^jddlZddlmZejeZ ddej dej dej dej dej d ed eed e d e d ej fdZ ddZ y)N)logging token_logitsduration_logitstargets logit_lengthstarget_lengthsblank_token_id durationssigma reductionreturnc 8 d} || vr(td|ddjd| Dd|j} |j\} } } }|j }|j }t j |d|z }t j |d}t j| | | ft d | }d |d d d d f<|d d d d d d |f}| dkDrj|jdjd| d}t j|d d d d d | dz d d fd|jdjd}t jt d | }td| | zdz D]}td || z dz}t|dz| }t j ||| }||z }g}t#|D]4\}}||z }|d k\}|j%s"|j'd } |d kDrX|d d | |f|d d | |fz|d d | ||fz}!t j(|jd |!|}!|j+|!|d kD}"||"z}#|#j%s|dz j'd }$| dkDr|$j'| dz n|$}%|d d | |$fd d | |%fz|d d | |$|fz}!t j(|#jd |!|}!|j+|!7|st j,|d }&t j.|&d |d d ||f<t j | | }'t j| ft d | }(t#|D]\}}|d k(r ||z })|)d k\}*|*j%s'|)j'd }+||'|+|f||'|+||fz||'|+||fz},t j,|(|,gd }-t j(|*t j.|-d |(}(|( }.|j }|dk(r!|.j1|j1z S|dk(r|.j3S|dk(r|.|z j3S|dk(r|.j1S|.S)ab Compute TDT (Token-and-Duration Transducer) loss (https://arxiv.org/abs/2304.06795). Ported from NeMo's `TDTLossPytorch` with anti-diagonal processing. Unlike standard RNNT loss, this loss trains both the token prediction head and the duration prediction head. It uses vectorized anti-diagonal processing for efficiency: all (t, u) pairs on each anti-diagonal t+u=n are computed in parallel as batched tensor operations. Args: token_logits: Token logits of shape `(batch, T, U+1, vocab_size+1)`. duration_logits: Duration logits of shape `(batch, T, U+1, num_durations)`. targets: Target labels of shape `(batch, U)`. logit_lengths: Encoder output lengths of shape `(batch,)`. target_lengths: Target lengths of shape `(batch,)`. blank_token_id: Blank token id. durations: List of duration values (e.g., `[0, 1, 2, 3, 4]`). sigma: Logit undernormalization constant (see TDT paper). Defaults to `0.0`. reduction: Loss reduction method. One of `"mean_volume"`, `"mean_batch"`, `"mean"`, `"sum"`, or `"none"`, mirroring NeMo's `RNNTLoss` (TDT shares the same reduction knob as RNN-T). Defaults to `"mean"`, the `rnnt_reduction` of the released Parakeet TDT checkpoints. Returns: Scalar loss tensor (or per-example losses if `reduction="none"`). ) mean_volume mean_batchmeansumnonezInvalid reduction mode "z". Expected one of z, c32K|]}t|yw)N)repr).0rs e/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/transformers/loss/loss_tdt.py ztdt_loss..>sNq[\tTUwNqs.)dimz-inf)deviceNr)rindex)minr)maxrrrr) ValueErrorjoinrshapefloattorch log_softmaxfull unsqueezeexpandgathersqueezetensorranger$r#arange enumerateanyclampwhereappendstack logsumexprr)/rrrrr r r r r valid_reductionsr batch_sizemax_tmax_u_token_log_probsduration_log_probs log_alphablank_log_probstargets_expandedlabel_log_probsneg_infnu_startu_end u_indices t_indicesall_candidatesidurt_prevvalid_tt_srccontribvalid_u valid_bothu_src u_src_labelstacked batch_idx log_probst_finalvalid t_clampedterminalcombinedlossess/ rtdt_lossr_sHL((&yk1DTYYNq`pNqEqDrrs t   F"."4"4Jua%%'L%++-O'' "=EO**?C Ju5uV}VTIIaAg&aA~&=>O qy",,Q/66r5"E,, Aq+EAI+q0 1",,R0  '"+  ll5=8G1eema' ()QaUQ'AE5!LL%?  M  * /FAs_FkG;;=LLQL'EQwa 12%a &9:;(E9a)?@A  ++g&7&7&:GWM%%g. !mG 7*J~~"Q--!-4rps    H %P,,P\\P\\P<< P LL P  PCyP PP \\Pv r`