Memory & Quantization¶

Run Sana on smaller GPUs. The biggest model (Sana-1.5 4.8B) normally needs 16 GB; with mode='low' it fits in 8 GB.

Memory modes¶

from strands_sana import sana_set_memory_mode

sana_set_memory_mode(mode="low", model="sana-1.6b-1024")

mode	What it enables	When to use
`"fast"`	Nothing extra — full GPU residence	A100/H100, plenty of VRAM
`"balanced"`	`model_cpu_offload`	Default for CUDA
`"low"`	`sequential_cpu_offload` + `vae_tiling` + `attention_slicing`	8-12 GB VRAM
`"ultra-low"`	low + future int4 quant	4-8 GB VRAM (with `[quantization]`)

Quantization¶

from strands_sana import sana_quantize

# Auto-probe: tries quanto → bitsandbytes → nunchaku
sana_quantize(bits=4, backend="auto", model="sana-1.6b-1024")

Backend	Install	Status
`quanto`	`pip install optimum-quanto`	runtime quant
`bitsandbytes`	`pip install bitsandbytes`	requires reload-with-config
`nunchaku`	`pip install nunchaku` (SVDQuant 4-bit)	pre-quantized weights

# Bundle install:
pip install 'strands-sana[quantization]'

VAE swap (DC-AE-Lite)¶

Switch to a more efficient VAE without retraining the diffusion backbone:

from strands_sana import sana_swap_vae

sana_swap_vae(
    vae_repo="mit-han-lab/dc-ae-lite-f32c32-sana-1.1-diffusers",
    model="sana-1.6b-1024",
)

DC-AE-Lite is ~30% faster decode with similar reconstruction quality.

Cache management¶

Free GPU memory between runs:

from strands_sana import sana_clear_cache

sana_clear_cache()
# {"freed_pipelines": 3}

Hardware suggestions¶

GPU	Recommended config
RTX 4090 (24 GB)	`mode="fast"` + Sana-1.5 4.8B
RTX 4070 Ti (12 GB)	`mode="balanced"` + Sana-1.6B
RTX 3060 (12 GB)	`mode="low"` + Sana-0.6B or Sprint
RTX 2060 (8 GB)	`mode="low"` + 4-bit quant + Sprint
Apple M-series	`mode="balanced"`, MPS auto-detected
CPU only	`mode="balanced"`, expect ~5min/image