diff --git a/aiter/ops/triton/_triton_kernels/flash_attn_triton_amd/interface_v2.py b/aiter/ops/triton/_triton_kernels/flash_attn_triton_amd/interface_v2.py index dd14476b0c..4e958849da 100644 --- a/aiter/ops/triton/_triton_kernels/flash_attn_triton_amd/interface_v2.py +++ b/aiter/ops/triton/_triton_kernels/flash_attn_triton_amd/interface_v2.py @@ -578,6 +578,14 @@ def varlen_bwd( "softcap is not supported in varlen_bwd (expected 0.0)." ) + is_sliding_window = (window_size_left >= 0) or (window_size_right >= 0) + if is_sliding_window: + raise NotImplementedError( + f"Sliding window attention is not yet supported in the AMD Triton backward pass " + f"(window_size_left={window_size_left}, window_size_right={window_size_right}). " + f"Use window_size=(-1, -1) for full attention." + ) + if DEBUG: print() print("varlen_bwd")