ComfyUI Fork — Bypass LoRA Chain Fix

Fork of comfyanonymous/ComfyUI with fixes for bypass LoRA chaining on INT8 quantized models.

What This Fork Changes

Two files modified from upstream:

comfy/sd.py — Shared Bypass Manager via Attachments

Problem: When multiple LoraLoaderBypass nodes are chained (e.g. distilled LoRA + style LoRA), the original code creates unique injection keys (bypass_lora_1, bypass_lora_2). This causes separate inject/eject cycles on the same modules. On second generation (execution cache hit), the eject order corrupts the forward chain — hook1 restores real_forward, then hook2 restores hook1._bypass_forward, leaving module.forward = hook1._bypass_forward. Next inject, hook1 saves its own _bypass_forward as original_forward and calls itself recursively.

Fix: Single "bypass_lora" injection key. A shared BypassInjectionManager is stored as a model attachment via get_attachment("bypass_lora_manager") / set_attachments(). Subsequent LoRA loaders add adapters to the existing manager instead of creating new injection sets. create_injections() rebuilds all hooks fresh each time.

comfy/weight_adapter/bypass.py — Per-Layer Adapter Offloading

Problem: With --normalvram, adapter weights permanently on GPU waste VRAM. On a 24GB card running Wan 2.2 I2V 14B INT8, every MB counts.

Fix: Adapter weights move to GPU at the start of each _bypass_forward() call, then back to CPU in a finally block. Only one layer's LoRA weights are on GPU at any time, mirroring --normalvram behavior for base model weights.

def _bypass_forward(self, x, *args, **kwargs):
    self._move_adapter_weights_to_device(x.device, x.dtype)
    try:
        base_out = self.original_forward(x, *args, **kwargs)
        h_out = self.adapter.h(x, base_out)
        return self.adapter.g(base_out + h_out)
    finally:
        self._move_adapter_weights_to_device(torch.device("cpu"))

Use Case

This fork is designed for:

• INT8 quantized models (W8A8) where bypass LoRA avoids re-quantization artifacts
• Multiple stacked LoRAs (distilled/turbo + style/NSFW) without infinite recursion
• Low VRAM setups (--normalvram) where per-layer offloading is critical

Staying Up to Date

This fork tracks upstream comfyanonymous/ComfyUI master. To update:

git fetch origin
git merge origin/master

The two modified files may conflict on merge — resolve by reapplying the changes described above.