Loading checkpoint before fabric.setup(model) gets abnormal loss when using fabric.init_module()

[kobenaxie]

from pathlib import Path

import torch
import lightning as L
from lightning.fabric.strategies import FSDPStrategy

from litgpt.args import TrainArgs
from litgpt.config import Config
from litgpt.model import GPT, Block
from litgpt.data import Alpaca2k
from litgpt.tokenizer import Tokenizer
from litgpt.utils import (
    chunked_cross_entropy,
    load_checkpoint,
    num_parameters,
    get_default_supported_precision,
)


def get_lr_scheduler(optimizer, warmup_steps: int, max_steps: int):
    # linear warmup followed by cosine annealing
    scheduler1 = torch.optim.lr_scheduler.LambdaLR(optimizer, lambda step: step / warmup_steps)
    scheduler2 = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=(max_steps - warmup_steps))
    return torch.optim.lr_scheduler.SequentialLR(optimizer, [scheduler1, scheduler2], milestones=[warmup_steps])


def main(
    checkpoint_dir: Path,
    devices: int = 8,
    num_nodes: int = 1,
    precision: str = "bf16-true",
    seed: int = 1337,
) -> None:
    torch.set_float32_matmul_precision("high")

    train_args = TrainArgs(
        save_interval = 1000,
        log_interval = 1,
        global_batch_size = 64,
        micro_batch_size = 4,
        lr_warmup_steps = 1000,
        epochs = 10,
        max_steps = 10000,
    )

    strategy = FSDPStrategy(
        auto_wrap_policy={Block},
        activation_checkpointing_policy={Block},
        state_dict_type="full",
        limit_all_gathers=True,
        cpu_offload=False,
    )
    
    fabric = L.Fabric(
        accelerator="cuda",
        devices=devices,
        num_nodes=num_nodes,
        strategy=strategy,
        precision=precision,
    )
    fabric.launch()
    fabric.seed_everything(seed)  # same seed for every process to init model (FSDP)
    
    dataset = Alpaca2k()
    tokenizer = Tokenizer(str(checkpoint_dir))
    dataset.connect(tokenizer, batch_size=train_args.micro_batch_size, max_seq_length=512)
    with fabric.rank_zero_first():
        dataset.prepare_data()
    dataset.setup()
    dataloader = dataset.train_dataloader()
    dataloader = fabric.setup_dataloaders(dataloader)

    checkpoint_path = str(checkpoint_dir / "lit_model.pth")
    config = Config.from_file(checkpoint_dir / "model_config.yaml")
    with fabric.init_module(empty_init=(fabric.world_size > 1)):
        model = GPT(config)
    fabric.print(f"Number of trainable parameters: {num_parameters(model, requires_grad=True):,}")
    # load_checkpoint(fabric, model, checkpoint_path)
    model = fabric.setup(model)
    load_checkpoint(fabric, model, checkpoint_path)

    optimizer = torch.optim.AdamW(model.parameters(), lr=0.0002, weight_decay=0.0, betas=(0.9, 0.95))
    optimizer = fabric.setup_optimizers(optimizer)
    scheduler = get_lr_scheduler(optimizer, warmup_steps=train_args.lr_warmup_steps, max_steps=train_args.max_steps)

    model.train()
    for epoch in range(train_args.epochs):
        for step, batch in enumerate(dataloader, 1):
            input, target = batch["input_ids"], batch["labels"]
            logits = model(input)
            loss = chunked_cross_entropy(logits[..., :-1, :], target[..., 1:])
            fabric.backward(loss)

            optimizer.step()
            optimizer.zero_grad()
            scheduler.step()
            fabric.print(f"{step = } | loss train: {loss.detach().item()}")


if __name__ == "__main__":
    checkpoint_dir = Path("./Qwen2.5-1.5B/")

    main(checkpoint_dir)

Init model with fabric.init_module(True) and load checkpoint after model = fabric.setup(model), the training loss is normal

with fabric.init_module(empty_init=(fabric.world_size > 1)):
    model = GPT(config)
model = fabric.setup(model)
load_checkpoint(fabric, model, checkpoint_path

step = 1 | loss train: 0.8448048233985901
step = 2 | loss train: 1.3229767084121704
step = 3 | loss train: 1.2647839784622192
step = 4 | loss train: 1.287076711654663
step = 5 | loss train: 1.0357563495635986

but when loading checkpoint before model = fabric.setup(model), get loss

with fabric.init_module(empty_init=(fabric.world_size > 1)):
    model = GPT(config)
load_checkpoint(fabric, model, checkpoint_path
model = fabric.setup(model)

step = 1 | loss train: 12.027938842773438
step = 2 | loss train: 12.051375389099121
step = 3 | loss train: 12.112957954406738
step = 4 | loss train: 12.08558177947998
step = 5 | loss train: 12.089488983154297

Another phenomenon is that, if not using fabric.init_module(), I can get normal loss when loading checkpoint before fabric.setup(model),

# with fabric.init_module(empty_init=(fabric.world_size > 1)):
if True:
    model = GPT(config)
load_checkpoint(fabric, model, checkpoint_path)
model = fabric.setup(model)

step = 1 | loss train: 0.8447667956352234
step = 2 | loss train: 1.3229438066482544
step = 3 | loss train: 1.2663335800170898
step = 4 | loss train: 1.2902932167053223
step = 5 | loss train: 1.035811185836792

So how to load hf models converted by litgpt.scripts.convert_hf_checkpoint in a correct way?

[Andrei-Aksionov]

Hello @kobenaxie

This code addresses the issue of high memory consumption when loading weights into a model.

In the traditional approach, two sets of weights exist simultaneously:

1. Randomly initialized weights when the model is created.
2. Pretrained weights that need to be loaded into the model.

model = GPT(...) # model with random weights
weights = torch.load(...) # pretrained weights
model.load_state_dict(weight)

To mitigate this, the process is split into multiple steps:

1. Model creation on a meta device:
   Using fabric.init_module, the model is created on a meta device. On this device, memory usage is minimal because the weight matrices remain "empty" until explicitly materialized. (Refer to the meta device documentation).

2. Target device setup:
   The fabric.setup(model) call specifies the target device (e.g., GPU) where the model will be placed.

3. Loading pretrained weights:
   Finally, load_checkpoint(fabric, model, checkpoint_path) loads the pretrained weights into the model, materializing it on the target device with minimal memory overhead.

---

but when loading checkpoint before model = fabric.setup(model), get loss

This happens because the model is materialized with random weights, as load_checkpoint was called before fabric.setup for the model on meta device.
load_checkpoint function uses lazy_load from PyTorch that cannot do materialization.

So, when you run fabric.init_module (placing on meta device) and then load_checkpoint, nothing really happens here, the model stays on meta device. And when the model is materialized on the target device, weights values are totally random.

When you commented out fabric.init_module the model was created on a CPU with random weights, then load_checkpoint loaded pretrained weights into it and fabric.setup moved the model to the target device.

The loss value provides a hint.
With a vocabulary size of approximately 151k (for Qwen2.5-1.5B) and randomly initialized weights, the expected loss is around 12.

import torch
import torch.nn.functional as F

batch_size = 32
vocab_size = 151_643
logits = torch.randn(batch_size, vocab_size)
targets = torch.randint(0, 2, (batch_size, vocab_size)).float()

loss_ce = F.cross_entropy(logits, targets.argmax(dim=1))
print(loss_ce)

>> tensor(12.1440)