pypto-case-norm-layernorm

name: pypto-case-norm-layernorm description: "模式 C 示例：2D Norm + Loop — LayerNorm，展示 forward 降维为 2D、kernel 内 sum 归约 + 归一化" category: example version: "1.0.0" metadata: backend: ascend dsl: pypto operator_patterns: "norm,reduction,loop"

模式 C-1：2D Norm — LayerNorm

forward 中 reshape(batch, -1) 降为 2D，kernel 沿 batch 维 loop。

BASIC_BATCH = 4

def create_layernorm_kernel(batch, hidden, eps):
    @pypto.frontend.jit(runtime_options=..., debug_options=...)
    def kernel(
        x: pypto.Tensor((batch, hidden), pypto.DT_FP32),
    ) -> pypto.Tensor((batch, hidden), pypto.DT_FP32):
        output = pypto.tensor([batch, hidden], pypto.DT_FP32)
        num_iters = ceil_div(batch, BASIC_BATCH)
        pypto.set_vec_tile_shapes(1, 16384)
        inv_h = 1.0 / hidden
        for bi in pypto.loop(0, num_iters, 1, name="LOOP_LN", idx_name="bi"):
            offset = bi * BASIC_BATCH
            x_chunk = pypto.view(x, [BASIC_BATCH, hidden], [offset, 0])
            mean = pypto.sum(x_chunk, dim=1, keepdim=True) * inv_h
            var = pypto.sum(x_chunk * x_chunk, dim=1, keepdim=True) * inv_h - mean * mean
            normed = (x_chunk - mean) / pypto.sqrt(var + eps)
            pypto.assemble(normed, [offset, 0], output)
        return output
    return kernel

forward：reshape(B, -1) → kernel → reshape(x.shape) GroupNorm 同模式：forward 中 reshape(B*G, -1) → 2D kernel。

模式要点

• pypto.sum(dim=int) — dim 只能传单个 int
• mean = sum * (1/size) — 没有 mean API
• 方差 = E[x²] - E[x]² — 两次 sum 实现
• set_vec_tile_shapes(1, 16384) — 2D，第一维小批量，第二维大 tile