FLUX.2-klein-4B-pouw

A pouw shaping repo for omnipearl/FLUX.2-klein-4B. It contains no weights — only the metadata that makes omnipearl/FLUX.2-klein-4B mine OmniPearl Proof-of-Useful-Work while it serves. The base weights are pulled from the base repo on load.

OmniPearl's mining is output-preserving: generation is bit-identical to the base model. The eligible transformer matmuls (in_features == common_dim = 3072) are reused as PoW lottery tickets — you serve real images and mine on the same compute, no second matmul.

It is GPU-agnostic (portable Triton/PyTorch kernels, no CUDA build): RTX 3090 (sm86) → 5090 → H100 → B200, same code.

Mining shape

field	value
base model	omnipearl/FLUX.2-klein-4B
modality	image
common_dim	3072
rank	32
mine_layers	16 (overhead dial; layer count)
pipeline	diffusers

Use

# install the OmniPearl miner into your serving venv (see the OmniPearl repo)
#   uv pip install --no-deps <omnipearl_mining wheel> -e miner-base -e vllm-omnipearl ...
from vllm_omnipearl import omnipearl_load
b = omnipearl_load("omnipearl/FLUX.2-klein-4B-pouw", gateway=False)   # gateway=True for the live chain
b["pipe"]("a single pearl on a clean white desk, studio light")   # serves AND mines
print("wrapped", b["wrapped"], "mining linears; common_dim", b["common_dim"])

gateway=False attaches an idle local job (for testing the mining path); gateway=True connects to a running OmniPearl gateway for the live block template / target.

Notes

• The live PoW job + difficulty target always come from the chain at runtime — never baked into this repo. GPU kernels compile per-arch on first run (one-time, cached on disk).
• Published under the omnipearl organization. Base weights are the apache-2.0 mirror at omnipearl/FLUX.2-klein-4B; original model attribution is preserved there.

Generated by OmniPearl publish_pouw_models.py. License: MIT.