* improvements to tool calling logic (merged changes from old PR branch)
* added some tweaks for improved tool calls to reuse old ctx, but needs testing. refer to PR.
* fixes to some stuff that concedo's modifications broke
* fixed error in reasoning
* extremely hacky way to cache tool list please fix
* oops forgot to add this
* slightly less hacky way to preserve the tool list in context
* prevented unintended toolcalls from happening when LLM states something irrelevant to toolcall decision
* fixed something that broke koboldlite
* fixed bug added by concedo that broke jinja tools
* experimental further compression of tools array, needs testing
* reverted experimental further compression of tools array
* final cleanup
* add newline after memory insert
* changed tool reasoning to always be in json format to enforce including final decision
* used new json format to skip extra llm call when not necessary
* more catching of possible bad llm output
* further cleanup
* got it down to just one llm call!
* better json format
* even better json format
* further refinement to json format
* further refinement to json format
* fixed broken tool calling
* single-call enforced json method now seems to work well. removed fallbacks as they are no longer required.
---------
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Squashed commits:
[621ff6392] sdcpp: optimize the handling of the FeedForward precision fix (+1 squashed commits)
Squashed commits:
[05b16906c] sdcpp: optimize the handling of the FeedForward precision fix
* sd: add backend support for choosing the default sampler
* use the default sampler on the API
* sd: add backend support for the scheduler
* sd: add backend support for distilled guidance
* sd: add backend support for timestep-shift
* sd: add a config field to set default image gen options
