Feat: add checkpoint loading mechanism#146
Open
JYMiracle305 wants to merge 5 commits intomasterfrom
Open
Conversation
format: use clang-format-16 instead
remove redundent arguments
format files
JYMiracle305
force-pushed
the
feature/add_checkpoint
branch
from
e8c5dd5 to
0a3deb2
Compare
JYMiracle305
changed the title
chen2021673
reviewed
May 8, 2026
| DEFINE_string(checkpoint_format, "pth", | ||
| "checkpoint format: bin|pth. " | ||
| "'bin' generates model.bin/optimizer.bin (bin supports LLMC model format via callbacks); " | ||
| "'pth' generates model.pth/optimizer.pth (native StateDict binary)."); |
Contributor
There was a problem hiding this comment.
这里 pth 和 bin是不是存的是同一种 native StateDict binary 格式啊，能不能去掉 pth
| std::unordered_map<std::string, std::shared_ptr<Tensor>> Adam::StateDict() const { | ||
| std::unordered_map<std::string, std::shared_ptr<Tensor>> state; | ||
| for (size_t i = 0; i < m_.size(); ++i) { | ||
| state.emplace(std::format("adam.m.{}", i), m_[i]); |
Contributor
There was a problem hiding this comment.
这里m_来自Module::Parameters()，是个vector，不是强保序的，可能有点风险
| return state; | ||
| } | ||
|
|
||
| void Module::LoadStateDict(const std::unordered_map<std::string, std::shared_ptr<Tensor>> &state_dict) { |
Contributor
There was a problem hiding this comment.
这里需不需要检查一下 checkpoint 中是否有多余 key
| @@ -0,0 +1,1060 @@ | |||
| #include "example/common/checkpoint_loader.h" | |||
Contributor
There was a problem hiding this comment.
这个文件太重了，有一千多行，checkpoint相关的基建和 llama / gpt 的 save / load 都混在一起了。要不要拆分一个example/common/checkpoint_utils.h/.cc，然后保留 gpt2 和 llama3 各自的特化调用？这个可以再讨论一下
chen2021673
reviewed
May 8, 2026
|
|
||
| ResumeFromCheckpointResult ResumeFromCheckpoint(const ResumeFromCheckpointArgs &args) { | ||
| ResumeFromCheckpointResult result; | ||
| int ddp_world_size = nn::parallel::global::GetDataParallelSize(); |
Contributor
There was a problem hiding this comment.
这里还需要检查 TP / PP / SP size 吗？
chen2021673
reviewed
May 8, 2026
| #include "gflags/gflags.h" | ||
| #include "gflags/gflags_declare.h" | ||
| #include "glog/logging.h" | ||
| #include "infini_train/include/nn/parallel/global.h" |
Contributor
There was a problem hiding this comment.
这几个 include 还需要吗，是不是可以删掉
| return DataLoaderIterator(*dataset_, batch_size_, max_batch_idx_, max_batch_idx_); | ||
| } | ||
|
|
||
| DataLoaderIterator DataLoader::IteratorAtBatchIndex(size_t batch_idx) const { |
Contributor
There was a problem hiding this comment.
检查一下batch_idx % ddp_world_size == ddp_rank
| int64_t global_step = 0; | ||
| int64_t data_batch_idx = 0; | ||
| int64_t data_batch_stride = 1; | ||
| float best_loss = 0.0f; |
Contributor
There was a problem hiding this comment.
best_loss 不应该是0.0，应该是infinity
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
From ArcaLunar：
Checkpoint 读取工具主要参数：
--checkpoint_dir训练过程中的保存目录--save_steps每 N 次保存一次，设置为 0 则不保存--max_checkpoint_keep最多保留 K 个 checkpoint--save_optimizer_state是否保存优化器的状态--resume_from从指定 checkpoint 目录恢复训练Checkpoint 文件可以通过从
/data/shared/....../llmc/gpt2(orllama3) 的原始模型参数训练而来，例子可见仓库中的 REPORT.md（Experiment 实际上也测试了llama3，但是命令只记录了 GPT2 训练），model.bin, optimizer.bin, trainer_state.json都可以从训练中获取．因此不在附件中提供Experiment
CUDA_VISIBLE_DEVICES=5,6,7 ./gpt2 --input_bin ../../data/llmc/gpt2/tinyshakespeare/tiny_shakespeare_train.bin --llmc_filepath ../../data/llmc/gpt2/gpt2_124M.bin --checkpoint_dir ../ckpt2/gpt2-noresume/ --num_iteration 100 --save_steps 20 --save_optimizer_state true --max_checkpoint_keep 10（以上两条训练命令同样用 llama3 也运行了）
运行 compare_loss.py，对于 llama3 模型，由于从 step 40 恢复训练，所以 step 1~40 数据缺失，而其余 60 步的 loss 在 FP32, BF16 下均吻合
对于 GPT2，模型保存的逻辑有误：训练中 lm_head 与 wte 并非真共享，而 LLMC 存取又按“共享”假设处理，resume 后 lm_head 很容易和 no resume 不一致。解决方法是把训练用 checkpoint 从 LLMC 回调路径切到原生 StateDict 二进制路径，并在加载后显式重建权重绑定语义 (
example/gpt2/main.cc)．经过修复后，也可以通过．