v3.4.0

Name: Paddle
Author: PaddlePaddle
PaddlePaddle/Paddlev3.4.0Sep 11, 2026by sneaxiy
Upgrading?

What changed from mineru-3.4.5-released → v3.4.0?
AI Summary

PaddlePaddle 3.4.0 focuses on enhancing large model training efficiency, stability, and domestic hardware compatibility. Key updates include the introduction of the Muon optimizer, VMM Allocator v2 for better memory management, and enhanced FP8 quantization capabilities. This version also brings significant improvements to Kunlun XPU and Hygon DCU support, alongside fixes for large shape tensor and core operator precision issues.
Key Highlights

Introduction of the Muon optimizer for high-performance large-scale distributed training.
VMM Allocator v2 for improved memory management and reduced fragmentation in MoE scenarios.
Significant improvements in domestic hardware support, including Kunlun XPU and Hygon DCU.
Enhanced core operator precision and fixes for large shape/0-size tensor stability issues.
SOT (Dynamic to Static) enhancements supporting Python 3.11–3.14.
Breaking Changes

This is the last version to provide Windows system support.
New Features

Muon Optimizer with distributed sharding and batched Newton-Schulz optimization.
VMM Allocator v2 with best-fit allocation and memory pool separation.
FP8 quantization support extended to ue8m0 scale and non-128 aligned blocks.
Kunlun XPU support for batched_gemm, CUDA Graph, and complex128 data types.
Hygon DCU support for HIP native API and bfloat16 convolution kernels.
SOT enhancements supporting Python 3.11–3.14 and PyTorch-aligned traversal.
FlexCheckpoint enhancements for AOA, safetensors, and DCP loading.
Auto-parallel capabilities for MoE FSDP and GLM5 EP inference.
Full Release Notes

### 重要更新
飞桨框架 3.4 版本围绕大模型训练的训练效率、算子正确性、稳定性以及国产硬件适配等关键领域持续迭代。本版本新增面向大规模分布式训练的高性能 Muon 优化器；持续打磨核心算子的数值精度，并针对大模型训练中高频出现的大 Shape Tensor、0-Size Tensor 场景做了系统性的正确性排查与修复，进一步夯实大模型训练的稳定性与正确性。同时，生态兼容能力与昆仑芯 XPU 适配持续深化，开发体验和推理部署能力同步提升。需要说明的是，本版本是最后一个提供 Windows 系统支持的版本。

**训练效率提升**
* Muon 优化器：新增 Muon 优化器，基于 batched GEMM、对称 GEMM 等技术对 Newton-Schulz 正交化过程做了高性能优化，并具备大规模并行训练能力。
* 基于 VMM 的 Allocator v2：新增基于 CUDA 虚拟内存（VMM）的 best-fit 分配器 v2，支持大/小内存池分离查询与全 block 信息查询，改善 MoE 等动态显存场景下的碎片管理，显著减少显存碎片，并提供高性能的碎片整理能力。
* FP8 量化能力增强：fp8_quant_blockwise 等 FP8 量化算子新增 ue8m0 scale 数据类型支持，并将适用范围扩展到非 128 对齐 block、大 Tensor 等场景。

**正确性增强**
* 核心算子精度提升：系统性提升 matmul、linear、reduce 系列、elementwise 系列、activation 系列算子，以及 AdamW、layer_norm、rms_norm、group_norm、conv 等核心算子的计算精度。
* 大 Shape Tensor 与 0-Size 正确性增强：系统性修复大 Shape Tensor 上的 int32 索引溢出，以及 0-Size Tensor 引发的 CUDA error(9)、段错误等正确性问题，全面保障大 Shape Tensor 与 0-Size 场景下的训练稳定性。

**开发体验优化**
* 动转静（SOT）持续增强：SOT 全面支持 Python 3.11–3.14，新增多种字节码指令的处理能力，并实现与 PyTorch 对齐的反向 DFS 遍历调度。

**国产硬件深度适配**
* 昆仑芯 XPU：新增 batched_gemm 与 CUDA Graph 支持，深度适配 DeepEP buffer 及 bkcl 通信库（支持 w4a8/w8a8 量化），并新增 complex128 数据类型支持。
* 海光 DCU：将 CUDA native API 适配为 HIP native API，并在 ROCm 上为 conv 系列算子注册 bfloat16 kernel，进一步完善国产芯片适配并提升推理性能。

### 1. 训练效率提升
**分布式策略优化（61 项）**
* 通信与计算重叠：新增 p2p 通信与数据并行权重更新/重计算之间的重叠，并支持上下文并行、FlashMaskV3 场景下的通信计算重叠（5 项）：[#79709](https://github.com/PaddlePaddle/Paddle/pull/79709), [#79522](https://github.com/PaddlePaddle/Paddle/pull/79522), [#79357](https://github.com/PaddlePaddle/Paddle/pull/79357), [#78783](https://github.com/PaddlePaddle/Paddle/pull/78783),[#78184](https://github.com/PaddlePaddle/Paddle/pull/78184)
* 流水线并行：合入 PaddleFleet 流水线并行实现，支持 MTP 权重共享、参数 offload 与显存优化（11 项）：[#79692](https://github.com/PaddlePaddle/Paddle/pull/79692), [#79551](https://github.com/PaddlePaddle/Paddle/pull/79551), [#79394](https://github.com/PaddlePaddle/Paddle/pull/79394), [#78688](https://github.com/PaddlePaddle/Paddle/pull/78688), [#78680](https://github.com/PaddlePaddle/Paddle/pull/78680), [#78449](https://github.com/PaddlePaddle/Paddle/pull/78449), [#78437](https://github.com/PaddlePaddle/Paddle/pull/78437), [#78230](https://github.com/PaddlePaddle/Paddle/pull/78230), [#78229](https://github.com/PaddlePaddle/Paddle/pull/78229), [#77852](https://github.com/PaddlePaddle/Paddle/pull/77852), [#77812](https://github.com/PaddlePaddle/Paddle/pull/77812)
* Muon 优化器：新增基于 Newton-Schulz 正交化的优化器，并完善其分布式 sharding 支持、通信优化与 bf16/fp32 混合参数处理，同时通过 batched Newton-Schulz 迭代、对称 GEMM 等手段优化性能（16 项）：[#79688](https://github.com/PaddlePaddle/Paddle/pull/79688), [#79538](https://github.com/PaddlePaddle/Paddle/pull/79538), [#79512](https://github.com/PaddlePaddle/Paddle/pull/79512), [#79485](https://github.com/PaddlePaddle/Paddle/pull/79485), [#79020](https://github.com/PaddlePaddle/Paddle/pull/79020), [#78987](https://github.com/PaddlePaddle/Paddle/pull/78987), [#78953](https://github.com/PaddlePaddle/Paddle/pull/78953), [#78856](https://github.com/PaddlePaddle/Paddle/pull/78856), [#78828](https://github.com/PaddlePaddle/Paddle/pull/78828), [#78748](https://github.com/PaddlePaddle/Paddle/pull/78748), [#78335](https://github.com/PaddlePaddle/Paddle/pull/78335), [#78928](https://github.com/PaddlePaddle/Paddle/pull/78928),[#79741](https://github.com/PaddlePaddle/Paddle/pull/79741), [#79715](https://github.com/PaddlePaddle/Paddle/pull/79715), [#79495](https://github.com/PaddlePaddle/Paddle/pull/79495), [#79048](https://github.com/PaddlePaddle/Paddle/pull/79048)
* 分布式训练其它优化与增强：Sharding 重建、显存/内存优化、启动脚本健壮性提升等（6 项）：[#79634](https://github.com/PaddlePaddle/Paddle/pull/79634), [#78522](https://github.com/PaddlePaddle/Paddle/pull/78522), [#78310](https://github.com/PaddlePaddle/Paddle/pull/78310), [#77459](https://github.com/PaddlePaddle/Paddle/pull/77459), [#77152](https://github.com/PaddlePaddle/Paddle/pull/77152), [#76591](https://github.com/PaddlePaddle/Paddle/pull/76591)
* FlexCheckpoint：增强权重转换与重组能力，覆盖 AOA、safetensors、DCP 加载、3D CommGroup 重分片等场景（10 项）：[#78818](https://github.com/PaddlePaddle/Paddle/pull/78818), [#78290](https://github.com/PaddlePaddle/Paddle/pull/78290), [#78208](https://github.com/PaddlePaddle/Paddle/pull/78208), [#77702](https://github.com/PaddlePaddle/Paddle/pull/77702), [#77405](https://github.com/PaddlePaddle/Paddle/pull/77405), [#77180](https://github.com/PaddlePaddle/Paddle/pull/77180), [#77065](https://github.com/PaddlePaddle/Paddle/pull/77065),[#79734](https://github.com/PaddlePaddle/Paddle/pull/79734), [#77266](https://github.com/PaddlePaddle/Paddle/pull/77266),[#79632](https://github.com/PaddlePaddle/Paddle/pull/79632)
* 自动并行与 MoE / 专家并行：新增 MoE FSDP、GLM5 EP 推理、Qwen3-VL 适配与 SPMD 推导规则等能力（4 项）：[#78392](https://github.com/PaddlePaddle/Paddle/pull/78392), [#78108](https://github.com/PaddlePaddle/Paddle/pull/78108), [#77722](https://github.com/PaddlePaddle/Paddle/pull/77722), [#77308](https://github.com/PaddlePaddle/Paddle/pull/77308)
* FSDP / 自动数据并行：新增 tensor fusion 与通信重叠能力，并修复 ACC 显存占用等问题（4 项）：[#78247](https://github.com/PaddlePaddle/Paddle/pull/78247), [#78074](https://github.com/PaddlePaddle/Paddle/pull/78074), [#77460](https://github.com/PaddlePaddle/Paddle/pull/77460), [#77147](https://github.com/PaddlePaddle/Paddle/pull/77147)
* 分布式训练 Bug 修复（5 项）：[#78805](https://github.com/PaddlePaddle/Paddle/pull/78805), [#78575](https://github.com/PaddlePaddle/Paddle/pull/78575), [#78564](https://github.com/PaddlePaddle/Paddle/pull/78564), [#77256](https://github.com/PaddlePaddle/Paddle/pull/77256), [#77050](https://github.com/PaddlePaddle/Paddle/pull/77050)

**基础功能增强（30 项）**
* 序列化与加载：paddle.save 支持 protocol=5 以实现零拷贝，并修复 dataclass 加载报错（2 项）：[#79675](https://github.com/PaddlePaddle/Paddle/pull/79675), [#78452](https://github.com/PaddlePaddle/Paddle/pull/78452)
* 调试工具：新增 GPU tensor 调试与 CUDA error 检查能力，并增强 api_tracer（6 项）：[#78842](https://github.com/PaddlePaddle/Paddle/pull/78842), [#78543](https://github.com/PaddlePaddle/Paddle/pull/78543), [#78481](https://github.com/PaddlePaddle/Paddle/pull/78481), [#78331](https://github.com/PaddlePaddle/Paddle/pull/78331), [#78116](https://github.com/PaddlePaddle/Paddle/pull/78116), [#77669](https://github.com/PaddlePaddle/Paddle/pull/77669)
* 动转静（SOT/Dy2St）：新增字节码指令与 SizeVariable 支持，增强图编译能力与缓存命中校验（7 项）：[#77839](https://github.com/PaddlePaddle/Paddle/pull/77839), [#77571](https://github.com/PaddlePaddle/Paddle/pull/77571), [#77420](https://github.com/PaddlePaddle/Paddle/pull/77420), [#77354](https://github.com/PaddlePaddle/Paddle/pull/77354), [#77173](https://github.com/PaddlePaddle/Paddle/pull/77173), [#77020](https://github.com/PaddlePaddle/Paddle/pull/77020), [#76766](https://github.com/PaddlePaddle/Paddle/pull/76766)
* CustomDevice：支持自定义设备的混合编译（1 项）：[#77727](https://github.com/PaddlePaddle/Paddle/pull/77727)
* CINN：支持在 CustomDevice 上运行（1 项）：[#77158](https://github.com/PaddlePaddle/Paddle/pull/77158)
* DataType：优化哈希计算性能，并修复枚举顺序问题（1 项）：[#78076](https://github.com/PaddlePaddle/Paddle/pull/78076)
* PyLayer / recompute：优化 forward 输入处理，并修复 save_for_backward 与浅拷贝相关问题（2 项）：[#77899](https://github.com/PaddlePaddle/Paddle/pull/77899), [#78053](https://github.com/PaddlePaddle/Paddle/pull/78053)
* 其它功能增强（10 项）：[#78521](https://github.com/PaddlePaddle/Paddle/pull/78521), [#78055](https://github.com/PaddlePaddle/Paddle/pull/78055), [#77879](https://github.com/PaddlePaddle/Paddle/pull/77879), [#77829](https://github.com/PaddlePaddle/Paddle/pull/77829), [#77675](https://github.com/PaddlePaddle/Paddle/pull/77675), [#77651](https://github.com/PaddlePaddle/Paddle/pull/77651), [#77533](https://github.com/PaddlePaddle/Paddle/pull/77533), [#77089](https://github.com/PaddlePaddle/Paddle/pull/77089), [#76773](https://github.com/PaddlePaddle/Paddle/pull/76773), [#77898](https://github.com/PaddlePaddle/Paddle/pull/77898)

**显存分配器（9 项）**
* 显存分配器：新增基于 CUDA 虚拟内存的 VMM v2 best-fit 分配器及内存池查询 API，并修复相关 core dump 问题（9 项）：[#79598](https://github.com/PaddlePaddle/Paddle/pull/79598), [#79453](https://github.com/PaddlePaddle/Paddle/pull/79453), [#79307](https://github.com/PaddlePaddle/Paddle/pull/79307), [#79261](https://github.com/PaddlePaddle/Paddle/pull/79261), [#78465](https://github.com/PaddlePaddle/Paddle/pull/78465), [#78344](https://github.com/PaddlePaddle/Paddle/pull/78344), [#78317](https://github.com/PaddlePaddle/Paddle/pull/78317), [#77796](https://github.com/PaddlePaddle/Paddle/pull/77796), [#77196](https://github.com/PaddlePaddle/Paddle/pull/77196)

