"test/srt/vscode:/vscode.git/clone" did not exist on "fe6cdf8972368d72fe79878e38ba310ef2140163"
- 16 Jul, 2025 1 commit
-
-
Mauro Bisson authored
* Replaced PyTorch's slow permutation. * Split kernel into general and specialized versions (for num_channel <= 8192) * Enabled float4-based vectorized memory access, when possible. * Added runtime dispatch logic for kernel specialization. Aligned attention_fwd_cuda.cu with attention_bwd_cuda.cu in terms of naming conventions and kernel parameters. Extracted shared host/device functions and declarations into a separate module: * attention_utils.cuh * attention_utils.cu
-
