[bghira]

cuDNN SDPA can return NaN gradients from finite inputs, including with cuDNN 9.27 (NVIDIA #1072, NVIDIA #160). Disable cuDNN SDPA during attention initialization, backend selection and context cleanup so native attention and automatic Kohaku delegation cannot select it during training or validation.

Reject explicit cuDNN aliases, remove the CLI/WebUI choice and MiniMax-H3 context-parallel allowance, and remove the obsolete cuDNN backward-layout override. Switch the Krea2 example and quickstart to Kohaku-FA and update OPTIONS, Krea2 and MiniMax-H3 documentation in all five translations.

Validation:

• .venv/bin/python -m unittest -v -f: 7,095 tests passed, 323 skipped.
• Black, staged diff checks and GitHub CI passed.
• L40S Krea2 Raw LoRA: 12 optimizer updates per cell, batch 1, rank 64, INT8 SDNQ weights, BF16 compute, gradient checkpointing and regional reduce-overhead compilation. All gradients were finite. Kohaku profiles confirmed CuTe kernels with cuDNN SDPA disabled. The cuDNN baseline uses the preceding main-branch policy in an isolated process.

Resolution	Backend	Peak Torch allocation	Observed GPU memory peak	Warm step time
512px	Kohaku-FA	16.2 GiB	17.9 GiB	0.70 s
512px	cuDNN baseline	16.3 GiB	17.6 GiB	0.71 s
1024px	Kohaku-FA	17.8 GiB	19.9 GiB	2.20 s
1024px	cuDNN baseline	17.9 GiB	19.9 GiB	2.13 s
1536px	Kohaku-FA	20.7 GiB	23.6 GiB	5.88 s
1536px	cuDNN baseline	20.8 GiB	23.9 GiB	5.32 s
2048px	Kohaku-FA	24.7 GiB	27.8 GiB	13.28 s
2048px	cuDNN baseline	24.8 GiB	28.4 GiB	11.37 s

2048px uses vae_enable_tiling=true for both backends: untiled VAE cache creation runs out of memory on the L40S before denoiser training begins. The smaller resolutions use untiled VAE preprocessing. Peak training memory includes first-use compilation; warm timing excludes the first three updates and the profiled update.