**Bug 修复（30 项）**
* 内存安全：修复 PyObject 内存泄漏与引用计数错误（4 项）：[#79344](https://github.com/PaddlePaddle/Paddle/pull/79344), [#78965](https://github.com/PaddlePaddle/Paddle/pull/78965), [#77448](https://github.com/PaddlePaddle/Paddle/pull/77448), [#77234](https://github.com/PaddlePaddle/Paddle/pull/77234)
* 修复 inplace API 误覆盖已有 Tensor 名称的问题（1 项）：[#79218](https://github.com/PaddlePaddle/Paddle/pull/79218)
* 修复 PyLayer / recompute 的数据提前释放、梯度节点属性拷贝等问题（3 项）：[#78796](https://github.com/PaddlePaddle/Paddle/pull/78796), [#78531](https://github.com/PaddlePaddle/Paddle/pull/78531), [#77150](https://github.com/PaddlePaddle/Paddle/pull/77150)
* 修复 CUDA Graph 兼容性问题（3 项）：[#78568](https://github.com/PaddlePaddle/Paddle/pull/78568), [#78535](https://github.com/PaddlePaddle/Paddle/pull/78535), [#77226](https://github.com/PaddlePaddle/Paddle/pull/77226)
* 修复 paddle.device / paddle.cuda.device 相关问题（3 项）：[#78512](https://github.com/PaddlePaddle/Paddle/pull/78512), [#78104](https://github.com/PaddlePaddle/Paddle/pull/78104), [#77463](https://github.com/PaddlePaddle/Paddle/pull/77463)
* 修复 DataType 相关问题（1 项）：[#78086](https://github.com/PaddlePaddle/Paddle/pull/78086)
* 修复 SOT 相关问题（6 项）：[#77290](https://github.com/PaddlePaddle/Paddle/pull/77290), [#77249](https://github.com/PaddlePaddle/Paddle/pull/77249), [#77227](https://github.com/PaddlePaddle/Paddle/pull/77227), [#77083](https://github.com/PaddlePaddle/Paddle/pull/77083), [#77009](https://github.com/PaddlePaddle/Paddle/pull/77009), [#77005](https://github.com/PaddlePaddle/Paddle/pull/77005)
* 修复 PIR 相关问题（1 项）：[#77016](https://github.com/PaddlePaddle/Paddle/pull/77016)
* 其它 Bug 修复（8 项）：[#78844](https://github.com/PaddlePaddle/Paddle/pull/78844), [#78276](https://github.com/PaddlePaddle/Paddle/pull/78276), [#78119](https://github.com/PaddlePaddle/Paddle/pull/78119), [#77710](https://github.com/PaddlePaddle/Paddle/pull/77710), [#77626](https://github.com/PaddlePaddle/Paddle/pull/77626), [#77478](https://github.com/PaddlePaddle/Paddle/pull/77478), [#77075](https://github.com/PaddlePaddle/Paddle/pull/77075), [#78478](https://github.com/PaddlePaddle/Paddle/pull/78478)

### 2. 算子优化完善
**精度优化（46 项）**
* 优化索引/切片类算子的计算精度与形状校验（1 项）：[#79727](https://github.com/PaddlePaddle/Paddle/pull/79727)
* 矩阵乘法系列精度优化：涵盖 matmul、linear_v2、addmm、baddbmm 等（5 项）：[#79515](https://github.com/PaddlePaddle/Paddle/pull/79515), [#78961](https://github.com/PaddlePaddle/Paddle/pull/78961), [#78750](https://github.com/PaddlePaddle/Paddle/pull/78750), [#77039](https://github.com/PaddlePaddle/Paddle/pull/77039), [#76922](https://github.com/PaddlePaddle/Paddle/pull/76922)
* reduce 系列算子精度优化：涵盖 layer_norm、rms_norm、group_norm、var/std、reduce 等（14 项）：[#79651](https://github.com/PaddlePaddle/Paddle/pull/79651), [#79530](https://github.com/PaddlePaddle/Paddle/pull/79530), [#79529](https://github.com/PaddlePaddle/Paddle/pull/79529), [#79407](https://github.com/PaddlePaddle/Paddle/pull/79407), [#79390](https://github.com/PaddlePaddle/Paddle/pull/79390), [#78622](https://github.com/PaddlePaddle/Paddle/pull/78622), [#78338](https://github.com/PaddlePaddle/Paddle/pull/78338), [#78321](https://github.com/PaddlePaddle/Paddle/pull/78321), [#78246](https://github.com/PaddlePaddle/Paddle/pull/78246), [#78225](https://github.com/PaddlePaddle/Paddle/pull/78225), [#78219](https://github.com/PaddlePaddle/Paddle/pull/78219), [#78201](https://github.com/PaddlePaddle/Paddle/pull/78201), [#76950](https://github.com/PaddlePaddle/Paddle/pull/76950), [#76831](https://github.com/PaddlePaddle/Paddle/pull/76831)
* 超越函数精度优化：涵盖 sin/cos/exp/pow/log 等（7 项）：[#79557](https://github.com/PaddlePaddle/Paddle/pull/79557), [#78265](https://github.com/PaddlePaddle/Paddle/pull/78265), [#78241](https://github.com/PaddlePaddle/Paddle/pull/78241), [#78186](https://github.com/PaddlePaddle/Paddle/pull/78186), [#78095](https://github.com/PaddlePaddle/Paddle/pull/78095), [#77855](https://github.com/PaddlePaddle/Paddle/pull/77855), [#77564](https://github.com/PaddlePaddle/Paddle/pull/77564)
* 复数运算精度优化（2 项）：[#79471](https://github.com/PaddlePaddle/Paddle/pull/79471), [#79322](https://github.com/PaddlePaddle/Paddle/pull/79322)
* 优化器精度优化：深度优化 AdamW/SGD GPU kernel，精度对齐 torch（3 项）：[#79389](https://github.com/PaddlePaddle/Paddle/pull/79389), [#78899](https://github.com/PaddlePaddle/Paddle/pull/78899), [#78752](https://github.com/PaddlePaddle/Paddle/pull/78752)
* 随机/初始化精度优化：涵盖 randperm、kaiming、gaussian、truncated_gaussian 等（3 项）：[#79309](https://github.com/PaddlePaddle/Paddle/pull/79309), [#79271](https://github.com/PaddlePaddle/Paddle/pull/79271), [#78614](https://github.com/PaddlePaddle/Paddle/pull/78614)
* 优化 cumprod/cumsum/lerp/interpolate 等算子的计算精度（4 项）：[#78147](https://github.com/PaddlePaddle/Paddle/pull/78147), [#78146](https://github.com/PaddlePaddle/Paddle/pull/78146), [#77149](https://github.com/PaddlePaddle/Paddle/pull/77149), [#77091](https://github.com/PaddlePaddle/Paddle/pull/77091)
* 卷积精度优化：conv2d/conv3d/conv2d_transpose 精度对齐 torch（1 项）：[#76980](https://github.com/PaddlePaddle/Paddle/pull/76980)
* 新增 cudnn_allow_tf32 / cublas_allow_tf32 开关，用于控制是否允许使用 TF32 Tensor Core（1 项）：[#77781](https://github.com/PaddlePaddle/Paddle/pull/77781)
* 修复 Tensor 转 Scalar 过程中的数值溢出问题（1 项）：[#77013](https://github.com/PaddlePaddle/Paddle/pull/77013)
* 其它算子精度优化（4 项）：[#79451](https://github.com/PaddlePaddle/Paddle/pull/79451), [#79308](https://github.com/PaddlePaddle/Paddle/pull/79308), [#78044](https://github.com/PaddlePaddle/Paddle/pull/78044),[#78798](https://github.com/PaddlePaddle/Paddle/pull/78798)

**性能优化（22 项）**
* permute/transpose：重新设计实现算法，针对解码场景优化性能（4 项）：[#79742](https://github.com/PaddlePaddle/Paddle/pull/79742), [#78063](https://github.com/PaddlePaddle/Paddle/pull/78063), [#77524](https://github.com/PaddlePaddle/Paddle/pull/77524), [#77371](https://github.com/PaddlePaddle/Paddle/pull/77371)
* 优化 cuBLAS workspace 的管理与复用性能（4 项）：[#79442](https://github.com/PaddlePaddle/Paddle/pull/79442), [#79406](https://github.com/PaddlePaddle/Paddle/pull/79406), [#79374](https://github.com/PaddlePaddle/Paddle/pull/79374), [#77874](https://github.com/PaddlePaddle/Paddle/pull/77874)
* 归一化 kernel 性能优化：涵盖 rms_norm、layer_norm、squared_l2_norm 等（4 项）：[#79047](https://github.com/PaddlePaddle/Paddle/pull/79047), [#77709](https://github.com/PaddlePaddle/Paddle/pull/77709), [#77337](https://github.com/PaddlePaddle/Paddle/pull/77337), [#77098](https://github.com/PaddlePaddle/Paddle/pull/77098)
* 优化器 kernel 性能优化：优化 Adam/AdamW GPU kernel，并将学习率的计算类型提升为 float64（2 项）：[#79044](https://github.com/PaddlePaddle/Paddle/pull/79044), [#78949](https://github.com/PaddlePaddle/Paddle/pull/78949)
* TopK：基于 radix-select 与多层排序重写 GPU kernel，并修复相关问题（2 项）：[#78703](https://github.com/PaddlePaddle/Paddle/pull/78703), [#79704](https://github.com/PaddlePaddle/Paddle/pull/79704)
* 其它算子 kernel 性能优化（6 项）：[#78601](https://github.com/PaddlePaddle/Paddle/pull/78601), [#77818](https://github.com/PaddlePaddle/Paddle/pull/77818), [#77233](https://github.com/PaddlePaddle/Paddle/pull/77233), [#77071](https://github.com/PaddlePaddle/Paddle/pull/77071), [#76988](https://github.com/PaddlePaddle/Paddle/pull/76988),[#79150](https://github.com/PaddlePaddle/Paddle/pull/79150)

**大 Shape Tensor 与 0-Size 正确性（55 项）**
* 大 Shape Tensor 正确性：系统性修复 int32 索引溢出与越界访问问题（27 项）：[#79702](https://github.com/PaddlePaddle/Paddle/pull/79702), [#79662](https://github.com/PaddlePaddle/Paddle/pull/79662), [#79661](https://github.com/PaddlePaddle/Paddle/pull/79661), [#79649](https://github.com/PaddlePaddle/Paddle/pull/79649), [#79597](https://github.com/PaddlePaddle/Paddle/pull/79597), [#79595](https://github.com/PaddlePaddle/Paddle/pull/79595), [#79583](https://github.com/PaddlePaddle/Paddle/pull/79583), [#79560](https://github.com/PaddlePaddle/Paddle/pull/79560), [#79548](https://github.com/PaddlePaddle/Paddle/pull/79548), [#79440](https://github.com/PaddlePaddle/Paddle/pull/79440), [#79212](https://github.com/PaddlePaddle/Paddle/pull/79212), [#79010](https://github.com/PaddlePaddle/Paddle/pull/79010), [#78978](https://github.com/PaddlePaddle/Paddle/pull/78978), [#78779](https://github.com/PaddlePaddle/Paddle/pull/78779), [#78494](https://github.com/PaddlePaddle/Paddle/pull/78494), [#78454](https://github.com/PaddlePaddle/Paddle/pull/78454), [#78451](https://github.com/PaddlePaddle/Paddle/pull/78451), [#77729](https://github.com/PaddlePaddle/Paddle/pull/77729), [#77383](https://github.com/PaddlePaddle/Paddle/pull/77383), [#77368](https://github.com/PaddlePaddle/Paddle/pull/77368), [#77154](https://github.com/PaddlePaddle/Paddle/pull/77154), [#77143](https://github.com/PaddlePaddle/Paddle/pull/77143), [#77074](https://github.com/PaddlePaddle/Paddle/pull/77074), [#79536](https://github.com/PaddlePaddle/Paddle/pull/79536), [#79480](https://github.com/PaddlePaddle/Paddle/pull/79480), [#77090](https://github.com/PaddlePaddle/Paddle/pull/77090), [#77191](https://github.com/PaddlePaddle/Paddle/pull/77191)
* 0-Size Tensor 正确性：系统性修复各算子对 0-Size 输入的处理，消除 CUDA error(9)、段错误等异常（28 项）：[#79658](https://github.com/PaddlePaddle/Paddle/pull/79658), [#79591](https://github.com/PaddlePaddle/Paddle/pull/79591), [#79408](https://github.com/PaddlePaddle/Paddle/pull/79408), [#79283](https://github.com/PaddlePaddle/Paddle/pull/79283), [#78565](https://github.com/PaddlePaddle/Paddle/pull/78565), [#78486](https://github.com/PaddlePaddle/Paddle/pull/78486), [#78482](https://github.com/PaddlePaddle/Paddle/pull/78482), [#78453](https://github.com/PaddlePaddle/Paddle/pull/78453), [#78446](https://github.com/PaddlePaddle/Paddle/pull/78446), [#78442](https://github.com/PaddlePaddle/Paddle/pull/78442), [#78440](https://github.com/PaddlePaddle/Paddle/pull/78440), [#78439](https://github.com/PaddlePaddle/Paddle/pull/78439), [#78432](https://github.com/PaddlePaddle/Paddle/pull/78432), [#78371](https://github.com/PaddlePaddle/Paddle/pull/78371), [#78316](https://github.com/PaddlePaddle/Paddle/pull/78316), [#78245](https://github.com/PaddlePaddle/Paddle/pull/78245), [#78243](https://github.com/PaddlePaddle/Paddle/pull/78243), [#78238](https://github.com/PaddlePaddle/Paddle/pull/78238), [#78237](https://github.com/PaddlePaddle/Paddle/pull/78237), [#78097](https://github.com/PaddlePaddle/Paddle/pull/78097), [#78096](https://github.com/PaddlePaddle/Paddle/pull/78096), [#78094](https://github.com/PaddlePaddle/Paddle/pull/78094), [#78075](https://github.com/PaddlePaddle/Paddle/pull/78075), [#78065](https://github.com/PaddlePaddle/Paddle/pull/78065), [#77593](https://github.com/PaddlePaddle/Paddle/pull/77593), [#77481](https://github.com/PaddlePaddle/Paddle/pull/77481), [#78064](https://github.com/PaddlePaddle/Paddle/pull/78064), [#79589](https://github.com/PaddlePaddle/Paddle/pull/79589)

