62 Pull requests merged by 25 people

• [RL] fix bug in qwen fuse qkv

  #10821 merged Jul 6, 2025

• Move setup_fp8.py to FleetY

  #10820 merged Jul 6, 2025

• [RL] disable aistudio download and fix qwen bug

  #10819 merged Jul 5, 2025

• Move FP8 ops to FleetY branch

  #10803 merged Jul 5, 2025

• support dispatch both bf16 & fp8

  #10817 merged Jul 4, 2025

• [AutoParallel] close dynamic sharding CI test

  #10791 merged Jul 4, 2025

• [LLM] fix_qwen3_moe

  #10801 merged Jul 4, 2025

• Set FP8Linear weight update by inplace add

  #10813 merged Jul 4, 2025

• Fix nccl ut

  #10812 merged Jul 4, 2025

• [tests] migrate to unittest.assertEqual

  #10670 merged Jul 4, 2025

• optimize scale transpose

  #10810 merged Jul 4, 2025

• optimizer_dual_pp_post_node_memory

  #10806 merged Jul 3, 2025

• lock gemm sm 112

  #10805 merged Jul 3, 2025

• Add fp8 opt config for dualpipe

  #10804 merged Jul 3, 2025

• Add expert num 32 support for FP8 ops and fix ut

  #10802 merged Jul 3, 2025

• Refine fp8 node

  #10798 merged Jul 3, 2025

• Support expert number 16 in FP8 operators

  #10799 merged Jul 3, 2025

• Adapt dispatch_quant_node to fp8 fusion moe node

  #10794 merged Jul 2, 2025

• [Auto-parallel] Fix sharding all_gather overlap in auto_dy

  #10782 merged Jul 2, 2025

• [Auto-parallel] Add llama13b benchmark fast_ln

  #10785 merged Jul 2, 2025

• [Auto-parallel] Add llama7b benchmark fast_ln

  #10786 merged Jul 2, 2025

• optimize dual pp memory

  #10787 merged Jul 2, 2025

• Reset EMA state dict when distributed strategy not matched

  #10790 merged Jul 1, 2025

• [CherryPick]Adapt new nccl version (#10768)

  #10784 merged Jul 1, 2025

• [Auto-Parallel] optimize the perfermance of GPT-3

  #10780 merged Jul 1, 2025

• adapt new nccl version

  #10768 merged Jul 1, 2025

• [Auto Parallel] Add enable_linear_fused_grad_add in qwen benchmark

  #10779 merged Jun 30, 2025

• [Auto Parallel] Adapt mp_async_allreduce optimize in auto paralle

  #10770 merged Jun 30, 2025

• disable use_cinn for intermediate_api_meta-llama-Llama-2-7b

  #10778 merged Jun 27, 2025

• [CI] update auto_parallel loss_base for paddle#73408

  #10771 merged Jun 26, 2025

• Fix compile and 0size problem

  #10773 merged Jun 25, 2025

• Add gpt dyanmic auto benchamrk

  #10759 merged Jun 25, 2025

• [AutoParallel] llama-13b dynamic auto json add model_type

  #10766 merged Jun 25, 2025

• [Auto Paralle] open llama test

  #10762 merged Jun 25, 2025

• Change custom op to paddle api

  #10748 merged Jun 25, 2025

• Fix the import issues caused by Paddle version

  #10760 merged Jun 24, 2025

• add zcc_ema_loss_threshold args to avoid merging models with loss spike

  #10763 merged Jun 24, 2025

• open gpt dp CI test

  #10740 merged Jun 24, 2025

• 【Hackathon 8th No.32】 Adam-mini 精调算法复现

  #10413 merged Jun 24, 2025

• Add expert subbatch & inplace fused_swiglu_probs_bwd ops

  #10757 merged Jun 23, 2025

• [Auto Parallel] add llama with auto pp

  #10751 merged Jun 23, 2025

• Remove restrictions on the use of allgather_overlap (#10741)

  #10756 merged Jun 23, 2025

• Run llama dyanmic pp perf

  #10753 merged Jun 20, 2025

• 0size fix in unzip-zip op

  #10755 merged Jun 20, 2025

• [Auto Parallel] close llama dynamic dp test temporarily

  #10750 merged Jun 19, 2025

• Integrate DataProto into the GRPO

  #10597 merged Jun 18, 2025

• [AutoParalle] fix fused layers fp32 multi precision

  #10743 merged Jun 18, 2025

• Remove restrictions on the use of allgather_overlap

  #10741 merged Jun 18, 2025

• Fix tp model load

  #10739 merged Jun 17, 2025

• Update architectures saving

  #10738 merged Jun 17, 2025

• close gpt dynamic dp test temporarily

  #10736 merged Jun 13, 2025

• fix big_tensor issue in swiglu_probs_bwd

  #10735 merged Jun 13, 2025

• fix wint4/8lora

  #10728 merged Jun 12, 2025

• Add cuda stream for fused quant kernel

  #10716 merged Jun 12, 2025

• aistudio 使用 SDK 进行下载

  #10678 merged Jun 12, 2025

• [inference]Add support for splitting Paddle tensors.

  #10721 merged Jun 12, 2025

• [CI] skip unit case for hang

  #10642 merged Jun 11, 2025

• [AutoParallel] Fix fast_rms_norm

  #10711 merged Jun 11, 2025

• update fp8 training

  #10720 merged Jun 11, 2025

• [LLM] use safer grad sync method when enabling sp

  #10714 merged Jun 9, 2025

• refine fp8 code

  #10669 merged Jun 9, 2025

• 【Inference Optimize】update moe_preprocess for wint2.x

  #10702 merged Jun 7, 2025

39 Pull requests opened by 25 people

• [Auto-Parallel] Add benchmark for fast_rms_norm in llama13b N4C32 dy_auto

  #10713 opened Jun 9, 2025

• Add forward unittest for fused_transpose_split_quant

  #10717 opened Jun 10, 2025

• [CI]add workflow yml

  #10718 opened Jun 10, 2025

• add auto_parallel context_parallel strategy

  #10722 opened Jun 10, 2025

• 使用新的aistudio sdk下载模型文件，上传功能不受影响

  #10723 opened Jun 10, 2025

• update deep_gemm

  #10724 opened Jun 10, 2025

• [DeepGEMM] Print tuning and compilation time stat

  #10725 opened Jun 11, 2025

• [x86 CPU] fix cpu GQA issue

  #10729 opened Jun 12, 2025

• [Autoparallel] support llama sep

  #10730 opened Jun 12, 2025

• [Auto Parallel] Add gpt pp model in dynamic auto

  #10731 opened Jun 12, 2025

• formers

  #10732 opened Jun 12, 2025

• Add big tensor tests for fused kernels

  #10734 opened Jun 13, 2025

• Refine Unzip V1: Reduce size of cudaMemset

  #10742 opened Jun 17, 2025

• Refine Unzip v2: Parallelism in Expert Dimensions

  #10745 opened Jun 17, 2025

• 'nola提交'

  #10747 opened Jun 18, 2025

• Support 3d input in dequant op

  #10752 opened Jun 19, 2025

• update deep_gemm

  #10754 opened Jun 20, 2025

• Fp8 gemm & quant operators replacement

  #10761 opened Jun 23, 2025

• [CI 不review]test a100

  #10764 opened Jun 24, 2025

• add strategies for user2new_group

  #10765 opened Jun 24, 2025

• [Auto-parallel] Adapt ernie pp for auto_dy paddlenlp

  #10767 opened Jun 25, 2025

• support subbatch for long context

  #10769 opened Jun 25, 2025

• 0size fix of unzip & zip

  #10772 opened Jun 25, 2025

• [fix] add parameters arg into AdamWMini

  #10774 opened Jun 25, 2025

• [Auto Parallel] Align the meaning of configuration items, decouple auto-parallel pipeline networking and trainer

  #10775 opened Jun 25, 2025

• Add ShortGPT integration

  #10777 opened Jun 26, 2025

• [AutoParallel] init sync param

  #10783 opened Jul 1, 2025

• Fix the _save function so that it can save the optimizer parameters.

