Functions — github.com/Narsil/smelte-rs

hub / github.com/Narsil/smelte-rs / functions

Functions212 in github.com/Narsil/smelte-rs

The shape of the tensor ``` use smelte_rs::gpu::f32::{Tensor, Device}; let device = Device::new(0).unwrap(); let tensor = Tensor::zeros(vec![2, 2], &

src/gpu/f32/tensor.rs:44

↓ 48 callersMethodshape

The shape of the tensor ``` use smelte_rs::cpu::f32::Tensor; let tensor = Tensor::zeros(vec![2, 2]); assert_eq!(tensor.shape(), vec![2, 2]); ```

src/cpu/f32/tensor.rs:23

↓ 40 callersMethodshape

The shape of the tensor ``` use smelte-rs::cpu::f16::Tensor; let tensor = Tensor::zeros(vec![2, 2]); assert_eq!(tensor.shape(), vec![2, 2]); ```

src/cpu/f16/tensor.rs:20

↓ 36 callersMethoddevice_id

The device id

src/gpu/f32/tensor.rs:74

↓ 34 callersMethoddata

The [CudaSlice] holding the data

src/gpu/f32/tensor.rs:49

↓ 34 callersMethodzeros

Creates a new nulled tensor with given shape ``` use smelte_rs::gpu::f32::{Tensor, Device}; let device = Device::new(0).unwrap(); let tensor = Tensor

src/gpu/f32/tensor.rs:85

↓ 27 callersMethoddata_mut

A mutable borrow of [CudaSlice] holding the data

src/gpu/f32/tensor.rs:54

↓ 16 callersMethodcuda

The device of the device

src/gpu/f32/tensor.rs:64

↓ 13 callersMethoddata

A slice to the underlying tensor data ``` use smelte_rs::cpu::f32::Tensor; let tensor = Tensor::zeros(vec![2, 2]); assert_eq!(tensor.data(), vec![0.0

src/cpu/f32/tensor.rs:34

↓ 13 callersMethoddata_mut

A mutable slice to the underlying tensor data ``` use smelte_rs::cpu::f32::Tensor; let mut tensor = Tensor::zeros(vec![2, 2]); tensor.data_mut().iter

src/cpu/f32/tensor.rs:58

↓ 11 callersFunctionmatmul

Regular matrix multiplication

src/gpu/f32/ops.rs:87

↓ 11 callersFunctionmatmul_t

Matrix multiplication matmul(A, B.transposed())

src/gpu/f32/ops.rs:92

↓ 10 callersFunctionadd

tensor elementwise addition. b += a.

src/gpu/f32/ops.rs:221

↓ 9 callersFunctiondevice

()

src/gpu/f32/ops.rs:474

↓ 8 callersMethoddata_mut

A mutable slice to the underlying tensor data ``` use smelte-rs::cpu::f16::Tensor; let mut tensor = Tensor::zeros(vec![2, 2]); tensor.data_mut().iter

src/cpu/f16/tensor.rs:43

↓ 8 callersFunctionto_tensor

(view: TensorView<'data>, device: &Device)

examples/gpt2.rs:47

↓ 8 callersMethodweight

TODO

src/nn/layers/linear.rs:25

↓ 7 callersMethodforward

TODO

src/nn/models/bert.rs:357

↓ 7 callersFunctionto_tensor

(view: TensorView<'data>, device: &Device)

examples/bert.rs:63

↓ 6 callersFunctioncausal_softmax

Causal softmax on the last dimension for tensor `x`. The causality is determined by the shape of `x` and `past_sequence_length` which defines how big

src/gpu/f32/ops.rs:411

↓ 6 callersMethoddata

A slice to the underlying tensor data ``` use smelte-rs::cpu::f16::Tensor; let tensor = Tensor::zeros(vec![2, 2]); assert_eq!(tensor.data(), vec![0.0

src/cpu/f16/tensor.rs:31

↓ 6 callersFunctionlinear_from_prefix

( prefix: &str, tensors: &'a SafeTensors<'a>, device: &Device, )

examples/bert.rs:108

↓ 6 callersFunctionmatmul

Regular matrix multiplication

src/cpu/f32/ops.rs:48

↓ 6 callersFunctionsoftmax

Softmax on the last dimension for tensor `x`

src/gpu/f32/ops.rs:404

↓ 5 callersFunctionbroadcast_add

broacasted tensor elementwise addition. b += a.

