[iamwavecut]

Changes

• orbitquant.fused.families.flux3 fuses Black Forest Labs' native flux3.dit.DiT (FLUX.3 Image Klein): per-mode and single-stream blocks, Q|K|V and SwiGLU GEMMs over one set of activation codes, Q/K RMSNorm and RoPE in one kernel, gated residuals in the GEMM epilogues. A mode whose tokens share one time embedding computes its AdaLN modulation on one row.
• Fused groups built on the host kept their codebook surrogate codes on the CPU, and every decode copied them to the GPU: one blocking host-to-device transfer per GEMM. The constants now move to the weight device once. This affects every fused family that is fused after the model is on the GPU.
• The activation quantizer no longer copies a plain bf16 input before the RPBH rotation; the prologue pass writes only the token norms.
• Version 0.13.0.

Measurements

One RTX 5090, Torch 2.12.1 CUDA 13.0, FLUX.3 Klein W4A4 release, 1024x1024, warm median:

Runtime	Text to image	Edit, 1 reference	GPU peak
0.12.0 + bundled family	2.53 s	5.32 s	16.9 GiB
this branch	2.06 s	4.72 s	12.4 GiB*

The images are bit-identical to the 0.12.0 runtime for both workloads. *The peak also includes loader changes in the model bundle (text encoder trimming, expandable allocator segments).

Checks

GPU test run and kernel checks on the RTX 5090 are in progress; results will be added here before merge.

[iamwavecut]

GPU checks, RTX 5090, Torch 2.12.1 CUDA 13.0:

• Full test suite with CUDA/Triton: all fused, kernel and FLUX.3 tests pass. Five failures are outside this change: four tests read files that the synced test tree did not contain (.github workflows, sdist layout), and test_real_cuda_projection_reuse_is_bit_identical fails for main's src too when the native-kernels source tree sits next to it; this branch's src passes it in a bare tree.
• scripts/run_cuda_kernel_checks.sh: all stages completed.
• FLUX.3 Klein release runtime: images bit-identical to the 0.12.0 runtime for text to image and one-reference edit.