**Bug 修复（38 项）**
* 修复 view/stride 类算子的相关问题（8 项）：[#79655](https://github.com/PaddlePaddle/Paddle/pull/79655), [#79163](https://github.com/PaddlePaddle/Paddle/pull/79163), [#78988](https://github.com/PaddlePaddle/Paddle/pull/78988), [#78612](https://github.com/PaddlePaddle/Paddle/pull/78612), [#78348](https://github.com/PaddlePaddle/Paddle/pull/78348), [#78089](https://github.com/PaddlePaddle/Paddle/pull/78089), [#77555](https://github.com/PaddlePaddle/Paddle/pull/77555), [#77244](https://github.com/PaddlePaddle/Paddle/pull/77244)
* 修复算子输入与形状校验问题：涵盖 baddbmm、matmul、segment_*、correlation 等（4 项）：[#79574](https://github.com/PaddlePaddle/Paddle/pull/79574), [#78506](https://github.com/PaddlePaddle/Paddle/pull/78506), [#77642](https://github.com/PaddlePaddle/Paddle/pull/77642), [#77637](https://github.com/PaddlePaddle/Paddle/pull/77637)
* 修复归一化算子问题：涵盖 layer_norm/rms_norm 反向计算等（8 项）：[#79528](https://github.com/PaddlePaddle/Paddle/pull/79528), [#79241](https://github.com/PaddlePaddle/Paddle/pull/79241), [#78792](https://github.com/PaddlePaddle/Paddle/pull/78792), [#78755](https://github.com/PaddlePaddle/Paddle/pull/78755), [#78567](https://github.com/PaddlePaddle/Paddle/pull/78567), [#77577](https://github.com/PaddlePaddle/Paddle/pull/77577), [#77335](https://github.com/PaddlePaddle/Paddle/pull/77335), [#77274](https://github.com/PaddlePaddle/Paddle/pull/77274)
* 修复 FP8 量化算子的相关问题（2 项）：[#79100](https://github.com/PaddlePaddle/Paddle/pull/79100), [#77398](https://github.com/PaddlePaddle/Paddle/pull/77398)
* 修复注意力算子问题：涵盖 memory-efficient attention 的 mask 广播、fused_rotary 等（2 项）：[#78530](https://github.com/PaddlePaddle/Paddle/pull/78530), [#77014](https://github.com/PaddlePaddle/Paddle/pull/77014)
* 修复其它算子问题：涵盖 full_like、slice、log、arange/range 等（4 项）：[#78191](https://github.com/PaddlePaddle/Paddle/pull/78191), [#77140](https://github.com/PaddlePaddle/Paddle/pull/77140), [#78788](https://github.com/PaddlePaddle/Paddle/pull/78788), [#78149](https://github.com/PaddlePaddle/Paddle/pull/78149)
* 其它 Bug 修复（10 项）：[#79590](https://github.com/PaddlePaddle/Paddle/pull/79590), [#79525](https://github.com/PaddlePaddle/Paddle/pull/79525), [#78786](https://github.com/PaddlePaddle/Paddle/pull/78786), [#78286](https://github.com/PaddlePaddle/Paddle/pull/78286), [#78253](https://github.com/PaddlePaddle/Paddle/pull/78253), [#77501](https://github.com/PaddlePaddle/Paddle/pull/77501), [#77445](https://github.com/PaddlePaddle/Paddle/pull/77445), [#77257](https://github.com/PaddlePaddle/Paddle/pull/77257), [#77134](https://github.com/PaddlePaddle/Paddle/pull/77134), [#76829](https://github.com/PaddlePaddle/Paddle/pull/76829)

**算子其它优化（24 项）**
* 算子功能扩展：cumsum 支持 bool、truncated_gaussian 支持 bf16，新增 div_scale、mm out_dtype 等能力（7 项）：[#79285](https://github.com/PaddlePaddle/Paddle/pull/79285), [#78360](https://github.com/PaddlePaddle/Paddle/pull/78360), [#78199](https://github.com/PaddlePaddle/Paddle/pull/78199), [#77782](https://github.com/PaddlePaddle/Paddle/pull/77782), [#77740](https://github.com/PaddlePaddle/Paddle/pull/77740), [#77646](https://github.com/PaddlePaddle/Paddle/pull/77646), [#77507](https://github.com/PaddlePaddle/Paddle/pull/77507)
* 卷积：新增 SlowConvDilated/SlowConv3DDilated kernel，并在 ROCm 上注册 bf16 支持（2 项）：[#78587](https://github.com/PaddlePaddle/Paddle/pull/78587), [#77807](https://github.com/PaddlePaddle/Paddle/pull/77807)
* FP8 量化：fp8_quant_blockwise 支持 ue8m0 scale、非 128 对齐 block 与大 Tensor 场景（8 项）：[#78315](https://github.com/PaddlePaddle/Paddle/pull/78315), [#77433](https://github.com/PaddlePaddle/Paddle/pull/77433), [#77351](https://github.com/PaddlePaddle/Paddle/pull/77351), [#77284](https://github.com/PaddlePaddle/Paddle/pull/77284), [#77165](https://github.com/PaddlePaddle/Paddle/pull/77165), [#77153](https://github.com/PaddlePaddle/Paddle/pull/77153), [#77133](https://github.com/PaddlePaddle/Paddle/pull/77133), [#77652](https://github.com/PaddlePaddle/Paddle/pull/77652)
* FlashAttention：放开 fa3 的参数限制（1 项）：[#77811](https://github.com/PaddlePaddle/Paddle/pull/77811)
* MoE：扩展 moe_permute/moe_unpermute 的推理特性，并支持 ue8m0 scale（1 项）：[#77547](https://github.com/PaddlePaddle/Paddle/pull/77547)
* 其它算子优化（5 项）：[#78430](https://github.com/PaddlePaddle/Paddle/pull/78430), [#78412](https://github.com/PaddlePaddle/Paddle/pull/78412), [#78039](https://github.com/PaddlePaddle/Paddle/pull/78039), [#77466](https://github.com/PaddlePaddle/Paddle/pull/77466),  [#77321](https://github.com/PaddlePaddle/Paddle/pull/77321)

### 3. 用户体验升级
本版本重点推进开发体验优化。C++ 侧新增 libtorch 兼容层（提供 torch/torch.h 头文件，并补齐 ScalarType、TypeMeta、SymInt、TensorAccessor、DLPack v1.3 等能力）；Python 侧系统性对齐参数别名、Tensor 方法与 inplace 语义，并引入 inplace API 下沉机制。同时新增 index_fill、paddle.audio.functional.resample、assert_close 等实用 API，并修复 pickle 反序列化等安全漏洞。
**Python API 优化（100 项）**
* Tensor 方法补齐：新增 nelement、is_cpu、sparse_mask、retain_grad 等方法，并对齐其行为（10 项）：[#78615](https://github.com/PaddlePaddle/Paddle/pull/78615), [#78593](https://github.com/PaddlePaddle/Paddle/pull/78593), [#78516](https://github.com/PaddlePaddle/Paddle/pull/78516), [#78491](https://github.com/PaddlePaddle/Paddle/pull/78491), [#78301](https://github.com/PaddlePaddle/Paddle/pull/78301), [#78264](https://github.com/PaddlePaddle/Paddle/pull/78264), [#78224](https://github.com/PaddlePaddle/Paddle/pull/78224), [#78164](https://github.com/PaddlePaddle/Paddle/pull/78164), [#78083](https://github.com/PaddlePaddle/Paddle/pull/78083), [#77367](https://github.com/PaddlePaddle/Paddle/pull/77367)
* 参数别名兼容：为大量 API 增加 torch 风格的参数别名（如 input→x、dim→axis、target→label）并支持 out 参数（52 项）：[#78589](https://github.com/PaddlePaddle/Paddle/pull/78589), [#78539](https://github.com/PaddlePaddle/Paddle/pull/78539), [#78475](https://github.com/PaddlePaddle/Paddle/pull/78475), [#78472](https://github.com/PaddlePaddle/Paddle/pull/78472), [#78298](https://github.com/PaddlePaddle/Paddle/pull/78298), [#78294](https://github.com/PaddlePaddle/Paddle/pull/78294), [#78204](https://github.com/PaddlePaddle/Paddle/pull/78204), [#78140](https://github.com/PaddlePaddle/Paddle/pull/78140), [#78134](https://github.com/PaddlePaddle/Paddle/pull/78134), [#78133](https://github.com/PaddlePaddle/Paddle/pull/78133), [#78132](https://github.com/PaddlePaddle/Paddle/pull/78132), [#78128](https://github.com/PaddlePaddle/Paddle/pull/78128), [#78114](https://github.com/PaddlePaddle/Paddle/pull/78114), [#78092](https://github.com/PaddlePaddle/Paddle/pull/78092), [#78077](https://github.com/PaddlePaddle/Paddle/pull/78077), [#78069](https://github.com/PaddlePaddle/Paddle/pull/78069), [#78048](https://github.com/PaddlePaddle/Paddle/pull/78048), [#78047](https://github.com/PaddlePaddle/Paddle/pull/78047), [#78046](https://github.com/PaddlePaddle/Paddle/pull/78046), [#78004](https://github.com/PaddlePaddle/Paddle/pull/78004), [#77981](https://github.com/PaddlePaddle/Paddle/pull/77981), [#77973](https://github.com/PaddlePaddle/Paddle/pull/77973), [#77923](https://github.com/PaddlePaddle/Paddle/pull/77923), [#77922](https://github.com/PaddlePaddle/Paddle/pull/77922), [#77920](https://github.com/PaddlePaddle/Paddle/pull/77920), [#77919](https://github.com/PaddlePaddle/Paddle/pull/77919), [#77916](https://github.com/PaddlePaddle/Paddle/pull/77916), [#77891](https://github.com/PaddlePaddle/Paddle/pull/77891), [#77877](https://github.com/PaddlePaddle/Paddle/pull/77877), [#77801](https://github.com/PaddlePaddle/Paddle/pull/77801), [#77766](https://github.com/PaddlePaddle/Paddle/pull/77766), [#77715](https://github.com/PaddlePaddle/Paddle/pull/77715), [#77681](https://github.com/PaddlePaddle/Paddle/pull/77681), [#77657](https://github.com/PaddlePaddle/Paddle/pull/77657), [#77591](https://github.com/PaddlePaddle/Paddle/pull/77591), [#77573](https://github.com/PaddlePaddle/Paddle/pull/77573), [#77561](https://github.com/PaddlePaddle/Paddle/pull/77561), [#77497](https://github.com/PaddlePaddle/Paddle/pull/77497), [#77451](https://github.com/PaddlePaddle/Paddle/pull/77451), [#77391](https://github.com/PaddlePaddle/Paddle/pull/77391), [#77355](https://github.com/PaddlePaddle/Paddle/pull/77355), [#77300](https://github.com/PaddlePaddle/Paddle/pull/77300), [#77218](https://github.com/PaddlePaddle/Paddle/pull/77218), [#77211](https://github.com/PaddlePaddle/Paddle/pull/77211), [#77194](https://github.com/PaddlePaddle/Paddle/pull/77194), [#77170](https://github.com/PaddlePaddle/Paddle/pull/77170), [#77168](https://github.com/PaddlePaddle/Paddle/pull/77168), [#77079](https://github.com/PaddlePaddle/Paddle/pull/77079), [#77012](https://github.com/PaddlePaddle/Paddle/pull/77012), [#77006](https://github.com/PaddlePaddle/Paddle/pull/77006), [#76578](https://github.com/PaddlePaddle/Paddle/pull/76578), [#77495](https://github.com/PaddlePaddle/Paddle/pull/77495)
* 数据加载兼容：Dataset/Sampler/BatchSampler 的别名与 PyTorch 对齐（2 项）：[#78401](https://github.com/PaddlePaddle/Paddle/pull/78401), [#78382](https://github.com/PaddlePaddle/Paddle/pull/78382)
* 修复 paddle.device 相关兼容性问题（4 项）：[#78350](https://github.com/PaddlePaddle/Paddle/pull/78350), [#77490](https://github.com/PaddlePaddle/Paddle/pull/77490), [#77467](https://github.com/PaddlePaddle/Paddle/pull/77467), [#77431](https://github.com/PaddlePaddle/Paddle/pull/77431)
* 新增兼容 API：paddle._assert、assert_close、pad_sequence/unpad_sequence、log_softmax 等（4 项）：[#78342](https://github.com/PaddlePaddle/Paddle/pull/78342), [#78220](https://github.com/PaddlePaddle/Paddle/pull/78220), [#77749](https://github.com/PaddlePaddle/Paddle/pull/77749), [#76847](https://github.com/PaddlePaddle/Paddle/pull/76847)
* 算子实现下沉 C++：real、inverse、allclose、diag、atan2 等 API 由 Python 实现下沉至 C++，降低调度开销（9 项）：[#78212](https://github.com/PaddlePaddle/Paddle/pull/78212), [#78198](https://github.com/PaddlePaddle/Paddle/pull/78198), [#78138](https://github.com/PaddlePaddle/Paddle/pull/78138), [#77294](https://github.com/PaddlePaddle/Paddle/pull/77294), [#77161](https://github.com/PaddlePaddle/Paddle/pull/77161), [#77078](https://github.com/PaddlePaddle/Paddle/pull/77078), [#77064](https://github.com/PaddlePaddle/Paddle/pull/77064), [#76939](https://github.com/PaddlePaddle/Paddle/pull/76939), [#76736](https://github.com/PaddlePaddle/Paddle/pull/76736)
* 原地操作：为 paddle.nn.* 系列 API 补充 inplace 支持（2 项）：[#77103](https://github.com/PaddlePaddle/Paddle/pull/77103), [#76873](https://github.com/PaddlePaddle/Paddle/pull/76873)
* 新增 rms_norm 公开接口（1 项）：[#76930](https://github.com/PaddlePaddle/Paddle/pull/76930)
* 其它 Python API 兼容性调整（16 项）：[#78438](https://github.com/PaddlePaddle/Paddle/pull/78438), [#78406](https://github.com/PaddlePaddle/Paddle/pull/78406), [#78235](https://github.com/PaddlePaddle/Paddle/pull/78235), [#78082](https://github.com/PaddlePaddle/Paddle/pull/78082), [#77892](https://github.com/PaddlePaddle/Paddle/pull/77892), [#77889](https://github.com/PaddlePaddle/Paddle/pull/77889), [#77869](https://github.com/PaddlePaddle/Paddle/pull/77869), [#77849](https://github.com/PaddlePaddle/Paddle/pull/77849), [#77824](https://github.com/PaddlePaddle/Paddle/pull/77824), [#77789](https://github.com/PaddlePaddle/Paddle/pull/77789), [#77778](https://github.com/PaddlePaddle/Paddle/pull/77778), [#77751](https://github.com/PaddlePaddle/Paddle/pull/77751), [#77632](https://github.com/PaddlePaddle/Paddle/pull/77632), [#77506](https://github.com/PaddlePaddle/Paddle/pull/77506), [#77010](https://github.com/PaddlePaddle/Paddle/pull/77010), [#76915](https://github.com/PaddlePaddle/Paddle/pull/76915)

