[trainer] fix: reproducible problem when resume training

[wlhgtc]

What does this PR do?

Add concise overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review.

Fix problem in #3457
When resuming training, we didn't calculate the correct epoch number to restart from. The StatefulDataLoader only saves the training step(in a epoch), not the specific step within a particular epoch.

This caused the sampler to always default to epoch 0 upon resuming, which breaks reproducibility.

This PR fixes this by making two changes:

1. Calculating and restoring the correct current epoch number.
2. Restoring the sampler's state correctly.

As mentioned by @wuxibin89 , we should use torchdata.stateful_dataloader.sampler instead of torch.utils.data.RandomSampler

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ……
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ……
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)

[gemini-code-assist]

Code Review

This pull request aims to fix a reproducibility issue when resuming training by correctly restoring the sampler's state. The changes correctly calculate the epoch to resume from and modify the training loop to start at that epoch. However, the implementation for restoring the sampler's state is highly inefficient and contains a correctness bug. My review includes a critical comment with a suggested code change to address both the inefficiency by using set_epoch where available, and to fix an off-by-one error in the loop logic that would cause incorrect data sampling upon resuming.

[gemini-code-assist]

This approach to restoring the sampler state is very inefficient and seems to have a correctness issue.

1. Correctness: The loop for _ in range(current_epoch - 1): appears to have an off-by-one error. If training resumes at a global_step corresponding to the start of current_epoch=1, this loop (range(0)) will not execute. The main training loop will then start at epoch 1, but the sampler will still be in its state for epoch 0. This would lead to incorrect data sampling and break the reproducibility you're trying to fix. To correctly advance the sampler to be ready for current_epoch, the loop should run current_epoch times.

2. Inefficiency: The nested loop iterates through all the samples in the dataset for each epoch you want to skip. For large datasets, this can add a significant delay to your training startup time.

A more robust and efficient solution would be to use the set_epoch() method if the sampler supports it (which is standard for torch.utils.data.distributed.DistributedSampler). This avoids iterating through the dataset entirely.

I suggest replacing this block with a more efficient and correct implementation that handles both points.

Suggested change

	for _ in range(current_epoch - 1):
	for _ in iter(self.train_dataloader.sampler):
	pass
	if hasattr(self.train_dataloader.sampler, "set_epoch"):
	# More efficient and standard way for DistributedSampler
	self.train_dataloader.sampler.set_epoch(current_epoch)
	else:
	# Fallback for samplers without set_epoch, with correction.
	# This is inefficient and should be avoided if possible.
	for _ in range(current_epoch):
	for _ in self.train_dataloader.sampler:
	pass

[wuxibin89]

Please address gemini review, set_epoch is more efficient.

[wlhgtc]

The sampler appears to be one epoch ahead after resuming from _load_checkpoint (e.g., it resumes at step 3 instead of step 1). As a workaround, we only advance the sampler current_epoch - 1 times.

[wlhgtc]

Please address gemini review, set_epoch is more efficient.
@wuxibin89 Sorry for the miss.
For the inefficiency problem, the only sampler that currently matches the condition hasattr(self.train_dataloader.sampler, "set_epoch") is DistributedSampler. We do not use it in create_rl_sampler. This will only have an effect if we call set_epoch() between epochs, which we are not doing currently.

For Correctness Problem, you could see the comment below.

[wuxibin89]

It's torch.utils.data.RandomSampler duty to skip sample consumed in this epoch so far, there's no need to skip it manually.
https://github.com/meta-pytorch/data/blob/main/torchdata/stateful_dataloader/sampler.py#L73-L74

[wlhgtc]

https://github.com/meta-pytorch/data/blob/main/torchdata/stateful_dataloader/sampler.py#L73-L74

Sorry for miss this comment again(magic network).
We used torch.utils.data.sampler.RandomSampler rather than torchdata.stateful_dataloader.sampler.RandomSampler. I am trying the second one, seems it will fit StatefulDataloader

[wlhgtc]

@eric-haibin-lin @vermouth1992 @PeterSH6 @tongyx361 Could anyone help me review this?

