Load LoRA INT8

Stock Load LoRA and INT8 weights don't naturally get along - ComfyUI's regular LoRA patching is generic, and applying it to an already-quantized model can quietly disable the fast INT8 path on whatever layers it touches. This node is the fix: a LoRA loader built specifically to work with INT8 weights, with three different strategies for how the LoRA actually gets applied.

How it works

The mode input is the whole story here. Stochastic applies the LoRA delta directly into the INT8 weights using stochastic rounding - rounding with a bit of randomized dither so the rounding error doesn't systematically bias in one direction, an approach the README credits to dxqb's INT8 work on OneTrainer (the same dxqb who's the second-most-active contributor to that project). It's the README's own "preferred speed-oriented" choice for LoRAs applied after INT8 is already active. Dynamic keeps a compatible LoRA as a separate runtime addition instead of baking it into the INT8 weights - more flexible if you're toggling LoRAs on and off, at some extra compute cost per step. Standard skips INT8-specific handling entirely and goes through ComfyUI's ordinary MODEL patch path - mainly useful for A/B testing a LoRA before you've committed to quantizing.

Where you put this node in the graph matters as much as which mode you pick. If you already ran stock Load LoRA nodes before converting to INT8 (and baked them in with bake_loaded_loras on Enable INT8 on MODEL), this node is for anything you want to add after the model is already INT8 - a new LoRA, a swap, an experiment - without re-running the whole stock-loader chain.

The inputs and outputs that matter

• mode - Stochastic, Dynamic, or Standard. Start with Stochastic for post-INT8 use.
• model - your INT8 model, from Enable INT8 on MODEL or Load Diffusion Model INT8 (W8A8).
• lora_name - the LoRA file, same dropdown you'd see on any LoRA loader.
• strength (default 1, range -10 to 10) - the usual LoRA weight dial.

Output is a single MODEL, ready for a sampler or another Load LoRA INT8 if you're stacking manually (though for several LoRAs at once, Load LoRA Stack INT8 is the cleaner tool - see that node's own page).

How to install it

• ComfyUI Manager - search "ComfyUI-INT8-Fast-Fork", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/SparknightLLC/ComfyUI-INT8-Fast-Fork, then restart ComfyUI.

No extra model downloads for this node itself - just your LoRA files in the usual loras folder, plus whatever GPU and PyTorch setup the rest of the pack needs (recent ComfyUI, an NVIDIA GPU with real INT8 throughput, a matching PyTorch build).

Common issues & troubleshooting

LoRA seems to do nothing, or the model doesn't feel any faster than before. Check your mode. Standard deliberately skips the INT8-aware handling - it's a testing mode, not the fast path. If you want speed, use Stochastic.

Applying multiple LoRAs one at a time and quality degrades more than expected. Each Stochastic application does its own rounding pass, and rounding loses a little precision each time. If you're layering three or more LoRAs, switch to Load LoRA Stack INT8 instead - in Stochastic mode it combines compatible LoRAs first and rounds once, rather than once per LoRA.

Getting the order backwards. Stock Load LoRA belongs before Enable INT8 on MODEL (baked in via bake_loaded_loras); this node belongs after your model is already INT8. Running stock Load LoRA on an INT8 model instead of this node is exactly the case that can silently break the fast path.

Not sure if Dynamic or Stochastic is right for you. Default to Stochastic - it's what the README calls preferred for post-INT8 work. Reach for Dynamic specifically when a LoRA doesn't play well with the stochastic weight-bake, or when you want to toggle a LoRA at runtime without re-quantizing anything.