**新增 API（2 项）**
* 新增 index_fill/index_fill_、paddle.audio.functional.resample、paddle.testing.assert_close、Tensor.reset、Tensor.layout 等 API（2 项）：[#78041](https://github.com/PaddlePaddle/Paddle/pull/78041), [#78032](https://github.com/PaddlePaddle/Paddle/pull/78032)

**C++ API 兼容 Torch（自定义算子）（64 项）**
* CUDA 上下文与设备：补齐 Stream/Event、pin_memory、record_stream 等 C++ 实现（9 项）：[#79593](https://github.com/PaddlePaddle/Paddle/pull/79593), [#78652](https://github.com/PaddlePaddle/Paddle/pull/78652), [#78631](https://github.com/PaddlePaddle/Paddle/pull/78631), [#78584](https://github.com/PaddlePaddle/Paddle/pull/78584), [#78553](https://github.com/PaddlePaddle/Paddle/pull/78553), [#78255](https://github.com/PaddlePaddle/Paddle/pull/78255), [#78143](https://github.com/PaddlePaddle/Paddle/pull/78143), [#78060](https://github.com/PaddlePaddle/Paddle/pull/78060), [#77436](https://github.com/PaddlePaddle/Paddle/pull/77436)
* C++ API 补齐：新增 Generator、Philox、CUDABlas、TensorAccessor、storage、Sparse、Tensor.reset/layout 等接口（9 项）：[#79562](https://github.com/PaddlePaddle/Paddle/pull/79562), [#78072](https://github.com/PaddlePaddle/Paddle/pull/78072), [#78070](https://github.com/PaddlePaddle/Paddle/pull/78070), [#77649](https://github.com/PaddlePaddle/Paddle/pull/77649), [#77581](https://github.com/PaddlePaddle/Paddle/pull/77581), [#77498](https://github.com/PaddlePaddle/Paddle/pull/77498), [#77319](https://github.com/PaddlePaddle/Paddle/pull/77319), [#77185](https://github.com/PaddlePaddle/Paddle/pull/77185), [#77176](https://github.com/PaddlePaddle/Paddle/pull/77176)
* 头文件与 schema：新增 torch/torch.h 头文件，并引入 schema 解析器与声明规范化机制（7 项）：[#78770](https://github.com/PaddlePaddle/Paddle/pull/78770), [#78590](https://github.com/PaddlePaddle/Paddle/pull/78590), [#78282](https://github.com/PaddlePaddle/Paddle/pull/78282), [#78266](https://github.com/PaddlePaddle/Paddle/pull/78266), [#78248](https://github.com/PaddlePaddle/Paddle/pull/78248), [#77938](https://github.com/PaddlePaddle/Paddle/pull/77938), [#77854](https://github.com/PaddlePaddle/Paddle/pull/77854)
* 构建与平台：支持 Windows 自动链接、XPU 测试与编译兼容性完善（4 项）：[#78769](https://github.com/PaddlePaddle/Paddle/pull/78769), [#78647](https://github.com/PaddlePaddle/Paddle/pull/78647), [#78641](https://github.com/PaddlePaddle/Paddle/pull/78641), [#78215](https://github.com/PaddlePaddle/Paddle/pull/78215)
* 类型系统：补齐 ScalarType、TypeMeta、SymInt、DispatchKey 等类型定义，与 PyTorch 对齐（5 项）：[#78581](https://github.com/PaddlePaddle/Paddle/pull/78581), [#78525](https://github.com/PaddlePaddle/Paddle/pull/78525), [#78257](https://github.com/PaddlePaddle/Paddle/pull/78257), [#77303](https://github.com/PaddlePaddle/Paddle/pull/77303), [#77301](https://github.com/PaddlePaddle/Paddle/pull/77301)
* DLPack：升级至 v1.3 并对齐转换行为（2 项）：[#77523](https://github.com/PaddlePaddle/Paddle/pull/77523), [#77052](https://github.com/PaddlePaddle/Paddle/pull/77052)
* 其它 C++ API 兼容性调整（28 项）：[#78982](https://github.com/PaddlePaddle/Paddle/pull/78982), [#78653](https://github.com/PaddlePaddle/Paddle/pull/78653), [#78633](https://github.com/PaddlePaddle/Paddle/pull/78633), [#78609](https://github.com/PaddlePaddle/Paddle/pull/78609), [#78591](https://github.com/PaddlePaddle/Paddle/pull/78591), [#78576](https://github.com/PaddlePaddle/Paddle/pull/78576), [#78555](https://github.com/PaddlePaddle/Paddle/pull/78555), [#78554](https://github.com/PaddlePaddle/Paddle/pull/78554), [#78552](https://github.com/PaddlePaddle/Paddle/pull/78552), [#78551](https://github.com/PaddlePaddle/Paddle/pull/78551), [#78550](https://github.com/PaddlePaddle/Paddle/pull/78550), [#78244](https://github.com/PaddlePaddle/Paddle/pull/78244), [#78182](https://github.com/PaddlePaddle/Paddle/pull/78182), [#78099](https://github.com/PaddlePaddle/Paddle/pull/78099), [#78037](https://github.com/PaddlePaddle/Paddle/pull/78037), [#78027](https://github.com/PaddlePaddle/Paddle/pull/78027), [#78026](https://github.com/PaddlePaddle/Paddle/pull/78026), [#77713](https://github.com/PaddlePaddle/Paddle/pull/77713), [#77614](https://github.com/PaddlePaddle/Paddle/pull/77614), [#77544](https://github.com/PaddlePaddle/Paddle/pull/77544), [#77542](https://github.com/PaddlePaddle/Paddle/pull/77542), [#77540](https://github.com/PaddlePaddle/Paddle/pull/77540), [#77514](https://github.com/PaddlePaddle/Paddle/pull/77514), [#77388](https://github.com/PaddlePaddle/Paddle/pull/77388), [#77270](https://github.com/PaddlePaddle/Paddle/pull/77270), [#77182](https://github.com/PaddlePaddle/Paddle/pull/77182), [#77156](https://github.com/PaddlePaddle/Paddle/pull/77156), [#77051](https://github.com/PaddlePaddle/Paddle/pull/77051)

**安全修复（3 项）**
* 安全加固：修复 pickle 反序列化远程代码执行漏洞（CWE-502）与 Zip Slip 路径穿越问题，并改用 yaml.safe_load 加载配置（3 项）：[#78508](https://github.com/PaddlePaddle/Paddle/pull/78508), [#78085](https://github.com/PaddlePaddle/Paddle/pull/78085), [#77113](https://github.com/PaddlePaddle/Paddle/pull/77113)

### 4. 国产硬件适配
昆仑芯 XPU 持续完善大模型场景适配，新增 batched_gemm、CUDA Graph、DeepEP 等能力，并更新 bkcl 通信库以支持 w4a8/w8a8 量化。海光 DCU 侧则通过在 ROCm 上注册 bf16 conv kernel 与适配 HIP native API，进一步提升推理性能。
昆仑芯 XPU（30 项）
* batched_gemm：新增 XPU 上的 batched_gemm kernel 实现（3 项）：[#78500](https://github.com/PaddlePaddle/Paddle/pull/78500), [#78289](https://github.com/PaddlePaddle/Paddle/pull/78289), [#77840](https://github.com/PaddlePaddle/Paddle/pull/77840)
* DeepEP / bkcl：适配 DeepEP buffer，bkcl 通信库新增 w4a8/w8a8 量化支持（3 项）：[#77837](https://github.com/PaddlePaddle/Paddle/pull/77837), [#77742](https://github.com/PaddlePaddle/Paddle/pull/77742), [#77017](https://github.com/PaddlePaddle/Paddle/pull/77017)
* 算子补齐：新增/对齐 top_p_sampling、linspace、bicubic 插值、masked_select 等算子（4 项）：[#77737](https://github.com/PaddlePaddle/Paddle/pull/77737), [#77699](https://github.com/PaddlePaddle/Paddle/pull/77699), [#77677](https://github.com/PaddlePaddle/Paddle/pull/77677), [#77023](https://github.com/PaddlePaddle/Paddle/pull/77023)
* 设备属性：新增 get_device_properties 与 set_xpu_current_device_id 接口（2 项）：[#77568](https://github.com/PaddlePaddle/Paddle/pull/77568), [#76877](https://github.com/PaddlePaddle/Paddle/pull/76877)
* 0-Size Tensor：为 XPU 算子补齐 0-Size 输入支持（7 项）：[#77557](https://github.com/PaddlePaddle/Paddle/pull/77557), [#77494](https://github.com/PaddlePaddle/Paddle/pull/77494), [#77455](https://github.com/PaddlePaddle/Paddle/pull/77455), [#77387](https://github.com/PaddlePaddle/Paddle/pull/77387), [#77363](https://github.com/PaddlePaddle/Paddle/pull/77363), [#77348](https://github.com/PaddlePaddle/Paddle/pull/77348), [#77332](https://github.com/PaddlePaddle/Paddle/pull/77332)
* CUDA Graph：XPU 新增 CUDA Graph 支持，并正确处理 stream 语义（2 项）：[#77421](https://github.com/PaddlePaddle/Paddle/pull/77421), [#77311](https://github.com/PaddlePaddle/Paddle/pull/77311)
* complex128：为 XPU 新增 complex128 数据类型支持（2 项）：[#76976](https://github.com/PaddlePaddle/Paddle/pull/76976), [#74161](https://github.com/PaddlePaddle/Paddle/pull/74161)
* 其它 XPU 适配与 Bug 修复（7 项）：[#78626](https://github.com/PaddlePaddle/Paddle/pull/78626), [#78563](https://github.com/PaddlePaddle/Paddle/pull/78563), [#78227](https://github.com/PaddlePaddle/Paddle/pull/78227), [#78177](https://github.com/PaddlePaddle/Paddle/pull/78177), [#77217](https://github.com/PaddlePaddle/Paddle/pull/77217), [#77188](https://github.com/PaddlePaddle/Paddle/pull/77188), [#77135](https://github.com/PaddlePaddle/Paddle/pull/77135)

