Moshi
Overview
The Moshi model was proposed in Moshi: a speech-text foundation model for real-time dialogue by Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Hervé Jégou, Edouard Grave and Neil Zeghidour.
Moshi is a speech-text foundation model that casts spoken dialogue as speech-to-speech generation. Starting from a text language model backbone, Moshi generates speech as tokens from the residual quantizer of a neural audio codec, while modeling separately its own speech and that of the user into parallel streams. This allows for the removal of explicit speaker turns, and the modeling of arbitrary conversational dynamics. Moshi also predicts time-aligned text tokens as a prefix to audio tokens. This “Inner Monologue” method significantly improves the linguistic quality of generated speech and provides streaming speech recognition and text-to-speech. As a result, Moshi is the first real-time full-duplex spoken large language model, with a theoretical latency of 160ms, 200ms in practice.
The abstract from the paper is the following:
We introduce Moshi, a speech-text foundation model and full-duplex spoken dialogue framework. Current systems for spoken dialogue rely on pipelines of independent components, namely voice activity detection, speech recognition, textual dialogue and text-to-speech. Such frameworks cannot emulate the experience of real conversations. First, their complexity induces a latency of several seconds between interactions. Second, text being the intermediate modality for dialogue, non-linguistic information that modifies meaning— such as emotion or non-speech sounds— is lost in the interaction. Finally, they rely on a segmentation into speaker turns, which does not take into account overlapping speech, interruptions and interjections. Moshi solves these independent issues altogether by casting spoken dialogue as speech-to-speech generation. Starting from a text language model backbone, Moshi generates speech as tokens from the residual quantizer of a neural audio codec, while modeling separately its own speech and that of the user into parallel streams. This allows for the removal of explicit speaker turns, and the modeling of arbitrary conversational dynamics. We moreover extend the hierarchical semantic-to-acoustic token generation of previous work to first predict time-aligned text tokens as a prefix to audio tokens. Not only this “Inner Monologue” method significantly improves the linguistic quality of generated speech, but we also illustrate how it can provide streaming speech recognition and text-to-speech. Our resulting model is the first real-time full-duplex spoken large language model, with a theoretical latency of 160ms, 200ms in practice, and is available at github.com/kyutai-labs/moshi.
Moshi deals with 3 streams of information:
- The user’s audio
- Moshi’s audio
- Moshi’s textual output
Similarly to ~MusicgenModel, audio is represented with audio codebooks, which can be interpreted like tokens. The main difference between text tokens and audio codebooks is that audio codebooks introduce an additional dimension of information.
Text tokens are typically of dim (batch_size, sequence_length)
but audio tokens are of dim (batch_size, num_codebooks, sequence_length)
.
Moshi’s made of 3 components:
1. The main decoder (Helium in the paper)
It corresponds to MoshiForCausalLM. It is strictly a classic text LLM, that uses an architecture similar to ~GemmaForCausalLM
. In other words, it takes text tokens, embeds them, pass them through the decoder and a language head, to get text logits.
2. The depth decoder
On its own, it’s also a classic LLM, but this time, instead of generating over the time dimension, it generates over the codebook dimension.
It also means that its context length is num_codebooks
, thus it can’t generate more than num_codebooks
.
Note that each timestamp - i.e each codebook - gets its own set of Linear Layers and Embeddings.
3. MimiModel
It’s the audio encoder from Kyutai, that has recently been integrated to transformers, which is used to “tokenize” audio. It has the same use that ~EncodecModel has in ~MusicgenModel.
Tips:
The original checkpoints can be converted using the conversion script src/transformers/models/moshi/convert_moshi_transformers.py
How to use the model:
This implementation has two main aims:
- quickly test model generation by simplifying the original API
- simplify training. A training guide will come soon, but user contributions are welcomed!
It is designed for intermediate use. We strongly recommend using the original implementation to infer the model in real-time streaming.
1. Model generation
Moshi is a streaming auto-regressive model with two streams of audio. To put it differently, one audio stream corresponds to what the model said/will say and the other audio stream corresponds to what the user said/will say.
