Accelerate x Trainer issue tracker:

[ArthurZucker]

A bunch of issues are a bit stale, and @SunMarc + @muellerzr are a bit short on bandwidth!
Thus we would love to have community support to solve the following:

Help needed

• [2023-12-04 11:52:08,378] [INFO] [autotuner.py:1110:run_after_tuning] No optimal DeepSpeed configuration found by autotuning. #27830
• Training hangs at the first gradient syncing of an MoE model while using deepspeed  #30911
• DDP error with load_best_model_at_end enabled #30702
• Trainer do not move the model to GPU when doing evaluation with FSDP #30239
• Please correct the following DeepSpeed config values that mismatch TrainingArguments values: scheduler.params.total_num_steps=0 vs hf num_training_steps (calculated)= 260 #29348
• Failed to load universal_checkpoint with deepspeed integreation #33157
• dataloader_persistent_workers=True causes fork-bomb due to repeated creation of eval_dataloader #28469
• Question about quantized model with zero3 #30663
• Cannot restore FSDP checkpoint with LOCAL_STATE_DICT #30811
• torchrun breaks with load_model_at_end and with metric_for_best_model=eval_f1 on question_answering example #30819
• Galore finetuning #stopped #31313
• Trainer having issues with DataLoaderShard when running with torchrun #31457
• Cache problem while runing on multiple nodes with GPU #30859
• RuntimeError: Expected a 'mps:0' generator device but found 'cpu' #31897
• [Finetuning OneFormer] Seems not to use multiple GPUs, with both DataParallel and Accelerate #30340
• Load fsdp+lora checkpoint error #31892
• Observed_masks not behaving as expected #28914
• resume_from_checkpoint may still fail with auto_find_batch_size #29518
• Jamba-v01 Model + Deepspeed Zero3 lead to "RuntimeError: Detected mismatch between collectives on ranks." #30277
• Weird text encoder NaNs specifically for FSDP + multi GPU #33376

Feature request

• Support for Multiple Datasets and Domain-Specific Loss Calculation in Trainer #30725

Replied with potential fix and following

• Cannot find the best model after training #31734 followed by @irislin1006
• DeepSpeed ZeRO stage3+Qwen2/Qwen2-57B-A14B-Instruct: RuntimeError: indices should be either on cpu or on the same device as the indexed tensor (cpu) #32312 followed by @irislin1006
• [Bug - I think] data_seed in TrainingArguments is unused #31818 followed by @MekkCyber
• Memory leak when using CLIPTextModel #31439 followed by @nnilayy
• [Trainer.train] learning rate logging inconsistency: learning rate for the future step is logged #28124 followed by @muellerz and @WizKnight
• Issues occuring during parallel evaluation (using Trainer.evaluate) #30767 followed by @SunMarc
• Multi-GPU setup: indices should be either on cpu or on the same device as the indexed tensor (cuda:1) #33147 followed by @SunMarc
• Encounter error when loading checkpoint generated by latest accelerate>=0.34.0 #33400 followed by @SunMarc
• resume_from_checkpoint function fails because "There seems to be not a single sample in your epoch_iterator" #26413 followed by @muupan, @muellerzr and @SunMarc
• It's an AlignModel or Deepspeed Zero3 bug. #28808 followed by @Ben-Schneider-code
• Batch is empty when fine-tuning flan-t5 using LoRA #31357 followed by @MekkCyber
• Trainer doesn't save evaluation metrics. #33733 resolved by the author
• Error on TPU: Invalid --2a886c8_slice_builder_worker_addresses specified. Expected 4 worker addresses, got 1. #30330 followed by @SunMarc
• Title: CUDA RuntimeError: Unspecified Launch Failure during Training #30913 followed by @MekkCyber
• ERROR in run_hp_search_optuna when trying to use multi-GPU #27487 followed by @SunMarc
• Training Resumes with Increased Loss Despite Checkpoint Loading #33336 followed by @MekkCyber
• tensor size mismatch with larger gradient_accumulation_steps and fewer training data #25695 followed by @MekkCyber
• Stuck on Initializing Transformers Model with FSDP (Fully Sharded Data Parallel) using meta device #31278 followed by @muellerzr
• error need either state_dict or a save folder #32427 followed by @muellerzr and @SunMarc
• Using accelerate launch FDSP cause weight saved after 2nd time onwards to be incomplete #31034 followed by @muellerzr
• Resuming from checkpoint runs into OOM #30822 followed by @muellerzr
• Transformer.Trainer fails in creating optimizer for optim adamw_torch_fused when launched with deepspeed. #31867 followed by @Ben-Schneider-code and @SunMarc

