MOSS-Image fused family, fused single projections, optional GPTQ rounding (0.14.0)

[iamwavecut]

Changes

• MOSS-Image fused family (orbitquant.fused.families.moss_image) for OpenMOSS MOSSImageTransformer:
  • both RMSNorms and the shared AdaLN modulation run in the activation prologue;
  • Q|K|V is one GEMM, and Q/K RMSNorm with rotate-half RoPE is one kernel;
  • gate|up is one GEMM with the new SiTU epilogue (EPILOGUE_SITU);
  • the gated residuals run in the output GEMM epilogues.
• fused.fuse_linears(model) runs the OrbitQuant projections of a component without a block family (for example a text encoder) as single INT8 GEMMs (FusedLinear). Activations are 8-bit absmax by default.
• orbitquant.gptq is opt-in calibration:
  • it chooses GPTQ codes in the same codebook, bit width, row-norm scale and packed layout, so artifacts load unchanged;
  • groups of one shape round in a batch on the GPU;
  • Hessians and updates run in full FP32 (TF32 noise broke the Cholesky factorization of a Qwen3.5 text encoder).
• README: a GPTQ section, fuse_linears usage, MOSS-Image in the fused-family table, and a Method Conformance note that GPTQ is outside the paper's method.
• Version: 0.14.0.

Verification

• macOS: ruff check, the full pytest suite, run_paper_methodology_checks.sh and run_hf_compat_checks.sh --mode all pass.
• RTX 5090 (Torch 2.12.1, CUDA 13.0): the full suite gives 746 passed, 37 skipped, 3 failed. All 3 failures come from the test setup:
  • 2 tests run git ls-files, and the copy on the pod was not a git checkout;
  • 1 test needs the native packed-matmul package, which was not provisioned; with ORBITQUANT_KERNELS_AUTOFETCH=1 the file passes (5 passed).
• MOSS-Image 4-bit transformer with GPTQ rounding, 8 images at 1024×1024 with 50 steps: mean LPIPS against BF16 is 0.209 (0.371 with nearest rounding).