Add tensor operation utilities and performance enhancements#165

Merged

LoserCheems merged 8 commits intomainfrom

fix-bwd-launch-templates

Sep 11, 2025

Commits on Sep 11, 2025

Add static switch utility and utility functions for tensor operations
LoserCheems
committed
Refine backward kernel launch parameters for improved performance on sm86 and sm89 architectures
LoserCheems
committed
Refactor file paths in setup.py for CUDA extensions to align with new directory structure
LoserCheems
committed
Refactor backward pass gradient computation to use sum() instead of random tensor
LoserCheems
committed
Update test configurations for CUDA forward equivalence to include additional head dimensions and adjust keep_window_size
LoserCheems
committed
Refactor backward pass in dynamic mask attention functions to remove dout parameter and use sum() for gradient computation
LoserCheems
committed
Refactor mask processor initialization to remove causal parameter in compute_attn functions
LoserCheems
committed
Remove unused causal parameter from Mask struct definition
LoserCheems
committed