Paddle · Issues· 1488 open

NotImplementedError in oneDNN PIR conversion breaks CPU inference in PaddlePaddle 3.3.0

使用PPOCRv5训练模型，出现错误：Segmentation fault，再次训练会出现系统崩溃，无任何终端输出错误，直接CPU卡死。

Segmentation fault

Paddle Inference 3.3.1 OneDNN 执行时 ConvertPirAttribute2RuntimeAttribute 不支持 ArrayAttribute<DoubleAttribute>

[Performance] cummax/cummin 的 CUDA kernel 硬编码 dim3(16,32)，长 innermost 维扫描比 torch.cummax 慢 19 倍、比 paddle.cumsum 慢 87 倍

【启航计划】PaddlePaddle API兼容性增强

Conv2D 默认初始化(`Normal(0, √(2/fan_in))`)在无归一化深层/共享权重结构下激活指数放大，导致训练发散与 NaN 梯度

容器内 GPU MMU Fault: ENGINE GRAPHICS GPC1 GPCCLIENT_T1_3，导致推理回退到 CPU，重启容器后问题解决

paddle.i0 prematurely overflows to Inf for finite float64 result

Enable CI Support for ppc64le (IBM Power) Architecture

在windows上为rocm编译发生错误 Paddle only supports Python version >= 3.10 now

paddle.nn.functional.silu backward returns -Inf for finite float64 gradient at x=-709

paddle.nn.functional.selu backward returns wrong finite gradient at x=-700

paddle.nn.functional.elu backward returns 0.0 for finite float64 gradient at x=-40

paddle.expm1 backward returns 0.0 for finite float64 gradient at x=-40