  #10789 opened Jul 1, 2025

• [AutoParallel] Open dynamic sharding CI test

  #10793 opened Jul 2, 2025

• [AutoParallel] Remove dynamic pipeline import adaption

  #10795 opened Jul 2, 2025

• N1C8 dsv3 config

  #10796 opened Jul 2, 2025

• [Auto Parallel] change fused_layers.py to support fused_linear_grad add/sync_mp_allreduce/sp_async_reduce_scatter in same time

  #10797 opened Jul 2, 2025

• [Auto-parallel] temp adjust the tensor_fusion config

  #10800 opened Jul 3, 2025

• support quant return transpose only

  #10811 opened Jul 4, 2025

• dispatch support fp8

  #10814 opened Jul 4, 2025

• Add recompute for post_norm and moe_gate

  #10815 opened Jul 4, 2025

• [Auto-parallel] Fix sp_async and mp_async used together

  #10816 opened Jul 4, 2025

• 修复qwen3moe的一系列报错(justin-0704)

  #10818 opened Jul 4, 2025

• Add tokens_zip_unique_add_subbatch and merge_subbatch_cast ops

  #10822 opened Jul 6, 2025

24 Issues closed by 5 people

• [Question]: slm/model_zoo/ernie-3.0-tiny 多个意图的话 应该怎么改

  #10478 closed Jul 7, 2025

• [Question]: 使用text_matching的predict_pointwise每次检测结果不一样

  #10459 closed Jul 6, 2025

• [Question]: UIE大模型下载地址

  #10414 closed Jul 2, 2025

• [Bug]: export静态图无报错但没有pdmodel文件

  #10415 closed Jul 2, 2025

• DocVQA-ZH数据集链接失效

  #6470 closed Jul 1, 2025

• [Bug]: ImportError: cannot import name 'download' from 'aistudio_sdk.hub'

  #10781 closed Jul 1, 2025

• [Bug]: 服务化部署失败，ModuleNotFoundError: No module named 'predict'

  #10265 closed Jun 30, 2025

• [Question]: pp_uie微调之后, 推理的output.json没有输出

  #10406 closed Jun 30, 2025

• [Question]: 微调之后的UIE结果输出

  #10420 closed Jun 30, 2025

• [Question]: UIE微调之后导出静态模型报错

  #10421 closed Jun 30, 2025

• [Bug]: 训练后评估阶段报错

  #9908 closed Jun 29, 2025

• [Question]: 训练好的pp-uie0.5B模型只能抽取开头到一半,抽取不到结尾的文字。

  #10316 closed Jun 29, 2025

• [Bug]: UIE导出onnx后模型可以正常推理，但是指标全0

  #10392 closed Jun 29, 2025

• [Question]: paddlepaddle-gpu paddlenlp UIE版本问题

  #10405 closed Jun 29, 2025

• [Bug]: Qwen2Tokenizer中add_special_tokens方法存在bug

  #10391 closed Jun 28, 2025

• [Question]: 进行BERT类模型修改special token实验时发现一个问题,如果tokenizer自动padding会造成显存占用很大提升

  #10371 closed Jun 25, 2025

• [Question]: pp-uie是否支持标注的内容是成多句或是成段的?

  #10388 closed Jun 25, 2025

• [Question]: 科学计算的工业应用场景Agent，需要100%控制随机性

  #10346 closed Jun 20, 2025

• [Question]: Will open dataset for ppuie model

  #10556 closed Jun 19, 2025

• 完成实测、提交测评，赢取奖金！——DeepSeek-R1-MTP 单机部署实战

  #10166 closed Jun 18, 2025

• [Question]:MTP原理

  #10258 closed Jun 13, 2025

• [Question]: pp-uie推理返回格式

  #10245 closed Jun 12, 2025

• [Question]: 部署语义检索系统后端报错 : search_phase_execution_exception

  #6535 closed Jun 10, 2025

• [Question]: ernie-doc模型ernie-doc-base-en，ernie-doc-base-zh无法下载

  #10287 closed Jun 10, 2025

9 Issues opened by 9 people

• paddlenlp库中的aistudio_sdk版本问题

  #10809 opened Jul 3, 2025

• [Bug]: RuntimeError: (PreconditionNotMet) Tensor's dimension is out of bound.Tensor's dimension must be equal or less than the size of its memory.But received Tensor's dimension is 8, memory's size is 0.

