Nj! hddlZddlmZddlZddlmZddlmZddlmZddl m Z ddl m Z ddl mZdd lmZej"eZeej(ej(ej(gej(fZdd ed ej(d ed ej(fdZGddej0Zy)N)partial)Callable)nn)Self)LUTConditioner)StreamingTransformer)SimpleMLPAdaLN) FlowLMConfigv_tx_0 num_stepsreturnc |}t|D][}||z }|dz|z }||tj|dddfz|tj|dddfz|}|||z z }]|S)atRebuilds the data sample from starting point x_0. Lagrangian Self Distillation (https://arxiv.org/pdf/2505.18825) Args: v_t: Function taking t and x_t as input and returning the flow. x_0: Starting point from the known distribution. num_steps: Number of steps to take. Returns: x_1_hat: (B, D) Reconstructed data sample. .N)rangetorch ones_like)r r r currentistflow_dirs j/Users/ahmed/devFolder/Ultron/claude-voice/.venv/lib/python3.12/site-packages/pocket_tts/models/flow_lm.py lsd_decodersG 9  M Ui  C!G - -q5??3sBQBw<3P/PRY  8i''  NceZdZdZ ddedededededed ed e ffd Z e d e fd Z dejdejdededededzded eejejffdZdejded ejfdZdejdejdededededzded eejejffdZededed e d efdZxZS) FlowLMModelaTransformer-based flow language model on multiple streams of latents. Args: conditioner (LUTConditioner): Text conditioner for processing text inputs. flow: Flow module that defines the flow loss and sampling strategy. flow_net: Trainable function (cond, t, x_t) -> u_t. dim (int): Dimension of the transformer encoder. norm (str): Normalization method. attribute_dropouts (dict): Attribute dropout probabilities. ldim (int): Latent dimension. stats_ema_decay (float): Decay for the EMA of the latent statistics. **kwargs: Additional parameters for the transformer encoder. N conditionerflow_net transformerdimldimstats_ema_decaytext_padding_weightinsert_bos_before_voicec t |||_||_||_||_||_||_||_|jdtj|||jdtj||tjjtj|||_| |_|j"rFtjjtjdd|j f||_tj&|j|d||_||_tj,|d|_tj&|d||_y) Nemb_std)dtypeemb_meanrF)biasr(gh㈵>)eps)super__init__rr"r#r!r$r(rregister_bufferroneszerosr Parameterrandnbos_embr%bos_before_voiceLinear input_linearr LayerNormout_normout_eos) selfrrr r!r"r#r$r(r% __class__s rr-zFlowLMModel.__init__:s# & .#6     Y 4u(EF ZT)GHxx))%++d%*HI '>$  ' '$)HH$6$6u{{Aq$((CS[`7a$bD !IIdii5N& Sd3 yyau5 rrc\t|jjjS)N)next parametersdevicetype)r:s rr?zFlowLMModel.device\s DOO%&--222rsequencetext_embeddings model_statelsd_decode_stepstemp noise_clamp eos_thresholdctjtj||j|}|j |}|j ||||} | j tj} |dkDsJ| dddf} |j| |kD} | jdd|jfz} |dz} tj| | j| j} |-tjjj!| d| n/tjjj#| d| | | t%|j&| }t)|| || fS) aApply language model on sequence and conditions. Given a tensor of sequence of shape [B, S, ldim], returns the loss in training mode or the reconstructed latent in generation mode. Args: sequence (torch.Tensor): Latents to model. text_embeddings (torch.Tensor): Pre-computed conditioning tensor. lsd_decode_steps (int): Number of steps to decode when generating audio. If zero, the model computes the loss. Returns: (output, eos_output, metrics). If `lsd_decode_steps` is zero, `output` is the loss tensor of shape [B, S], otherwise it is the reconstructed latent. )rCrNg?)r(r?g)meanstd)rJrKab)rwhereisnanr3r6backbonetofloat32r9shaper"emptyr(r?rinitnormal_ trunc_normal_rrr)r:rArBrCrDrErFrGinput_transformer_outr9 noise_shaperKnoiseconditioned_flows rforwardzFlowLMModel.forward`s?2;;u{{84dllHM""8,--Wb-c),,U]];!###)!R%0,,/-?%++CR0DII<? Ci K/D/D_McMcd   HHMM ! !%cs ! ; HHMM ' 'CS[LT_ ' `"4==/B*E3CDgMMrctj||gd}|j||}|jr|j|}|dd|jd df}|S)Nr)r!)rcatr r8rS)r:rXrBrArCrYs rrPzFlowLMModel.backbonesdOV4!<**6;? =="mmOrs $"7D10   8 $ U\\5<<> L MH5<<C 0m "))m r