Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Docs] Add docs/zh_cn/preparation/prompt_template.rst #475

Merged
merged 9 commits into from
Apr 9, 2024

Conversation

LZHgrla
Copy link
Collaborator

@LZHgrla LZHgrla commented Mar 14, 2024

No description provided.

docs/zh_cn/preparation/prompt_template.md Outdated Show resolved Hide resolved
docs/zh_cn/preparation/prompt_template.md Outdated Show resolved Hide resolved

- 微调 base 模型

- 全量微调:任选对话模版,优先使用 chat 版模型所对应的对话模版或默认对话模版 `default`。
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

有些 base 模型的 tokenizer 和 chat 模型不一样,需要说一下怎么解决

@amulil 给点建议?

Copy link
Contributor

@amulil amulil Mar 14, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

目前我是手动解决的,拿 yi-34b-chat 和 yi-34-200k(想让 200k 和 chat 一样使用 chatml 的对话模版) 为例子:

  1. 对比 yi-34b-chat 和 yi-34b-200k 的 tokenizer.json 文件,看词表大小一样不,因为 yi 没提供 json 的格式,需要手动下载一下:
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("01-ai/Yi-34B-Chat")
tokenizer.save_pretrained("01-ai/Yi-34B-Chat")
  1. 对比之后,发现 Yi-34B-Chat 的 bos 和 eos 的 token_id 和 Yi-34B-200k 不一样,这部分以要微调的模型为准,所以保留 Yi-34B-200k 的 bos 和 eos,然后 special_tokens_map.jsontokenizer_config.json 覆盖原始的 Yi-34B-200K 的文件,然后把 eos 和 bos 改成 Yi-34-200k 使用的。
  2. tokenizer.json 里的 added_tokens 字段和 Yi-34B-Chat 同步,对齐 bos 和 eos。
  3. tokenizer.json 还有 normalizer,pre_tokenizer,post_processor, decoder chat 和 base 也会有差别,都以 chat 的配置为准。(这个不知道会对效果有影响不)
  4. 还有就是如果直接用 Yi-34B-200k 使用 xtuner 里的 qwen_chat(chatml) 模版:
    qwen_chat=dict(
        SYSTEM=('<|im_start|>system\n{system}<|im_end|>\n'),
        INSTRUCTION=('<|im_start|>user\n{input}<|im_end|>\n'
                     '<|im_start|>assistant\n'),
        SUFFIX='<|im_end|>',
        SUFFIX_AS_EOS=True,
        SEP='\n',
        STOP_WORDS=['<|im_end|>', '<|endoftext|>']),

SUFFIX_AS_EOS 设置为 False,SUFFIX 还是存在的,数据量少的时候比较难训练,得把 SUFFIX 干掉。
6. 有没有比较方便的自动化更新的脚本,还在思考。
7. 还有就是现在 base 都使用 default 模版是有问题的,因为如果没有加 specical token 的配置,类似 <|system|> 一个 token 会拆成很多个 token,并没有实现对话模版想要有的效果。

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

有些 base 模型的 tokenizer 和 chat 模型不一样,需要说一下怎么解决

@amulil 给点建议?

我有点不明白,在微调 base 和 微调 chat 时候不是应该选择对应的 tokenizer 吗? 会存在跨 tokenizer 使用的场景吗?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

有些 base 模型的 tokenizer 和 chat 模型不一样,需要说一下怎么解决
@amulil 给点建议?

我有点不明白,在微调 base 和 微调 chat 时候不是应该选择对应的 tokenizer 吗? 会存在跨 tokenizer 使用的场景吗?

直接用 base 的 tokenizer 可能会把对话模版的 一个 token 编码成 几个 token,虽然看起来 text 还是一样,但和使用对话模版的初衷不一样了,就得改 base 的 tokenizer 的配置,让 tokenizer 正确编码对话模版。

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

有些 base 模型的 tokenizer 和 chat 模型不一样,需要说一下怎么解决
@amulil 给点建议?

我有点不明白,在微调 base 和 微调 chat 时候不是应该选择对应的 tokenizer 吗? 会存在跨 tokenizer 使用的场景吗?

