This issue tracks features and bugs in PyTorch's torch.special module, which is modeled after scipy.special.

Aliases to add

• Alias softmax
• Alias kl_div

Ops to add (NOTE: We are not adding ops until we can reduce their impact on build time and CUDA context size.)

• Add iv
• Add ive
• Add kv
• Add logerfc (see Function request: logerfc, logerfcx special functions #31945)
• Add logerfcx (see Function request: logerfc, logerfcx special functions #31945)
• Add hyp2f1
• Add betainc (being added in [torch.special] Adding betainc #58700)
• Add log_ndtr (reserved for @krshrimali) #74630
• gammaincinv
• gammainccinv

Bugs

• Probable mathematical error in torch.functional.kl_div() KLDivLoss and F.kl_div compute KL(Q || P) rather than KL(P || Q) #57459
• [feature] torch.polygamma : Support Tensor for argument n [feature] torch.polygamma : Support Tensor for argument n #55299
• torch.polygamma inconsistent with scipy.special.polygamma for n >= 1 [bug] torch.polygamma inconsistent with scipy.special.polygamma for n >= 1 #55357

TODO

• Review Docs (especially logit)
• torch.special.expit should point to torch.sigmoid (currently it is opposite)

Notes
Reference PR for migrating ops to special namespace : #53260

cc @mruberry @rgommers @heitorschueroff @kshitij12345