Z-Image-Turbo, repackaged for H3ddle
The released weights rearranged into one file per subsystem so a native engine
can validate and demand-page each independently. No tensor was retrained,
merged, pruned, or quantized here.
| file | what |
|---|
transformer.safetensors | 30 S3-DiT layers plus 2+2 refiners, int8 ConvRot |
text_encoder.safetensors | Qwen3-4B |
vae_decoder.safetensors | AutoencoderKL's decoder half |
tokenizer.json | the released vocabulary, byte for byte |
Measured performance versus the regular layout
On an M1 Pro, a matched 4,128-token transformer forward measured:
| comparison | regular / baseline | optimized path | resulting gain |
|---|
| complete transformer forward | 23.9 s output-major | 22.5 s input-major | 5.9% (1.06x) |
| individual INT8 projection shapes | 1.00x | about 1.09x | about 9% |
The numerical reference error remained unchanged at 3.25e-02. Actual complete
image-generation gains vary because text encoding and VAE decoding are not
changed by this layout.
What was changed, and why
- One file per subsystem. Each can be checked against a reference and
loaded on its own.
- The autoencoder's encoder half is gone. Text to image never encodes an
image, so half of it is dead weight: 244 tensors down to 138, 167 MB to 99.
- The int8 matrices are stored input-major, transposed from the release.
A GPU tile reading two weights a lane one output column apart finds them
input_dim bytes apart in the original layout, so every lane of a simdgroup
takes its own cache line. Transposed, both reads are adjacent — worth about
9% on Apple silicon. Values are untouched, and ConvRot rotates the activation
rather than the weight, so the quantization is unaffected.
- The final layer's two small linears are back at bf16. Together they are
1.2M of 6.15B parameters — 2.5 MB against 1.2 — so the saving was never real,
and
final_layer.linear at [64, 3840] was the most aggressively quantized
tensor in the model. This reverses quantization rather than applying it.
recipe.json carries the constants that live only in the reference's
Python: the flow-match shift, the latent scaling and shift, the DiT's
hyperparameters. The engine reads them back instead of holding its own copy.
Provenance and licences
Z-Image-Turbo is by Alibaba Tongyi Lab under the Apache License 2.0. The
diffusion transformer is the INT8-ConvRot quantization published by Martin
Rizzo, also Apache 2.0, copied layer for layer. Both licences travel with this
repository; see NOTICE.