@hhaAndroid 很多模型的 base tokenizer 中不包含对话模板的 token
https://huggingface.co/internlm/internlm2-1_8b/blob/main/tokenizer_config.json#L19
https://huggingface.co/internlm/internlm2-chat-1_8b/blob/main/tokenizer_config.json#L40-L87

Copy link
Collaborator Author

@LZHgrla LZHgrla Mar 19, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@pppppM @amulil @hhaAndroid

最正确的做法确实是要修改 tokenizer,一种修改方式是直接应用 chat 模型的 tokenizer 到 base 的微调中。

但感觉如果放到 xtuner 的应用场景,会不会把问题搞复杂了?如果用户在微调 base 模型时,期望使用 chat 的模版,那依旧使用 base 的 tokenizer,直接把特殊词当普通 text 来训会不会更合适?
目前想到有两个优势(主要是1)

  1. 普通 text 是有语义的,而 special_tokens 没有。因此,使用普通 text 可以避免在数据量不够或数据质量不高的时候,special_tokens 表现不好的情况。
  2. 从整个流程来看,可以确保 tokenizer、llm 两者与模型开源仓库的对齐,从而避免后续应用时的混用。

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这个事还是要强调一下,实际答疑的过程中,还是因为不知道这件事而用错了 tokenizer 的居多,从 base 训完模型后发现原来chat 模型的工具链出问题了

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@pppppM 增加了一个 小贴士


- LoRA / QLoRA 微调:使用默认对话模版 `default`。这是由于 LoRA / QLoRA 微调默认会关闭 `embed_tokens` 和 `lm_head` 的训练,此时如果引入未学习过的特殊 token(如对话模版中的 `<|im_start|>`),则会影响模型的训练。

- 注:通过修改 `LoraConfig` 可以引入 `embed_tokens` 和 `lm_head` 的训练(会增大显存需求),进而支持任选对话模版
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这里任选对话模版后,可以用 https://github.com/InternLM/xtuner/blob/main/xtuner/tools/log_dataset.py 这个工具打印下
text 和 input_ids 信息,看下任选的对话模版是否编码正确,编码正确,在进行训练

@LZHgrla LZHgrla changed the title [Docs] Add docs/zh_cn/preparation/prompt_template.md [Docs] Add docs/zh_cn/preparation/prompt_template.rst Apr 3, 2024
@pppppM pppppM merged commit 636004d into InternLM:docs Apr 9, 2024
0 of 3 checks passed
pppppM added a commit that referenced this pull request Jun 12, 2024
* [Docs] Readthedocs (#304)

* init readthedocs

* add en docs

* add zh docs

* fix lint

* [Fix] Support ZH Readthedocs (#305)

* add zh yaml

* test zh cn

* test yaml path

* pass

* update conf.py

* [Docs] Document optimization (#362)

Document optimization

* [Docs] Update Docs docs/en/get_started/installation.md  (#364)

* 更新中文 installation.md

完成中文的 安装-安装流程-最佳实践 & 安装-验证安装

* Update installation.md en

* Update installation.md zh

typo

* [Docs] Refine Quick Start (#378)

* [Docs] Add zh_cn quickstart

* [Fix] Fix color rendering logic for github

* [Fix] Fix comments

* [Fix] Add hyperlinks

* [Docs] Add en quickstart

* [Fix] Fix comments

* Update overview.md (#412)

* Update overview.md

* Update overview.md

已根据要求进行修改,请查阅

* Update overview.md

进一步的修正

* Update overview.md

根据要求的完善

* Merge branch 'main' into 'docs' (#463)

* [Improve] Redesign the `prompt_template` (#294)

* update

* update cfgs

* update

* fix bugs

* upload docs

* rename

* update

* Revert "update cfgs"

This reverts commit 93966aa.

* update cfgs

* update

* rename

* rename

* fix bc

* fix stop_word

* fix

* fix

* Update prompt_template.md

* [Fix] Fix errors about `stop_words`  (#313)

* fix bugs

* Update mmbench.py

* [Fix] Fix Mixtral LoRA setting (#312)

set target_modules

* [Feature] Support DeepSeek-MoE (#311)

* support deepseek moe

* update docs

* update

* update

* [Fix] Set `torch.optim.AdamW` as the default optimizer (#318)

fix

* [FIx] Fix `pth_to_hf` for LLaVA model (#316)

Update pth_to_hf.py

* [Improve] Add `demo_data` examples (#278)

* update examples

* add examples

* add json template config

* rename

* update

* update

* update

* [Feature] Support InternLM2 (#321)

* add cfgs

* add internlm2 template

* add dispatch

* add docs

* update readme

* update

* [Fix] Fix the resume of seed (#309)

* fix

* Update utils.py

* [Feature] Accelerate `xtuner xxx`  (#307)

* accelerate cli

* Update entry_point.py

* Update entry_point.py

---------

Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>

* [Fix] Fix InternLM2 url (#325)

* fix

* update

* Update README.md

* Update README_zh-CN.md

* [Fix] Limit the version of python, `>=3.8, <3.11` (#327)

update

* [Fix] Add `trust_remote_code=True` for AutoModel  (#328)

update

* [Docs] Improve README  (#326)

* update

* Update README.md

* Update README.md

* Update README.md

* Update README_zh-CN.md

* update

* update

* fix pre-commit

* update

* bump verion to v0.1.12 (#323)

bump v0.1.12

* set dev version (#329)

Update version.py

* [Docs] Add LLaVA-InternLM2 results (#332)

* update results

* update

* Update internlm2_chat template (#339)

Update internlm2 template

* [Fix] Fix examples demo_data configs (#334)

fix

* bump version to v0.1.13 (#340)

update

* set dev version (#341)

update

* [Feature] More flexible `TrainLoop` (#348)

* add new loop

* rename

* fix pre-commit

* add max_keep_ckpts

* fix

* update cfgs

* update examples

* fix

* update

* update llava

* update

* update

* update

* update

* [Feature]Support CEPH (#266)

* support petrelfs

* fix deepspeed save/load/resume

* add ENV to toggle petrelfs

* support hf save_pretrained

* patch deepspeed engine

* [Improve] Add `--repetition-penalty` for `xtuner chat` (#351)

fix

* [Feature] Support MMBench DDP Evaluate (#300)

* support ddp mmbench evaluate

* Update xtuner/tools/mmbench.py

Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>

* Update xtuner/tools/mmbench.py

Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>

* update minimum version of mmengine

* Update runtime.txt

---------

Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>

* [Fix] `KeyError` of `encode_fn` (#361)

fix

* [Fix] Fix `batch_size` of full fine-tuing LLaVA-InternLM2 (#360)

fix

* [Fix] Remove `system` for `alpaca_map_fn` (#363)

update

* [Fix] Use `DEFAULT_IMAGE_TOKEN` instead of `'<image>'` (#353)

Update utils.py

* [Feature] Efficient SFT (#302)

* add local_attn_args_to_messagehub_hook

* add internlm repo sampler

* add internlm repo dataset and collate_fn

* dispatch internlm1 and internlm2 local attn

* add internlm2 config

* add internlm1 and intenrlm2 config

* add internlm2 template

* fix replace_internlm1_rote bugs

* add internlm1 and internlm2 config templates

* change priority of EvaluateChatHook

* fix docs

* fix config

* fix bug

* set rotary_base according the latest internlm2 config

* add llama local attn

* add llama local attn

* update intern_repo_dataset docs when using aliyun

* support using both hf load_dataset and intern_repo packed_dataset

* add configs

* add opencompass doc

* update opencompass doc

* use T data order

* use T data order

* add config

* add a tool to get data order

* support offline processing untokenized dataset

* add docs

* add doc about only saving model weights

* add doc about only saving model weights

* dispatch mistral

* add mistral template

* add mistral template

* fix torch_dtype

* reset pre-commit-config

* fix config

* fix internlm_7b_full_intern_repo_dataset_template

* update local_attn to varlen_attn

* rename local_attn

* fix InternlmRepoSampler and train.py to support resume

* modify Packer to support varlen attn

* support varlen attn in default pipeline

* update mmengine version requirement to 0.10.3

* Update ceph.md

* delete intern_repo_collate_fn

* delete intern_repo_collate_fn

* delete useless files

* assert pack_to_max_length=True if use_varlen_attn=True

* add varlen attn doc

* add varlen attn to configs

* delete useless codes

* update

* update

* update configs

* fix priority of ThroughputHook and flake8 ignore W504

* using map_fn to set length attr to dataset

* support split=None in process_hf_dataset

* add dataset_format_mapping

* support preprocess ftdp and normal dataset

* refactor process_hf_dataset

* support pack dataset in process_untokenized_datasets

* add xtuner_dataset_timeout

* using gloo backend for monitored barrier

* set gloo timeout

* fix bugs

* fix configs

* refactor intern repo dataset docs

* fix doc

* fix lint

---------

Co-authored-by: pppppM <67539920+pppppM@users.noreply.github.com>
Co-authored-by: pppppM <gjf_mail@126.com>