MoshiForConditionalGeneration.generate() thus needs 3 inputs:
input_ids
- corresponding to the text token historymoshi_input_values
ormoshi_audio_codes
- corresponding to the model audio historyuser_input_values
oruser_audio_codes
- corresponding to the user audio history
These three inputs must be synchronized. Meaning that their lengths must correspond to the same number of tokens.
You can dynamically use the 3 inputs depending on what you want to test:
- Simply check the model response to an user prompt - in that case,
input_ids
can be filled with pad tokens anduser_input_values
can be a zero tensor of the same shape than the user prompt. - Test more complex behaviour - in that case, you must be careful about how the input tokens are synchronized with the audios.
The original model is synchronized text with audio by padding the text in between each token enunciation.
To follow the example of the following image, "Hello, I'm Moshi"
could be transformed to "Hello,<pad><unk>I'm Moshi"
.
MoshiForConditionalGeneration.generate() then auto-regressively feeds to itself its own audio stream, but since it doesn’t have access to the user input stream while using transformers
, it will thus assume that the user is producing blank audio.
>>> from datasets import load_dataset, Audio
>>> import torch, math
>>> from transformers import MoshiForConditionalGeneration, AutoFeatureExtractor, AutoTokenizer
>>> librispeech_dummy = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> # prepare user input audio
>>> librispeech_dummy = librispeech_dummy.cast_column("audio", Audio(sampling_rate=feature_extractor.sampling_rate))
>>> audio_sample = librispeech_dummy[-1]["audio"]["array"]
>>> user_input_values = feature_extractor(raw_audio=audio_sample, sampling_rate=feature_extractor.sampling_rate, return_tensors="pt").to(device=device, dtype=dtype)
>>> # prepare moshi input values - we suppose moshi didn't say anything while the user spoke
>>> moshi_input_values = torch.zeros_like(user_input_values.input_values)
>>> # prepare moshi input ids - we suppose moshi didn't say anything while the user spoke
>>> num_tokens = math.ceil(moshi_input_values.shape[-1] * waveform_to_token_ratio)
>>> input_ids = torch.ones((1, num_tokens), device=device, dtype=torch.int64) * tokenizer.encode("<pad>")[0]
>>> # generate 25 new tokens (around 2s of audio)
>>> output = model.generate(input_ids=input_ids, user_input_values=user_input_values.input_values, moshi_input_values=moshi_input_values, max_new_tokens=25)
>>> text_tokens = output.sequences
>>> audio_waveforms = output.audio_sequences
2. Model training
Most of the work has to be done during data creation/pre-processing, because of the need to align/synchronize streams.
Once it’s done, you can simply forward text_labels
and audio_labels
to MoshiForConditionalGeneration.forward(), alongside the usual inputs, to get the model loss.
A training guide will come soon, but user contributions are welcomed!
How does the model forward the inputs / generate:
The input streams are embedded and combined into
inputs_embeds
.inputs_embeds
is passed through the main decoder, which processes it like a normal LLM would.The main decoder outputs
text logits
but also itslast hidden state
which is calledtemporal context
in the paper.The depth decoder switches the dimension on which we forward / generate (codebooks instead of time). It uses the token generated from
text logits
and thetemporal context
to auto-regressively generate audio codebooks.
This model was contributed by Yoach Lacombe (ylacombe).
The original code can be found here.
MoshiConfig
class transformers.MoshiConfig
< source >( vocab_size = 32000 hidden_size = 4096 num_hidden_layers = 32 num_attention_heads = 32 num_key_value_heads = None audio_vocab_size = None max_position_embeddings = 3000 rope_theta = 10000.0 hidden_act = 'silu' head_dim = None initializer_range = 0.02 use_cache = True sliding_window = 3000 attention_dropout = 0.0 ffn_dim = 22528 rms_norm_eps = 1e-08 num_codebooks = 8 tie_word_embeddings = False **kwargs )
Parameters
- vocab_size (
int
, optional, defaults to 32000) — Vocabulary size of the MoshiDecoder model. Defines the number of different tokens that can be represented by theinputs_ids
passed when callingMoshiDecoder
. - hidden_size (
int
, optional, defaults to 4096) — Dimensionality of the layers and the pooler layer of the main decoder. - num_hidden_layers (
int
, optional, defaults to 32) — Number of decoder layers. - num_attention_heads (
int
, optional, defaults to 32) — Number of attention heads for each attention layer in the main decoder block. - num_key_value_heads (
int
, optional) — This is the number of key_value heads that should be used to implement Grouped Query Attention. Ifnum_key_value_heads=num_attention_heads
, the model will use Multi Head Attention (MHA), ifnum_key_value_heads=1
the model will use Multi Query Attention (MQA) otherwise GQA is used. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details checkout this paper. If it is not specified, will default tonum_attention_heads
. - audio_vocab_size (
int
, optional) — Vocabulary size of the audio part of model. Defines the number of different tokens that can be represented by theaudio_codes
passed when calling the Moshi models. - max_position_embeddings (
int
, optional, defaults to 3000) — The maximum sequence length that this model might ever be used with. Typically, set this to something large just in case (e.g., 512 or 1024 or 2048). - rope_theta (
float
, optional, defaults to 10000.0) — The base period of the RoPE embeddings. - hidden_act (
str
orfunction
, optional, defaults to"silu"
) — The non-linear activation function (function or string) in the decoder. - head_dim (
int
, optional, defaults tohidden_size // num_attention_heads
) — The attention head dimension. - initializer_range (
float
, optional, defaults to 0.02) — The standard deviation of the truncated_normal_initializer for initializing all weight matrices. - use_cache (
bool
, optional, defaults toTrue
) — Whether or not the model should return the last key/values attentions (not used by all models). Only relevant ifconfig.is_decoder=True
. - sliding_window (
int
, optional, defaults to 3000) — Sliding window attention window size. If not specified, will default to3000
. - attention_dropout (
float
, optional, defaults to 0.0) — The dropout ratio for the attention probabilities. - ffn_dim (
int
, optional, defaults to 22528) — Dimensionality of the “intermediate” (often named feed-forward) layer in the main decoder block. Must be even. - rms_norm_eps (
float
, optional, defaults to 1e-08) — The epsilon used by the rms normalization layers. - num_codebooks (
int
, optional, defaults to 8) — The number of audio codebooks for each audio channels. - tie_word_embeddings (
bool
, optional, defaults toFalse
) — Whether to tie weight embeddings - kwargs (optional) —
Dictionary of keyword arguments. Notably:
- audio_encoder_config (PretrainedConfig, optional) — An instance of a configuration object that defines the audio encoder config.
- depth__config (PretrainedConfig, optional) — An instance of a configuration object that defines the depth decoder config.
This is the configuration class to store the configuration of a MoshiModel. It is used to instantiate a Moshi model according to the specified arguments, defining the audio encoder, Moshi depth decoder and Moshi decoder configs. Instantiating a configuration with the defaults will yield a similar configuration to that of the Moshiko model, e.g. kmhf/hf-moshiko
Configuration objects inherit from PretrainedConfig and can be used to control the model outputs. Read the documentation from PretrainedConfig for more information.
Example:
>>> from transformers import (
... MoshiConfig,
... MoshiForConditionalGeneration,
... )
>>> configuration = MoshiConfig()
>>> # Initializing a MoshiForConditionalGeneration (with random weights) from the kmhf/hf-moshiko style configuration
>>> model = MoshiForConditionalGeneration(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
>>> # Saving the model, including its configuration
>>> model.save_pretrained("kmhf/hf-moshiko")
>>> # loading model and config from pretrained folder
>>> moshi_config = MoshiConfig.from_pretrained("kmhf/hf-moshiko")
>>> model = MoshiForConditionalGeneration.from_pretrained("kmhf/hf-moshiko", config=moshi_config)
from_audio_encoder_config
< source >( audio_encoder_config: PretrainedConfig **kwargs ) → MoshiConfig
Instantiate a MoshiConfig (or a derived class) from an audio encoder configuration.
MoshiDepthConfig
class transformers.MoshiDepthConfig
< source >( vocab_size = 32000 hidden_size = 1024 input_size = 4096 num_hidden_layers = 6 num_attention_heads = 16 num_key_value_heads = None audio_vocab_size = 2048 max_position_embeddings = 9 hidden_act = 'silu' head_dim = None initializer_range = 0.02 use_cache = True sliding_window = 8 attention_dropout = 0.0 ffn_dim = 5632 rms_norm_eps = 1e-08 num_codebooks = 8 tie_word_embeddings = False **kwargs )
Parameters
- vocab_size (
int
, optional, defaults to 32000) — Vocabulary size of the MoshiDepthDecoder model. Defines the number of different tokens that can be represented by theinputs_ids
passed when callingMoshiDepthDecoder
. - hidden_size (
int
, optional, defaults to 1024) — Dimensionality of the layers and the pooler layer of the depth decoder. - input_size (
int
, optional, defaults to 4096) — Dimensionality of the input hidden states. Used to connect the main decoder to the depth decoder. - num_hidden_layers (
int
, optional, defaults to 6) — Number of depth decoder layers. - num_attention_heads (
int
, optional, defaults to 16) — Number of attention heads for each attention layer in the depth decoder block. - num_key_value_heads (
int
, optional) — This is the number of key_value heads that should be used to implement Grouped Query Attention. Ifnum_key_value_heads=num_attention_heads
, the model will use Multi Head Attention (MHA), ifnum_key_value_heads=1
the model will use Multi Query Attention (MQA) otherwise GQA is used. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details checkout this paper. If it is not specified, will default tonum_attention_heads
. - audio_vocab_size (
int
, optional, defaults to 2048) — Vocabulary size of the audio part of model. Defines the number of different tokens that can be represented by theaudio_codes
passed when calling the Moshi models. - max_position_embeddings (
int
, optional, defaults to 9) — The maximum sequence length that this model might ever be used with. Typically, set this to something large just in case (e.g., 512 or 1024 or 2048). - hidden_act (
str
orfunction
, optional, defaults to"silu"
) — The non-linear activation function (function or string) in the depth decoder. - head_dim (
int
, optional, defaults tohidden_size // num_attention_heads
) — The attention head dimension. - initializer_range (
float
, optional, defaults to 0.02) — The standard deviation of the truncated_normal_initializer for initializing all weight matrices. - use_cache (
bool
, optional, defaults toTrue
) — Whether or not the model should return the last key/values attentions (not used by all models). Only relevant ifconfig.is_decoder=True
. - sliding_window (
int
, optional, defaults to 8) — Sliding window attention window size. If not specified, will default to8
. - attention_dropout (
float
, optional, defaults to 0.0) — The dropout ratio for the attention probabilities. - ffn_dim (
int
, optional, defaults to 5632) — Dimensionality of the “intermediate” (often named feed-forward) layer in the depth decoder block. Must be even. - rms_norm_eps (
float
, optional, defaults to 1e-08) — The epsilon used by the rms normalization layers. - num_codebooks (
int
, optional, defaults to 8) — The number of audio codebooks for each audio channels. - tie_word_embeddings (
bool
, optional, defaults toFalse
) — Whether to tie weight embeddings - kwargs (optional) —
Dictionary of keyword arguments. Notably:
- audio_encoder_config (PretrainedConfig, optional) — An instance of a configuration object that defines the audio encoder config.
This is the configuration class to store the configuration of a MoshiDepthDecoder
. It is used to instantiate a
Moshi depth decoder model according to the specified arguments, defining the Moshi depth decoder config.
Configuration objects inherit from PretrainedConfig and can be used to control the model outputs. Read the documentation from PretrainedConfig for more information.
