[iamwavecut]

Adds the fused block family for Vimoo 1.0 and Vimoo 1.0 Lite (VimooTransformer3DModel, the Vimoo-AI Diffusers fork) and sets the version to 0.17.0.

Changes

• orbitquant.fused.families.vimoo: the 40 dual-stream blocks. Per stream, LayerNorm and the (1 + scale) * x + shift modulation run in the 8-bit activation prologue, Q|K|V is one GEMM, Q/K RMSNorm and the interleaved video RoPE are one kernel, and both output projections and the feed-forward output add their gated residual in the GEMM epilogue. The feed-forward input GEMM applies GELU in its epilogue: tanh for standard checkpoints, exact (erf) for Lite.
• The transformer sorts the valid text tokens of each sample first. The fused blocks attend over [video | valid text] only, and padded text rows leave the block unchanged.
• Lite (numerical_profile="lite"): residual streams stay in the dtype the transformer hands over (FP32), the separately stored output and feed-forward biases are added in the epilogues, and the observed first frame of image-to-video sampling reads row 1 of two-row modulation tables (mod_index/gate_index).
• Standard: the direct patch embedding hands the blocks a column-major video stream. The blocks work on row-major copies.
• EPILOGUE_GELU_ERF (exact GELU) in the INT8 GEMM and in the BF16 fallback.
• FusedRuntime.linear(..., bias=...) adds a bias stored outside the group.
• matmul_int8_scaled_with_triton rejects an out without unit column stride. Before, such an output was written as if it were row-major.
• Feed-forward rows run in chunks of 32768, which bounds the [rows, 12288] intermediate of long videos.
• GPTQ reports an error of 0 for a projection whose calibration inputs are all zero. Before, the relative error was 0/0 (NaN), and a NaN in a report makes invalid JSON. The codes do not change: GPTQ already sets the weights of such inputs to 0. Vimoo-1.0 image-to-video has one such projection (the feed-forward output of the first block's video stream: every GELU input is below -5.7).
• The version is 0.17.0.

Checks

Parity on captured transformer calls, nearest rounding, against the eager OrbitQuant model and the BF16 original:

Model, GPU	Video tokens, batch	Fused vs eager W4A8	Eager W4A8 vs BF16	Fused vs BF16	Time per forward: BF16 / fused cuDNN / fused INT8 attention
Lite T2V+I2V, RTX PRO 6000	8.3k, 2-3	1.8-2.6 %	15-26 %	15-26 %	2.34 / 0.73 / 0.68 s (step 0)
Vimoo-1.0 T2V, RTX 5090	14.3k, 2	1.5 %	6.5-6.8 %	6.6-6.9 %	5.59 / 1.49 / 1.09 s
Vimoo-1.0 I2V, RTX 5090	14.3k, 2	2.7-4.6 %	15-20 %	15-21 %	5.63 / 1.59 / 1.20 s

GPTQ rounding lowers the quantization error further (release artifacts: mean group error 0.0094 -> 0.00054 for the Lite transformer).

• tests/test_fused_vimoo.py: groups, layouts, padding and both profiles against the eager block (CUDA), with a column-major video input.
• tests/test_gptq.py::test_projection_without_input_reports_zero_error.
• ruff check ., pytest -ra, scripts/run_paper_methodology_checks.sh, scripts/run_hf_compat_checks.sh --mode all: pass locally.
• scripts/run_cuda_kernel_checks.sh: pass on an RTX 5090 (Torch 2.12.1 + CUDA 13.0, Triton 3.7.1).