Skip to content

Stabilize Qwen3 MoE LoRA mixed-scale test #1640

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
taivu1998 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Stabilize Qwen3 MoE LoRA mixed-scale test #1640

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
38 changes: 36 additions & 2 deletions tests/tx/models/test_qwen3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,6 +65,32 @@ def load_moe_base_weights(jax_moe_layer: Qwen3MoeSparseMoeBlock, hf_moe_layer: H
)


def assert_allclose_mixed_scale(
actual,
expected,
*,
rtol: float = 1e-3,
base_atol: float = 1e-3,
scale_atol: float = 1e-6,
err_msg: str = "",
) -> None:
actual_arr = np.asarray(actual)
expected_arr = np.asarray(expected)
if not np.isfinite(actual_arr).all() or not np.isfinite(expected_arr).all():
raise AssertionError(f"{err_msg}\nNon-finite values found in MoE LoRA comparison.")

max_abs = max(float(np.max(np.abs(actual_arr))), float(np.max(np.abs(expected_arr))), 1.0)
atol = max(base_atol, scale_atol * max_abs)
np.testing.assert_allclose(
actual_arr,
expected_arr,
rtol=rtol,
atol=atol,
equal_nan=False,
err_msg=f"{err_msg}\nmax_abs={max_abs:.6g}, atol={atol:.6g}",
)


@pytest.mark.parametrize("ep,tp", [(1, 1), (1, 2), (2, 1)])
def test_qwen3_moe_layer(ep: int, tp: int):
model_name = "trl-internal-testing/tiny-Qwen3MoeForCausalLM"
Expand Down Expand Up @@ -130,7 +156,8 @@ def test_qwen3_moe_layer_lora(ep: int, tp: int):
config = Qwen3Config(base_config, max_lora_adapters=3, max_lora_rank=4, shard_attention_heads=True)

hf_moe_layer = hf_model.model.layers[0].mlp
x = torch.randn(3, 4, config.hidden_size)
torch_rng = torch.Generator().manual_seed(0)
x = torch.randn(3, 4, config.hidden_size, generator=torch_rng)

mesh = jax.make_mesh((1, ep, tp), ("fsdp", "ep", "tp"), axis_types=(jax.sharding.AxisType.Auto,) * 3)
with jax.set_mesh(mesh):
Expand Down Expand Up @@ -177,7 +204,14 @@ def test_qwen3_moe_layer_lora(ep: int, tp: int):
x_sample = x[sample_idx : sample_idx + 1].numpy()
output_merged, _ = moe_layer_merged(x_sample, return_router_logits=True)

assert np.allclose(output_with_lora[sample_idx : sample_idx + 1], output_merged, rtol=1e-3, atol=1e-3)
assert_allclose_mixed_scale(
output_with_lora[sample_idx : sample_idx + 1],
output_merged,
rtol=1e-3,
base_atol=1e-3,
scale_atol=1e-6,
err_msg=f"MoE LoRA merged-weight mismatch for sample_idx={sample_idx}, adapter_idx={adapter_idx}",
)


def test_qwen3_lora():
Expand Down
Loading