[WizKnight]

@ArthurZucker Hey there!👋 I'm new to this repository and excited to learn and contribute. Please let me know if there are any good starting points or tasks where I can be of assistance.

[ArthurZucker]

Any of these issue that have the Good First Issues should be fairly easy! 🤗

[irislin1006]

Hi @ArthurZucker, I'm a first time contributor, but I would love to take issue #31734 as a start 👍

[Update on 202409/07] Handled and replied in the issue

[nnilayy]

Hi there👋 @ArthurZucker, Handled issue #31439, hope that helps🤗.

[WizKnight]

Hi there👋 @ArthurZucker, I'll handle the issue #28124

[irislin1006]

Hi there👋 @ArthurZucker, I would like to take #32312 😀

[Update on 202409/09] Handled and replied in the issue

[SunMarc]

cc @matthewdouglas

[godspeed5]

I had opened PR #31268 as a fix for issue #30819. I think some discussion is needed on there @amyeroberts

[WizKnight]

Hey @amyeroberts, just wanted to check in on issue #28124. It seems like @muellerzr already tackled it with his fix in #30169.
Should I still work on this further, or is it good to go as is?

Thanks!

[amyeroberts]

Hi @WizKnight - best to ask @muellerzr (ideally on the relevant PR / issue to avoid pinging everyone here) on the status of those. I can see in #30169 the PR wasn't merged in due to inactivity -- pending a response to these questions..

In general, if something has just been closed by the github stale bot and not because of a clear decision not to pursue the PR / a clear rejection from the review process you're free to pick up the work :)

[SunMarc]

cc @MekkCyber

[P-Potdar]

Hey @SunMarc and @muellerzr,

I'd love to contribute to this project and help resolve some of the issues mentioned here, especially the DeepSpeed Zero3-related bugs. I’ve already gone through some of the issues and identified potential starting points for solutions. I'll be focusing on these:

Training hangs at the first gradient syncing of an MoE model while using DeepSpeed (#30911)
Trainer doesn't save evaluation metrics (#33733)
CUDA RuntimeError: Unspecified Launch Failure during Training (#30913)
I'll submit PRs with proposed fixes and updates soon. Thank you for the opportunity to contribute!

Also, if there are any specific guidelines or areas where help is most needed, feel free to point me in the right direction!

Looking forward to collaborating on this during Hacktoberfest 🎉

[b423016]

Hey @SunMarc and @muellerzr I would be happy to contribute the issue
Trainer doesn't save evaluation metrics (#33733 )

[ArthurZucker]

Awesome! Just added the tag to make sure it works for everyone! 🥳

[Thejaggeddevil]

i want to work on #29348 please assign this to me

[eeshan15]

hey @ArthurZucker Will this be counted in hacktoberfest?

[ArthurZucker]

Yes, given that there is the tag!
We don't assign issue, first PR that is up will be reviewed, if stale anyone can take it, if no PR is linked, you can also create one 🤗

[naba89]

Hi,
I don't think #28469 has been fixed yet. Facing this even in 4.46.3.

I have a hacky workaround here: custom_trainer
It works on the repro-script, but might not cover all cases.

[SunMarc]

I don't think #28469 has been fixed yet. Facing this even in 4.46.3.

Just reopened ! If you have a fix, would you like to open a PR so that we can have a look ? Thanks !