MVSplit-DiT (1000 layers)

Paper: https://arxiv.org/abs/2605.06169
Released weights: https://huggingface.co/StableKirito/mvsplit-dit-1000l

Files

dit.py # DiT model (FusedMVSplitNorm1 residual + RoPE + QK-Norm)
text_encoder.py # Qwen3 text encoder wrapper
vae.py # FLUX.2 AutoEncoder
sample.py # Inference / image sampling entry point
sample_prompts.txt # Example prompts (one per line)
kernels/
fused_mvsplit_rmsnorm.py # Triton kernel: MVSplit + RMSNorm
rmsnorm.py # Triton RMSNorm + QK-Norm
rope.py # Triton RoPE
swiglu.py # Triton packed SwiGLU
_common.py

All Triton kernels have PyTorch fallbacks, so the model runs on machines without Triton — just slower.

Installation

pip install -r requirements.txt
# Triton (required for the fast path; ships with PyTorch on Linux+CUDA):# pip install triton

Tested with PyTorch 2.x on CUDA. CPU works for the fallback path but is impractical at this depth.

Weights

Download three artifacts:

1. DiT checkpoint — model.pt from the HF repo above.
2. FLUX.2 AE — flux2_ae.safetensors (available as ae.safetensors in https://huggingface.co/black-forest-labs/FLUX.2-dev; rename or pass the original name via --flux_vae_path).
3. Qwen3 text encoder — Qwen/Qwen3-0.6B is fetched automatically by transformers on first run; or set --qwen_model_path to a local snapshot.

Place them anywhere and pass the paths via CLI flags (the defaults assume /workspace/...).

Sampling

# Custom prompt
python sample.py \
--checkpoint_path /path/to/model.pt \
--flux_vae_path /path/to/flux2_ae.safetensors \
--qwen_model_path Qwen/Qwen3-0.6B \
--prompt "a red panda climbing a bamboo stalk" \
--output_dir ./samples
# Pick a specific line from the prompts file (1-indexed)
python sample.py --line 5
# Randomly sample N prompts (reproducible with --seed)
python sample.py --num_samples 4 --seed 42
# Generate every prompt in the file
python sample.py --all

Outputs <name>.png plus a metadata.jsonl log under --output_dir.

Key sampling flags

Flag	Default	Meaning
--image_size	256	Square output side in pixels (must be a multiple of 16).
--num_inference_steps	35	Euler steps for the flow-matching ODE.
--cfg_scale	2.0	Classifier-free guidance; >1.0 enables CFG.
--time_shift_alpha	4.0	Time-shift in the flow schedule. Must match training.
--batch_size	4	Prompts per forward pass.

Architecture flags (must match the released checkpoint)

Flag	Default	Notes
--model_width	1024	Hidden size.
--model_head_dim	128	→ 8 attention heads (num_kv_heads = 8, no GQA in this config).
--depth_stages	50	Total depth = depth_stages * blocks_per_stage * 2.
--blocks_per_stage	10	Default config gives 1000 transformer blocks.
--rope_base	10000	2-D RoPE base.
--train_bias_and_rms	off	Toggle if the checkpoint was trained with QKV bias + trainable RMSNorm.

The default flags reproduce the released 1000-layer checkpoint.

Third-party code

The Triton kernels under kernels/ (rmsnorm.py, swiglu.py, rope.py) are derived from Unsloth. Each file carries the upstream copyright header, license text, and a list of modifications. See THIRD_PARTY_LICENSES.md for a top-level summary.

Citation

Paper: https://arxiv.org/abs/2605.06169

@article{lu2026mms,
title = {Mean Mode Screaming: Mean--Variance Split Residuals for 1000-Layer Diffusion Transformers},
author = {Lu, Pengqi},
journal = {arXiv preprint arXiv:2605.06169},
year = {2026},
}