Gigachad Checkpoint Loader
Two toggles that make checkpoint loading a Sage Attention on-ramp — just install the package first
- model
- clip
- vae
Gigachad Checkpoint Loader is the stock ComfyUI checkpoint loader with two extra toggles bolted on: Sage Attention and Triton. If you've been meaning to try sage-attention to squeeze a bigger model onto your card but don't want to learn the ComfyUI extension dance, this is the friendliest on-ramp - with one honest caveat about what those toggles need.
The node loads a checkpoint the exact same way the built-in CheckpointLoaderSimple does, via comfy.sd.load_checkpoint_guess_config. That means it reads the model type from the checkpoint itself - SD1.5, SDXL, Flux, whatever - and pops out model, clip, and vae as three separate outputs ready to wire into your graph. Nothing novel there, and that's fine. The difference is what happens after load.
The two knobs that matter
ckpt_name- dropdown of everything in yourmodels/checkpointsfolder. The same list you're used to.sage_attention- the interesting one. The tooltip says it best: "Enable Sage Attention for reduced VRAM usage. Requires sageattn package." When on, the node patches the model's attention to route throughsageattn(int8 quantized attention kernels), scoped per-model viatransformer_optionsrather than stomping on global ComfyUI state. That scoping is genuinely good engineering - older implementations patched attention globally and made nodes fight each other; this one travels with the model through the graph.triton- "Enable Triton-accelerated kernels where available." A best-effort enable; if Triton isn't available it degrades gracefully.
The catch you'll hit first
If you flip sage_attention on without the package installed, the node doesn't error - it logs sageattention not installed - Sage Attention skipped and quietly returns your unpatched model. That's the kind of silent no-op that sends people chasing phantom slowdowns. So before you trust the toggle, run:
pip install sageattention
and restart ComfyUI, then check the console after loading for the confirmation line. Sage attention isn't a magic bullet on every card either - it shines on RTX 40-series and newer where the fused kernels map well to the hardware. On older cards you may see nothing but compatibility friction.
Install
Same as every node in this pack:
cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/comfyui-gigachad.git
or install comfyui-gigachad from ComfyUI Manager, then restart. There's no requirements.txt, so the pack itself needs nothing - the optional sageattn install above is on you.
Should you reach for it?
If you already run a sage-attention extension you don't need this node - your setup is doing the same thing with more flexibility. If you're on a fresh install and want a one-tap sage toggle without touching ComfyUI internals, it's a legitimately convenient drop-in. Just remember the baseline truth: this is still a checkpoint loader. It doesn't generate, doesn't sample, doesn't guess your hardware's happy place. It loads, optionally accelerates, and gets out of the way.
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| ckpt_name | COMBO | 0 options: | |
| sage_attention | BOOLEAN | false | Enable Sage Attention for reduced VRAM usage. Requires sageattn package. |
| triton | BOOLEAN | false | Enable Triton-accelerated kernels where available. |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |
| clip | CLIP | — |
| vae | VAE | — |