  #10792 opened Jul 2, 2025

• ernie-1.0 支持padding 吗？我试验了padding, 结果不对呢

  #10788 opened Jul 1, 2025

• pp-uie sft训练时报错

  #10776 opened Jun 26, 2025

• [Question]: 在paddlenlp部署推理过程中加载Qwen/Qwen2-1.5B等模型报错

  #10758 opened Jun 23, 2025

• [Docs]: llama xpu部署相关文档部分下载链接失效

  #10737 opened Jun 13, 2025

• [Question]: 多机部署支持

  #10727 opened Jun 12, 2025

• [Question]: PaddleNLP 训练好的模型如何部署推理服务？

  #10726 opened Jun 12, 2025

• [Bug]: 安装完paddlenlp_ops运行不起来

  #10719 opened Jun 10, 2025

56 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Llama auto 13b benchmark with fuse linear

  #10695 commented on Jun 19, 2025 • 1 new comment

• Deepseek xpu

  #10340 commented on Jun 22, 2025 • 0 new comments

• Cleaning up and enhancing act_quant & act_dequant optest.

  #10351 commented on Jun 11, 2025 • 0 new comments

• Optimize fused expert

  #10355 commented on Jun 9, 2025 • 0 new comments

• [LLM] modify zcc names back for test

  #10356 commented on Jun 11, 2025 • 0 new comments

• optmize fp8 expert gemm

  #10366 commented on Jun 9, 2025 • 0 new comments

• add_pybind for dyGraph predictor running

  #10374 commented on Jun 11, 2025 • 0 new comments

• update

  #10418 commented on Jun 16, 2025 • 0 new comments

• [Dont Merge] Simple Append Attention impl for TPA & MFA benchmark

  #10419 commented on Jun 23, 2025 • 0 new comments

• 【Inference】add moe gemm 1

  #10436 commented on Jun 18, 2025 • 0 new comments

• int8 train test

  #10451 commented on Jun 9, 2025 • 0 new comments

• Sharding reshard supports mismatch parameter name

  #10467 commented on Jun 25, 2025 • 0 new comments

• Add forward_backward_overlap_scheduler in pipeline_parallel_config

  #10474 commented on Jun 23, 2025 • 0 new comments

• [Not Review] Support the use of pcc.

  #10477 commented on Jun 23, 2025 • 0 new comments

• Add arbitrary expert_num and topk support for unzip and zip.

  #10511 commented on Jun 27, 2025 • 0 new comments

• [LLM] add fuse attention options to LlmMetaConfig

  #10542 commented on Jul 6, 2025 • 0 new comments

• Add gpt3 13b dynamic auto benchmark

  #10548 commented on Jul 7, 2025 • 0 new comments

• Add llama-13b dynamic auto benchmark

  #10549 commented on Jul 7, 2025 • 0 new comments

• [LLM] Modify fuse layout

  #10555 commented on Jul 7, 2025 • 0 new comments

• [Inference] Add new wint2.75/wint2.5 quant type and support DeepseekV3

  #10578 commented on Jul 4, 2025 • 0 new comments

• Apply merge and split function for mp_reshard

  #10637 commented on Jun 17, 2025 • 0 new comments

• Support fastsafetensors to load model

  #10667 commented on Jun 13, 2025 • 0 new comments

• [feat] Integrate Galvatron (an automatic parallel system integrating …

  #10680 commented on Jun 23, 2025 • 0 new comments

• 【Hackathon 8th No.28】在 PaddleNLP 中复现 Phi3

  #10688 commented on Jun 9, 2025 • 0 new comments

• Tmp llama13b benchmark

  #10694 commented on Jun 10, 2025 • 0 new comments

• add register for auto model

  #10699 commented on Jun 11, 2025 • 0 new comments

• add generate_expert_indices op

  #10705 commented on Jun 11, 2025 • 0 new comments

• Adapt modeling_pp to fp8 fusion moe node

  #10712 commented on Jun 17, 2025 • 0 new comments

• [Question]: pp-uie使用taskflow无法加载自定义路径模型，只能加载默认路径：C:\Users\.paddlenlp\models\paddlenlp/PP-UIE-0.5B

  #10409 commented on Jun 9, 2025 • 0 new comments

• [Question]: 我需要在离线环境下部署pp-uie模型,我把训练好的模型放到了uer/.paddlenlp/model/paddlenlp/PP-uie-7B,但是每次信息抽取仍然需要下载shards,我应该如何解决呢?

