Repository navigation

Issues

New issue

[Call for Bench] SWE-Paddle 社区共建
#79447 · sunzhongkai588 opened on Jul 9, 2026
【Hackathon 10th Overview】第十期飞桨黑客松活动总览
#78292 · mattheliu opened on Mar 13, 2026
【HACKATHON 10th ERNIE with Partners】文心合作伙伴赛道
#78485 · onecatcn opened on Mar 25, 2026
300

3.4.0 官网 CPU 源缺少 macOS arm64 wheel
PaddlePaddle/Paddle#79835 · opened on Sep 29, 2026
The cu129 wheel leaks ~4 KB of host memory per paddle.inference run on GPU; the cu130 wheel on the same machine and driver does not
PaddlePaddle/Paddle#79828 · opened on Sep 28, 2026
paddle.inference leaks ~4 KB of host memory per run() on GPU (one page per inference); the same model on CPU does not
PaddlePaddle/Paddle#79827 · opened on Sep 28, 2026
gloo: send/recv draw from the same per-group tag counter as collectives, so an intervening send/recv desyncs the next broadcast/all_reduce → TIMEOUT
PaddlePaddle/Paddle#79815 · opened on Sep 27, 2026
gloo: ProcessGroupGloo's own (unused) connectFullMesh collides with the default group's rendezvous keys → Connection refused / hang in new_group()
PaddlePaddle/Paddle#79814 · opened on Sep 27, 2026
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。
PaddlePaddle/Paddle#79758 · opened on Sep 10, 2026
Paddle Inference 3.3.1 OneDNN 执行时 ConvertPirAttribute2RuntimeAttribute 不支持 ArrayAttribute<DoubleAttribute>
PaddlePaddle/Paddle#79749 · opened on Sep 8, 2026
1
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
PaddlePaddle/Paddle#79706 · opened on Aug 27, 2026
1
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
PaddlePaddle/Paddle#79701 · opened on Aug 26, 2026
1
[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
PaddlePaddle/Paddle#79670 · opened on Aug 18, 2026
1
在windows上为rocm编译发生错误 Paddle only supports Python version >= 3.10 now
PaddlePaddle/Paddle#79637 · opened on Aug 8, 2026
paddle.nn.functional.silu backward returns -Inf for finite float64 gradient at x=-709
PaddlePaddle/Paddle#79630 · opened on Aug 6, 2026
paddle.nn.functional.selu backward returns wrong finite gradient at x=-700
PaddlePaddle/Paddle#79629 · opened on Aug 6, 2026
paddle.nn.functional.elu backward returns 0.0 for finite float64 gradient at x=-40
PaddlePaddle/Paddle#79628 · opened on Aug 6, 2026
paddle.expm1 backward returns 0.0 for finite float64 gradient at x=-40
PaddlePaddle/Paddle#79627 · opened on Aug 6, 2026
When will PaddlePaddle be available/kompatible for Python v3.14?
PaddlePaddle/Paddle#79626 · opened on Aug 6, 2026
paddle.slice_scatter segfaults instead of raising for length-mismatched control args
PaddlePaddle/Paddle#79621 · opened on Aug 6, 2026
paddle.hypot overflows to Inf for large finite float64 inputs
PaddlePaddle/Paddle#79620 · opened on Aug 6, 2026
paddle.i1 prematurely overflows to Inf for finite float64 result
PaddlePaddle/Paddle#79619 · opened on Aug 6, 2026
paddle.i0 prematurely overflows to Inf for finite float64 result
PaddlePaddle/Paddle#79618 · opened on Aug 6, 2026
1
[Dy2St][CINN] Tensor kwargs 在第二次执行时导致 Embedding 读取错误 ID
PaddlePaddle/Paddle#79580 · opened on Jul 29, 2026
[Dy2St] full_graph=True 下方法名与全局函数同名时调用被错误绑定
PaddlePaddle/Paddle#79579 · opened on Jul 29, 2026
[PIR][CINN] Tensor.view() backward fails with view_shape_grad kernel output mismatch
PaddlePaddle/Paddle#79559 · opened on Jul 28, 2026
Support Python 3.14 for PaddlePaddle
PaddlePaddle/Paddle#79527 · opened on Jul 21, 2026
1
paddle.sparse.subtract() with CSR tensor containing negative column indices triggers heap-buffer-overflow
PaddlePaddle/Paddle#79506 · opened on Jul 18, 2026

Navigation Menu

Repository navigation

Issues

Issues

3.4.0 官网 CPU 源缺少 macOS arm64 wheel

The cu129 wheel leaks ~4 KB of host memory per paddle.inference run on GPU; the cu130 wheel on the same machine and driver does not

paddle.inference leaks ~4 KB of host memory per run() on GPU (one page per inference); the same model on CPU does not

gloo: `send`/`recv` draw from the same per-group tag counter as collectives, so an intervening send/recv desyncs the next `broadcast`/`all_reduce` → `TIMEOUT`

gloo: `ProcessGroupGloo`'s own (unused) `connectFullMesh` collides with the default group's rendezvous keys → `Connection refused` / hang in `new_group()`

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。

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

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

[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

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

When will PaddlePaddle be available/kompatible for Python v3.14?

paddle.slice_scatter segfaults instead of raising for length-mismatched control args

paddle.hypot overflows to Inf for large finite float64 inputs

paddle.i1 prematurely overflows to Inf for finite float64 result

paddle.i0 prematurely overflows to Inf for finite float64 result

[Dy2St][CINN] Tensor kwargs 在第二次执行时导致 Embedding 读取错误 ID

[Dy2St] `full_graph=True` 下方法名与全局函数同名时调用被错误绑定

[PIR][CINN] `Tensor.view()` backward fails with `view_shape_grad` kernel output mismatch

Support Python 3.14 for PaddlePaddle

paddle.sparse.subtract() with CSR tensor containing negative column indices triggers heap-buffer-overflow