Function at::_flash_attention_backward¶

Function Documentation¶

inline ::std::tuple<at::Tensor, at::Tensor, at::Tensor> at::_flash_attention_backward(const at::Tensor &grad_out, const at::Tensor &query, const at::Tensor &key, const at::Tensor &value, const at::Tensor &out, const at::Tensor &logsumexp, const at::Tensor &cum_seq_q, const at::Tensor &cum_seq_k, int64_t max_q, int64_t max_k, double dropout_p, bool is_causal, const at::Tensor &philox_seed, const at::Tensor &philox_offset, ::std::optional<double> scale = ::std::nullopt)¶