Skip to content

fix(merge_lora): use CPU-safe precision to avoid false AMP warning #2322

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.

Sign up for GitHub

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

Open
ShamikOfficial wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(merge_lora): use CPU-safe precision to avoid false AMP warning #2322

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 4 additions & 1 deletion litgpt/scripts/merge_lora.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,7 +53,10 @@ def merge_lora(
pretrained_checkpoint_dir = meta_pretrained_checkpoint_dir
pretrained_checkpoint_dir = extend_checkpoint_dir(pretrained_checkpoint_dir)

fabric = L.Fabric(devices=1, precision=precision, accelerator="cpu")
# Merge only cares about weights (dtype is set from the LoRA checkpoint below). Using the
# training precision (often "16-mixed") on CPU makes Fabric warn about unsupported AMP fp16
# and switch to bf16-mixed, which is a false positive for this script (#1242).
fabric = L.Fabric(devices=1, precision="32-true", accelerator="cpu")
config = Config.from_file(checkpoint_dir / "model_config.yaml", **lora_params)

with fabric.init_module(), torch.device("meta"):
Expand Down
32 changes: 32 additions & 0 deletions tests/test_merge_lora.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -74,6 +74,38 @@ def test_merge_lora(tmp_path, fake_checkpoint_dir, pretrained_dtype, lora_dtype)
assert "LoRA weights have already been merged" in stdout.getvalue()


@mock.patch.dict(os.environ, {"LT_ACCELERATOR": "cpu"})
def test_merge_lora_avoids_16_mixed_cpu_warning(tmp_path, fake_checkpoint_dir, caplog):
"""Training precision 16-mixed must not trigger Fabric's CPU AMP warning (#1242)."""
pretrained_checkpoint_dir = tmp_path / "pretrained"
lora_checkpoint_dir = tmp_path / "lora"
shutil.copytree(fake_checkpoint_dir, pretrained_checkpoint_dir)
shutil.copytree(fake_checkpoint_dir, lora_checkpoint_dir)
(lora_checkpoint_dir / "lit_model.pth").unlink()
shutil.rmtree(tmp_path / "checkpoints")

config = dict(block_size=128, padded_vocab_size=256, n_layer=3, n_head=8, n_embd=16)
with open(pretrained_checkpoint_dir / "model_config.yaml", "w", encoding="utf-8") as fp:
yaml.dump(config, fp)
torch.save(GPT.from_name("pythia-14m", **config).state_dict(), pretrained_checkpoint_dir / "lit_model.pth")

lora_kwargs = dict(lora_r=8, lora_alpha=16, lora_dropout=0.05, lora_query=True, lora_value=True)
lora_model = LoRAGPT.from_name("pythia-14m", **config, **lora_kwargs)
state_dict = {k: v for k, v in lora_model.state_dict().items() if lora_filter(k, v)}
torch.save(state_dict, lora_checkpoint_dir / "lit_model.pth.lora")
hparams = dict(checkpoint_dir=str(pretrained_checkpoint_dir), precision="16-mixed", **lora_kwargs)
with open(lora_checkpoint_dir / "hyperparameters.yaml", "w", encoding="utf-8") as file:
yaml.dump(hparams, file)
shutil.copyfile(pretrained_checkpoint_dir / "model_config.yaml", lora_checkpoint_dir / "model_config.yaml")

with caplog.at_level("WARNING"):
merge_lora(lora_checkpoint_dir)

joined = "\n".join(record.getMessage() for record in caplog.records)
assert "AMP with fp16 is not supported on CPU" not in joined
assert (lora_checkpoint_dir / "lit_model.pth").is_file()


def test_load_lora_metadata(fake_checkpoint_dir):
assert not (fake_checkpoint_dir / "hyperparameters.yaml").is_file()
with pytest.raises(FileNotFoundError, match="missing a `hyperparameters.yaml` file"):
Expand Down