* [Fix] Add `attention_mask` for `default_collate_fn` (#371)

fix

* [Fix] Update requirements (#369)

Update runtime.txt

* [Fix] Fix rotary_base, add `colors_map_fn` to `DATASET_FORMAT_MAPPING` and rename 'internlm_repo' to 'intern_repo' (#372)

* fix

* rename internlm_repo to intern_repo

* add InternlmRepoSampler for preventing bc break

* add how to install flash_attn to doc

* update (#377)

* Delete useless codes and refactor process_untokenized_datasets (#379)

* delete useless codes

* refactor process_untokenized_datasets: add ftdp to dataset-format

* fix lint

* [Feature] support flash attn 2 in internlm1, internlm2 and llama (#381)

support flash attn 2 in internlm1, internlm2 and llama

* [Fix] Fix installation docs of mmengine in `intern_repo_dataset.md` (#384)

update

* [Fix] Update InternLM2 `apply_rotary_pos_emb` (#383)

update

* [Feature] support saving eval output before save checkpoint (#385)

* support saving eval output before save checkpoint

* refactor

* [Fix] lr scheduler setting (#394)

* fix lr scheduler setting

* fix more

---------

Co-authored-by: zilong.guo <zilong.guo@zeron.ai>
Co-authored-by: LZHgrla <linzhihao@pjlab.org.cn>

* [Fix] Remove pre-defined `system` of `alpaca_zh_map_fn` (#395)

fix

* [Feature] Support `Qwen1.5` (#407)

* rename

* update docs

* update template

* update

* add cfgs

* update

* update

* [Fix] Fix no space in chat output using InternLM2. (#357) (#404)

* [Fix] Fix no space in chat output using InternLM2. (#357)

* Update chat.py

* Update utils.py

* Update utils.py

* fix pre-commit

---------

Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>
Co-authored-by: LZHgrla <linzhihao@pjlab.org.cn>

* [Fix] typo: `--system-prompt` to `--system-template` (#406)

fix

* [Improve] Add `output_with_loss` for dataset process (#408)

update

* [Fix] Fix dispatch to support transformers>=4.36 & Add USE_TRITON_KERNEL environment variable (#411)

* dispatch support transformers>=4.36

* add USE_TRITON_KERNEL environment variable

* raise RuntimeError use triton kernels on cpu

* fix lint

* [Feature]Add InternLM2-1_8b configs (#396)

* [Feature]Add InternLM2-Chat-1_8b full config

* [Feature]Add InternLM2-Chat-1_8b full config

* update

---------

Co-authored-by: LZHgrla <linzhihao@pjlab.org.cn>
Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>

* [Fix] Fix `extract_json_objects` (#419)

* [Fix] Fix pth_to_hf error (#426)

fix

* [Feature] Support `Gemma` (#429)

* added gemma config and template

* check config and make sure the consistancy

* Update xtuner/configs/gemma/gemma_2b_base/gemma_2b_base_qlora_alpaca_e3.py

Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>

* Update xtuner/configs/gemma/gemma_2b_base/gemma_2b_base_full_alpaca_e3.py

Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>

* Update xtuner/configs/gemma/gemma_7b_base/gemma_7b_base_full_alpaca_e3.py

Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>

* Update xtuner/configs/gemma/gemma_7b_base/gemma_7b_base_qlora_alpaca_e3.py

Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>

* Update xtuner/utils/templates.py

Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>

* update

* added  required version

* update

* update

---------

Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>
Co-authored-by: LZHgrla <linzhihao@pjlab.org.cn>

* add refcoco to llava (#425)

* add base dataset

* update dataset generation

* update refcoco

* add convert refcooc

* add eval_refcoco

* add config

* update dataset

* fix bug

* fix bug

* update data prepare

* fix error

* refactor eval_refcoco

* fix bug

* fix error

* update readme

* add entry_point

* update config

* update config

* update entry point

* update

* update doc

* update

---------

Co-authored-by: jacky <jacky@xx.com>

* [Fix] Inconsistent BatchSize of `LengthGroupedSampler` (#436)

update

* bump version to v0.1.14 (#431)

update

* set dev version (#437)

* Update version.py

* Update version.py

* [Bugs] Fix bugs when using EpochBasedRunner (#439)

fix bugs when using epochbasedrunner

* [Feature] Support processing ftdp dataset and custom dataset offline (#410)

