ComfyUI Node
H3MD LowVRAM LoRA Loader
Same math and inputs as 'Load LoRA (Bypass, Model Only)', but the LoRA term is computed in token slices and added in place, so no full-size temporaries are allocated.
H3MD LowVRAM LoRA Loader
- model
- MODEL
◄lora_name▾►
◄strength_model1.00►
◄stack_modereplace_like_comfy►
◄chunk_tokens8192►
CategoryMiniMax H3/H3MD
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| lora_name | COMBO | 0 options: | |
| strength_model | FLOAT | 1.00-100–100 | — |
| stack_modeopt | COMBO | replace_like_comfy | replace_like_comfy (default): identical to ComfyUI's bypass loader - every bypass loader REPLACES the previous bypass LoRAs, so in a chain only the LAST bypass LoRA is active (normal/merged LoRAs are unaffected). Swapping nodes keeps your current output. stack: keep all upstream bypass LoRAs active too (also absorbs ones from ComfyUI's own bypass loader). This changes the result if your chain had several bypass loaders. |
| chunk_tokensopt | INT | 8192256–1048576 | Tokens per LoRA slice. Peak extra VRAM per layer ~ chunk_tokens x layer width x 2 bytes (8192 x 21504 qkv in bf16 = 0.34 GiB). Lower = less VRAM, slightly slower. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |