20 const std::optional<array>& weight,
21 const std::optional<array>& bias,
29 std::optional<float> base,
32 const std::optional<array>& freqs = std::nullopt,
41 const std::optional<array>& mask = std::nullopt,
42 const std::optional<int>& memory_efficient_threshold = std::nullopt,
72 const std::string& name,
73 const std::string& source,
74 bool ensure_row_contiguous)
77 ensure_row_contiguous_(ensure_row_contiguous) {}
80 std::map<std::string, array>& inputs,
81 std::map<std::string, std::vector<int>> output_shapes,
82 std::map<std::string, Dtype> output_dtypes,
83 std::tuple<int, int, int> grid,
84 std::tuple<int, int, int> threadgroup,
85 std::optional<std::map<std::string, TemplateArg>> template_args =
93 bool ensure_row_contiguous_ =
true;
array layer_norm(const array &x, const std::optional< array > &weight, const std::optional< array > &bias, float eps, StreamOrDevice s={})
array affine_dequantize(const array &w, const array &scales, const array &biases, int group_size=64, int bits=4, StreamOrDevice s={})
array rope(const array &x, int dims, bool traditional, std::optional< float > base, float scale, int offset, const std::optional< array > &freqs=std::nullopt, StreamOrDevice s={})
array scaled_dot_product_attention(const array &queries, const array &keys, const array &values, const float scale, const std::optional< array > &mask=std::nullopt, const std::optional< int > &memory_efficient_threshold=std::nullopt, StreamOrDevice s={})
Computes: O = softmax(Q @ K.T) @ V.
std::variant< int, bool, Dtype > TemplateArg
Definition fast.h:67
std::tuple< array, array, array > affine_quantize(const array &w, int group_size=64, int bits=4, StreamOrDevice s={})
array rms_norm(const array &x, const array &weight, float eps, StreamOrDevice s={})
std::variant< std::monostate, Stream, Device > StreamOrDevice
Definition utils.h:14