Example:
>>> from transformers import (
... MoshiDepthConfig,
... MoshiDepthDecoder,
... )
>>> configuration = MoshiDepthConfig()
>>> # Initializing a MoshiDepthDecoder (with random weights) from the kmhf/hf-moshiko style configuration
>>> model = MoshiDepthDecoder(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
MoshiModel
class transformers.MoshiModel
< source >( config: MoshiConfig )
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 from_pretrained() method to load the model weights. config — MoshiConfig
The bare Moshi Model outputting raw hidden-states without any specific head on top. 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 subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.
Transformer decoder consisting of config.num_hidden_layers layers. Each layer is a MoshiDecoderLayer
forward
< source >( input_ids: LongTensor = None attention_mask: Optional = None position_ids: Optional = None past_key_values: Union = None inputs_embeds: Optional = None use_cache: Optional = None output_attentions: Optional = None output_hidden_states: Optional = None return_dict: Optional = None cache_position: Optional = None )
Parameters
- 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.
- 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.
Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
If
past_key_values
is used, optionally only the lastinput_ids
have to be input (seepast_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 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]
. - past_key_values (
Cache
ortuple(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 thepast_key_values
returned by the model at a previous stage of decoding, whenuse_cache=True
orconfig.use_cache=True
.Two formats are allowed:
- a Cache instance, see our kv cache guide;
- Tuple of
tuple(torch.FloatTensor)
of lengthconfig.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 lastinput_ids
(those that don’t have their past key value states given to this model) of shape(batch_size, 1)
instead of allinput_ids
of shape(batch_size, sequence_length)
. - inputs_embeds (
torch.FloatTensor
of shape(batch_size, sequence_length, hidden_size)
, optional) — Optionally, instead of passinginput_ids
you can choose to directly pass an embedded representation. This is useful if you want more control over how to convertinput_ids
indices into associated vectors than the model’s internal embedding lookup matrix. - use_cache (
bool
, optional) — If set toTrue
,past_key_values
key value states are returned and can be used to speed up decoding (seepast_key_values
). - output_attentions (
bool
, optional) — Whether or not to return the attentions tensors of all attention layers. Seeattentions
under returned tensors for more detail. - output_hidden_states (
bool
, optional) — Whether or not to return the hidden states of all layers. Seehidden_states
under returned tensors for more detail. - return_dict (
bool
, optional) — Whether or not to return a 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 toposition_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.
The MoshiModel forward method, overrides the __call__
special method.
Although the recipe for forward pass needs to be defined within this function, one should call the Module
instance afterwards instead of this since the former takes care of running the pre and post processing steps while
the latter silently ignores them.
MoshiForCausalLM
class transformers.MoshiForCausalLM
< source >( config )
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 from_pretrained() method to load the model weights.
The Moshi decoder model with a text language modelling head on top. Only usable for text. 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 subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.
forward
< source >( input_ids: LongTensor = None attention_mask: Optional = None position_ids: Optional = None past_key_values: Union = None inputs_embeds: Optional = None use_cache: Optional = None output_attentions: Optional = None output_hidden_states: Optional = None return_dict: Optional = None cache_position: Optional = None labels: Optional = None num_logits_to_keep: int = 0 ) → transformers.models.moshi.modeling_moshi.MoshiCausalLMOutputWithPast
or tuple(torch.FloatTensor)
Parameters
- 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.
- 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.
Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
If
past_key_values
is used, optionally only the lastinput_ids
have to be input (seepast_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 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]
. - past_key_values (
Cache
ortuple(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 thepast_key_values
returned by the model at a previous stage of decoding, whenuse_cache=True
orconfig.use_cache=True
.Two formats are allowed:
- a Cache instance, see our kv cache guide;
- Tuple of
tuple(torch.FloatTensor)
of lengthconfig.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 lastinput_ids
(those that don’t have their past key value states given to this model) of shape(batch_size, 1)
instead of allinput_ids
of shape(batch_size, sequence_length)
. - inputs_embeds (
torch.FloatTensor
of shape(batch_size, sequence_length, hidden_size)
, optional) — Optionally, instead of passinginput_ids
you can choose to directly pass an embedded representation. This is useful if you want more control over how to convertinput_ids
indices into associated vectors than the model’s internal embedding lookup matrix. - use_cache (
bool
, optional) — If set toTrue
,past_key_values
key value states are returned and can be used to speed up decoding (seepast_key_values
). - output_attentions (
bool
, optional) — Whether or not to return the attentions tensors of all attention layers. Seeattentions
under returned tensors for more detail. - output_hidden_states (
bool
, optional) — Whether or not to return the hidden states of all layers. Seehidden_states
under returned tensors for more detail. - return_dict (
bool
, optional) — Whether or not to return a 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 toposition_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.Args — 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 (seeinput_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]
.num_logits_to_keep (
int
, optional): Calculate logits for the lastnum_logits_to_keep
tokens. If0
, calculate logits for allinput_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.
