423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2026-05-05 07:38:55 +08:00
Code Issues Packages Projects Releases Wiki Activity

[v1] fix device_mesh and sp for fsdp2 (#10429)

Browse Source
This commit is contained in:
sunyi0505
2026-04-28 11:20:11 +08:00
committed by GitHub
parent 2f0bef207a
commit 50945ef850
3 changed files with 23 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
tests_v1/plugins/model_plugins/test_ulysses_cp.py
View File

@@ -27,7 +27,9 @@ from llamafactory.v1.utils.env import find_available_port
from llamafactory.v1.utils.pytest import dist_env
def _test_sequence_parallel_loss(local_rank: int, world_size: int, master_port: int, cp_size: int, dp_size: int):
def _test_sequence_parallel_loss(
local_rank: int, world_size: int, master_port: int, cp_size: int, dp_size: int, batch_size: int
):
with dist_env(local_rank, world_size, master_port):
model_args = ModelArguments(model="llamafactory/tiny-random-qwen3")
@@ -41,12 +43,13 @@ def _test_sequence_parallel_loss(local_rank: int, world_size: int, master_port:
# Apply sequence parallel plugin
SequenceParallelModelPlugin(dist_config.get("cp_mode", "ulysses"))(model_engine.model, dist_config)
input_ids = torch.arange(1, batch_size * 5 + 1, dtype=torch.long).view(batch_size, 5)
model_inputs = {
"input_ids": torch.tensor([[1, 2, 3, 4, 5]]),
"labels": torch.tensor([[1, 2, 3, 4, 5]]),
"attention_mask": torch.tensor([[1, 1, 1, 1, 1]]),
"position_ids": torch.tensor([[1, 2, 3, 4, 5]]),
"loss_weights": torch.tensor([[1.0, 1.0, 1.0, 1.0, 1.0]]),
"input_ids": input_ids,
"labels": input_ids.clone(),
"attention_mask": torch.ones_like(input_ids),
"position_ids": torch.arange(1, 6, dtype=torch.long).repeat(batch_size, 1),
"loss_weights": torch.ones(batch_size, 5),
}
loss = sequence_parallel_loss(model_engine.model, model_inputs)
@@ -55,8 +58,10 @@ def _test_sequence_parallel_loss(local_rank: int, world_size: int, master_port:
@pytest.mark.runs_on(["cuda", "npu"])
@pytest.mark.require_distributed(2)
@pytest.mark.parametrize("cp_size, dp_size", [(2, 1)])
def test_sequence_parallel_loss(cp_size, dp_size):
@pytest.mark.parametrize(("cp_size", "dp_size", "batch_size"), [(2, 1, 1), (2, 1, 2)])
def test_sequence_parallel_loss(cp_size, dp_size, batch_size):
master_port = find_available_port()
world_size = cp_size * dp_size
mp.spawn(_test_sequence_parallel_loss, args=(world_size, master_port, cp_size, dp_size), nprocs=world_size)
mp.spawn(
_test_sequence_parallel_loss, args=(world_size, master_port, cp_size, dp_size, batch_size), nprocs=world_size
)