ó ²—ýhZãó•SrS\4Sjrg)a#Computes cross-entropy loss using the logits generated by performing the matrix multiplication between the embeddings (e) and classifier (c). This method saves GPU memory by not materializing the logits into GPU main memory. Specifically, this computes ```python loss = F.cross_entropy((e @ c.T).float(), targets) ``` without allocating the intermediary (e @ c.T).float() matrix. :param e: Embedding of the inputs used to compute the logits. Shape (..., D) :param c: Classifier matrix. Shape (NumClasses, D) :param targets: The target class for each input. Values must be in [0, NumClasses). Shape (...) :param ignore_index: If an input as a target of this value, it is ignored in the loss computation. :param softcap: The value for logit softcapping. :param reduction: The reduction to perform over the loss. Supports "mean", "sum", and "none". :param shift: If true, the embedding and targets are assumed to require a shift along the temporal axis to perform next token prediction. Specifically, setting this to true will efficiently compute ```python shift_e = e[..., :-1, :].flatten(0, -2) shift_targets = targets[..., 1:].flatten() loss = F.cross_entropy((shift_e @ c.T), targets) ``` Údocstrcó^•U4SjnU$)Ncól>•SRT5URb UROS-UlU$)NÚ)ÚjoinÚ__doc__)Úfnrs €ÚO/home/james-whalen/.local/lib/python3.13/site-packages/cut_cross_entropy/doc.pyÚadd_docÚadd_doc_start..add_doc&s,ø€Ø—W‘W˜V“_°b·j±jÑ6L¨¯ ª ÐRTÑUˆŒ àˆ ó©)rr s` r Ú add_doc_startr%sø€õð €Nr N)ÚLINEAR_CROSS_ENTROPY_DOCÚstrrr r r Úrsðð ÐðF˜3õr