* support smart_tokenizer_and_embedding_resize

* replace ast with json.loads

* support list_dataset_format cli

* add doc about ftdp and custom dataset

* add custom dataset template

* add args name to process_hf_dataset

* use new process_untokenized_datasets

* support tokenize_ftdp_datasets

* add mistral_7b_w_tokenized_dataset config

* update doc

* update doc

* add comments

* fix data save path

* smart_tokenizer_and_embedding_resize support zero3

* fix lint

* add data format to internlm2_7b_full_finetune_custom_dataset_e1.py

* add a data format example to configs associated with finetuning custom dataset

* add a data format example to configs associated with finetuning custom dataset

* fix lint

* Update prompt_template.md (#441)

修改了一个错别字

* [Doc] Split finetune_custom_dataset.md to 6 parts (#445)

* split finetune_custom_dataset.md to 6 parts

* refactor custom_dataset and ftdp_dataset related docs

* fix comments

* fix pre-commit

---------

Co-authored-by: pppppM <67539920+pppppM@users.noreply.github.com>
Co-authored-by: RangiLyu <lyuchqi@gmail.com>
Co-authored-by: whcao <41630003+HIT-cwh@users.noreply.github.com>
Co-authored-by: pppppM <gjf_mail@126.com>
Co-authored-by: gzlong96 <30570937+gzlong96@users.noreply.github.com>
Co-authored-by: zilong.guo <zilong.guo@zeron.ai>
Co-authored-by: Ko Sung <34935911+KooSung@users.noreply.github.com>
Co-authored-by: 不要葱姜蒜 <77671993+KMnO4-zx@users.noreply.github.com>
Co-authored-by: fanqiNO1 <75657629+fanqiNO1@users.noreply.github.com>
Co-authored-by: PommesPeter <54879512+PommesPeter@users.noreply.github.com>
Co-authored-by: LKJacky <108643365+LKJacky@users.noreply.github.com>
Co-authored-by: jacky <jacky@xx.com>
Co-authored-by: xzw <62385492+aJupyter@users.noreply.github.com>

* [Docs] Add `docs/zh_cn/preparation/pretrained_model.md` (#462)

* fix pre-commit

* update

* Update pretrained_model.md

* Update pretrained_model.md

* fix pre-commit

* Update pretrained_model.md

* update

* update

* update

* update

* Update pretrained_model.md

* [Docs] Add `docs/zh_cn/training/multi_modal_dataset.md` (#503)

* update

* update

* [Docs] Improve readthedocs style (#545)

* update style

* update style

* fix requirements

* fix

* fix

* add logo

* update

* update

* update

* [Docs] `.md` to `.rst` (#544)

* update rst

* update rst

* update rst

* [Docs] Add `docs/zh_cn/training/custom_pretrain_dataset.rst` (#535)

* update

* update

* update rst

* [Docs] Add docs about training on large scale dataset (#517)

* add train_on_large_scale_dataset doc

* refine doc

* add llava offline doc

* refine doc

* replace md with rst

* refine rst

* refine rst

* [Docs] Add internevo migration related documents (#506)

* add internevo related

* fix comments

* refine doc

* rename internlm2_7b_w_tokenized_dataset.py to internlm2_7b_w_internevo_dataset.py

* refine doc

* replace md with rst

* refine rst

* refine rst

* [Docs] Add `docs/zh_cn/training/modify_settings.rst` (#490)

* update

* update

* update

* update

* update

* update

* Update modify_settings.md

* Update modify_settings.md

* update

* Update docs/zh_cn/training/modify_settings.md

Co-authored-by: Haian Huang(深度眸) <1286304229@qq.com>

* update deepspeed

* update rst

* update rst

---------

Co-authored-by: Haian Huang(深度眸) <1286304229@qq.com>

* [Docs] Add `length_grouped_sampler.rst` (#511)

* update

* update

* update

* Update length_grouped_sampler.md

* update rst

* Update length_grouped_sampler.rst

Co-authored-by: whcao <41630003+HIT-cwh@users.noreply.github.com>

---------

Co-authored-by: whcao <41630003+HIT-cwh@users.noreply.github.com>

* [Docs] Add accelerate related (#504)