### 5. 编译安装与平台支持
本版本完善编译安装与平台支持，新增 CUDA 13/13.2 的编译与运行支持、Python 3.14 适配、arm64 平台 FlashAttention 构建等能力，并升级 NCCL、protobuf、Sleef 等第三方依赖。
**编译与安装（79 项）**
* CUDA/NCCL：支持 CUDA 13/13.2 的编译与运行，NCCL 升级至 2.30.7，并修复相关编译兼容性问题（12 项）：[#79644](https://github.com/PaddlePaddle/Paddle/pull/79644), [#79378](https://github.com/PaddlePaddle/Paddle/pull/79378), [#79183](https://github.com/PaddlePaddle/Paddle/pull/79183), [#79154](https://github.com/PaddlePaddle/Paddle/pull/79154), [#78606](https://github.com/PaddlePaddle/Paddle/pull/78606), [#78168](https://github.com/PaddlePaddle/Paddle/pull/78168), [#77821](https://github.com/PaddlePaddle/Paddle/pull/77821), [#77324](https://github.com/PaddlePaddle/Paddle/pull/77324), [#77287](https://github.com/PaddlePaddle/Paddle/pull/77287), [#76999](https://github.com/PaddlePaddle/Paddle/pull/76999), [#77229](https://github.com/PaddlePaddle/Paddle/pull/77229), [#77358](https://github.com/PaddlePaddle/Paddle/pull/77358)
* ARM：支持 arm64 平台上的 FlashAttention 构建、liblapack 打包与 ARM Dockerfile（6 项）：[#79518](https://github.com/PaddlePaddle/Paddle/pull/79518), [#79496](https://github.com/PaddlePaddle/Paddle/pull/79496), [#79484](https://github.com/PaddlePaddle/Paddle/pull/79484), [#78730](https://github.com/PaddlePaddle/Paddle/pull/78730), [#78702](https://github.com/PaddlePaddle/Paddle/pull/78702), [#78619](https://github.com/PaddlePaddle/Paddle/pull/78619)
* 第三方依赖：protobuf 升级至 7、启用 Sleef、开启 OpenVINO 等（10 项）：[#79422](https://github.com/PaddlePaddle/Paddle/pull/79422), [#79257](https://github.com/PaddlePaddle/Paddle/pull/79257), [#78513](https://github.com/PaddlePaddle/Paddle/pull/78513), [#78480](https://github.com/PaddlePaddle/Paddle/pull/78480), [#78285](https://github.com/PaddlePaddle/Paddle/pull/78285), [#78236](https://github.com/PaddlePaddle/Paddle/pull/78236), [#78153](https://github.com/PaddlePaddle/Paddle/pull/78153), [#78088](https://github.com/PaddlePaddle/Paddle/pull/78088), [#78059](https://github.com/PaddlePaddle/Paddle/pull/78059), [#77705](https://github.com/PaddlePaddle/Paddle/pull/77705)
* 扩展构建：优化 CustomDevice/CUDAExtension/CppExtension 的构建流程（8 项）：[#78668](https://github.com/PaddlePaddle/Paddle/pull/78668), [#78356](https://github.com/PaddlePaddle/Paddle/pull/78356), [#77863](https://github.com/PaddlePaddle/Paddle/pull/77863), [#77328](https://github.com/PaddlePaddle/Paddle/pull/77328), [#77259](https://github.com/PaddlePaddle/Paddle/pull/77259), [#77116](https://github.com/PaddlePaddle/Paddle/pull/77116), [#76961](https://github.com/PaddlePaddle/Paddle/pull/76961), [#76870](https://github.com/PaddlePaddle/Paddle/pull/76870)
* DCU / Iluvatar：适配 HIP native API，并完善 CI 环境配置（5 项）：[#78595](https://github.com/PaddlePaddle/Paddle/pull/78595), [#77813](https://github.com/PaddlePaddle/Paddle/pull/77813), [#77587](https://github.com/PaddlePaddle/Paddle/pull/77587), [#77574](https://github.com/PaddlePaddle/Paddle/pull/77574), [#77345](https://github.com/PaddlePaddle/Paddle/pull/77345)
* Python 版本：新增 Python 3.13/3.14 适配，XPU Dockerfile 支持 Python 3.9（3 项）：[#78558](https://github.com/PaddlePaddle/Paddle/pull/78558), [#78515](https://github.com/PaddlePaddle/Paddle/pull/78515), [#77703](https://github.com/PaddlePaddle/Paddle/pull/77703)
* Windows：修复 Unicode 路径下的 DLL 加载失败、win32 变量与 phi.lib 打包等问题（3 项）：[#78416](https://github.com/PaddlePaddle/Paddle/pull/78416), [#78210](https://github.com/PaddlePaddle/Paddle/pull/78210), [#77690](https://github.com/PaddlePaddle/Paddle/pull/77690)
* XPU 工具链：更新 xhpc/xccl/xpufft 版本，并支持 zcc 编译器（13 项）：[#77823](https://github.com/PaddlePaddle/Paddle/pull/77823), [#77688](https://github.com/PaddlePaddle/Paddle/pull/77688), [#77673](https://github.com/PaddlePaddle/Paddle/pull/77673), [#77629](https://github.com/PaddlePaddle/Paddle/pull/77629), [#77530](https://github.com/PaddlePaddle/Paddle/pull/77530), [#77518](https://github.com/PaddlePaddle/Paddle/pull/77518), [#77477](https://github.com/PaddlePaddle/Paddle/pull/77477), [#77260](https://github.com/PaddlePaddle/Paddle/pull/77260), [#77219](https://github.com/PaddlePaddle/Paddle/pull/77219), [#77199](https://github.com/PaddlePaddle/Paddle/pull/77199), [#77069](https://github.com/PaddlePaddle/Paddle/pull/77069), [#77028](https://github.com/PaddlePaddle/Paddle/pull/77028), [#76978](https://github.com/PaddlePaddle/Paddle/pull/76978)
* 其它编译安装调整（19 项）：[#79434](https://github.com/PaddlePaddle/Paddle/pull/79434), [#78734](https://github.com/PaddlePaddle/Paddle/pull/78734), [#78705](https://github.com/PaddlePaddle/Paddle/pull/78705), [#78655](https://github.com/PaddlePaddle/Paddle/pull/78655), [#78650](https://github.com/PaddlePaddle/Paddle/pull/78650), [#78594](https://github.com/PaddlePaddle/Paddle/pull/78594), [#78511](https://github.com/PaddlePaddle/Paddle/pull/78511), [#78502](https://github.com/PaddlePaddle/Paddle/pull/78502), [#78389](https://github.com/PaddlePaddle/Paddle/pull/78389), [#78318](https://github.com/PaddlePaddle/Paddle/pull/78318), [#78239](https://github.com/PaddlePaddle/Paddle/pull/78239), [#77728](https://github.com/PaddlePaddle/Paddle/pull/77728), [#77685](https://github.com/PaddlePaddle/Paddle/pull/77685), [#77625](https://github.com/PaddlePaddle/Paddle/pull/77625), [#77549](https://github.com/PaddlePaddle/Paddle/pull/77549), [#77471](https://github.com/PaddlePaddle/Paddle/pull/77471), [#77125](https://github.com/PaddlePaddle/Paddle/pull/77125), [#77045](https://github.com/PaddlePaddle/Paddle/pull/77045), [#76998](https://github.com/PaddlePaddle/Paddle/pull/76998)

**CI/构建（75 项）**
* 优化 CI 超时设置与缓存策略（7 项）：[#79674](https://github.com/PaddlePaddle/Paddle/pull/79674), [#78045](https://github.com/PaddlePaddle/Paddle/pull/78045), [#77890](https://github.com/PaddlePaddle/Paddle/pull/77890), [#77786](https://github.com/PaddlePaddle/Paddle/pull/77786), [#77701](https://github.com/PaddlePaddle/Paddle/pull/77701), [#77680](https://github.com/PaddlePaddle/Paddle/pull/77680), [#77640](https://github.com/PaddlePaddle/Paddle/pull/77640)
* 升级 GitHub Actions 及 workflow 基础设施（10 项）：[#79149](https://github.com/PaddlePaddle/Paddle/pull/79149), [#78648](https://github.com/PaddlePaddle/Paddle/pull/78648), [#78645](https://github.com/PaddlePaddle/Paddle/pull/78645), [#78424](https://github.com/PaddlePaddle/Paddle/pull/78424), [#78080](https://github.com/PaddlePaddle/Paddle/pull/78080), [#78079](https://github.com/PaddlePaddle/Paddle/pull/78079), [#77250](https://github.com/PaddlePaddle/Paddle/pull/77250), [#77174](https://github.com/PaddlePaddle/Paddle/pull/77174), [#77167](https://github.com/PaddlePaddle/Paddle/pull/77167), [#77160](https://github.com/PaddlePaddle/Paddle/pull/77160)
* 完善 CI 审批与 PR 检查流程（14 项）：[#78588](https://github.com/PaddlePaddle/Paddle/pull/78588), [#78569](https://github.com/PaddlePaddle/Paddle/pull/78569), [#78498](https://github.com/PaddlePaddle/Paddle/pull/78498), [#78479](https://github.com/PaddlePaddle/Paddle/pull/78479), [#78471](https://github.com/PaddlePaddle/Paddle/pull/78471), [#78422](https://github.com/PaddlePaddle/Paddle/pull/78422), [#78411](https://github.com/PaddlePaddle/Paddle/pull/78411), [#78386](https://github.com/PaddlePaddle/Paddle/pull/78386), [#78242](https://github.com/PaddlePaddle/Paddle/pull/78242), [#78023](https://github.com/PaddlePaddle/Paddle/pull/78023), [#77928](https://github.com/PaddlePaddle/Paddle/pull/77928), [#77873](https://github.com/PaddlePaddle/Paddle/pull/77873), [#77159](https://github.com/PaddlePaddle/Paddle/pull/77159), [#77008](https://github.com/PaddlePaddle/Paddle/pull/77008)
* 升级静态检查工具 typos/ruff/yamlfmt（4 项）：[#78145](https://github.com/PaddlePaddle/Paddle/pull/78145), [#77761](https://github.com/PaddlePaddle/Paddle/pull/77761), [#77700](https://github.com/PaddlePaddle/Paddle/pull/77700), [#77171](https://github.com/PaddlePaddle/Paddle/pull/77171)
* 引入并优化 L2 ccache 以加速构建（3 项）：[#77911](https://github.com/PaddlePaddle/Paddle/pull/77911), [#77402](https://github.com/PaddlePaddle/Paddle/pull/77402), [#77282](https://github.com/PaddlePaddle/Paddle/pull/77282)
* 修复 fleet 相关的 CI 依赖与安装问题（7 项）：[#77800](https://github.com/PaddlePaddle/Paddle/pull/77800), [#77658](https://github.com/PaddlePaddle/Paddle/pull/77658), [#77346](https://github.com/PaddlePaddle/Paddle/pull/77346), [#77343](https://github.com/PaddlePaddle/Paddle/pull/77343), [#77323](https://github.com/PaddlePaddle/Paddle/pull/77323), [#77267](https://github.com/PaddlePaddle/Paddle/pull/77267), [#77192](https://github.com/PaddlePaddle/Paddle/pull/77192)
* 其它 CI 优化（30 项）：[#79635](https://github.com/PaddlePaddle/Paddle/pull/79635), [#79409](https://github.com/PaddlePaddle/Paddle/pull/79409), [#79380](https://github.com/PaddlePaddle/Paddle/pull/79380), [#78700](https://github.com/PaddlePaddle/Paddle/pull/78700), [#78592](https://github.com/PaddlePaddle/Paddle/pull/78592), [#78573](https://github.com/PaddlePaddle/Paddle/pull/78573), [#78548](https://github.com/PaddlePaddle/Paddle/pull/78548), [#78458](https://github.com/PaddlePaddle/Paddle/pull/78458), [#78431](https://github.com/PaddlePaddle/Paddle/pull/78431), [#78343](https://github.com/PaddlePaddle/Paddle/pull/78343), [#78329](https://github.com/PaddlePaddle/Paddle/pull/78329), [#78312](https://github.com/PaddlePaddle/Paddle/pull/78312), [#78188](https://github.com/PaddlePaddle/Paddle/pull/78188), [#78034](https://github.com/PaddlePaddle/Paddle/pull/78034), [#77901](https://github.com/PaddlePaddle/Paddle/pull/77901), [#77767](https://github.com/PaddlePaddle/Paddle/pull/77767), [#77668](https://github.com/PaddlePaddle/Paddle/pull/77668), [#77635](https://github.com/PaddlePaddle/Paddle/pull/77635), [#77553](https://github.com/PaddlePaddle/Paddle/pull/77553), [#77535](https://github.com/PaddlePaddle/Paddle/pull/77535), [#77521](https://github.com/PaddlePaddle/Paddle/pull/77521), [#77505](https://github.com/PaddlePaddle/Paddle/pull/77505), [#77468](https://github.com/PaddlePaddle/Paddle/pull/77468), [#77432](https://github.com/PaddlePaddle/Paddle/pull/77432), [#77292](https://github.com/PaddlePaddle/Paddle/pull/77292), [#77289](https://github.com/PaddlePaddle/Paddle/pull/77289), [#77258](https://github.com/PaddlePaddle/Paddle/pull/77258), [#77081](https://github.com/PaddlePaddle/Paddle/pull/77081), [#77044](https://github.com/PaddlePaddle/Paddle/pull/77044), [#77000](https://github.com/PaddlePaddle/Paddle/pull/77000)

