diff --git a/README.md b/README.md index 1fd4d59f..cb135c25 100644 --- a/README.md +++ b/README.md @@ -10,7 +10,6 @@ [![Twitter](https://img.shields.io/twitter/follow/llamafactory_ai)](https://twitter.com/llamafactory_ai) [![Discord](https://dcbadge.vercel.app/api/server/rKfvV9r9FK?compact=true&style=flat)](https://discord.gg/rKfvV9r9FK) -[![GitCode](https://gitcode.com/zhengyaowei/LLaMA-Factory/star/badge.svg)](https://gitcode.com/zhengyaowei/LLaMA-Factory) [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1eRTPn37ltBbYsISy9Aw2NuI2Aq5CQrD9?usp=sharing) [![Open in DSW](https://gallery.pai-ml.com/assets/open-in-dsw.svg)](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory) @@ -101,7 +100,7 @@ Choose your path: ## Blogs -- [Fine-tune GPT-OSS for Role-Playing using LLaMA-Factory](https://docs.llamafactory.com.cn/docs/documents/best-practice/gptoss/?utm_source=LLaMA-Factory) (Chinese) +- [Fine-tune GPT-OSS for Role-Playing using LLaMA-Factory](https://docs.llamafactory.com.cn/docs/documents/best-practice/gptroleplay/?utm_source=LLaMA-Factory) (Chinese) - [Fine-tune Llama3.1-70B for Medical Diagnosis using LLaMA-Factory](https://docs.alayanew.com/docs/documents/bestPractice/bigModel/llama70B/?utm_source=LLaMA-Factory) (Chinese) - [A One-Stop Code-Free Model Reinforcement Learning and Deployment Platform based on LLaMA-Factory and EasyR1](https://aws.amazon.com/cn/blogs/china/building-llm-model-hub-based-on-llamafactory-and-easyr1/) (Chinese) - [How Apoidea Group enhances visual information extraction from banking documents with multimodal models using LLaMA-Factory on Amazon SageMaker HyperPod](https://aws.amazon.com/cn/blogs/machine-learning/how-apoidea-group-enhances-visual-information-extraction-from-banking-documents-with-multimodal-models-using-llama-factory-on-amazon-sagemaker-hyperpod/) (English) @@ -266,7 +265,7 @@ Choose your path: | [Falcon](https://huggingface.co/tiiuae) | 7B/11B/40B/180B | falcon | | [Falcon-H1](https://huggingface.co/tiiuae) | 0.5B/1.5B/3B/7B/34B | falcon_h1 | | [Gemma/Gemma 2/CodeGemma](https://huggingface.co/google) | 2B/7B/9B/27B | gemma/gemma2 | -| [Gemma 3/Gemma 3n](https://huggingface.co/google) | 1B/4B/6B/8B/12B/27B | gemma3/gemma3n | +| [Gemma 3/Gemma 3n](https://huggingface.co/google) | 270M/1B/4B/6B/8B/12B/27B | gemma3/gemma3n | | [GLM-4/GLM-4-0414/GLM-Z1](https://huggingface.co/zai-org) | 9B/32B | glm4/glmz1 | | [GLM-4.1V](https://huggingface.co/zai-org) | 9B | glm4v | | [GLM-4.5/GLM-4.5V](https://huggingface.co/zai-org)* | 106B/355B | glm4_moe/glm4v_moe | diff --git a/README_zh.md b/README_zh.md index 18a4e74b..6c7480ec 100644 --- a/README_zh.md +++ b/README_zh.md @@ -10,7 +10,6 @@ [![Twitter](https://img.shields.io/twitter/follow/llamafactory_ai)](https://twitter.com/llamafactory_ai) [![Discord](https://dcbadge.vercel.app/api/server/rKfvV9r9FK?compact=true&style=flat)](https://discord.gg/rKfvV9r9FK) -[![GitCode](https://gitcode.com/zhengyaowei/LLaMA-Factory/star/badge.svg)](https://gitcode.com/zhengyaowei/LLaMA-Factory) [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1d5KQtbemerlSDSxZIfAaWXhKr30QypiK?usp=sharing) [![Open in DSW](https://gallery.pai-ml.com/assets/open-in-dsw.svg)](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory) @@ -103,7 +102,7 @@ https://github.com/user-attachments/assets/43b700c6-a178-41db-b1f8-8190a5d3fcfc ## 官方博客 -- [使用 LLaMA-Factory 构建 GPT-OSS 角色扮演模型](https://docs.llamafactory.com.cn/docs/documents/best-practice/gptoss/?utm_source=LLaMA-Factory)（中文） +- [使用 LLaMA-Factory 构建 GPT-OSS 角色扮演模型](https://docs.llamafactory.com.cn/docs/documents/best-practice/gptroleplay/?utm_source=LLaMA-Factory)（中文） - [使用 LLaMA-Factory 微调 Llama3.1-70B 医学诊断模型](https://docs.alayanew.com/docs/documents/bestPractice/bigModel/llama70B/?utm_source=LLaMA-Factory)（中文） - [基于 LLaMA-Factory 和 EasyR1 打造一站式无代码大模型强化学习和部署平台 LLM Model Hub](https://aws.amazon.com/cn/blogs/china/building-llm-model-hub-based-on-llamafactory-and-easyr1/)（中文） - [通过亚马逊 SageMaker HyperPod 上的 LLaMA-Factory 增强多模态模型银行文档的视觉信息提取](https://aws.amazon.com/cn/blogs/machine-learning/how-apoidea-group-enhances-visual-information-extraction-from-banking-documents-with-multimodal-models-using-llama-factory-on-amazon-sagemaker-hyperpod/)（英文） @@ -268,7 +267,7 @@ https://github.com/user-attachments/assets/43b700c6-a178-41db-b1f8-8190a5d3fcfc | [Falcon](https://huggingface.co/tiiuae) | 7B/11B/40B/180B | falcon | | [Falcon-H1](https://huggingface.co/tiiuae) | 0.5B/1.5B/3B/7B/34B | falcon_h1 | | [Gemma/Gemma 2/CodeGemma](https://huggingface.co/google) | 2B/7B/9B/27B | gemma/gemma2 | -| [Gemma 3/Gemma 3n](https://huggingface.co/google) | 1B/4B/6B/8B/12B/27B | gemma3/gemma3n | +| [Gemma 3/Gemma 3n](https://huggingface.co/google) | 270M/1B/4B/6B/8B/12B/27B | gemma3/gemma3n | | [GLM-4/GLM-4-0414/GLM-Z1](https://huggingface.co/zai-org) | 9B/32B | glm4/glmz1 | | [GLM-4.1V](https://huggingface.co/zai-org) | 9B | glm4v | | [GLM-4.5/GLM-4.5V](https://huggingface.co/zai-org)* | 106B/355B | glm4_moe/glm4v_moe | diff --git a/assets/wechat.jpg b/assets/wechat.jpg index 0bcc198d..857b5560 100644 Binary files a/assets/wechat.jpg and b/assets/wechat.jpg differ diff --git a/assets/wechat_npu.jpg b/assets/wechat_npu.jpg index 7cdcba5d..086992b8 100644 Binary files a/assets/wechat_npu.jpg and b/assets/wechat_npu.jpg differ diff --git a/pyproject.toml b/pyproject.toml index c9d753cb..a42faf8a 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -4,11 +4,11 @@ build-backend = "setuptools.build_meta" [project] name = "llamafactory" +requires-python = ">=3.9.0" dynamic = [ "version", "dependencies", "optional-dependencies", - "requires-python", "scripts", "authors", "description", diff --git a/requirements.txt b/requirements.txt index d39d3042..bea4fec2 100644 --- a/requirements.txt +++ b/requirements.txt @@ -6,7 +6,7 @@ peft>=0.14.0,<=0.15.2 trl>=0.8.6,<=0.9.6 tokenizers>=0.19.0,<=0.21.1 # gui -gradio>=4.38.0,<=5.31.0 +gradio>=4.38.0,<=5.42.0 matplotlib>=3.7.0 tyro<0.9.0 # ops diff --git a/src/llamafactory/extras/constants.py b/src/llamafactory/extras/constants.py index 3bec8474..e96f9fa8 100644 --- a/src/llamafactory/extras/constants.py +++ b/src/llamafactory/extras/constants.py @@ -776,6 +776,10 @@ register_model_group( register_model_group( models={ + "Gemma-3-270M": { + DownloadSource.DEFAULT: "google/gemma-3-270m", + DownloadSource.MODELSCOPE: "LLM-Research/gemma-3-270m", + }, "Gemma-3-4B": { DownloadSource.DEFAULT: "google/gemma-3-4b-pt", DownloadSource.MODELSCOPE: "LLM-Research/gemma-3-4b-pt", @@ -788,6 +792,10 @@ register_model_group( DownloadSource.DEFAULT: "google/gemma-3-27b-pt", DownloadSource.MODELSCOPE: "LLM-Research/gemma-3-27b-pt", }, + "Gemma-3-270M-Instruct": { + DownloadSource.DEFAULT: "google/gemma-3-270m-it", + DownloadSource.MODELSCOPE: "LLM-Research/gemma-3-270m-it", + }, "Gemma-3-4B-Instruct": { DownloadSource.DEFAULT: "google/gemma-3-4b-it", DownloadSource.MODELSCOPE: "LLM-Research/gemma-3-4b-it", @@ -1669,8 +1677,8 @@ register_model_group( }, "MiMo-VL-7B-RL-2508": { DownloadSource.DEFAULT: "XiaomiMiMo/MiMo-VL-7B-RL-2508", - DownloadSource.MODELSCOPE: "XiaomiMiMo/MiMo-VL-7B-RL-2508" - } + DownloadSource.MODELSCOPE: "XiaomiMiMo/MiMo-VL-7B-RL-2508", + }, }, template="mimo_vl", multimodal=True, @@ -1685,7 +1693,7 @@ register_model_group( }, "MiMo-VL-7B-SFT-2508": { DownloadSource.DEFAULT: "XiaomiMiMo/MiMo-VL-7B-SFT-2508", - DownloadSource.DEFAULT: "XiaomiMiMo/MiMo-VL-7B-SFT-2508" + DownloadSource.DEFAULT: "XiaomiMiMo/MiMo-VL-7B-SFT-2508", }, }, template="qwen2_vl", diff --git a/src/llamafactory/hparams/parser.py b/src/llamafactory/hparams/parser.py index 0f4e632f..2a2fc2ee 100644 --- a/src/llamafactory/hparams/parser.py +++ b/src/llamafactory/hparams/parser.py @@ -32,6 +32,7 @@ from transformers.utils import is_torch_bf16_gpu_available, is_torch_npu_availab from ..extras import logging from ..extras.constants