  #10425 commented on Jun 13, 2025 • 0 new comments

• [Question]: paddlenlp 3.0.0b1版本使用uie-m-base报错找不到static/inference.pdmodel

  #9646 commented on Jun 17, 2025 • 0 new comments

• [Bug]: 在用Taskflow推理的时候，指定本地模型路径没生效

  #10660 commented on Jun 20, 2025 • 0 new comments

• [Question]: xpu镜像下，安装paddle和paddlenlp（beta3/4）后跑Qwen2.5-0.5B报错

  #10435 commented on Jun 24, 2025 • 0 new comments

• [Question]: text_to_image训练显存消耗过大

  #5717 commented on Jun 24, 2025 • 0 new comments

• Task类里部分代码是不是写错了？

  #10499 commented on Jun 25, 2025 • 0 new comments

• 文本分类中的多分类，模型预测能否返回多个lebel和score？

  #10495 commented on Jun 25, 2025 • 0 new comments

• [Question]: qa = Taskflow("document_intelligence") ，找不到 inference.json文件

  #10512 commented on Jun 27, 2025 • 0 new comments

• [Question]: docker --read-only运行镜像时报错：OSError: [Errno 30] Read-only file system: '/root/.paddlenlp'

  #10503 commented on Jun 27, 2025 • 0 new comments

• [Question]: text_classification 模型训练完成导出的文件夹export中缺少model.pdmodel文件

  #10416 commented on Jun 29, 2025 • 0 new comments

• [Bug]: 按照PP-UIE的新版本说明文档，调用大模型精调脚本 llm/run-finetune.py时，报错 No module named 'paddlenlp.datasets.json'

  #10543 commented on Jul 6, 2025 • 0 new comments

• [Question]: 使用GPU运行时报错：terminate called after throwing an instance of 'thrust::system::system_error'

  #10535 commented on Jul 6, 2025 • 0 new comments

• [Question]: 层次分类有些推理结果到不了最底层

  #10557 commented on Jul 7, 2025 • 0 new comments

• implemention of lqlora

  #8820 commented on Jun 28, 2025 • 0 new comments

• [FastTokenizer] split_special_tokens same as slow tokenizer

  #9610 commented on Jun 14, 2025 • 0 new comments

• [Tokenizer] Upgrade tokenizer

  #9683 commented on Jun 20, 2025 • 0 new comments

• [LLM] Add fused attention in Qwen2MoE

  #9767 commented on Jun 9, 2025 • 0 new comments

• Test deep seek v3 single

  #9849 commented on Jun 15, 2025 • 0 new comments

• opt xpu perf for deepseek

  #9916 commented on Jun 24, 2025 • 0 new comments

• 快速加载

  #9934 commented on Jun 28, 2025 • 0 new comments

• [Feature] Sageattn write 8 bit kv-cache

  #10032 commented on Jun 9, 2025 • 0 new comments

• fix rng_state checkpoint error when tp8

  #10187 commented on Jun 27, 2025 • 0 new comments

• [DCU]support dcu PagedAttention prefix in compute_mla_absorb

  #10229 commented on Jun 15, 2025 • 0 new comments

• [LLM] fix openai client and stream output bug

  #10267 commented on Jun 8, 2025 • 0 new comments

• Dsv3 dev

  #10273 commented on Jul 4, 2025 • 0 new comments

• Bf16 batch gemm dual gemm

  #10281 commented on Jun 9, 2025 • 0 new comments

• update pybind_H

  #10299 commented on Jun 11, 2025 • 0 new comments