1 Release published by 1 person

• v2.6.1

  published Nov 29, 2024

3 Pull requests merged by 2 people

• fix loss func

  #2528 merged Dec 1, 2024

• support glm-edge & glm-edge-v

  #2526 merged Nov 29, 2024

• support qwq

  #2520 merged Nov 27, 2024

7 Issues closed by 7 people

• "Too many 500 error responses" when trying to download model weights

  #2539 closed Dec 4, 2024

• 实际微调Qwen2-vl 所用显存和最佳实践差异显著-导致OOM

  #2536 closed Dec 3, 2024

• qwen2-vl微调报错

  #2529 closed Dec 2, 2024

• Does it support fine-tuning the PaliGemma model for object detection and segmentation?

  #2508 closed Nov 29, 2024

• support for Experiment tracking

  #2523 closed Nov 29, 2024

• qwen2-vl微调报错

  #2524 closed Nov 29, 2024

• glm4v-9b deploy后在输入image时无法通过text指定问题

  #2518 closed Nov 28, 2024

14 Issues opened by 12 people

• 请问支持保存优化器状态吗？

  #2541 opened Dec 4, 2024

• How to add special tokens and finetune a new model?

  #2540 opened Dec 4, 2024

• DPO微调报错，老是出现Storage size calculation overflowed with sizes。

  #2538 opened Dec 3, 2024

• 请问目前使用deepspeed进行 sft 支持 pipe parallel 和 tensor parallel 配置么，是否考虑支持一下？

  #2537 opened Dec 3, 2024

• Qwen2-vl data format

  #2535 opened Dec 2, 2024

• 为什么对自我认知微调后的模型进行推理，得到的推理结果与文档教程不符。

  #2534 opened Dec 2, 2024

• qwen2 vl微调报错

  #2533 opened Dec 2, 2024

• qwen2vl-7B swift infer和swift deploy推理结果不一致

  #2532 opened Dec 2, 2024

• Qwen2-audio-instruct vllm推理request数据结构与vllm所需未对应，推理报错。

  #2531 opened Dec 2, 2024

• qwenvl使用swift命令和sdk推理结果不一致

  #2530 opened Dec 2, 2024

• Animatediff Gradient Accumulation Adjustment on Loss and Learning Rate

  #2527 opened Nov 29, 2024

• any plan to support flash attention 3?

  #2525 opened Nov 28, 2024

• 910B多卡deep speed训练在下载模型时报错，单卡训练就可以。

  #2522 opened Nov 28, 2024

• SFT之后的模型再进行fine tune？

  #2521 opened Nov 28, 2024

12 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.

• Error occurs in lazy tokenize: unsupported operand type(s) for /: 'list' and 'int'

  #2512 commented on Nov 27, 2024 • 0 new comments

• Florence fine tuning

  #2513 commented on Nov 27, 2024 • 0 new comments

• 对qwen2-7b model 进行sft loss 异常

  #2516 commented on Nov 27, 2024 • 0 new comments

• Using multimodal datasets to train ovis1_6-gemma2-9b, an error occurred: RuntimeError: self and mat2 must have the same dtype, but got BFloat16 and Char

  #2514 commented on Nov 28, 2024 • 0 new comments

• Fine-tuning best practices for qwen2.5-72b-instruct and qwen2-vl-72b-instruct.

  #2064 commented on Nov 29, 2024 • 0 new comments

• Converting base models to instruct models

  #2420 commented on Nov 29, 2024 • 0 new comments

• mplug-owl3-7b-chat fine-tuning document

  #1969 commented on Dec 1, 2024 • 0 new comments

• ms-swift==3.0 Suggestion Box

  #2217 commented on Dec 2, 2024 • 0 new comments

• SFT/DPO可以支持自定义和special_tokens和chat_template吗

  #2109 commented on Dec 2, 2024 • 0 new comments

• 请问在微调Qwen2-Audio-Instruct的时候支持对音频文件做数据增强吗？

  #2395 commented on Dec 3, 2024 • 0 new comments

• Best practice for Qwen2-Audio

  #1653 commented on Dec 4, 2024 • 0 new comments

• [WIP]Feat/refactor3

  #2030 commented on Dec 4, 2024 • 0 new comments