423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2026-01-12 00:50:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

[v1] Refactor kernel plugin (#9669)

Browse Source 
Co-authored-by: frozenleaves <frozen@Mac.local>
This commit is contained in:
浮梦
2025-12-31 18:26:48 +08:00
committed by GitHub
parent 4e1d69579a
commit 16735b9e35
19 changed files with 777 additions and 433 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
tests_v1/plugins/model_plugins/test_kernel_plugin.py
View File

@@ -12,16 +12,13 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import sys
from unittest.mock import MagicMock, patch
import pytest
from transformers import AutoModelForCausalLM
from llamafactory.v1.accelerator.helper import get_current_accelerator
from llamafactory.v1.plugins.model_plugins.kernels.mlp import npu_swiglu
from llamafactory.v1.plugins.model_plugins.kernels.registry import apply_available_kernels, apply_kernel
from llamafactory.v1.plugins.model_plugins.kernels.rms_norm import npu_rms_norm
from llamafactory.v1.plugins.model_plugins.kernels.rope import npu_rope
@pytest.fixture(autouse=True)
@@ -29,24 +26,29 @@ def clear_accelerator_cache():
get_current_accelerator.cache_clear()
def reload_kernels():
"""Helper to reload kernel modules to respect mocked accelerator."""
# Unload kernel interface and registry
keys_to_remove = [k for k in sys.modules if k.startswith("llamafactory.v1.plugins.model_plugins.kernels")]
for k in keys_to_remove:
del sys.modules[k]
@patch("torch.accelerator.current_accelerator")
def test_apply_kernel(mock_get_accelerator: MagicMock):
mock_device = MagicMock()
setattr(mock_device, "type", "npu")
mock_get_accelerator.return_value = mock_device
# Force reload of kernels with mocked accelerator
reload_kernels()
from llamafactory.v1.plugins.model_plugins.kernels.interface import apply_default_kernels
model = AutoModelForCausalLM.from_pretrained("llamafactory/tiny-random-qwen2.5")
original_rmsnorm_forward = model.model.layers[0].input_layernorm.forward
original_swiglu_forward = model.model.layers[0].mlp.forward
apply_kernel(model, npu_rope.NpuRoPEKernel)
model = apply_kernel(model, npu_rms_norm.NpuRMSNormKernel)
assert model.model.layers[0].input_layernorm is not original_rmsnorm_forward
model = apply_kernel(model, npu_swiglu.NpuSwiGluKernel)
assert model.model.layers[0].mlp.forward is not original_swiglu_forward
model = apply_default_kernels(model=model, include_kernels="npu_fused_rmsnorm")
assert model.model.layers[0].input_layernorm.forward.__func__ is not original_rmsnorm_forward.__func__
assert model.model.layers[0].mlp.forward.__func__ is original_swiglu_forward.__func__
@patch("torch.accelerator.current_accelerator")
@@ -56,12 +58,15 @@ def test_apply_all_kernels(mock_get_accelerator: MagicMock):
setattr(mock_device, "type", "npu")
mock_get_accelerator.return_value = mock_device
# Force reload of kernels with mocked accelerator
reload_kernels()
from llamafactory.v1.plugins.model_plugins.kernels.interface import apply_default_kernels
model = AutoModelForCausalLM.from_pretrained("llamafactory/tiny-random-qwen2.5")
original_rmsnorm_forward = model.model.layers[0].input_layernorm.forward
original_swiglu_forward = model.model.layers[0].mlp.forward
model = apply_available_kernels(model)
assert model.model.layers[0].input_layernorm is not original_rmsnorm_forward
assert model.model.layers[0].mlp.forward is not original_swiglu_forward
model = apply_default_kernels(model=model, include_kernels=True)
assert model.model.layers[0].input_layernorm.forward.__func__ is not original_rmsnorm_forward.__func__
assert model.model.layers[0].mlp.forward.__func__ is not original_swiglu_forward.__func__