Returns
transformers.models.moshi.modeling_moshi.MoshiCausalLMOutputWithPast
or tuple(torch.FloatTensor)
A transformers.models.moshi.modeling_moshi.MoshiCausalLMOutputWithPast
or a tuple of
torch.FloatTensor
(if return_dict=False
is passed or when config.return_dict=False
) comprising various
elements depending on the configuration (MoshiConfig) and inputs.
-
loss (
torch.FloatTensor
of shape(1,)
, optional, returned whenlabels
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 whenuse_cache=True
is passed or whenconfig.use_cache=True
) — Tuple oftuple(torch.FloatTensor)
of lengthconfig.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 whenoutput_hidden_states=True
is passed or whenconfig.output_hidden_states=True
) — Tuple oftorch.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 whenoutput_attentions=True
is passed or whenconfig.output_attentions=True
) — Tuple oftorch.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.
The MoshiForCausalLM forward method, overrides the __call__
special method.
Although the recipe for forward pass needs to be defined within this function, one should call the Module
instance afterwards instead of this since the former takes care of running the pre and post processing steps while
the latter silently ignores them.
Example:
>>> 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?"
MoshiForConditionalGeneration
class transformers.MoshiForConditionalGeneration
< source >( config: MoshiConfig )
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 from_pretrained() method to load the model weights.
The original Moshi model with an audio encoder, a Moshi depth decoder and a Moshi decoder, for speech-to-speech. 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 subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.
forward
< source >( input_ids: Optional = None attention_mask: Optional = None user_input_values: Optional = None user_audio_codes: Optional = None moshi_input_values: Optional = None moshi_audio_codes: Optional = None past_key_values: Tuple = None inputs_embeds: Optional = None text_labels: Optional = None audio_labels: Optional = None use_cache: Optional = None output_attentions: Optional = None output_hidden_states: Optional = None return_dict: Optional = None **kwargs ) → transformers.modeling_outputs.Seq2SeqLMOutput or tuple(torch.FloatTensor)
Parameters
- 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.
- 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.
- 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_valuesand 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_valuesand 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 passinginput_ids
you can choose to directly pass an embedded representation. Ifpast_key_values
is used, optionally only the lastinputs_embeds
have to be input (seepast_key_values
). This is useful if you want more control over how to convertinput_ids
indices into associated vectors than the model’s internal embedding lookup matrix.If
input_ids
andinputs_embeds
are both unset,inputs_embeds
takes the value ofinputs_embeds
. - past_key_values (
Cache
ortuple(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 thepast_key_values
returned by the model at a previous stage of decoding, whenuse_cache=True
orconfig.use_cache=True
.Two formats are allowed:
- a Cache instance;
- Tuple of
tuple(torch.FloatTensor)
of lengthconfig.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 setlabels = 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 setlabels = 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 toTrue
,past_key_values
key value states are returned and can be used to speed up decoding (seepast_key_values
). - output_attentions (
bool
, optional) — Whether or not to return the attentions tensors of all attention layers. Seeattentions
under returned tensors for more detail. - output_hidden_states (
bool
, optional) — Whether or not to return the hidden states of all layers. Seehidden_states
under returned tensors for more detail. - return_dict (
bool
, optional) — Whether or not to return a ModelOutput instead of a plain tuple.
