+ QV-jX^RIt^RIHt]P!]4tRRRlltRRltR#)N)loggingcV^8dQhR\PR\PR\PR\PR\PR\R\\,R\R \ R \P/ #)  token_logitsduration_logitstargets logit_lengthstarget_lengthsblank_token_id durationssigma reductionreturn)torchTensorintlistfloatstr)formats"k/Volumes/fast/ai/experiments/ui-tars-smoke/.venv/lib/python3.14/site-packages/transformers/loss/loss_tdt.py __annotate__rsPP,,P\\P\\P<< P LL P  PCyP PP \\Pc v Rp W9d)\RV RRPRV 44 R 24hVPp VPwrrVP 4pVP 4p\ P !VRR 7V, p\ P !VRR 7p\ P!WV 3\ R 4V R 7pR VR&VRRRV3,pV ^8dnVP^4PRV R4p\ P!VRRRV ^, 1R3,^VPR4R7PR4p\ P!\ R 4V R 7p\^W,^, 4EFp\^VV , ^,4p\V^,V 4p\ P !VVV R 7pVV, p.p\#V4EFuwppVV, pV^8pVP%4'gK-VP'^R7p V^8dhVRV V3,VRV V3,,VRV VV3,,p!\ P(!VP^4V!V4p!VP+V!4V^8p"VV",p#V#P%4'gKV^, P'^R7p$V ^8dV$P'V ^, R7MT$p%VRV V$3,XRV V%3,,VRV V$V3,,p!\ P(!V#P^4V!V4p!VP+V!4EKx V'gEK\ P,!V^R 7p&\ P.!V&^R 7VRVV3&EK \ P !WR 7p'\ P!V 3\ R 4V R 7p(\#V4FwppV^8XdKVV, p)V)^8p*V*P%4'gK5V)P'^R7p+VV'V+V3,VV'V+WE3,,VV'V+VV3,,p,\ P,!V(V,.^R 7p-\ P(!V*\ P.!V-^R 7V(4p(K V()p.VP 4pVR8Xd&V.P14VP14, #VR8XdV.P34#VR8XdV.V, P34#VR8XdV.P14#V.#)a 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_batchmeansumzInvalid reduction mode "z". Expected one of z, c38"TFp\V4xK R#5i)N)repr).0rs& r tdt_loss..>sNq`p[\tTUww`ps.)dimz-inf)deviceNNNN)r&index)min)max)rrrrnone)r)rr) ValueErrorjoinr'shaperr log_softmaxfull unsqueezeexpandgathersqueezetensorranger,r+arange enumerateanyclampwhereappendstack logsumexprr)/rrrr r r r r rvalid_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_lossrgsHL(&yk1DTYYNq`pNqEqDrrs t   F"."4"4Ju%%'L%++-O'' "=EO**?C Ju5uV}VTIIg&aA~&=>O qy",,Q/66r5"E,, Aq+EAI+q0 1",,R0  '"+  ll5=8G1ema' (aUQ'AE5!LL%?  M  *FAs_FkG;;==LLQL'EQwa 12%a &9:;(E9a)?@A  ++g&7&7&:GWM%%g. !mG 7*J~~"Q--!-4kk.a8G16a1PIaI- .S)X Z7I J=%-GII&3 !8 #%1 yy{{ MMaM( iN: ;iNRS T I~q!HI J  ;; 84!<KKuxQ'GS '"ZF#))+NM!zz|n00222 l "{{} f '--// e zz| Mrc VPp \VVVPV 4P4VPV 4P4VPV 4P4VVVVR7 #)) rrrr r r r r r)r'rgtor) rrlabelsr label_lengthsr r r rkwargsr's &&&&&&&&&, rParakeetForTDTLossrmsq F !' &!%%'#&&v.224$''/335%  r)r(r)rutilsr get_logger__name__loggerrgrmrrrss+    H %Pfr