Lyra2 Step 1 inference fails with TransformerEngine fused attention: CUDNN_STATUS_SUBLIBRARY_LOADING_FAILED

[destiny718]

Hello, I'm running lyra2 step1 preset trajectory but fails during the first autoregressive denoising step. The model and checkpoints load successfully. The environment can successfully pass the verification step in INSTALL.md (# 8. Verify installation), so the basic installation check appears to be fine. However, real Lyra2 inference still crashes when execution reaches the TransformerEngine fused attention path.

Error

RuntimeError: /TransformerEngine/transformer_engine/common/fused_attn/fused_attn_f16_arbitrary_seqlen.cu:417 in function operator():
cuDNN Error: CUDNN_BACKEND_TENSOR_DESCRIPTOR cudnnFinalize failed
ptrDesc->finalize() cudnn_status: CUDNN_STATUS_SUBLIBRARY_LOADING_FAILED

Relevant traceback

File "lyra_2/_src/networks/wan2pt1.py", line 347, in forward
x = self.attn_op(rope_apply(q, video_size, freqs), rope_apply(k, video_size, freqs), v, video_size)
...
File ".../transformer_engine/pytorch/cpp_extensions/fused_attn.py", line 307, in fused_attn_fwd
output_tensors = tex.fused_attn_fwd(
RuntimeError: ... CUDNN_STATUS_SUBLIBRARY_LOADING_FAILED

Environment

• GPU: NVIDIA H100 80GB
• Driver: 535.129.03
• PyTorch: 2.7.1+cu126
• torchvision: 0.22.1+cu126
• TransformerEngine: 2.13.0
• cuDNN: 9.5.1.17
• CUDA version: 12.6

Question

Can Lyra2 be expected to run correctly in a CUDA 12.6 environment?

[junaoshen]

Same error with you. Have you find the solution?