[wuxibin89]

From https://docs.pytorch.org/docs/stable/data.html#torch.utils.data.distributed.DistributedSampler

In distributed mode, calling the set_epoch() method at the beginning of each epoch before creating the DataLoader iterator is necessary to make shuffling work properly across multiple epochs. Otherwise, the same ordering will be always used.

I think, we miss set_epoch() here.

[wlhgtc]

It works only when we use DistributedSampler, should I change sampler from RandomSampler/SequentialSampler to DistributedSampler?

[wuxibin89]

No, it works for both RandomSampler/SequentialSampler and DistributedSampler, see comment above.

[wlhgtc]

https://docs.pytorch.org/docs/stable/data.html#torch.utils.data.distributed.DistributedSampler

You could see the code of DistributedSampler; it set generator by g.manual_seed(self.seed + self.epoch)

But in RandomSampler, it only based on seed, this sampler doesn't has the property(self.epoch)

In verl colde: creat_rl_sampler get generator(seed) for data_config.

[wuxibin89]

@wlhgtc I think we should use torchdata.stateful_dataloader.sampler.RandomSampler instead of torch.utils.data.RandomSampler

import os
import torch

from torch.utils.data import TensorDataset
from torchdata.stateful_dataloader import StatefulDataLoader

if os.environ.get("USE_TORCHDATA", "0") == "1":
    from torchdata.stateful_dataloader.sampler import RandomSampler
    print(f"Using torchdata.stateful_dataloader.sampler.RandomSampler")
else:
    from torch.utils.data import RandomSampler
    print(f"Using torch.utils.data.RandomSampler")

# Set seed for reproducibility
torch.manual_seed(42)

dataset = TensorDataset(torch.arange(10))
sampler = RandomSampler(dataset)
dataloader = StatefulDataLoader(dataset, num_workers=2, batch_size=4, drop_last=True, sampler=sampler)

global_step = 0
total_epoch = 3

for epoch in range(total_epoch):
    should_break = False
    for i, batch in enumerate(dataloader):
        global_step += 1
        print(f"{epoch=}, {i=}, {global_step=}: {batch}")

        # Save state_dict at step 0 of epoch 1
        if epoch == 1 and i == 0:
            state_dict = dataloader.state_dict()
            should_break = True
            break

    if should_break:
        break

start_epoch = global_step // len(dataloader)
print(f"recover from epoch {start_epoch}")

del sampler, dataloader

# Training run resumes with the previous checkpoint
sampler2 = RandomSampler(dataset)
dataloader2 = StatefulDataLoader(dataset, num_workers=2, batch_size=4, drop_last=True, sampler=sampler2)

# Resume state with DataLoader
dataloader2.load_state_dict(state_dict)
for epoch in range(start_epoch, total_epoch):
    for i, batch in enumerate(dataloader2):
        global_step += 1
        print(f"{epoch=}, {i=}, {global_step=}: {batch}")

The torch.utils.data.RandomSampler generator is not being recovered properly:

<Trial 66568621 worker_0> open_verl [wuxibin/rl_model_engine] $ python3 test_dataloader.py 
Using torch.utils.data.RandomSampler
epoch=0, i=0, global_step=1: [tensor([8, 0, 9, 1])]
epoch=0, i=1, global_step=2: [tensor([4, 5, 6, 3])]
epoch=1, i=0, global_step=3: [tensor([4, 6, 9, 3])]
recover from epoch 1
epoch=1, i=0, global_step=4: [tensor([3, 8, 5, 9])]
epoch=2, i=0, global_step=5: [tensor([4, 3, 8, 9])]
epoch=2, i=1, global_step=6: [tensor([2, 0, 5, 1])]
<Trial 66568621 worker_0> open_verl [wuxibin/rl_model_engine] $ USE_TORCHDATA=1 python3 test_dataloader.py 
Using torchdata.stateful_dataloader.sampler.RandomSampler
epoch=0, i=0, global_step=1: [tensor([8, 3, 0, 1])]
epoch=0, i=1, global_step=2: [tensor([7, 2, 9, 5])]
epoch=1, i=0, global_step=3: [tensor([1, 8, 2, 7])]
recover from epoch 1
epoch=1, i=0, global_step=4: [tensor([5, 0, 9, 3])]
epoch=2, i=0, global_step=5: [tensor([6, 3, 8, 0])]
epoch=2, i=1, global_step=6: [tensor([1, 5, 4, 7])]