### 6. 代码质量与文档
本版本系统性推进代码质量提升，完成 phi_ops 命名空间重构，并进行了大量拼写修正与文档示例格式统一，同时补充了 API 文档与测试用例。
**代码风格与重构（265 项）**
* 修正代码中的拼写错误与不规范命名（8 项）：[#78637](https://github.com/PaddlePaddle/Paddle/pull/78637), [#78421](https://github.com/PaddlePaddle/Paddle/pull/78421), [#78139](https://github.com/PaddlePaddle/Paddle/pull/78139), [#78067](https://github.com/PaddlePaddle/Paddle/pull/78067), [#77650](https://github.com/PaddlePaddle/Paddle/pull/77650), [#77615](https://github.com/PaddlePaddle/Paddle/pull/77615), [#77414](https://github.com/PaddlePaddle/Paddle/pull/77414), [#77255](https://github.com/PaddlePaddle/Paddle/pull/77255)
* 系统性统一 phi::DenseTensor 的用法（phi_ops 重构）（28 项）：[#78566](https://github.com/PaddlePaddle/Paddle/pull/78566), [#77959](https://github.com/PaddlePaddle/Paddle/pull/77959), [#77531](https://github.com/PaddlePaddle/Paddle/pull/77531), [#77426](https://github.com/PaddlePaddle/Paddle/pull/77426), [#77400](https://github.com/PaddlePaddle/Paddle/pull/77400), [#77271](https://github.com/PaddlePaddle/Paddle/pull/77271), [#77243](https://github.com/PaddlePaddle/Paddle/pull/77243), [#77242](https://github.com/PaddlePaddle/Paddle/pull/77242), [#77241](https://github.com/PaddlePaddle/Paddle/pull/77241), [#77240](https://github.com/PaddlePaddle/Paddle/pull/77240), [#77239](https://github.com/PaddlePaddle/Paddle/pull/77239), [#77238](https://github.com/PaddlePaddle/Paddle/pull/77238), [#77237](https://github.com/PaddlePaddle/Paddle/pull/77237), [#77124](https://github.com/PaddlePaddle/Paddle/pull/77124), [#77121](https://github.com/PaddlePaddle/Paddle/pull/77121), [#77104](https://github.com/PaddlePaddle/Paddle/pull/77104), [#77088](https://github.com/PaddlePaddle/Paddle/pull/77088), [#77087](https://github.com/PaddlePaddle/Paddle/pull/77087), [#77086](https://github.com/PaddlePaddle/Paddle/pull/77086), [#77085](https://github.com/PaddlePaddle/Paddle/pull/77085), [#77084](https://github.com/PaddlePaddle/Paddle/pull/77084), [#77082](https://github.com/PaddlePaddle/Paddle/pull/77082), [#77058](https://github.com/PaddlePaddle/Paddle/pull/77058), [#77057](https://github.com/PaddlePaddle/Paddle/pull/77057), [#77056](https://github.com/PaddlePaddle/Paddle/pull/77056), [#77055](https://github.com/PaddlePaddle/Paddle/pull/77055), [#77054](https://github.com/PaddlePaddle/Paddle/pull/77054), [#77049](https://github.com/PaddlePaddle/Paddle/pull/77049)
* 优化 phi::make_ddim 性能（11 项）：[#78420](https://github.com/PaddlePaddle/Paddle/pull/78420), [#78415](https://github.com/PaddlePaddle/Paddle/pull/78415), [#78403](https://github.com/PaddlePaddle/Paddle/pull/78403), [#78395](https://github.com/PaddlePaddle/Paddle/pull/78395), [#78394](https://github.com/PaddlePaddle/Paddle/pull/78394), [#78374](https://github.com/PaddlePaddle/Paddle/pull/78374), [#77562](https://github.com/PaddlePaddle/Paddle/pull/77562), [#77541](https://github.com/PaddlePaddle/Paddle/pull/77541), [#77281](https://github.com/PaddlePaddle/Paddle/pull/77281), [#77115](https://github.com/PaddlePaddle/Paddle/pull/77115), [#77114](https://github.com/PaddlePaddle/Paddle/pull/77114)
* 优化 common::vectorize / phi::vectorize 的实现（11 项）：[#78254](https://github.com/PaddlePaddle/Paddle/pull/78254), [#77759](https://github.com/PaddlePaddle/Paddle/pull/77759), [#77711](https://github.com/PaddlePaddle/Paddle/pull/77711), [#77415](https://github.com/PaddlePaddle/Paddle/pull/77415), [#77313](https://github.com/PaddlePaddle/Paddle/pull/77313), [#77280](https://github.com/PaddlePaddle/Paddle/pull/77280), [#77279](https://github.com/PaddlePaddle/Paddle/pull/77279), [#77254](https://github.com/PaddlePaddle/Paddle/pull/77254), [#77195](https://github.com/PaddlePaddle/Paddle/pull/77195), [#77187](https://github.com/PaddlePaddle/Paddle/pull/77187), [#77059](https://github.com/PaddlePaddle/Paddle/pull/77059)
* 统一 phi::DataType 的用法（4 项）：[#78109](https://github.com/PaddlePaddle/Paddle/pull/78109), [#77871](https://github.com/PaddlePaddle/Paddle/pull/77871), [#77695](https://github.com/PaddlePaddle/Paddle/pull/77695), [#77532](https://github.com/PaddlePaddle/Paddle/pull/77532)
* 统一 phi::dynload 动态库加载的用法（3 项）：[#77768](https://github.com/PaddlePaddle/Paddle/pull/77768), [#77676](https://github.com/PaddlePaddle/Paddle/pull/77676), [#77674](https://github.com/PaddlePaddle/Paddle/pull/77674)
* 统一 phi::Place / phi::CustomPlace 的用法（3 项）：[#77566](https://github.com/PaddlePaddle/Paddle/pull/77566), [#77440](https://github.com/PaddlePaddle/Paddle/pull/77440), [#77384](https://github.com/PaddlePaddle/Paddle/pull/77384)
* 统一 pir:: 命名空间的用法（10 项）：[#77511](https://github.com/PaddlePaddle/Paddle/pull/77511), [#77503](https://github.com/PaddlePaddle/Paddle/pull/77503), [#77458](https://github.com/PaddlePaddle/Paddle/pull/77458), [#77425](https://github.com/PaddlePaddle/Paddle/pull/77425), [#77404](https://github.com/PaddlePaddle/Paddle/pull/77404), [#77403](https://github.com/PaddlePaddle/Paddle/pull/77403), [#77401](https://github.com/PaddlePaddle/Paddle/pull/77401), [#77395](https://github.com/PaddlePaddle/Paddle/pull/77395), [#77341](https://github.com/PaddlePaddle/Paddle/pull/77341), [#77253](https://github.com/PaddlePaddle/Paddle/pull/77253)
* 系统性纯代码清理与重构（27 项）：[#77316](https://github.com/PaddlePaddle/Paddle/pull/77316), [#77315](https://github.com/PaddlePaddle/Paddle/pull/77315), [#77295](https://github.com/PaddlePaddle/Paddle/pull/77295), [#77245](https://github.com/PaddlePaddle/Paddle/pull/77245), [#77232](https://github.com/PaddlePaddle/Paddle/pull/77232), [#77228](https://github.com/PaddlePaddle/Paddle/pull/77228), [#77137](https://github.com/PaddlePaddle/Paddle/pull/77137), [#77132](https://github.com/PaddlePaddle/Paddle/pull/77132), [#77131](https://github.com/PaddlePaddle/Paddle/pull/77131), [#77067](https://github.com/PaddlePaddle/Paddle/pull/77067), [#77060](https://github.com/PaddlePaddle/Paddle/pull/77060), [#77038](https://github.com/PaddlePaddle/Paddle/pull/77038), [#77037](https://github.com/PaddlePaddle/Paddle/pull/77037), [#77022](https://github.com/PaddlePaddle/Paddle/pull/77022), [#77018](https://github.com/PaddlePaddle/Paddle/pull/77018), [#77015](https://github.com/PaddlePaddle/Paddle/pull/77015), [#77004](https://github.com/PaddlePaddle/Paddle/pull/77004), [#77003](https://github.com/PaddlePaddle/Paddle/pull/77003), [#77002](https://github.com/PaddlePaddle/Paddle/pull/77002), [#76997](https://github.com/PaddlePaddle/Paddle/pull/76997), [#76996](https://github.com/PaddlePaddle/Paddle/pull/76996), [#76995](https://github.com/PaddlePaddle/Paddle/pull/76995), [#76993](https://github.com/PaddlePaddle/Paddle/pull/76993), [#76992](https://github.com/PaddlePaddle/Paddle/pull/76992), [#76910](https://github.com/PaddlePaddle/Paddle/pull/76910), [#76715](https://github.com/PaddlePaddle/Paddle/pull/76715), [#76665](https://github.com/PaddlePaddle/Paddle/pull/76665)
* 统一 phi::CPUPlace 与设备类型判断的用法（14 项）：[#77296](https://github.com/PaddlePaddle/Paddle/pull/77296), [#77224](https://github.com/PaddlePaddle/Paddle/pull/77224), [#77223](https://github.com/PaddlePaddle/Paddle/pull/77223), [#77222](https://github.com/PaddlePaddle/Paddle/pull/77222), [#77221](https://github.com/PaddlePaddle/Paddle/pull/77221), [#77209](https://github.com/PaddlePaddle/Paddle/pull/77209), [#77208](https://github.com/PaddlePaddle/Paddle/pull/77208), [#77207](https://github.com/PaddlePaddle/Paddle/pull/77207), [#77204](https://github.com/PaddlePaddle/Paddle/pull/77204), [#77203](https://github.com/PaddlePaddle/Paddle/pull/77203), [#77202](https://github.com/PaddlePaddle/Paddle/pull/77202), [#77198](https://github.com/PaddlePaddle/Paddle/pull/77198), [#77197](https://github.com/PaddlePaddle/Paddle/pull/77197), [#77179](https://github.com/PaddlePaddle/Paddle/pull/77179)
* 清理与 DataLayout::kNHWC 字符串相关的历史代码（6 项）：[#77061](https://github.com/PaddlePaddle/Paddle/pull/77061), [#77036](https://github.com/PaddlePaddle/Paddle/pull/77036), [#77035](https://github.com/PaddlePaddle/Paddle/pull/77035), [#77026](https://github.com/PaddlePaddle/Paddle/pull/77026), [#76769](https://github.com/PaddlePaddle/Paddle/pull/76769), [#76609](https://github.com/PaddlePaddle/Paddle/pull/76609)
* 其它代码重构与清理（140 项）：[#79519](https://github.com/PaddlePaddle/Paddle/pull/79519), [#78825](https://github.com/PaddlePaddle/Paddle/pull/78825), [#78646](https://github.com/PaddlePaddle/Paddle/pull/78646), [#78583](https://github.com/PaddlePaddle/Paddle/pull/78583), [#78580](https://github.com/PaddlePaddle/Paddle/pull/78580), [#78559](https://github.com/PaddlePaddle/Paddle/pull/78559), [#78519](https://github.com/PaddlePaddle/Paddle/pull/78519), [#78510](https://github.com/PaddlePaddle/Paddle/pull/78510), [#78501](https://github.com/PaddlePaddle/Paddle/pull/78501), [#78433](https://github.com/PaddlePaddle/Paddle/pull/78433), [#78413](https://github.com/PaddlePaddle/Paddle/pull/78413), [#78400](https://github.com/PaddlePaddle/Paddle/pull/78400), [#78396](https://github.com/PaddlePaddle/Paddle/pull/78396), [#78390](https://github.com/PaddlePaddle/Paddle/pull/78390), [#78376](https://github.com/PaddlePaddle/Paddle/pull/78376), [#78322](https://github.com/PaddlePaddle/Paddle/pull/78322), [#78300](https://github.com/PaddlePaddle/Paddle/pull/78300), [#78256](https://github.com/PaddlePaddle/Paddle/pull/78256), [#78218](https://github.com/PaddlePaddle/Paddle/pull/78218), [#78217](https://github.com/PaddlePaddle/Paddle/pull/78217), [#78211](https://github.com/PaddlePaddle/Paddle/pull/78211), [#78200](https://github.com/PaddlePaddle/Paddle/pull/78200), [#78196](https://github.com/PaddlePaddle/Paddle/pull/78196), [#78190](https://github.com/PaddlePaddle/Paddle/pull/78190), [#78185](https://github.com/PaddlePaddle/Paddle/pull/78185), [#78170](https://github.com/PaddlePaddle/Paddle/pull/78170), [#78167](https://github.com/PaddlePaddle/Paddle/pull/78167), [#78166](https://github.com/PaddlePaddle/Paddle/pull/78166), [#78165](https://github.com/PaddlePaddle/Paddle/pull/78165), [#78142](https://github.com/PaddlePaddle/Paddle/pull/78142), [#78141](https://github.com/PaddlePaddle/Paddle/pull/78141), [#78137](https://github.com/PaddlePaddle/Paddle/pull/78137), [#78136](https://github.com/PaddlePaddle/Paddle/pull/78136), [#78120](https://github.com/PaddlePaddle/Paddle/pull/78120), [#78110](https://github.com/PaddlePaddle/Paddle/pull/78110), [#78051](https://github.com/PaddlePaddle/Paddle/pull/78051), [#78050](https://github.com/PaddlePaddle/Paddle/pull/78050), [#78049](https://github.com/PaddlePaddle/Paddle/pull/78049), [#78030](https://github.com/PaddlePaddle/Paddle/pull/78030), [#78025](https://github.com/PaddlePaddle/Paddle/pull/78025), [#77957](https://github.com/PaddlePaddle/Paddle/pull/77957), [#77936](https://github.com/PaddlePaddle/Paddle/pull/77936), [#77872](https://github.com/PaddlePaddle/Paddle/pull/77872), [#77870](https://github.com/PaddlePaddle/Paddle/pull/77870), [#77860](https://github.com/PaddlePaddle/Paddle/pull/77860), [#77815](https://github.com/PaddlePaddle/Paddle/pull/77815), [#77809](https://github.com/PaddlePaddle/Paddle/pull/77809), [#77808](https://github.com/PaddlePaddle/Paddle/pull/77808), [#77802](https://github.com/PaddlePaddle/Paddle/pull/77802), [#77769](https://github.com/PaddlePaddle/Paddle/pull/77769), [#77734](https://github.com/PaddlePaddle/Paddle/pull/77734), [#77733](https://github.com/PaddlePaddle/Paddle/pull/77733), [#77724](https://github.com/PaddlePaddle/Paddle/pull/77724), [#77723](https://github.com/PaddlePaddle/Paddle/pull/77723), [#77707](https://github.com/PaddlePaddle/Paddle/pull/77707), [#77706](https://github.com/PaddlePaddle/Paddle/pull/77706), [#77704](https://github.com/PaddlePaddle/Paddle/pull/77704), [#77683](https://github.com/PaddlePaddle/Paddle/pull/77683), [#77678](https://github.com/PaddlePaddle/Paddle/pull/77678), [#77627](https://github.com/PaddlePaddle/Paddle/pull/77627), [#77606](https://github.com/PaddlePaddle/Paddle/pull/77606), [#77605](https://github.com/PaddlePaddle/Paddle/pull/77605), [#77604](https://github.com/PaddlePaddle/Paddle/pull/77604), [#77603](https://github.com/PaddlePaddle/Paddle/pull/77603), [#77596](https://github.com/PaddlePaddle/Paddle/pull/77596), [#77595](https://github.com/PaddlePaddle/Paddle/pull/77595), [#77589](https://github.com/PaddlePaddle/Paddle/pull/77589), [#77579](https://github.com/PaddlePaddle/Paddle/pull/77579), [#77558](https://github.com/PaddlePaddle/Paddle/pull/77558), [#77534](https://github.com/PaddlePaddle/Paddle/pull/77534), [#77522](https://github.com/PaddlePaddle/Paddle/pull/77522), [#77509](https://github.com/PaddlePaddle/Paddle/pull/77509), [#77502](https://github.com/PaddlePaddle/Paddle/pull/77502), [#77456](https://github.com/PaddlePaddle/Paddle/pull/77456), [#77452](https://github.com/PaddlePaddle/Paddle/pull/77452), [#77443](https://github.com/PaddlePaddle/Paddle/pull/77443), [#77442](https://github.com/PaddlePaddle/Paddle/pull/77442), [#77435](https://github.com/PaddlePaddle/Paddle/pull/77435), [#77428](https://github.com/PaddlePaddle/Paddle/pull/77428), [#77427](https://github.com/PaddlePaddle/Paddle/pull/77427), [#77423](https://github.com/PaddlePaddle/Paddle/pull/77423), [#77422](https://github.com/PaddlePaddle/Paddle/pull/77422), [#77418](https://github.com/PaddlePaddle/Paddle/pull/77418), [#77413](https://github.com/PaddlePaddle/Paddle/pull/77413), [#77412](https://github.com/PaddlePaddle/Paddle/pull/77412), [#77411](https://github.com/PaddlePaddle/Paddle/pull/77411), [#77410](https://github.com/PaddlePaddle/Paddle/pull/77410), [#77409](https://github.com/PaddlePaddle/Paddle/pull/77409), [#77408](https://github.com/PaddlePaddle/Paddle/pull/77408), [#77407](https://github.com/PaddlePaddle/Paddle/pull/77407), [#77397](https://github.com/PaddlePaddle/Paddle/pull/77397), [#77394](https://github.com/PaddlePaddle/Paddle/pull/77394), [#77392](https://github.com/PaddlePaddle/Paddle/pull/77392), [#77386](https://github.com/PaddlePaddle/Paddle/pull/77386), [#77378](https://github.com/PaddlePaddle/Paddle/pull/77378), [#77377](https://github.com/PaddlePaddle/Paddle/pull/77377), [#77376](https://github.com/PaddlePaddle/Paddle/pull/77376), [#77375](https://github.com/PaddlePaddle/Paddle/pull/77375), [#77373](https://github.com/PaddlePaddle/Paddle/pull/77373), [#77327](https://github.com/PaddlePaddle/Paddle/pull/77327), [#77317](https://github.com/PaddlePaddle/Paddle/pull/77317), [#77312](https://github.com/PaddlePaddle/Paddle/pull/77312), [#77309](https://github.com/PaddlePaddle/Paddle/pull/77309), [#77307](https://github.com/PaddlePaddle/Paddle/pull/77307), [#77306](https://github.com/PaddlePaddle/Paddle/pull/77306), [#77299](https://github.com/PaddlePaddle/Paddle/pull/77299), [#77273](https://github.com/PaddlePaddle/Paddle/pull/77273), [#77246](https://github.com/PaddlePaddle/Paddle/pull/77246), [#77236](https://github.com/PaddlePaddle/Paddle/pull/77236), [#77230](https://github.com/PaddlePaddle/Paddle/pull/77230), [#77216](https://github.com/PaddlePaddle/Paddle/pull/77216), [#77206](https://github.com/PaddlePaddle/Paddle/pull/77206), [#77205](https://github.com/PaddlePaddle/Paddle/pull/77205), [#77184](https://github.com/PaddlePaddle/Paddle/pull/77184), [#77183](https://github.com/PaddlePaddle/Paddle/pull/77183), [#77142](https://github.com/PaddlePaddle/Paddle/pull/77142), [#77139](https://github.com/PaddlePaddle/Paddle/pull/77139), [#77130](https://github.com/PaddlePaddle/Paddle/pull/77130), [#77129](https://github.com/PaddlePaddle/Paddle/pull/77129), [#77128](https://github.com/PaddlePaddle/Paddle/pull/77128), [#77126](https://github.com/PaddlePaddle/Paddle/pull/77126), [#77123](https://github.com/PaddlePaddle/Paddle/pull/77123), [#77122](https://github.com/PaddlePaddle/Paddle/pull/77122), [#77120](https://github.com/PaddlePaddle/Paddle/pull/77120), [#77119](https://github.com/PaddlePaddle/Paddle/pull/77119), [#77118](https://github.com/PaddlePaddle/Paddle/pull/77118), [#77117](https://github.com/PaddlePaddle/Paddle/pull/77117), [#77112](https://github.com/PaddlePaddle/Paddle/pull/77112), [#77111](https://github.com/PaddlePaddle/Paddle/pull/77111), [#77110](https://github.com/PaddlePaddle/Paddle/pull/77110), [#77108](https://github.com/PaddlePaddle/Paddle/pull/77108), [#77106](https://github.com/PaddlePaddle/Paddle/pull/77106), [#77101](https://github.com/PaddlePaddle/Paddle/pull/77101), [#77100](https://github.com/PaddlePaddle/Paddle/pull/77100), [#77094](https://github.com/PaddlePaddle/Paddle/pull/77094), [#77092](https://github.com/PaddlePaddle/Paddle/pull/77092), [#77062](https://github.com/PaddlePaddle/Paddle/pull/77062), [#77041](https://github.com/PaddlePaddle/Paddle/pull/77041), [#77032](https://github.com/PaddlePaddle/Paddle/pull/77032), [#76764](https://github.com/PaddlePaddle/Paddle/pull/76764)

