ComfyUI Node
H3 SageAttention (Mem Eff)
EXPERIMENTAL! Replaces every MiniMax H3 self-attention with int8-QK / fp8-V SageAttention called at kernel level, so the float q/k/v are freed before the kernel runs. Reduces peak VRAM and overrides the attention mode. Requires a recent sageattention build with CUDA kernels for this GPU.
H3 SageAttention (Mem Eff)
- model
- model
◄head_chunks4►
CategoryPainter/MiniMax
Inputs (2)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| head_chunks | INT | 41–56 | Split the attention call into this many head groups. Heads are independent so the result is unchanged; the kernel's internal transients shrink by the chunk count. Measured on a 56-head/128-dim block at 32k tokens: 1 = no saving (slightly more than stock), 4 = ~7% below stock. 1 is the fastest, higher values trade speed for VRAM. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |