[ltx-pipelines] i2v at 1216x832 crashes DiffVAE decoder with illegal memory access (t2v same res OK, i2v 704/768 OK)

Summary

ltx_pipelines.distilled with image conditioning at 1216x832 deterministically crashes in the DiffVAE decoder with cudaErrorIllegalAddress. Same resolution without --image works, and the same conditioning at 704/768 works — it is only the combination.

832 is a legal size (64x13) and passes the pipeline's own divisible-by-64 check.

Repro

python -m ltx_pipelines.distilled \
  --transformer-path ltx-2.5-22b-distilled-transformer-nvfp4.safetensors \
  --quantization nvfp4-prequant \
  --text-encoder-path gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  --video-vae-path ltx-2.5-video-vae-bf16.safetensors \
  --audio-vae-path ltx-2.5-audio-vae-bf16.safetensors \
  --duration-head-path ltx-2.5-duration-head-bf16.safetensors \
  --spatial-upsampler-path ltx-2.3-spatial-upscaler-x2-1.1.safetensors \
  --image cond_1216x832.png 0 1.0 \
  --width 1216 --height 832 --num-frames 97 --seed 101 \
  --prompt "The camera slowly pushes in over the hillside city" \
  --output-path out.mp4

Isolation

Traceback (tail)

ltx_core/model/video_vae/transformer/chunked/block.py:51  in forward_x_ctx
ltx_core/model/video_vae/transformer/chunked/context.py:77 in deferred
ltx_core/model/video_vae/transformer/chunked/context.py:27 in _upsample_then_ctx
RuntimeError: CUDA error: CUBLAS_STATUS_INTERNAL_ERROR when calling cublasGemmStridedBatchedEx(...)
torch.AcceleratorError: CUDA error: an illegal memory access was encountered

Failure is in the decoder's chunked upsample path, not the sampler — looks like a tile-boundary issue at this specific latent shape. Possibly related to #277 (DiffVAE tiling at a boundary), different symptom.

Environment

• LTX-2 @ 400fd31, ltx-pipelines/ltx-core/ltx-kernels 1.2.0
• torch 2.13.0+cu132, natten 0.21.7+torch2130cu132 (pinned wheel), CUDA 13.2
• NVIDIA RTX PRO 6000 Blackwell (sm120), driver 595.45.04
• Distilled NVFP4 transformer, --quantization nvfp4-prequant
• GPU otherwise idle (ruled out VRAM starvation — a separate tenant on the card produced OOM/cuBLAS errors that looked similar; these runs were on a free card)

Aside

Model card says width/height must be divisible by 32; the two-stage pipeline enforces 64. Worth aligning the docs.

Workaround: keep i2v at 1216x704 or 1216x768.