**文档（187 项）**
* 修复 API 文档示例代码、补充 docstring 并更新文档链接（19 项）：[#78410](https://github.com/PaddlePaddle/Paddle/pull/78410), [#78405](https://github.com/PaddlePaddle/Paddle/pull/78405), [#78024](https://github.com/PaddlePaddle/Paddle/pull/78024), [#77935](https://github.com/PaddlePaddle/Paddle/pull/77935), [#77771](https://github.com/PaddlePaddle/Paddle/pull/77771), [#77369](https://github.com/PaddlePaddle/Paddle/pull/77369), [#77326](https://github.com/PaddlePaddle/Paddle/pull/77326), [#77298](https://github.com/PaddlePaddle/Paddle/pull/77298), [#77297](https://github.com/PaddlePaddle/Paddle/pull/77297), [#77252](https://github.com/PaddlePaddle/Paddle/pull/77252), [#77169](https://github.com/PaddlePaddle/Paddle/pull/77169), [#77164](https://github.com/PaddlePaddle/Paddle/pull/77164), [#77157](https://github.com/PaddlePaddle/Paddle/pull/77157), [#77138](https://github.com/PaddlePaddle/Paddle/pull/77138), [#77070](https://github.com/PaddlePaddle/Paddle/pull/77070), [#76952](https://github.com/PaddlePaddle/Paddle/pull/76952), [#76838](https://github.com/PaddlePaddle/Paddle/pull/76838), [#76668](https://github.com/PaddlePaddle/Paddle/pull/76668), [#76632](https://github.com/PaddlePaddle/Paddle/pull/76632)
* 系统性修正文档示例中的代码块标记与格式，统一为 pycon 标记（164 项）：[#78022](https://github.com/PaddlePaddle/Paddle/pull/78022), [#78021](https://github.com/PaddlePaddle/Paddle/pull/78021), [#78020](https://github.com/PaddlePaddle/Paddle/pull/78020), [#78019](https://github.com/PaddlePaddle/Paddle/pull/78019), [#78018](https://github.com/PaddlePaddle/Paddle/pull/78018), [#78017](https://github.com/PaddlePaddle/Paddle/pull/78017), [#78016](https://github.com/PaddlePaddle/Paddle/pull/78016), [#78015](https://github.com/PaddlePaddle/Paddle/pull/78015), [#78014](https://github.com/PaddlePaddle/Paddle/pull/78014), [#78013](https://github.com/PaddlePaddle/Paddle/pull/78013), [#78012](https://github.com/PaddlePaddle/Paddle/pull/78012), [#78011](https://github.com/PaddlePaddle/Paddle/pull/78011), [#78010](https://github.com/PaddlePaddle/Paddle/pull/78010), [#78009](https://github.com/PaddlePaddle/Paddle/pull/78009), [#78008](https://github.com/PaddlePaddle/Paddle/pull/78008), [#78007](https://github.com/PaddlePaddle/Paddle/pull/78007), [#78006](https://github.com/PaddlePaddle/Paddle/pull/78006), [#78005](https://github.com/PaddlePaddle/Paddle/pull/78005), [#78003](https://github.com/PaddlePaddle/Paddle/pull/78003), [#78002](https://github.com/PaddlePaddle/Paddle/pull/78002), [#78001](https://github.com/PaddlePaddle/Paddle/pull/78001), [#78000](https://github.com/PaddlePaddle/Paddle/pull/78000), [#77999](https://github.com/PaddlePaddle/Paddle/pull/77999), [#77998](https://github.com/PaddlePaddle/Paddle/pull/77998), [#77997](https://github.com/PaddlePaddle/Paddle/pull/77997), [#77996](https://github.com/PaddlePaddle/Paddle/pull/77996), [#77995](https://github.com/PaddlePaddle/Paddle/pull/77995), [#77994](https://github.com/PaddlePaddle/Paddle/pull/77994), [#77993](https://github.com/PaddlePaddle/Paddle/pull/77993), [#77992](https://github.com/PaddlePaddle/Paddle/pull/77992), [#77990](https://github.com/PaddlePaddle/Paddle/pull/77990), [#77989](https://github.com/PaddlePaddle/Paddle/pull/77989), [#77988](https://github.com/PaddlePaddle/Paddle/pull/77988), [#77987](https://github.com/PaddlePaddle/Paddle/pull/77987), [#77986](https://github.com/PaddlePaddle/Paddle/pull/77986), [#77985](https://github.com/PaddlePaddle/Paddle/pull/77985), [#77984](https://github.com/PaddlePaddle/Paddle/pull/77984), [#77983](https://github.com/PaddlePaddle/Paddle/pull/77983), [#77982](https://github.com/PaddlePaddle/Paddle/pull/77982), [#77980](https://github.com/PaddlePaddle/Paddle/pull/77980), [#77979](https://github.com/PaddlePaddle/Paddle/pull/77979), [#77978](https://github.com/PaddlePaddle/Paddle/pull/77978), [#77976](https://github.com/PaddlePaddle/Paddle/pull/77976), [#77975](https://github.com/PaddlePaddle/Paddle/pull/77975), [#77974](https://github.com/PaddlePaddle/Paddle/pull/77974), [#77972](https://github.com/PaddlePaddle/Paddle/pull/77972), [#77970](https://github.com/PaddlePaddle/Paddle/pull/77970), [#77969](https://github.com/PaddlePaddle/Paddle/pull/77969), [#77968](https://github.com/PaddlePaddle/Paddle/pull/77968), [#77967](https://github.com/PaddlePaddle/Paddle/pull/77967), [#77966](https://github.com/PaddlePaddle/Paddle/pull/77966), [#77965](https://github.com/PaddlePaddle/Paddle/pull/77965), [#77964](https://github.com/PaddlePaddle/Paddle/pull/77964), [#77963](https://github.com/PaddlePaddle/Paddle/pull/77963), [#77962](https://github.com/PaddlePaddle/Paddle/pull/77962), [#77961](https://github.com/PaddlePaddle/Paddle/pull/77961), [#77960](https://github.com/PaddlePaddle/Paddle/pull/77960), [#77958](https://github.com/PaddlePaddle/Paddle/pull/77958), [#77956](https://github.com/PaddlePaddle/Paddle/pull/77956), [#77955](https://github.com/PaddlePaddle/Paddle/pull/77955), [#77954](https://github.com/PaddlePaddle/Paddle/pull/77954), [#77953](https://github.com/PaddlePaddle/Paddle/pull/77953), [#77952](https://github.com/PaddlePaddle/Paddle/pull/77952), [#77951](https://github.com/PaddlePaddle/Paddle/pull/77951), [#77950](https://github.com/PaddlePaddle/Paddle/pull/77950), [#77949](https://github.com/PaddlePaddle/Paddle/pull/77949), [#77948](https://github.com/PaddlePaddle/Paddle/pull/77948), [#77947](https://github.com/PaddlePaddle/Paddle/pull/77947), [#77946](https://github.com/PaddlePaddle/Paddle/pull/77946), [#77945](https://github.com/PaddlePaddle/Paddle/pull/77945), [#77944](https://github.com/PaddlePaddle/Paddle/pull/77944), [#77943](https://github.com/PaddlePaddle/Paddle/pull/77943), [#77941](https://github.com/PaddlePaddle/Paddle/pull/77941), [#77940](https://github.com/PaddlePaddle/Paddle/pull/77940), [#77939](https://github.com/PaddlePaddle/Paddle/pull/77939), [#77937](https://github.com/PaddlePaddle/Paddle/pull/77937), [#77934](https://github.com/PaddlePaddle/Paddle/pull/77934), [#77933](https://github.com/PaddlePaddle/Paddle/pull/77933), [#77932](https://github.com/PaddlePaddle/Paddle/pull/77932), [#77931](https://github.com/PaddlePaddle/Paddle/pull/77931), [#77930](https://github.com/PaddlePaddle/Paddle/pull/77930), [#77929](https://github.com/PaddlePaddle/Paddle/pull/77929), [#77926](https://github.com/PaddlePaddle/Paddle/pull/77926), [#77914](https://github.com/PaddlePaddle/Paddle/pull/77914), [#77912](https://github.com/PaddlePaddle/Paddle/pull/77912), [#77910](https://github.com/PaddlePaddle/Paddle/pull/77910), [#77908](https://github.com/PaddlePaddle/Paddle/pull/77908), [#77906](https://github.com/PaddlePaddle/Paddle/pull/77906), [#77905](https://github.com/PaddlePaddle/Paddle/pull/77905), [#77904](https://github.com/PaddlePaddle/Paddle/pull/77904), [#77903](https://github.com/PaddlePaddle/Paddle/pull/77903), [#77902](https://github.com/PaddlePaddle/Paddle/pull/77902), [#77900](https://github.com/PaddlePaddle/Paddle/pull/77900), [#77888](https://github.com/PaddlePaddle/Paddle/pull/77888), [#77887](https://github.com/PaddlePaddle/Paddle/pull/77887), [#77886](https://github.com/PaddlePaddle/Paddle/pull/77886), [#77885](https://github.com/PaddlePaddle/Paddle/pull/77885), [#77884](https://github.com/PaddlePaddle/Paddle/pull/77884), [#77883](https://github.com/PaddlePaddle/Paddle/pull/77883), [#77882](https://github.com/PaddlePaddle/Paddle/pull/77882), [#77881](https://github.com/PaddlePaddle/Paddle/pull/77881), [#77865](https://github.com/PaddlePaddle/Paddle/pull/77865), [#77864](https://github.com/PaddlePaddle/Paddle/pull/77864), [#77862](https://github.com/PaddlePaddle/Paddle/pull/77862), [#77859](https://github.com/PaddlePaddle/Paddle/pull/77859), [#77858](https://github.com/PaddlePaddle/Paddle/pull/77858), [#77857](https://github.com/PaddlePaddle/Paddle/pull/77857), [#77856](https://github.com/PaddlePaddle/Paddle/pull/77856), [#77853](https://github.com/PaddlePaddle/Paddle/pull/77853), [#77850](https://github.com/PaddlePaddle/Paddle/pull/77850), [#77830](https://github.com/PaddlePaddle/Paddle/pull/77830), [#77828](https://github.com/PaddlePaddle/Paddle/pull/77828), [#77826](https://github.com/PaddlePaddle/Paddle/pull/77826), [#77822](https://github.com/PaddlePaddle/Paddle/pull/77822), [#77820](https://github.com/PaddlePaddle/Paddle/pull/77820), [#77817](https://github.com/PaddlePaddle/Paddle/pull/77817), [#77799](https://github.com/PaddlePaddle/Paddle/pull/77799), [#77798](https://github.com/PaddlePaddle/Paddle/pull/77798), [#77794](https://github.com/PaddlePaddle/Paddle/pull/77794), [#77791](https://github.com/PaddlePaddle/Paddle/pull/77791), [#77777](https://github.com/PaddlePaddle/Paddle/pull/77777), [#77774](https://github.com/PaddlePaddle/Paddle/pull/77774), [#77773](https://github.com/PaddlePaddle/Paddle/pull/77773), [#77764](https://github.com/PaddlePaddle/Paddle/pull/77764), [#77763](https://github.com/PaddlePaddle/Paddle/pull/77763), [#77760](https://github.com/PaddlePaddle/Paddle/pull/77760), [#77758](https://github.com/PaddlePaddle/Paddle/pull/77758), [#77748](https://github.com/PaddlePaddle/Paddle/pull/77748), [#77747](https://github.com/PaddlePaddle/Paddle/pull/77747), [#77745](https://github.com/PaddlePaddle/Paddle/pull/77745), [#77744](https://github.com/PaddlePaddle/Paddle/pull/77744), [#77721](https://github.com/PaddlePaddle/Paddle/pull/77721), [#77720](https://github.com/PaddlePaddle/Paddle/pull/77720), [#77698](https://github.com/PaddlePaddle/Paddle/pull/77698), [#77697](https://github.com/PaddlePaddle/Paddle/pull/77697), [#77693](https://github.com/PaddlePaddle/Paddle/pull/77693), [#77692](https://github.com/PaddlePaddle/Paddle/pull/77692), [#77686](https://github.com/PaddlePaddle/Paddle/pull/77686), [#77672](https://github.com/PaddlePaddle/Paddle/pull/77672), [#77671](https://github.com/PaddlePaddle/Paddle/pull/77671), [#77667](https://github.com/PaddlePaddle/Paddle/pull/77667), [#77666](https://github.com/PaddlePaddle/Paddle/pull/77666), [#77656](https://github.com/PaddlePaddle/Paddle/pull/77656), [#77645](https://github.com/PaddlePaddle/Paddle/pull/77645), [#77634](https://github.com/PaddlePaddle/Paddle/pull/77634), [#77633](https://github.com/PaddlePaddle/Paddle/pull/77633), [#77630](https://github.com/PaddlePaddle/Paddle/pull/77630), [#77620](https://github.com/PaddlePaddle/Paddle/pull/77620), [#77618](https://github.com/PaddlePaddle/Paddle/pull/77618), [#77617](https://github.com/PaddlePaddle/Paddle/pull/77617), [#77616](https://github.com/PaddlePaddle/Paddle/pull/77616), [#77613](https://github.com/PaddlePaddle/Paddle/pull/77613), [#77612](https://github.com/PaddlePaddle/Paddle/pull/77612), [#77611](https://github.com/PaddlePaddle/Paddle/pull/77611), [#77608](https://github.com/PaddlePaddle/Paddle/pull/77608), [#77607](https://github.com/PaddlePaddle/Paddle/pull/77607), [#77600](https://github.com/PaddlePaddle/Paddle/pull/77600), [#77570](https://github.com/PaddlePaddle/Paddle/pull/77570), [#77569](https://github.com/PaddlePaddle/Paddle/pull/77569), [#77513](https://github.com/PaddlePaddle/Paddle/pull/77513), [#77493](https://github.com/PaddlePaddle/Paddle/pull/77493), [#77472](https://github.com/PaddlePaddle/Paddle/pull/77472), [#77462](https://github.com/PaddlePaddle/Paddle/pull/77462), [#76794](https://github.com/PaddlePaddle/Paddle/pull/76794)
* 其它文档修改（4 项）：[#78448](https://github.com/PaddlePaddle/Paddle/pull/78448), [#77738](https://github.com/PaddlePaddle/Paddle/pull/77738), [#77177](https://github.com/PaddlePaddle/Paddle/pull/77177), [#76991](https://github.com/PaddlePaddle/Paddle/pull/76991)

