{"owner":"PaddlePaddle","repo":"PaddleFormers","hasSkills":true,"totalSkillsCount":4,"totalTokensCount":1029,"categories":["plugin-manifest","anthropic-skill"],"hasMcp":false,"mcpConfig":null,"found":["examples/tools/gpt-oss_weight_change/README.md","examples/experiments/ernie_pretrain/tools/sharded_to_uc/README_zh.md","examples/experiments/ernie_pretrain/tools/uc_to_sharded/README.md","examples/experiments/ernie_pretrain/tools/uc_to_sharded/README_zh.md"],"skills":{"examples/tools/gpt-oss_weight_change/README.md":"## 权重转换描述\n\ngpt-oss 官方权重包含 MXFP4数据类型，在进行后训练时需要先将权重转换为 bf16，可以通过 change_wieght_dtype.py 中的 fp4_to_bf16方法进行转换，同样的训练后得到 bf16权重也可以通过 bf16_to_fp4方法进行转换。\n\n\n示例：\n```python\ntempdir = \"./models/gpt-oss\"\n\n# fp4_to_bf16\nload_path = os.path.join(tempdir, \"gpt-oss-test-fp4\")\nsave_path = os.path.join(tempdir, \"gpt-oss-test-new-bf16\")\nfp4_to_bf16(load_path, save_path)\n\n# bf16_to_fp4\nload_path = os.path.join(tempdir, \"gpt-oss-test-bf16\")\nsave_path = os.path.join(tempdir, \"gpt-oss-test-new-fp4\")\nbf16_to_fp4(load_path, save_path)\n\n```\n### 依赖\nPaddleFormers 跟目录下执行\n```bash\npip install -r requirements.txt\n```\n\n### 使用说明\n\n当前转换脚本只支持单卡转换，按照原始文件个数依次转换。\n例如：\nload_path 下有3个 model-0000x-of-00003.safetensors.\nsave_path 下会得到3个 model-0000x-of-00003.safetensors 和一个 model.safetensors.index.json\n\n修改 change_weight_dtype.py 中的 load_path 和 save_path 后，执行\n\n```bash\npython change_wieght_dtype.py\n```\n即可\n","examples/experiments/ernie_pretrain/tools/sharded_to_uc/README_zh.md":"# 预训练权重转换工具\n这篇文档介绍如何将预训练产出的`sharded model`格式的多机权重转化为单卡视角的`unified checkpoint`格式的权重。当前转化脚本仅针对`ERNIE-4.5-21B-A3B`使用`pure sharding + ep`策略训练产生的权重。\n\n## 进行多机预训练并保存当前模型的 checkpoint\n运行预训练模型，并得到一份当前模型的多机 checkpoint。\n\n## 多机权重聚合\n将多机权重聚合到一台机器上：\n`python tools/sharded_to_uc/gather_all_ckpt.py --org_path <path_to_multi_nodes_weights> --tgt_path <path_to_gathered_ckpt> --hostfile_path <path_to_hostfile>`\n\n## 聚合 sharding、ep 切分的参数\n将 sharding stage1 v2切分的参数进行还原，将 ep 切分的参数进行还原：\n`python tools/sharded_to_uc/merge_sharding_ep.py --base_path <path_to_gathered_ckpt> --output_dir_path <path_to_single_card_ckpt>`\n\n## 将 sharded ckpt 转化为 unified checkpoint\n转化为 unified checkpoint，用来给 post training 使用：\n`python tools/sharded_to_uc/convert_sharded_to_uc.py --sharded_path <path_to_single_card_ckpt> --uc_path <path_to_unified_ckpt>`\n","examples/experiments/ernie_pretrain/tools/uc_to_sharded/README.md":"English | [简体中文](./README_zh.md)\n\n# Pretrained Weight Conversion Tool\n\nThis tool converts our released pretrained weights into formats compatible with current model.\n\n## Download Pretrained Weights\n\nDownload the released weights from our repository. For details.\n\n## Generate Model Checkpoint\n\nStart Pre-Training and save checkpoints at any step.\n\n## Convert Pretrained Weights\n\n`python tools/uc_to_sharded/convert_uc_to_sharded.py --org <path_to_pretrained_weights> --cur <path_to_checkpoint> --dst <path_to_converted_weights>`\n","examples/experiments/ernie_pretrain/tools/uc_to_sharded/README_zh.md":"[English](README.md) | 简体中文\n\n# 预训练权重转换工具\n这篇文档介绍如何将我们发布的预训练权重转换为当前模型可以加载的权重格式。\n\n## 下载预训练权重\n下载已发布的预训练权重。\n\n## 保存当前模型的 checkpoint\n运行预训练模型，并得到一份当前模型的 checkpoint。\n\n## 转换权重\n`python  tools/uc_to_sharded/convert_uc_to_sharded.py --org <path_to_pretrained_weights> --cur <path_to_checkpoint> --dst <path_to_converted_weights>`\n"},"files":{"examples/tools/gpt-oss_weight_change/README.md":"## 权重转换描述\n\ngpt-oss 官方权重包含 MXFP4数据类型，在进行后训练时需要先将权重转换为 bf16，可以通过 change_wieght_dtype.py 中的 fp4_to_bf16方法进行转换，同样的训练后得到 bf16权重也可以通过 bf16_to_fp4方法进行转换。\n\n\n示例：\n```python\ntempdir = \"./models/gpt-oss\"\n\n# fp4_to_bf16\nload_path = os.path.join(tempdir, \"gpt-oss-test-fp4\")\nsave_path = os.path.join(tempdir, \"gpt-oss-test-new-bf16\")\nfp4_to_bf16(load_path, save_path)\n\n# bf16_to_fp4\nload_path = os.path.join(tempdir, \"gpt-oss-test-bf16\")\nsave_path = os.path.join(tempdir, \"gpt-oss-test-new-fp4\")\nbf16_to_fp4(load_path, save_path)\n\n```\n### 依赖\nPaddleFormers 跟目录下执行\n```bash\npip install -r requirements.txt\n```\n\n### 使用说明\n\n当前转换脚本只支持单卡转换，按照原始文件个数依次转换。\n例如：\nload_path 下有3个 model-0000x-of-00003.safetensors.\nsave_path 下会得到3个 model-0000x-of-00003.safetensors 和一个 model.safetensors.index.json\n\n修改 change_weight_dtype.py 中的 load_path 和 save_path 后，执行\n\n```bash\npython change_wieght_dtype.py\n```\n即可\n","examples/experiments/ernie_pretrain/tools/sharded_to_uc/README_zh.md":"# 预训练权重转换工具\n这篇文档介绍如何将预训练产出的`sharded model`格式的多机权重转化为单卡视角的`unified checkpoint`格式的权重。当前转化脚本仅针对`ERNIE-4.5-21B-A3B`使用`pure sharding + ep`策略训练产生的权重。\n\n## 进行多机预训练并保存当前模型的 checkpoint\n运行预训练模型，并得到一份当前模型的多机 checkpoint。\n\n## 多机权重聚合\n将多机权重聚合到一台机器上：\n`python tools/sharded_to_uc/gather_all_ckpt.py --org_path <path_to_multi_nodes_weights> --tgt_path <path_to_gathered_ckpt> --hostfile_path <path_to_hostfile>`\n\n## 聚合 sharding、ep 切分的参数\n将 sharding stage1 v2切分的参数进行还原，将 ep 切分的参数进行还原：\n`python tools/sharded_to_uc/merge_sharding_ep.py --base_path <path_to_gathered_ckpt> --output_dir_path <path_to_single_card_ckpt>`\n\n## 将 sharded ckpt 转化为 unified checkpoint\n转化为 unified checkpoint，用来给 post training 使用：\n`python tools/sharded_to_uc/convert_sharded_to_uc.py --sharded_path <path_to_single_card_ckpt> --uc_path <path_to_unified_ckpt>`\n","examples/experiments/ernie_pretrain/tools/uc_to_sharded/README.md":"English | [简体中文](./README_zh.md)\n\n# Pretrained Weight Conversion Tool\n\nThis tool converts our released pretrained weights into formats compatible with current model.\n\n## Download Pretrained Weights\n\nDownload the released weights from our repository. For details.\n\n## Generate Model Checkpoint\n\nStart Pre-Training and save checkpoints at any step.\n\n## Convert Pretrained Weights\n\n`python tools/uc_to_sharded/convert_uc_to_sharded.py --org <path_to_pretrained_weights> --cur <path_to_checkpoint> --dst <path_to_converted_weights>`\n","examples/experiments/ernie_pretrain/tools/uc_to_sharded/README_zh.md":"[English](README.md) | 简体中文\n\n# 预训练权重转换工具\n这篇文档介绍如何将我们发布的预训练权重转换为当前模型可以加载的权重格式。\n\n## 下载预训练权重\n下载已发布的预训练权重。\n\n## 保存当前模型的 checkpoint\n运行预训练模型，并得到一份当前模型的 checkpoint。\n\n## 转换权重\n`python  tools/uc_to_sharded/convert_uc_to_sharded.py --org <path_to_pretrained_weights> --cur <path_to_checkpoint> --dst <path_to_converted_weights>`\n"},"items":[{"name":"README_zh.md","path":"examples/experiments/ernie_pretrain/tools/sharded_to_uc/README_zh.md","rawUrl":"https://raw.githubusercontent.com/PaddlePaddle/PaddleFormers/HEAD/examples/experiments/ernie_pretrain/tools/sharded_to_uc/README_zh.md","title":"Sharded_to_uc Skill","category":"anthropic-skill","format":"markdown","content":"# 预训练权重转换工具\n这篇文档介绍如何将预训练产出的`sharded model`格式的多机权重转化为单卡视角的`unified checkpoint`格式的权重。当前转化脚本仅针对`ERNIE-4.5-21B-A3B`使用`pure sharding + ep`策略训练产生的权重。\n\n## 进行多机预训练并保存当前模型的 checkpoint\n运行预训练模型，并得到一份当前模型的多机 checkpoint。\n\n## 多机权重聚合\n将多机权重聚合到一台机器上：\n`python tools/sharded_to_uc/gather_all_ckpt.py --org_path <path_to_multi_nodes_weights> --tgt_path <path_to_gathered_ckpt> --hostfile_path <path_to_hostfile>`\n\n## 聚合 sharding、ep 切分的参数\n将 sharding stage1 v2切分的参数进行还原，将 ep 切分的参数进行还原：\n`python tools/sharded_to_uc/merge_sharding_ep.py --base_path <path_to_gathered_ckpt> --output_dir_path <path_to_single_card_ckpt>`\n\n## 将 sharded ckpt 转化为 unified checkpoint\n转化为 unified checkpoint，用来给 post training 使用：\n`python tools/sharded_to_uc/convert_sharded_to_uc.py --sharded_path <path_to_single_card_ckpt> --uc_path <path_to_unified_ckpt>`\n","isInternal":false,"tokens":370,"sizeBytes":1143},{"name":"README_zh.md","path":"examples/experiments/ernie_pretrain/tools/uc_to_sharded/README_zh.md","rawUrl":"https://raw.githubusercontent.com/PaddlePaddle/PaddleFormers/HEAD/examples/experiments/ernie_pretrain/tools/uc_to_sharded/README_zh.md","title":"Uc_to_sharded Skill","category":"anthropic-skill","format":"markdown","content":"[English](README.md) | 简体中文\n\n# 预训练权重转换工具\n这篇文档介绍如何将我们发布的预训练权重转换为当前模型可以加载的权重格式。\n\n## 下载预训练权重\n下载已发布的预训练权重。\n\n## 保存当前模型的 checkpoint\n运行预训练模型，并得到一份当前模型的 checkpoint。\n\n## 转换权重\n`python  tools/uc_to_sharded/convert_uc_to_sharded.py --org <path_to_pretrained_weights> --cur <path_to_checkpoint> --dst <path_to_converted_weights>`\n","isInternal":false,"tokens":165,"sizeBytes":513},{"name":"README.md","path":"examples/experiments/ernie_pretrain/tools/uc_to_sharded/README.md","rawUrl":"https://raw.githubusercontent.com/PaddlePaddle/PaddleFormers/HEAD/examples/experiments/ernie_pretrain/tools/uc_to_sharded/README.md","title":"uc_to_sharded Documentation","category":"plugin-manifest","format":"markdown","content":"English | [简体中文](./README_zh.md)\n\n# Pretrained Weight Conversion Tool\n\nThis tool converts our released pretrained weights into formats compatible with current model.\n\n## Download Pretrained Weights\n\nDownload the released weights from our repository. For details.\n\n## Generate Model Checkpoint\n\nStart Pre-Training and save checkpoints at any step.\n\n## Convert Pretrained Weights\n\n`python tools/uc_to_sharded/convert_uc_to_sharded.py --org <path_to_pretrained_weights> --cur <path_to_checkpoint> --dst <path_to_converted_weights>`\n","isInternal":false,"tokens":117,"sizeBytes":537},{"name":"README.md","path":"examples/tools/gpt-oss_weight_change/README.md","rawUrl":"https://raw.githubusercontent.com/PaddlePaddle/PaddleFormers/HEAD/examples/tools/gpt-oss_weight_change/README.md","title":"gpt-oss_weight_change Documentation","category":"plugin-manifest","format":"markdown","content":"## 权重转换描述\n\ngpt-oss 官方权重包含 MXFP4数据类型，在进行后训练时需要先将权重转换为 bf16，可以通过 change_wieght_dtype.py 中的 fp4_to_bf16方法进行转换，同样的训练后得到 bf16权重也可以通过 bf16_to_fp4方法进行转换。\n\n\n示例：\n```python\ntempdir = \"./models/gpt-oss\"\n\n# fp4_to_bf16\nload_path = os.path.join(tempdir, \"gpt-oss-test-fp4\")\nsave_path = os.path.join(tempdir, \"gpt-oss-test-new-bf16\")\nfp4_to_bf16(load_path, save_path)\n\n# bf16_to_fp4\nload_path = os.path.join(tempdir, \"gpt-oss-test-bf16\")\nsave_path = os.path.join(tempdir, \"gpt-oss-test-new-fp4\")\nbf16_to_fp4(load_path, save_path)\n\n```\n### 依赖\nPaddleFormers 跟目录下执行\n```bash\npip install -r requirements.txt\n```\n\n### 使用说明\n\n当前转换脚本只支持单卡转换，按照原始文件个数依次转换。\n例如：\nload_path 下有3个 model-0000x-of-00003.safetensors.\nsave_path 下会得到3个 model-0000x-of-00003.safetensors 和一个 model.safetensors.index.json\n\n修改 change_weight_dtype.py 中的 load_path 和 save_path 后，执行\n\n```bash\npython change_wieght_dtype.py\n```\n即可\n","isInternal":false,"tokens":377,"sizeBytes":1143}],"systemPromptSnippet":"<yakaai_skills repo=\"PaddlePaddle/PaddleFormers\">\n<!-- File: examples/experiments/ernie_pretrain/tools/sharded_to_uc/README_zh.md (Tokens: ~370 | Category: anthropic-skill) -->\n# 预训练权重转换工具\n这篇文档介绍如何将预训练产出的`sharded model`格式的多机权重转化为单卡视角的`unified checkpoint`格式的权重。当前转化脚本仅针对`ERNIE-4.5-21B-A3B`使用`pure sharding + ep`策略训练产生的权重。\n\n## 进行多机预训练并保存当前模型的 checkpoint\n运行预训练模型，并得到一份当前模型的多机 checkpoint。