src/gpu/f32/ops.rs:252

↓ 5 callersFunctionmatmul_t

Matrix multiplication matmul(A, B.transposed())

src/cpu/f32/ops.rs:53

↓ 4 callersFunctioncausal_softmax

Causal softmax on the last dimension for tensor `x`. The causality is determined by the shape of `x` and `past_sequence_length` which defines how big

src/cpu/f32/ops.rs:362

↓ 4 callersFunctioncausal_softmax

Causal softmax on the last dimension for tensor `x`. The causality is determined by the shape of `x` and `past_sequence_length` which defines how big

src/cpu/f16/ops.rs:299

↓ 4 callersFunctioncuda_split_heads

( src: &F32CudaTensor, dst: &mut F32CudaTensor, )

src/nn/models/bert.rs:180

↓ 4 callersFunctionlinear_from_prefix

( prefix: &str, tensors: &'a SafeTensors<'a>, device: &Device, )

examples/gpt2.rs:96

↓ 4 callersFunctionmatmul

Regular matrix multiplication

src/cpu/f16/ops.rs:60

↓ 4 callersFunctionmatmul_t

Matrix multiplication matmul(A, B.transposed())

src/cpu/f16/ops.rs:65

↓ 4 callersFunctionnormalize

Basic operation for the layernorm. x = (x - x.mean()) / (x.var() + epsilon) `mean` and `var` do not have to be initialized, they are simply passed to

src/gpu/f32/ops.rs:355

↓ 4 callersFunctionselect

Operation for selecting entire rows within tensor `weights`. Each `id` is the index of the row.

src/gpu/f32/ops.rs:42

↓ 4 callersFunctionsplit_heads

(q: &F32Tensor, out_q: &mut F32Tensor)

src/nn/models/bert.rs:73

↓ 3 callersFunctionapply

Applies `func` to every item of the tensor

src/cpu/f32/ops.rs:432

↓ 3 callersFunctioncopy

Copy tensor into another tensor

src/gpu/f32/ops.rs:81

↓ 3 callersMethoddevice

The device of the device

src/gpu/f32/tensor.rs:59

↓ 3 callersFunctionembedding_from

(weights: TensorView<'a>, device: &Device)

examples/bert.rs:120

↓ 3 callersMethodforward

TODO

src/nn/models/gpt2.rs:458

↓ 3 callersFunctiongelu

(x: &mut Tensor)

src/gpu/f32/ops.rs:437

↓ 3 callersFunctionlayer_norm_from_prefix

( prefix: &str, tensors: &'a SafeTensors<'a>, device: &Device, )

examples/bert.rs:270

↓ 3 callersFunctionlayer_norm_from_prefix

( prefix: &str, tensors: &'a SafeTensors<'a>, device: &Device, )

examples/gpt2.rs:152

↓ 3 callersFunctionlinear_from

( weights: TensorView<'a>, bias: TensorView<'a>, device: &Device, )

examples/bert.rs:97

↓ 2 callersFunctionattention

( q_weights: &Linear<F32Tensor>, k_weights: &Linear<F32Tensor>, v_weights: &Linear<F32

src/nn/models/bert.rs:109

↓ 2 callersFunctionattention

( qkv_weights: &LinearT<F32Tensor>, ctx: &mut Gpt2Context<F32Tensor>, i: usize, )

src/nn/models/gpt2.rs:245

↓ 2 callersFunctionbroadcast_add

broacasted tensor elementwise addition. b += a.

src/cpu/f32/ops.rs:241

↓ 2 callersFunctionbroadcast_mul

broadcasted tensor elementwise multiplication. b *= a.

src/gpu/f32/ops.rs:317

↓ 2 callersFunctioncuda_unsplit_heads

( src: &F32CudaTensor, dst: &mut F32CudaTensor, )

src/nn/models/bert.rs:216

↓ 2 callersFunctiondevice

()

src/nn/models/bert.rs:668

↓ 2 callersFunctionembedding_from

(weights: TensorView<'a>, device: &Device)

examples/gpt2.rs:108

↓ 2 callersFunctionexp

(x: f32)

src/cpu/f32/ops.rs:398

↓ 2 callersFunctionmul

tensor elementwise multiplication. b *= a.

src/gpu/f32/ops.rs:285

↓ 2 callersFunctionmul_scalar

(x: &mut Tensor, factor: f32)

src/gpu/f32/ops.rs:453

↓ 2 callersFunctionnormalize

Basic operation for the layernorm. x = (x - x.mean()) / (x.var() + epsilon) `mean` and `var` do not have to be initialized, they are simply passed to

src/cpu/f32/ops.rs:297

↓ 2 callersMethodrun

TODO

src/nn/models/bert.rs:649

↓ 2 callersFunctionselect

Operation for selecting entire rows within tensor `weights`. Each `id` is the index of the row.

src/cpu/f32/ops.rs:18

↓ 2 callersFunctionsoftmax

Softmax on the last dimension for tensor `x`

src/cpu/f32/ops.rs:355

↓ 2 callersFunctionsplit_qkv

(ctx: &mut Gpt2Context<F32Tensor>, i: usize)

src/nn/models/gpt2.rs:193

↓ 2 callersFunctiontanh

(x: &mut Tensor)

src/gpu/f32/ops.rs:418

↓ 2 callersFunctionunsplit_heads

(src: &F32Tensor, dst: &mut F32Tensor)

src/nn/models/bert.rs:92

↓ 1 callersFunctionadd

tensor elementwise addition. b += a.

src/cpu/f32/ops.rs:226

↓ 1 callersFunctionbert_attention_from_tensors

( index: usize, tensors: &'a SafeTensors<'a>, device: &Device, )

examples/bert.rs:214

↓ 1 callersFunctionbert_layer_from_tensors

( index: usize, tensors: &'a SafeTensors<'a>, device: &Device, )

examples/bert.rs:205

↓ 1 callersFunctionbert_mlp_from_tensors

( index: usize, tensors: &'a SafeTensors<'a>, device: &Device, )

examples/bert.rs:247

↓ 1 callersMethodblas

The CudaBlas handle

src/gpu/f32/tensor.rs:69

↓ 1 callersFunctionbroadcast_mul

broacasted tensor elementwise multiplication. b += a.

src/cpu/f32/ops.rs:275

↓ 1 callersFunctionbuild_ptx

()

build.rs:111

↓ 1 callersFunctioncopy

Copy tensor into another tensor

src/cpu/f32/ops.rs:42

↓ 1 callersMethodcpu_data

Returns a cpu vec containing copied data from the device.

src/gpu/f32/tensor.rs:114

↓ 1 callersFunctioncuda_attention

( q_weights: &Linear<F32CudaTensor>, k_weights: &Linear<F32CudaTensor>, v_weights: &Li

src/nn/models/bert.rs:251

↓ 1 callersFunctioncuda_attention

( qkv: &LinearT<F32CudaTensor>, ctx: &mut Gpt2Context<F32CudaTensor>, i: usize, )

src/nn/models/gpt2.rs:384

↓ 1 callersFunctionfaster_tanh

utility function to use a faster but less precise tanh

src/cpu/f32/ops.rs:386

↓ 1 callersFunctionfaster_tanh

utility function to use a faster but less precise tanh

src/cpu/f16/ops.rs:323

↓ 1 callersMethodforward

Forward pass

src/nn/layers/linear.rs:18

↓ 1 callersMethodforward

TODO

src/nn/layers/embedding.rs:17

↓ 1 callersMethodforward

TODO

src/nn/layers/layer_norm.rs:23

↓ 1 callersMethodgenerate

TODO

src/nn/models/gpt2.rs:141

↓ 1 callersFunctionget_label

(id2label: Option<&HashMap<String, String>>, i: usize)

examples/bert.rs:51

↓ 1 callersFunctiongpt2_attention_from_tensors

( index: usize, tensors: &'a SafeTensors<'a>, device: &Device, )

examples/gpt2.rs:132

↓ 1 callersFunctiongpt2_from_tensors

(tensors: &SafeTensors, device: &Device, num_heads: usize)

examples/gpt2.rs:112

↓ 1 callersFunctiongpt2_layer_from_tensors

( index: usize, tensors: &'a SafeTensors<'a>, device: &Device, )

examples/gpt2.rs:121

↓ 1 callersFunctiongpt2_mlp_from_tensors

( index: usize, tensors: &'a SafeTensors<'a>, device: &Device, )

examples/gpt2.rs:142

↓ 1 callersMethodid2label

(&self)

examples/bert.rs:46

↓ 1 callersFunctioninline_tanh

(x: f32)

src/cpu/f32/ops.rs:410

↓ 1 callersFunctioninline_tanh

(x: f16)

src/cpu/f16/ops.rs:335

↓ 1 callersFunctionlinear_from

( weights: TensorView<'a>, bias: TensorView<'a>, device: &Device, )

examples/gpt2.rs:81

↓ 1 callersFunctionmul

tensor elementwise multiplication. b += a.

src/cpu/f32/ops.rs:260

↓ 1 callersMethodnew_context

TODO

src/nn/models/bert.rs:599

↓ 1 callersMethodnew_context

TODO

src/nn/models/gpt2.rs:622

↓ 1 callersMethodnew_tokens

TODO

src/nn/models/gpt2.rs:85

↓ 1 callersFunctionnormalize

Basic operation for the layernorm. x = (x - x.mean()) / (x.var() + epsilon) `mean` and `var` do not have to be initialized, they are simply passed to

src/cpu/f16/ops.rs:234

↓ 1 callersFunctionrun

()

examples/bert.rs:321

↓ 1 callersFunctionrun

()

examples/gpt2.rs:203

↓ 1 callersFunctionselect

Operation for selecting entire rows within tensor `weights`. Each `id` is the index of the row.

src/cpu/f16/ops.rs:40

↓ 1 callersMethodset_num_heads

TODO

src/nn/models/bert.rs:585

↓ 1 callersFunctionsoftmax

Softmax on the last dimension for tensor `x`

src/cpu/f16/ops.rs:292

↓ 1 callersFunctionto_f32

(view: TensorView)

examples/bert.rs:77

↓ 1 callersFunctionto_f32

(view: TensorView)

examples/gpt2.rs:61

↓ 1 callersFunctionunbiased_linear_from

(weights: TensorView<'a>, device: &Device)

examples/gpt2.rs:92

next →1–100 of 212, ranked by callers