LTX 2.5 repair/generation fails in SageAttention because input tensors reach sageattn2 as float32 instead of FP16/BF16.

What happened?
Tried generating in Director Mode. Used a compatible image LoRA and reference images for character and environment.
Using Flux2 9b and LTX 2.5. Generation fails. Repair will not work. Still says generation failed.

File "shared\attention.py", line 528, in pay_attention
x = sageattn2_wrapper(...)

File "shared\sage2_core.py", line 744, in sageattn_qk_int8_pv_fp8_cuda
assert dtype in [torch.float16, torch.bfloat16]

AssertionError: Input tensors must be in dtype of torch.float16 or torch.bfloat16