Fix the input_output_stride for 4-dim input.

This commit is contained in:
Satya Kumar Jandhyala 2025-01-21 14:32:49 -08:00
parent 1009fc9a63
commit a4d848277c

View file

@ -95,7 +95,7 @@ Status RotaryEmbedding::ComputeInternal(onnxruntime::webgpu::ComputeContext& con
input_shape.NumDimensions() == 3
? std::vector<uint32_t>({batch_stride, hidden_size, head_size, 1})
: (input_shape.NumDimensions() == 4
? std::vector<uint32_t>({batch_stride, head_size, sequence_length * head_size, 1})
? std::vector<uint32_t>({batch_stride, sequence_length * head_size, head_size, 1})
: std::vector<uint32_t>({}));
program