Returns
transformers.modeling_outputs.Seq2SeqLMOutput or tuple(torch.FloatTensor)
A transformers.modeling_outputs.Seq2SeqLMOutput or a tuple of
torch.FloatTensor
(if return_dict=False
is passed or when config.return_dict=False
) comprising various
elements depending on the configuration (MoshiConfig) and inputs.
-
loss (
torch.FloatTensor
of shape(1,)
, optional, returned whenlabels
is provided) — Language modeling loss. -
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). -
past_key_values (
tuple(tuple(torch.FloatTensor))
, optional, returned whenuse_cache=True
is passed or whenconfig.use_cache=True
) — Tuple oftuple(torch.FloatTensor)
of lengthconfig.n_layers
, with each tuple having 2 tensors of shape(batch_size, num_heads, sequence_length, embed_size_per_head)
) and 2 additional tensors of shape(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)
.Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used (see
past_key_values
input) to speed up sequential decoding. -
decoder_hidden_states (
tuple(torch.FloatTensor)
, optional, returned whenoutput_hidden_states=True
is passed or whenconfig.output_hidden_states=True
) — Tuple oftorch.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 decoder at the output of each layer plus the initial embedding outputs.
-
decoder_attentions (
tuple(torch.FloatTensor)
, optional, returned whenoutput_attentions=True
is passed or whenconfig.output_attentions=True
) — Tuple oftorch.FloatTensor
(one for each layer) of shape(batch_size, num_heads, sequence_length, sequence_length)
.Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads.
-
cross_attentions (
tuple(torch.FloatTensor)
, optional, returned whenoutput_attentions=True
is passed or whenconfig.output_attentions=True
) — Tuple oftorch.FloatTensor
(one for each layer) of shape(batch_size, num_heads, sequence_length, sequence_length)
.Attentions weights of the decoder’s cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads.
-
encoder_last_hidden_state (
torch.FloatTensor
of shape(batch_size, sequence_length, hidden_size)
, optional) — Sequence of hidden-states at the output of the last layer of the encoder of the model. -
encoder_hidden_states (
tuple(torch.FloatTensor)
, optional, returned whenoutput_hidden_states=True
is passed or whenconfig.output_hidden_states=True
) — Tuple oftorch.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 encoder at the output of each layer plus the initial embedding outputs.
-
encoder_attentions (
tuple(torch.FloatTensor)
, optional, returned whenoutput_attentions=True
is passed or whenconfig.output_attentions=True
) — Tuple oftorch.FloatTensor
(one for each layer) of shape(batch_size, num_heads, sequence_length, sequence_length)
.Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads.
The MoshiForConditionalGeneration forward method, overrides the __call__
special method.
Although the recipe for forward pass needs to be defined within this function, one should call the Module
instance afterwards instead of this since the former takes care of running the pre and post processing steps while
the latter silently ignores them.
Examples:
>>> 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])
generate
< source >( input_ids: Optional = None user_input_values: Optional = None user_audio_codes: Optional = None moshi_input_values: Optional = None moshi_audio_codes: Optional = None inputs_embeds: Optional = None return_audio_waveforms: Optional = True return_audio_codes: Optional = None concat_unconditional_inputs: Optional = True **kwargs )
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_valuesand 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_valuesand 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 passinginput_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 toTrue
) — IfFalse
, won’t generate the audio waveforms. - return_audio_codes (
bool
, optional) — IfTrue
, will also returns the generated audio codes, i.e the intermediate audio “tokens” which transforms toaudio_sequences
once passed through the audio decoder. - concat_unconditional_inputs (
bool
, optional, defaults toTrue
) — IfFalse
, won’t concatenate initial audio and text tokens. - kwargs (
Dict[str, Any]
, optional) — Remaining dictionary of keyword arguments that are passed to thegenerate
method. Refers to the originalgenerate
docstrings for more information on how to use them. Note that keywords with a depth_ prefix will be input for thegenerate
method of the depth decoder. Otherwise, the latter will use its default generation config.
Generates sequences of text token ids and audio tokens ids.
Helper function to get null inputs for unconditional generation, enabling the model to be used without the feature extractor or tokenizer.
Example:
>>> 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)