Feature: attention slicing + VAE tile decoding for low-memory Apple Silicon

[dahai80]

Context

fusion-mlx (local model serving on Apple Silicon, Metal/MLX) uses mflux-fusion 0.18.0 as its image-generation backend. For large diffusion models (Qwen-Image-2512, Flux2-Dev 12B, SDXL at 1024+) the per-step attention and the VAE decode peak memory can exceed available unified memory on 16GB MacBooks, causing Metal OOM.

diffusers solves this with two APIs:

• enable_attention_slicing() — chunked attention computation (trade speed for memory)
• enable_vae_slicing() / enable_tiling() — VAE decode in tiles

Current behavior

mflux-fusion 0.18.0 has no slicing/chunking API on the inference path. Grepping the package finds only dataset __getitem__(slice) and SeedVR2 window slicing — none relevant to diffusion inference memory. The transformer forward and VAE decode run fully materialized.

There is no way to reduce peak memory without lowering resolution or step count.

Requested

1. An opt-in attention slicing mode on the Flux/Qwen transformer forward (process the sequence dimension in chunks, accumulate).
2. A VAE tiling decode option (decode latent in spatial tiles, blend overlaps).
3. Exposed as kwargs on generate_image() (e.g. attention_slice_size, vae_tile_size) or engine-level setters.

Impact

Blocks fusion-mlx issue #848 (attention/vae slicing support). Without it, 16GB machines cannot run Qwen-Image-2512 at 1024x1024 without OOM. Currently fusion-mlx must warn-and-skip when slicing is requested.

Workaround on our side

Env-gated stub that logs slicing unavailable when a user sets FUSION_IMAGE_SLICING=1, then proceeds without slicing. Not a real fix — needs the upstream API.

Happy to contribute a PR if there's appetite for the feature.

[ianscrivener]

@dahai80 , yes pls contribute a PR

[IonDen]

Most of the VAE half already exists in current upstream mflux (0.19.1). --vae-tiling / --vae-tile-size (and --low-ram, which turns it on) do a real overlapping-tile, blended decode, and Qwen-Image already runs it on the decode path. It may not have surfaced in a grep for "slicing" (it's filed under "tiling"), so it's worth checking against a current upstream install.

Two real gaps remain. FLUX.2 (Klein/Dev) has the tiling config wired but its decode call still ignores it, so those flags don't tile FLUX.2's decode yet. And attention slicing doesn't exist anywhere; the transformer runs one full attention pass per step. That second one is the genuine new feature worth a PR.