Struct BatchExec 

pub struct BatchExec<'c, 'g> { /* private fields */ }

一次批量前向的执行器。一批一个，用完检查 BatchExec::finish。

Implementations

impl<'c, 'g> BatchExec<'c, 'g>

pub fn finish(&self) -> Result<(), String>

等已提交的工作跑完，报告这一批有没有出错。

pub fn dry_usage(&self) -> (u64, u64)

试算的结果：(工作内存峰值, 还没上显存的权重)，字节。真正执行时恒为 0。

pub fn upload(&self, x: &Tensor) -> DevTensor<'c, 'g>

上传一个样本（批大小必须是 1），有效宽度就是实际宽度。

pub fn placeholder(&self, shape: Vec<usize>) -> DevTensor<'c, 'g>

不带数据的张量（试算时当输入用）；有效宽度取最后一维。

pub fn upload_batch(&self, xs: &[&Tensor]) -> DevTensor<'c, 'g>

上传一批 [C, H, w_i]（C、H 相同），右边补 0 到最宽的宽度，并记下各样本的有效宽度。

pub fn download(&self, t: &DevTensor<'_, 'g>) -> Result<Vec<f32>, String>

下载整个批量张量（行优先，含补边）；出错或试算时返回错误或空。

Trait Implementations

impl Drop for BatchExec<'_, '_>

fn drop(&mut self)

Executes the destructor for this type.

fn pin_drop(self: Pin<&mut Self>)

🔬This is a nightly-only experimental API. (pin_ergonomics)

Execute the destructor for this type, but different to Drop::drop, it requires self to be pinned.

impl<'c, 'g> Exec for BatchExec<'c, 'g>

type T = DevTensor<'c, 'g>

张量句柄。

fn shape(&self, x: &Self::T) -> Vec<usize>

单个样本的形状。

fn conv(&self, l: &Conv2d, x: &Self::T) -> Self::T

fn conv_transpose(&self, l: &ConvTranspose2d, x: &Self::T) -> Self::T

fn act(&self, a: Activation, x: &mut Self::T)

fn add(&self, a: &Self::T, b: &Self::T) -> Self::T

fn add_assign(&self, a: &mut Self::T, b: &Self::T)

fn pad_right_bottom(&self, x: &Self::T, right: usize, bottom: usize) -> Self::T

fn max_pool2d( &self, x: &Self::T, k: (usize, usize), stride: (usize, usize), ceil_mode: bool, ) -> Self::T

fn avg_pool2d(&self, x: &Self::T, k: (usize, usize)) -> Self::T

fn global_avg(&self, x: &Self::T) -> Self::T

每个通道的全局平均，结果 [C, 1, 1]。

fn scale_channels(&self, x: &mut Self::T, s: &Self::T)

s 是 [C, 1, 1]，逐通道乘到 x 上。

fn se_gate(&self, s: &mut Self::T)

颈部 SE 的门：v ← 1 + clamp(0.2·v + 0.5, 0, 1)。

fn concat(&self, xs: &[&Self::T]) -> Self::T

fn upsample(&self, x: &Self::T, oh: usize, ow: usize) -> Self::T

最近邻插值到 oh × ow。

fn chw_to_seq(&self, x: &Self::T) -> Self::T

fn linear(&self, l: &Linear, x: &Self::T) -> Self::T

fn layer_norm(&self, l: &LayerNorm, x: &Self::T) -> Self::T

fn attention(&self, qkv: &Self::T, heads: usize) -> Self::T

多头自注意力的中段：输入线性层出来的 [T, 3C]（列 = s·C + h·hd + d，s 为 0 = q、1 = k、2 = v）， 返回拼回的 [T, C]，分数按 head_dim^−0.5 缩放。

fn ctc_digest(&self, logits: &Self::T) -> Vec<Vec<(u32, f32)>>

CTC 摘要：每个样本一个列表，每步一个（最大类下标，softmax 的最大概率）。