[wlhgtc]

@wlhgtc I think we should use torchdata.stateful_dataloader.sampler.RandomSampler instead of torch.utils.data.RandomSampler

import os
import torch

from torch.utils.data import TensorDataset
from torchdata.stateful_dataloader import StatefulDataLoader

if os.environ.get("USE_TORCHDATA", "0") == "1":
    from torchdata.stateful_dataloader.sampler import RandomSampler
    print(f"Using torchdata.stateful_dataloader.sampler.RandomSampler")
else:
    from torch.utils.data import RandomSampler
    print(f"Using torch.utils.data.RandomSampler")

# Set seed for reproducibility
torch.manual_seed(42)

dataset = TensorDataset(torch.arange(10))
sampler = RandomSampler(dataset)
dataloader = StatefulDataLoader(dataset, num_workers=2, batch_size=4, drop_last=True, sampler=sampler)

global_step = 0
total_epoch = 3

for epoch in range(total_epoch):
    should_break = False
    for i, batch in enumerate(dataloader):
        global_step += 1
        print(f"{epoch=}, {i=}, {global_step=}: {batch}")

        # Save state_dict at step 0 of epoch 1
        if epoch == 1 and i == 0:
            state_dict = dataloader.state_dict()
            should_break = True
            break

    if should_break:
        break

start_epoch = global_step // len(dataloader)
print(f"recover from epoch {start_epoch}")

del sampler, dataloader

# Training run resumes with the previous checkpoint
sampler2 = RandomSampler(dataset)
dataloader2 = StatefulDataLoader(dataset, num_workers=2, batch_size=4, drop_last=True, sampler=sampler2)

# Resume state with DataLoader
dataloader2.load_state_dict(state_dict)
for epoch in range(start_epoch, total_epoch):
    for i, batch in enumerate(dataloader2):
        global_step += 1
        print(f"{epoch=}, {i=}, {global_step=}: {batch}")

The torch.utils.data.RandomSampler generator is not being recovered properly:

<Trial 66568621 worker_0> open_verl [wuxibin/rl_model_engine] $ python3 test_dataloader.py 
Using torch.utils.data.RandomSampler
epoch=0, i=0, global_step=1: [tensor([8, 0, 9, 1])]
epoch=0, i=1, global_step=2: [tensor([4, 5, 6, 3])]
epoch=1, i=0, global_step=3: [tensor([4, 6, 9, 3])]
recover from epoch 1
epoch=1, i=0, global_step=4: [tensor([3, 8, 5, 9])]
epoch=2, i=0, global_step=5: [tensor([4, 3, 8, 9])]
epoch=2, i=1, global_step=6: [tensor([2, 0, 5, 1])]
<Trial 66568621 worker_0> open_verl [wuxibin/rl_model_engine] $ USE_TORCHDATA=1 python3 test_dataloader.py 
Using torchdata.stateful_dataloader.sampler.RandomSampler
epoch=0, i=0, global_step=1: [tensor([8, 3, 0, 1])]
epoch=0, i=1, global_step=2: [tensor([7, 2, 9, 5])]
epoch=1, i=0, global_step=3: [tensor([1, 8, 2, 7])]
recover from epoch 1
epoch=1, i=0, global_step=4: [tensor([5, 0, 9, 3])]
epoch=2, i=0, global_step=5: [tensor([6, 3, 8, 0])]
epoch=2, i=1, global_step=6: [tensor([1, 5, 4, 7])]

Thanks for your advice, I update my code and any other suggestions? @wuxibin89

[wuxibin89]

@wlhgtc Could you please verify the change on your task?

[wlhgtc]

@wlhgtc Could you please verify the change on your task?

It worked.