**测试（11 项）**
* 新增 AI 编辑相关的测试用例（4 项）：[#78634](https://github.com/PaddlePaddle/Paddle/pull/78634), [#78509](https://github.com/PaddlePaddle/Paddle/pull/78509), [#78459](https://github.com/PaddlePaddle/Paddle/pull/78459), [#78380](https://github.com/PaddlePaddle/Paddle/pull/78380)
* 其它测试调整（7 项）：[#78562](https://github.com/PaddlePaddle/Paddle/pull/78562), [#78561](https://github.com/PaddlePaddle/Paddle/pull/78561), [#77804](https://github.com/PaddlePaddle/Paddle/pull/77804), [#77750](https://github.com/PaddlePaddle/Paddle/pull/77750), [#77586](https://github.com/PaddlePaddle/Paddle/pull/77586), [#77446](https://github.com/PaddlePaddle/Paddle/pull/77446), [#77046](https://github.com/PaddlePaddle/Paddle/pull/77046)

### 7. 贡献者名单
[Nyakku Shigure](https://github.com/SigureMo), [Ryan](https://github.com/DrRyanHuang), [LiYuRio](https://github.com/LiYuRio), [wanghuancoder](https://github.com/wanghuancoder), [Shuhao Liang](https://github.com/lshpku), [adam-xiaoyao](https://github.com/adam-xiaoyao), [SUN Dong](https://github.com/DanielSun11), [Yohanna](https://github.com/YuhanXu), [Guoxia Wang](https://github.com/GuoxiaWang), [fxyfxy777](https://github.com/fxyfxy777), [Nyako Shigure](https://github.com/ShigureNyako), [zhengshengning](https://github.com/zhengshengning), [Zhan Rongrui](https://github.com/zrr1999), [Yuqiang Ge](https://github.com/YqGe585), [xingmingyyj](https://github.com/xingmingyyj), [xuanyuanminzheng](https://github.com/xuanyuanminzheng), [Wang Rui](https://github.com/wr-web), [gouzil](https://github.com/gouzil), [ALGO1832](https://github.com/algorithm1832), [Zx](https://github.com/ZhangX-21), [HydrogenSulfate](https://github.com/HydrogenSulfate), [Haze188 灏喆](https://github.com/Hz188), [Wennie396](https://github.com/Wennie396), [Difer](https://github.com/Difers), [AlAuAu](https://github.com/AlAuAu), [Xiangrui Yu](https://github.com/xxyux), [Yuang Liu](https://github.com/FeixLiu), [Eddie-Wang](https://github.com/Eddie-Wang1120), [XiangzheWang](https://github.com/Waynezee), [ShenLiang](https://github.com/ForFishes), [HU Shenwei](https://github.com/hushenwei2000), [Chen Kehong](https://github.com/whatyourname12345), [AIbin](https://github.com/chang-wenbin), [Nana](https://github.com/NKNaN), [ZhenxingLi](https://github.com/Xing-lil), [sevenan2](https://github.com/sevenan2), [tianhaodongbd](https://github.com/tianhaodongbd), [co63oc](https://github.com/co63oc), [ooo oo](https://github.com/ooooo-create), [sneaxiy](https://github.com/sneaxiy), [Gu Shiwei](https://github.com/swgu98), [wangbingguang](https://github.com/wangbingguang2026), [omoYang](https://github.com/omoYang), [baiyue](https://github.com/DongBaiYue), [zhanghonggeng](https://github.com/zhanghonggeng), [Zhaowu Pan](https://github.com/A-nnonymous), [zccjjj](https://github.com/zccjjj), [iLeGend](https://github.com/jinyouzhi), [Tianyu Zheng](https://github.com/zty-king), [yangguohao](https://github.com/yangguohao), [Zhou Xin](https://github.com/LittleHeroZZZX), [Zhang Ting](https://github.com/zhangting2020), [feixi](https://github.com/feixi139), [Bohao Fan](https://github.com/greenhandF), [baoqiwen](https://github.com/baoqiwen), [Qianyue He](https://github.com/Enigmatisms), [苍天荒](https://github.com/cangtianhuang), [zhwesky2010](https://github.com/zhwesky2010), [liuruyan](https://github.com/liuruyan), [Tofu](https://github.com/WHoutstanding), [Yichen Zhang](https://github.com/pkuzyc), [starcrown001](https://github.com/starcrown001), [lijin23](https://github.com/lj970926), [Fang Ru](https://github.com/G2uge), [onecatcn](https://github.com/onecatcn), [Jia Ningyu](https://github.com/jianingyu-ustc), [bigwhite37](https://github.com/bigwhite37), [lizan1999](https://github.com/lizan1999), [xinruiM](https://github.com/xinruiM), [Leo Guo](https://github.com/ZibinGuo), [Jingzong Liu](https://github.com/tjujingzong), [yinwei](https://github.com/iosmers), [Lucas](https://github.com/cqulilujia), [ddchenhao66](https://github.com/ddchenhao66), [方程](https://github.com/fchange), [ming1753](https://github.com/ming1753), [umiswing](https://github.com/umiswing), [Runming Xie](https://github.com/youge325), [Bingoo](https://github.com/BingooYang), [Xiaochun Yang](https://github.com/Le-soleile), [Wenfei (Charles) Qi](https://github.com/Manfredss), [Moon](https://github.com/cvbret), [AmeoInoru](https://github.com/OctoberPrayRain), [Aidenwu0209](https://github.com/Aidenwu0209), [cyberslack_lee](https://github.com/enkilee), [AuraWu](https://github.com/BayesianAura), [Merlin](https://github.com/MerlinSMQWQ), [Rui Huang](https://github.com/Khoray), [qingyun](https://github.com/xpislp), [Xianzheng Pang](https://github.com/pxzbest), [xunyoyo](https://github.com/xunyoyo), [fangfangssj](https://github.com/fangfangssj), [PlumBlossomMaid](https://github.com/PlumBlossomMaid), [risemeup1](https://github.com/risemeup1), [liuhao2638](https://github.com/liuhao2638), [risemeup1111](https://github.com/risemeup1111), [Yutian Rao](https://github.com/raoyutian), [cutetocute](https://github.com/Noai-oss), [tianyuzhou668](https://github.com/tianyuzhou668), [tianlef](https://github.com/tianlef), [Zhenghai Zhang](https://github.com/ccsuzzh), [zhengzhonghui](https://github.com/deepllz), [ZhouMinhao98](https://github.com/ZhouMinhao98), [jzhang533](https://github.com/jzhang533), [Ray961123](https://github.com/Ray961123), [Salman Chishti](https://github.com/salmanmkc), [paddle-xpu-bot](https://github.com/paddle-xpu-bot), [盈盈秋水](https://github.com/HZ1ovo), [liang](https://github.com/liangqi520), [yyxxddjj](https://github.com/yyxxddjj), [John Feng](https://github.com/JohnLeFeng), [SidusAntares](https://github.com/SidusAntares), [JiaViii](https://github.com/JiaViii), [Ricardo-shuo-liu](https://github.com/Ricardo-shuo-liu), [Alnair](https://github.com/A1nair), [DavidGu](https://github.com/DavidGu-Datong), [M4jupitercannon](https://github.com/M4jupitercannon), [Pandya Mayur](https://github.com/Pandya-mayur)
Related Repositories

ggml-org/llama.cpp★ 131,945
LLM inference in C/C++
LadybirdBrowser/ladybird★ 66,606
Truly independent web browser
ClickHouse/ClickHouse★ 50,528
ClickHouse® is a real-time analytics database management system
LizardByte/Sunshine★ 42,490
Self-hosted game stream host for Moonlight.
deskflow/deskflow★ 29,980
Share a single keyboard and mouse between multiple computers.