Nodes/ComfyUI-M8nodes/M8 · Local LLM Inference
ComfyUI Node

M8 · Local LLM Inference

Runs inference on a local GGUF model: no network, no API key.Models live in ComfyUI's models/LLM folder.Pair it with an mmproj file in the same folder to enable vision; a skill node can supply a knowledge pack.

By chenr5934-tech·Created 3 days ago·Updated a day ago· 0
M8 · Local LLM Inference
  • image
  • skill
  • text
  • status
◄model▾►
◄system_prompt►
◄user_prompt►
◄extra_text►
◄mmproj▾►
◄max_tokens512►
◄temperature0.70►
◄top_p0.95►
◄ctx4096►
◄gpu_layers-1►
◄thinkingfalse►
CategoryM8/LLM

Inputs (13)

NameTypeDefaultDescription
modelCOMBOThe main .gguf in models/LLM. If the list is empty, check that folder.
system_promptSTRINGSystem prompt. Combined with any skill into the system block, skill first.
user_promptSTRINGWhat you want to ask this round.
extra_textSTRINGExtra text appended after the prompt. Good for additional requirements, style constraints, fields to reference.
imageoptIMAGEImages for the model to look at. Without it the run is text only. Needs an mmproj file next to the model, otherwise vision does nothing.
skilloptM8_SKILLWire the output of a skill loader node here to use it as a knowledge pack.
mmprojoptCOMBOMultimodal projection file. Auto-paired by name; pick one manually if that guessed wrong.
max_tokensoptINT51216–8192—
temperatureoptFLOAT0.700–2—
top_poptFLOAT0.950.05–1—
ctxoptINT4096512–32768Context length. Larger uses more memory; only needed for long conversations or many images.
gpu_layersoptINT-1-1–200-1 = offload as many layers as the GPU allows (recommended; depends on whether llama-cpp-python was built with CUDA). 0 = CPU only.
thinkingoptBOOLEANfalseWhether the model should show its thinking. Off gives cleaner output; some models ignore it, in which case ask in the prompt too.

Outputs (2)

NameTypeDescription
textSTRING—
statusSTRING—