Bug Description

When using sageattn3_blackwell mode on RTX 5060 Ti (Blackwell architecture), it throws "CUDA error: misaligned address" error during inference.

Environment

• OS: Windows 11
• GPU: NVIDIA RTX 5060 Ti 16GB (Blackwell architecture, sm_120)
• Python: 3.14.0
• PyTorch: 2.12.0.dev20260318+cu130
• CUDA: 13.0
• SageAttention3: Latest code from GitHub (fresh clone and compile)

Steps to Reproduce

1. Clone SageAttention3 repository from GitHub
2. Compile with Visual Studio 2022 19.44 (added -allow-unsupported-compiler flag to nvcc_flags)
3. Install with pip install -e .
4. Use sageattn3_blackwell mode in ComfyUI with Patch Sage Attention node
5. Run inference

Expected Behavior

Attention computation should work without CUDA errors.

Actual Behavior

RuntimeError: CUDA error: misaligned address
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.

Additional Context

• SageAttention 2.2 works fine on the same system
• The issue appears to be specific to the Blackwell kernel implementation in SageAttention3
• Fresh compilation from latest GitHub code does not resolve the issue
• Related to existing issues change ComfyUI kj-nodes to support sageattn3_blackwell #270 and SageAttention fails to work on RTX 50xx series GPU (e.g. 5090) despite clean venv install #148 about Blackwell support

Notes

I have successfully compiled SageAttention3 after adding -allow-unsupported-compiler flag to handle Visual Studio version compatibility, but the runtime misaligned address error persists. This suggests the issue is in the Blackwell kernel code itself rather than a compilation problem.