* add accelerate related

* split accelerate docs

* fix comments

* add speed benchmark

* explain why qlora can not be used with zero3

* refine doc

* fix configs

* refine doc

* refine doc

* refine configs

* add benchmark to index.rst

* refine doc

* add hyper-param docs

* refine doc

* add explanation about memory cost optimization when using zero

* add figure to show the speed comparison

* refine figures

* refine doc

* fix figures

* refine figures

* update figures and benchmark configs

* add pack rst

* delete pack md

* replace md with rst

* replace md with rst

* replace md with rst

* replace md with rst

* refine rst

* refine rst

* refine rst

* refine rst

* refine rst

* refine rst

* refine rst

* refine rst

* refine rst

* refine rst

* refine rst

* refine rst

* refine rst

* refine rst

---------

Co-authored-by: pppppM <67539920+pppppM@users.noreply.github.com>

* [Docs] Add visualization docs (#516)

* add visualization docs

* delete other visualization tools and add explanation about how to use tensorboard

* replace md with rst

---------

Co-authored-by: pppppM <67539920+pppppM@users.noreply.github.com>

* [Docs] Add docs about SFT with custom dataset (#514)

* add custom sft dataset docs

* add custom dataset template configs

* add openai data format

* refine doc

* update (#2)

* replace md with rst

---------

Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>
Co-authored-by: pppppM <67539920+pppppM@users.noreply.github.com>

* [Docs] Add `docs/zh_cn/training/open_source_dataset.rst` (#502)

* update

* update

* update

* update

* format table

* fix typo

* update rst

---------

Co-authored-by: pppppM <67539920+pppppM@users.noreply.github.com>

* [Docs] Add `docs/zh_cn/preparation/prompt_template.rst` (#475)

* update

* update

* Update prompt_template.md

* Update prompt_template.md

* update

* add tips

* update

* update rst

---------

Co-authored-by: pppppM <67539920+pppppM@users.noreply.github.com>

* [Docs] Add Sequence Parallel documents (#505)

* add sp related

* add sequence parallel supported models

* refine doc

* Update docs/zh_cn/training/training_extreme_long_sequence.md

Co-authored-by: Haian Huang(深度眸) <1286304229@qq.com>

* refine doc

* refine doc

* test the capability boundary of zero3

* refine doc

* test rst

* test rst

* add training speed figure

* delete debug rst

* sp need flash_attn

* WIP

* replace md with rst

* refine rst

* refine rst

* add explanation about why pt 2.1 is not accepted

* refine rst

* refine rst

* add loss curve

---------

Co-authored-by: Haian Huang(深度眸) <1286304229@qq.com>
Co-authored-by: pppppM <67539920+pppppM@users.noreply.github.com>

* [Docs] Update `docs/zh_cn` outline (#556)

update

* [Docs] Update `docs/en` theme (#557)

* update

* update

* update

* update

* update

* update

* update

* update

* [Docs] Add tokenizer to sft in Case 2 (#584)

add tokenizer to sft in Case 2

* [Docs] Improve the Rendering Effect of Readthedocs (#664)

* refine get_start and training

* fix acceleration

* update maxdepth

* refine internevo migration

* refine internevo

* fix typos

* fix lint

---------

Co-authored-by: zhengjie.xu <jerryxuzhengjie@gmail.com>
Co-authored-by: Ma Zhiming <101508488+JimmyMa99@users.noreply.github.com>
Co-authored-by: fanqiNO1 <75657629+fanqiNO1@users.noreply.github.com>
Co-authored-by: Jianfeng777 <108343727+Jianfeng777@users.noreply.github.com>
Co-authored-by: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com>
Co-authored-by: RangiLyu <lyuchqi@gmail.com>
Co-authored-by: whcao <41630003+HIT-cwh@users.noreply.github.com>
Co-authored-by: gzlong96 <30570937+gzlong96@users.noreply.github.com>
Co-authored-by: zilong.guo <zilong.guo@zeron.ai>
Co-authored-by: Ko Sung <34935911+KooSung@users.noreply.github.com>
Co-authored-by: 不要葱姜蒜 <77671993+KMnO4-zx@users.noreply.github.com>
Co-authored-by: PommesPeter <54879512+PommesPeter@users.noreply.github.com>
Co-authored-by: LKJacky <108643365+LKJacky@users.noreply.github.com>
Co-authored-by: jacky <jacky@xx.com>
Co-authored-by: xzw <62385492+aJupyter@users.noreply.github.com>
Co-authored-by: Haian Huang(深度眸) <1286304229@qq.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

None yet

6 participants