape device = inputs_embeds.device else: batch_size, sequence_length = input_ids.shape device = input_ids.device attention_mask = self.decoder.model._prepare_4d_causal_attention_mask_with_cache_position( attention_mask, sequence_length=sequence_length, target_length=past_key_values.get_max_cache_shape(), dtype=self.decoder.lm_head.weight.dtype, device=device, cache_position=cache_position, batch_size=batch_size, config=self.config, past_key_values=past_key_values, ) model_inputs.update( { "position_ids": position_ids, "past_key_values": past_key_values, "use_cache": use_cache, "attention_mask": attention_mask, "cache_position": cache_position, } ) # 2. Now that everything is prepared, generate audio_codes using the depth decoder # we want to do it after a first token has been generated if model_inputs["input_ids"] is not None: last_hidden_state = kwargs.get("last_hidden_state") # (batch_size, sequence_length, dim) -> (batch_size * sequence_length, 1, dim) last_hidden_state = last_hidden_state.view(-1, 1, last_hidden_state.shape[-1]) input_ids = model_inputs.pop("input_ids") generated_audio_codes = self.depth_decoder.generate( last_hidden_state=last_hidden_state, input_ids=input_ids.view(-1, 1), **kwargs_depth_decoder, ) # the first tokens are text tokens generated_audio_codes = generated_audio_codes[:, 1:].unsqueeze(2) user_audio_codes = self.apply_delay_pattern_mask( torch.cat( [self.generated_audio_codes, blank_user_audio_codes.to(self.generated_audio_codes.device)], dim=2 ), user_delay_pattern_mask, )[:, :, -1:] self.generated_audio_codes = self.apply_delay_pattern_mask( torch.cat([self.generated_audio_codes, generated_audio_codes], dim=2), moshi_delay_pattern_mask ) inputs_embeds, _, _, _, _, _, _ = self._prepare_inputs_embeds_for_generation( input_ids, moshi_audio_codes=self.generated_audio_codes[:, :, -1:], user_audio_codes=user_audio_codes ) model_inputs["input_ids"] = None model_inputs["inputs_embeds"] = inputs_embeds return model_inputs def _update_model_kwargs_for_generation( self, outputs: ModelOutput, model_kwargs: Dict[str, Any], is_encoder_decoder: bool = False, num_new_tokens: int = 1, ) -> Dict[str, Any]: model_kwargs = super()._update_model_kwargs_for_generation( outputs, model_kwargs, is_encoder_decoder, num_new_tokens ) # update last_hidden_state that'll be used in the depth decoder model_kwargs["last_hidden_state"] = outputs.get("last_hidden_state")[:, -1:] # dirty, but we need to make a last depth_decoder.generate self.last_hidden_state = outputs.get("last_hidden_state")[:, -1:] return model_kwargs def get_input_embeddings(self): return self.decoder.get_input_embeddings() def set_input_embeddings(self, value): self.decoder.set_input_embeddings(value) def get_output_embeddings(self): return self.decoder.get_output_embeddings() def set_output_embeddings(self, new_embeddings): self.decoder.set_output_embeddings(new_embeddings) def freeze_audio_encoder(self): """ Freeze the audio encoder weights. """ for param in self.audio_encoder.parameters(): param.requires_grad = False self.audio_encoder._requires_grad = False def freeze_depth_decoder(self): """ Freeze the depth encoder weights. """ for param in self.depth_decoder.parameters(): param.requires_grad = False self.depth_decoder._requires_grad = False @staticmethod # Copied from transformers.models.musicgen.modeling_musicgen.MusicgenForCausalLM.apply_delay_pattern_mask def apply_delay_pattern_mask(input_ids, decoder_pad_token_mask): """Apply a delay pattern mask to the decoder input ids, only preserving predictions where the mask is set to -1, and otherwise setting to the value detailed in the mask.""" seq_len = input_ids.shape[-1] decoder_pad_token_mask = decoder_pad_token_mask[..., :seq_len] input_ids = torch.where(decoder_pad_token_mask == -1, input_ids, decoder_pad_token_mask) return input_ids def build_delay_pattern_mask( self, input_ids: torch.LongTensor, bos_token_id: int, pad_token_id: int, max_length: Optional[int] = None ): """Build a delayed pattern mask to the input_ids. Each codebook, except the first one, is offset by one, giving a delayed pattern mask at the start of sequence and end of sequence. Take the example where there are 4 codebooks and a max sequence length of 6, we have the delayed pattern mask of shape `(codebooks, seq_len)`: - [-1, -1, -1, -1, -1, P] - [ B, -1, -1, -1, -1, -1] - [ B, -1, -1, -1, -1, -1] - [ B, -1, -1, -1, -1, -1] where B is the beginning-of-sentence token, P is the special padding token id and -1 indicates that the token is valid for prediction. If we include a prompt (input ids), the -1 positions indicate where new tokens should be predicted. Otherwise, the mask is set to the value in the prompt: - [ a0, a1, -1, -1, -1, P] - [ B, b0, b1, -1, -1, -1] - [ B, c0, c1, -1, -1, -1] - [ B, d0, d1, -1, -1, -1] where a-d indicate the codebook channel and 0/1 indicates the temporality. Now, we only override the -1 tokens in our prediction. """ bsz, num_codebooks, seq_len = input_ids.shape max_length = max_length if max_length is not None else self.generation_config.max_length input_ids_shifted = ( torch.ones((bsz, num_codebooks, max_length), dtype=torch.long, device=input_ids.device) * -1 ) # the first codebook channel is not shifted seq_len_to_keep = min(seq_len, max_length - 1) input_ids_shifted[:, 0, :seq_len_to_keep] = input_ids[:, 0, :seq_len_to_keep] # fill the shifted ids with the prompt entries input_ids_shifted[:, 1:, 1 : seq_len_to_keep + 1] = input_ids[:, 1:, :seq_len_to_keep] # fill with BOS and PAD input_ids_shifted[:, 1:, 0] = bos_token_id input_ids_shifted[:, 0, -1] = pad_token_id # construct a pattern mask that indicates the positions of BOS and PAD tokens for each codebook pattern_mask = input_ids_shifted input_ids = input_ids_shifted[..., :seq_len_to_keep] return input_ids, pattern_mask def get_unconditional_inputs(self, num_samples=1): """ Helper function to get null inputs for unconditional generation, enabling the model to be used without the feature extractor or tokenizer. Args: num_samples (int, *optional*): Number of audio samples to unconditionally generate. max_new_tokens (int, *optional*): Number of tokens to generate for each sample. More tokens means longer audio samples, at the expense of longer inference (since more audio tokens need to be generated per sample). Example: ```python >>> from transformers import MoshiForConditionalGeneration >>> model = MoshiForConditionalGeneration.from_pretrained("kmhf/hf-moshiko-pytorch-bf16") >>> # get the unconditional (or 'null') inputs for the model >>> unconditional_inputs = model.get_unconditional_inputs(num_samples=1) >>> audio_samples = model.generate(**unconditional_inputs, max_new_tokens=256) ```""" input_ids = torch.ones((num_samples, 1), device=self.device, dtype=torch.int64) * self.config.vocab_size user_audio_codes = ( torch.ones((num_samples, self.num_codebooks, 1), device=self.device, dtype=torch.int64) * self.config.audio_vocab_size ) moshi_audio_codes = ( torch.ones((num_samples, self.num_codebooks, 1), device=self.device, dtype=torch.int64) * self.config.audio_vocab_size ) attention_mask = torch.ones((num_samples, 1), device=self.device, dtype=torch.long) return MoshiUnconditionalInput( input_ids=input_ids, user_audio_codes=user_audio_codes, moshi_audio_codes=moshi_audio_codes, attention_mask=attention_mask, ) def _check_and_maybe_initalize_inputs( self, input_ids=None, user_input_values=None, user_audio_codes=None, moshi_input_values=None, moshi_audio_codes=None, inputs_embeds=None, concat_unconditional_inputs=None, ): inputs = input_ids if inputs_embeds is None else inputs_embeds user_input = user_audio_codes if user_input_values is None else user_input_values moshi_input = moshi_audio_codes if moshi_input_values is None else moshi_input_values one_input_has_been_passed = (user_input is not None) or (moshi_input is not None) or (inputs is not None) # concat_unconditional_inputs will be False if inputs_embeds is used concat_unconditional_inputs = concat_unconditional_inputs and not ( inputs_embeds is not None and input_ids is None ) # if one or two of the three required inputs have been passed, throws an error if one_input_has_been_passed and (user_input is None): raise ValueError( "No user audio inputs have been passed alongside the other inputs. Make sure either `user_input_values` or `user_audio_codes` is passed or use `MoshiForConditionalGeneration.get_unconditional_inputs`. Check the `MoshiForConditionalGeneration` docstrings for more information." ) elif one_input_has_been_passed and (moshi_input is None): raise ValueError( "No Moshi audio inputs have been passed alongside the other inputs. Make sure either `moshi_input_values` or `moshi_audio_codes` is passed or use `MoshiForConditionalGeneration.get_unconditional_inputs`. Check the `MoshiForConditionalGeneration` docstrings for more information." ) elif one_input_has_been_passed and (inputs is None): raise ValueError( "No `input_ids` or `inputs_embeds` have been passed alongside the other inputs. Make sure `input_ids` is passed or use `MoshiForConditionalGeneration.get_unconditional_inputs`. Check the `MoshiForConditionalGeneration` docstrings for more information." ) elif not one_input_has_been_passed: # if no inputs have been passed, use default values unconditional_inputs = self.get_unconditional_inputs() input_ids = unconditional_inputs.input_ids user_audio_codes = unconditional_inputs.user_audio_codes moshi_audio_codes = unconditional_inputs.moshi_audio_codes # in that case, no need to concat unconditional inputs concat_unconditional_inputs = False else: # check if same sequence length user_seq_length = user_input.shape[-1] moshi_seq_length = moshi_input.shape[-1] tokens_seq_length = inputs.shape[1] ratio = self.config.audio_encoder_config.frame_rate / self.config.sampling_rate moshi_seq_length = math.ceil(moshi_seq_length * ratio) if moshi_audio_codes is None else moshi_seq_length user_seq_length = math.ceil(user_seq_length * ratio) if user_audio_codes is None else user_seq_length if tokens_seq_length != moshi_seq_length or tokens_seq_length != user_seq_length: raise ValueError( "At least one of the 3 inputs of `MoshiForConditionalGeneration` doesn't have the same sequence length as the others." "Make sure that they all have the same sequence length. Check the `MoshiForConditionalGeneration` docstrings for more information." ) return input_ids, user_audio_codes, moshi_audio_codes, concat_unconditional_inputs @staticmethod def _reorder_cache( past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor ) -> Tuple[Tuple[torch.Tensor]]: """ This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct beam_idx at every generation step. """ return tuple( tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past) for layer_past in past_key_values ) __all__ = ["MoshiForCausalLM", "MoshiForConditionalGeneration", "MoshiModel", "MoshiPreTrainedModel"]