leejet
fix: avoid crash on sdxl loras
Some SDXL LoRAs (eg. PCM) can exceed 12k nodes.
Also tested with DMD2 4 step:
stable-diffusion.cpp/ggml/src/ggml.c:5764: GGML_ASSERT(cgraph->n_nodes < cgraph->size) failed
(...)
#12 0x000055769307cfa3 in ggml_build_forward_expand ()
#13 0x0000557692f0ff56 in LoraModel::build_lora_graph(std::map<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ggml_tensor*, std::less<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > >, std::allocator<std::pair<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const, ggml_tensor*> > >, SDVersion) ()
Before the assertion, the LoRA graph reaches 12608 nodes.
Related: https://github.com/leejet/stable-diffusion.cpp/commit/1be2491dcf5e696e620840dc9149fb071047b500#r155465406
I ran into the problem too, and increasing the node limit didn't seem to have any ill effects (other than the increased memory usage i guess). I can't help but wonder if there's not some way to calculate the required buffer size based on the input instead of leaving it as a static define.
A couple ideas without looking at the code:
- Make it configurable from input
- Double the size if current size has reached the limit
I also agree that computing it from input would be a good idea too.
