ComfyUI Plugin Optimizes LoRA Loading, Slashing MiniMax-H3 VRAM by 38GB
marres · reddit · 2026-08-11
The ComfyUI plugin DoRA Dynamic LoRA Loader has released v1.0.39, introducing a new Runtime LoRA Bypass mode that effectively solves the VRAM bloat issue when applying LoRAs to massive models.
Optimization Principle
Under the traditional ComfyUI loading mechanism, applying a LoRA materializes a complete copy of the patched base weights. For a huge model like MiniMax-H3, this results in an additional 38GB of actual live VRAM allocation.
The Runtime Bypass mode changes the calculation logic: instead of creating a materialized copy of the weights, it evaluates the low-rank LoRA contribution directly during the forward pass. This achieves the exact same mathematical transformation but completely eliminates the memory duplication caused by copying massive base weights.
Benchmark Results & Use Cases
In tests with MiniMax-H3 (Ref2VA pruned BF16) under HIGHVRAM mode, enabling this feature completely eliminated the 38GB of duplicated VRAM usage, while LoRAs like Turbo remained fully effective.
It is important to note that the maximum 38GB saving specifically occurs in HIGHVRAM mode where the entire model is GPU-resident. In NORMALVRAM or LOWVRAM modes, the steady-state VRAM savings will be smaller since ComfyUI is already aggressively offloading, though it still reduces temporary memory pressure during patching.
More from Infra
- SK hynix to Boost NAND Production in China by 50% with Dalian Fab 2 — zephyr_z9 · 2026-08-11
- Reproducing NVIDIA's Inference Paper on 2xRTX 3090s Cuts TTFT by 2.2x at 32K Context — teortaxesTex · 2026-08-11
- Classic McEliece Crypto Scheme Under Attack, Shattering Quantum-Proof Illusions — jedisct1 · 2026-08-11
- Can a Single NVIDIA DGX Replace All Your AI Subscriptions? — jackedAJ · 2026-08-11
- Rumored 50-Series Super Bumps VRAM: 5070 Ti to 24GB — PROfil_Official · 2026-08-11
- MacBook + DGX Spark: Testing Heterogeneous Inference and KV Cache Shipping — HankYeomans · 2026-08-11