vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 17:14:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

reenable LCM sampler

Browse source
This commit is contained in:
Concedo 2024-03-05 17:39:21 +08:00
parent ac43e0115c
commit 4eb3a95cbb
3 changed files with 8 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

2
Makefile
View file

@ -143,7 +143,7 @@ endif
# it is recommended to use the CMAKE file to build for cublas if you can - will likely work better
ifdef LLAMA_CUBLAS
CUBLAS_FLAGS = -DGGML_USE_CUBLAS -DSD_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
CUBLASLD_FLAGS = -lcuda -lcublas -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L/usr/local/cuda/targets/aarch64-linux/lib -L/usr/lib/wsl/lib
CUBLASLD_FLAGS = -lcuda -lcublas -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L/usr/local/cuda/targets/aarch64-linux/lib -L/usr/local/cuda/targets/sbsa-linux/lib -L/usr/lib/wsl/lib
CUBLAS_OBJS = ggml-cuda.o ggml_v3-cuda.o ggml_v2-cuda.o ggml_v2-cuda-legacy.o
NVCC = nvcc
NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math

6
koboldcpp.py
View file

@ -982,7 +982,7 @@ Enter Prompt:<br>
if friendlysdmodelname=="inactive" or fullsdmodelpath=="":
response_body = (json.dumps([]).encode())
else:
response_body = (json.dumps([{"name":"Euler a","aliases":["k_euler_a","k_euler_ancestral"],"options":{}},{"name":"Euler","aliases":["k_euler"],"options":{}},{"name":"Heun","aliases":["k_heun"],"options":{}},{"name":"DPM2","aliases":["k_dpm_2"],"options":{}},{"name":"DPM++ 2M","aliases":["k_dpmpp_2m"],"options":{}}]).encode())
response_body = (json.dumps([{"name":"Euler a","aliases":["k_euler_a","k_euler_ancestral"],"options":{}},{"name":"Euler","aliases":["k_euler"],"options":{}},{"name":"Heun","aliases":["k_heun"],"options":{}},{"name":"DPM2","aliases":["k_dpm_2"],"options":{}},{"name":"DPM++ 2M","aliases":["k_dpmpp_2m"],"options":{}},{"name":"LCM","aliases":["k_lcm"],"options":{}}]).encode())
elif self.path.endswith('/sdapi/v1/latent-upscale-modes'):
response_body = (json.dumps([]).encode())
elif self.path.endswith('/sdapi/v1/upscalers'):
@ -1756,7 +1756,7 @@ def show_new_gui():
quick_gpuname_label.grid(row=3, column=1, padx=75, sticky="W")
quick_gpuname_label.configure(text_color="#ffff00")
quick_gpu_layers_entry,quick_gpu_layers_label = makelabelentry(quick_tab,"GPU Layers:", gpulayers_var, 6, 50,"How many layers to offload onto the GPU.\nVRAM intensive, usage increases with model and context size.\nRequires some trial and error to find the best fit value.")
quick_lowvram_box = makecheckbox(quick_tab, "Low VRAM", lowvram_var, 4,0,tooltiptxt="Low VRAM mode avoids offloading the KV cache to the GPU.")
quick_lowvram_box = makecheckbox(quick_tab, "Low VRAM (No KV offload)", lowvram_var, 4,0,tooltiptxt="Avoid offloading KV Cache or scratch buffers to VRAM.\nAllows more layers to fit, but may result in a speed loss.")
quick_mmq_box = makecheckbox(quick_tab, "Use QuantMatMul (mmq)", mmq_var, 4,1,tooltiptxt="Enable MMQ mode instead of CuBLAS for prompt processing. Read the wiki. Speed may vary.")
@ -1795,7 +1795,7 @@ def show_new_gui():
gpuname_label.configure(text_color="#ffff00")
gpu_layers_entry,gpu_layers_label = makelabelentry(hardware_tab,"GPU Layers:", gpulayers_var, 6, 50,"How many layers to offload onto the GPU.\nVRAM intensive, usage increases with model and context size.\nRequires some trial and error to find the best fit value.")
tensor_split_entry,tensor_split_label = makelabelentry(hardware_tab, "Tensor Split:", tensor_split_str_vars, 8, 80, tooltip='When using multiple GPUs this option controls how large tensors should be split across all GPUs.\nUses a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order.\nFor example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1.')
lowvram_box = makecheckbox(hardware_tab, "Low VRAM", lowvram_var, 4,0, tooltiptxt='Select lowvram to not allocate VRAM scratch buffer/ "K" & "V" Cache.\nCan save a decent amount of VRAM, but makes fully offloaded model processing a bit slower.')
lowvram_box = makecheckbox(hardware_tab, "Low VRAM (No KV offload)", lowvram_var, 4,0, tooltiptxt='Avoid offloading KV Cache or scratch buffers to VRAM.\nAllows more layers to fit, but may result in a speed loss.')
mmq_box = makecheckbox(hardware_tab, "Use QuantMatMul (mmq)", mmq_var, 4,1, tooltiptxt="Enable MMQ mode to use finetuned kernels instead of default CuBLAS/HipBLAS for prompt processing.\nRead the wiki. Speed may vary.")
splitmode_box = makecheckbox(hardware_tab, "Row-Split", rowsplit_var, 5,0, tooltiptxt="Split rows across GPUs instead of splitting layers and KV across GPUs.\nUses the main GPU for small tensors and intermediate results. Speed may vary.")

4
otherarch/sdcpp/sdtype_adapter.cpp
View file

@ -287,6 +287,10 @@ sd_generation_outputs sdtype_generate(const sd_generation_inputs inputs)
{
sd_params->sample_method = sample_method_t::DPM2;
}
else if(sampler=="lcm")
{
sd_params->sample_method = sample_method_t::LCM;
}
else if(sampler=="dpm++ 2m karras" || sampler=="dpm++ 2m")
{
sd_params->sample_method = sample_method_t::DPMPP2M;