import CHECKPOINT_NAMES, EngineName from ..extras.misc import check_dependencies, check_version, get_current_device, is_env_enabled +from ..extras.packages import is_transformers_version_greater_than from .data_args import DataArguments from .evaluation_args import EvaluationArguments from .finetuning_args import FinetuningArguments @@ -304,6 +305,9 @@ def get_train_args(args: Optional[Union[dict[str, Any], list[str]]] = None) -> _ if model_args.use_unsloth and is_deepspeed_zero3_enabled(): raise ValueError("Unsloth is incompatible with DeepSpeed ZeRO-3.") + if data_args.neat_packing and is_transformers_version_greater_than("4.53.0"): + raise ValueError("Neat packing is incompatible with transformers>=4.53.0.") + _set_env_vars() _verify_model_args(model_args, data_args, finetuning_args) _check_extra_dependencies(model_args, finetuning_args, training_args) diff --git a/src/llamafactory/model/adapter.py b/src/llamafactory/model/adapter.py index 07573266..d9522d39 100644 --- a/src/llamafactory/model/adapter.py +++ b/src/llamafactory/model/adapter.py @@ -16,11 +16,10 @@ import re from typing import TYPE_CHECKING import torch -from peft import LoraConfig, LoraModel, OFTConfig, OFTModel, PeftModel, TaskType, get_peft_model +from peft import LoraConfig, LoraModel, OFTConfig, PeftModel, TaskType, get_peft_model from transformers.integrations import is_deepspeed_zero3_enabled from ..extras import logging -from ..extras.misc import check_version from .model_utils.misc import find_all_linear_modules, find_expanded_modules from .model_utils.quantization import QuantizationMethod from .model_utils.unsloth import get_unsloth_peft_model, load_unsloth_peft_model diff --git a/src/llamafactory/train/dpo/trainer.py b/src/llamafactory/train/dpo/trainer.py index d1840fae..da593d8d 100644 --- a/src/llamafactory/train/dpo/trainer.py +++ b/src/llamafactory/train/dpo/trainer.py @@ -111,6 +111,7 @@ class CustomDPOTrainer(DPOTrainer): if self.bco_gemma >= 1e-6: from trl.trainer import RunningMoments + self.running = RunningMoments(self.accelerator) @override @@ -161,14 +162,14 @@ class CustomDPOTrainer(DPOTrainer): chosen_logps: "torch.Tensor", rejected_logps: "torch.Tensor", reference_chosen_logps: "torch.Tensor", - reference_rejected_logps: "torch.Tensor" + reference_rejected_logps: "torch.Tensor", ) -> "torch.Tensor": chosen_logratios = chosen_logps - reference_chosen_logps rejected_logratios = rejected_logps - reference_rejected_logps chosen_rewards = self.beta * chosen_logratios rejected_rewards = self.beta * rejected_logratios rewards = torch.cat((chosen_rewards, rejected_rewards), 0).mean().detach() - self.running.update(rewards) # update baseline + self.running.update(rewards) # update baseline delta = self.running.mean bco_loss = -F.logsigmoid((self.beta * chosen_logratios) - delta) - F.logsigmoid( -(self.beta * rejected_logratios - delta) @@ -195,15 +196,12 @@ class CustomDPOTrainer(DPOTrainer): rejected_rewards = self.beta * policy_rejected_logps.to(self.accelerator.device).detach() else: losses, chosen_rewards, rejected_rewards = self.dpo_loss( - policy_chosen_logps, policy_rejected_logps, reference_chosen_logps, reference_rejected_logps - ) + policy_chosen_logps, policy_rejected_logps, reference_chosen_logps, reference_rejected_logps + ) if self.bco_gemma > 1e-6: bco_losses = self.bco_loss( - policy_chosen_logps, - policy_rejected_logps, - reference_chosen_logps, - reference_rejected_logps + policy_chosen_logps, policy_rejected_logps, reference_chosen_logps, reference_rejected_logps ) losses += bco_losses * self.bco_gemma @@ -288,7 +286,7 @@ class CustomDPOTrainer(DPOTrainer): losses += self.ftx_gamma * sft_loss if self.bco_gemma > 1e-6: # re-weigthing for MPO - losses /= (self.ftx_gamma + self.bco_gemma + 1.0) + losses /= self.ftx_gamma + self.bco_gemma + 1.0 prefix = "eval_" if train_eval == "eval" else "" metrics[f"{prefix}rewards/chosen"] = chosen_rewards.mean().item()