\n\n## 多机权重聚合\n将多机权重聚合到一台机器上：\n`python tools/sharded_to_uc/gather_all_ckpt.py --org_path <path_to_multi_nodes_weights> --tgt_path <path_to_gathered_ckpt> --hostfile_path <path_to_hostfile>`\n\n## 聚合 sharding、ep 切分的参数\n将 sharding stage1 v2切分的参数进行还原，将 ep 切分的参数进行还原：\n`python tools/sharded_to_uc/merge_sharding_ep.py --base_path <path_to_gathered_ckpt> --output_dir_path <path_to_single_card_ckpt>`\n\n## 将 sharded ckpt 转化为 unified checkpoint\n转化为 unified checkpoint，用来给 post training 使用：\n`python tools/sharded_to_uc/convert_sharded_to_uc.py --sharded_path <path_to_single_card_ckpt> --uc_path <path_to_unified_ckpt>`\n\n<!-- File: examples/experiments/ernie_pretrain/tools/uc_to_sharded/README_zh.md (Tokens: ~165 | Category: anthropic-skill) -->\n[English](README.md) | 简体中文\n\n# 预训练权重转换工具\n这篇文档介绍如何将我们发布的预训练权重转换为当前模型可以加载的权重格式。\n\n## 下载预训练权重\n下载已发布的预训练权重。\n\n## 保存当前模型的 checkpoint\n运行预训练模型，并得到一份当前模型的 checkpoint。\n\n## 转换权重\n`python  tools/uc_to_sharded/convert_uc_to_sharded.py --org <path_to_pretrained_weights> --cur <path_to_checkpoint> --dst <path_to_converted_weights>`\n\n<!-- File: examples/experiments/ernie_pretrain/tools/uc_to_sharded/README.md (Tokens: ~117 | Category: plugin-manifest) -->\nEnglish | [简体中文](./README_zh.md)\n\n# Pretrained Weight Conversion Tool\n\nThis tool converts our released pretrained weights into formats compatible with current model.\n\n## Download Pretrained Weights\n\nDownload the released weights from our repository. For details.\n\n## Generate Model Checkpoint\n\nStart Pre-Training and save checkpoints at any step.\n\n## Convert Pretrained Weights\n\n`python tools/uc_to_sharded/convert_uc_to_sharded.py --org <path_to_pretrained_weights> --cur <path_to_checkpoint> --dst <path_to_converted_weights>`\n\n<!-- File: examples/tools/gpt-oss_weight_change/README.md (Tokens: ~377 | Category: plugin-manifest) -->\n## 权重转换描述\n\ngpt-oss 官方权重包含 MXFP4数据类型，在进行后训练时需要先将权重转换为 bf16，可以通过 change_wieght_dtype.py 中的 fp4_to_bf16方法进行转换，同样的训练后得到 bf16权重也可以通过 bf16_to_fp4方法进行转换。\n\n\n示例：\n```python\ntempdir = \"./models/gpt-oss\"\n\n# fp4_to_bf16\nload_path = os.path.join(tempdir, \"gpt-oss-test-fp4\")\nsave_path = os.path.join(tempdir, \"gpt-oss-test-new-bf16\")\nfp4_to_bf16(load_path, save_path)\n\n# bf16_to_fp4\nload_path = os.path.join(tempdir, \"gpt-oss-test-bf16\")\nsave_path = os.path.join(tempdir, \"gpt-oss-test-new-fp4\")\nbf16_to_fp4(load_path, save_path)\n\n```\n### 依赖\nPaddleFormers 跟目录下执行\n```bash\npip install -r requirements.txt\n```\n\n### 使用说明\n\n当前转换脚本只支持单卡转换，按照原始文件个数依次转换。\n例如：\nload_path 下有3个 model-0000x-of-00003.safetensors.\nsave_path 下会得到3个 model-0000x-of-00003.safetensors 和一个 model.safetensors.index.json\n\n修改 change_weight_dtype.py 中的 load_path 和 save_path 后，执行\n\n```bash\npython change_wieght_dtype.py\n```\n即可\n</yakaai_skills>"}