tjE *dZddlZddlmZddlmZddlmZddlZddlm Z ddl m Z dd l mZdd lmZmZdd lmZdd lmZdd lmZddlmZmZddlmZmZddlmZddl m!Z!m"Z"m#Z#m$Z$ddl%m&Z&ddl'm(Z(e$j)e*Z+e#eGdde!Z,GddZ-e#eGdde!Z.e#eGdde!Z/Gdde j0Z1Gdd e j0Z2Gd!d"e j0Z3Gd#d$e j0Z4Gd%d&e j0Z5Gd'd(e j0Z6d)Z7dQd*Z8Gd+d,e j0Z9d-ej:d.e;d/ej:fd0Z< dRd2e j0d3ej:d4ej:d5ej:d6ej:dzd7e=d8e=d9ee"fd:Z>Gd;deZ@Gd?d@e j0ZAGdAdBe j0ZBGdCdDe j0ZCGdEdFe j0ZDGdGdHe j0ZEGdIdJe j0ZFe#GdKdLeZGe#dMNGdOdPeGZHdPdLgZIdS)SzPyTorch Mimi model.N)Callable) dataclass)Optional)nn)initialization)ACT2FN)Cache DynamicCache)!create_sliding_window_causal_mask)GradientCheckpointingLayer)BaseModelOutputWithPast)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack) ModelOutputTransformersKwargsauto_docstringlogging)maybe_autocast) MimiConfigcxeZdZUdZdZejdzed<dZej dzed<dZ e dzed<dZ e dzed<dS) MimiOutputaV audio_codes (`torch.LongTensor` of shape `(batch_size, num_quantizers, codes_length)`, *optional*): Discrete code embeddings computed using `model.encode`. audio_values (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*): Decoded audio values, obtained using the decoder part of Mimi. encoder_past_key_values (`Cache`, *optional*): Pre-computed hidden-states (key and values in the self-attention blocks) that can be used to speed up sequential decoding of the encoder transformer. 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`. The model will output the same cache format that is fed as input. If `past_key_values` are used, the user can optionally input only the last `audio_values` or `audio_codes (those that don't have their past key value states given to this model). decoder_past_key_values (`Cache`, *optional*): Pre-computed hidden-states (key and values in the self-attention blocks) that can be used to speed up sequential decoding of the decoder transformer. 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`. The model will output the same cache format that is fed as input. If `past_key_values` are used, the user can optionally input only the last `audio_values` or `audio_codes (those that don't have their past key value states given to this model). N audio_codes audio_valuesencoder_past_key_valuesdecoder_past_key_values) __name__ __module__ __qualname____doc__rtorch LongTensor__annotations__r FloatTensorrr r d/var/www/html/CA-Chatbot/venv/lib/python3.11/site-packages/transformers/models/mimi/modeling_mimi.pyrr)sx.,0K!D(///-1L%#d*111,0UT\000,0UT\00000r*rc eZdZdZdedeedeedeefdZdej defd Z dej defd Z d S) MimiConv1dPaddingCachea Padding cache for MimiConv1d causal convolutions in order to support streaming via cache padding. See: https://huggingface.co/papers/2005.06720 & https://huggingface.co/papers/2204.07064 A padding cache is a list of cached partial hidden states for each convolution layer. Hidden states are cached from the previous call to the MimiConv1d forward pass, given the padding size. num_layersper_layer_paddingper_layer_padding_modeper_layer_in_channelsct|t|t|h}t|dks||krtd|d||_||_||_dg|z|_dS)NrzExpected `num_layers` (zU) values in `per_layer_padding`, `per_layer_padding_mode` and `per_layer_in_channels`)lenpop ValueErrorr/r0r1 padding_cache)selfr.r/r0r1from_args_num_layerss r+__init__zMimiConv1dPaddingCache.__init__Rs!$$5 6 6