EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Feat.][310P] addrmsnorm for 300I DUO (#6704)

Browse Source 
### What this PR does / why we need it?
This PR integrates the `npu_add_rms_norm` fused kernel for RMSNorm
operations with residual connections on 310P devices. This change
optimizes the computation by replacing a two-step process (manual
residual addition followed by RMSNorm) with a single, more efficient
fused operation. This is needed to improve the performance of models
utilizing RMSNorm with residual connections on the 310P architecture.

Fixes #

### Does this PR introduce _any_ user-facing change?
No, this PR introduces an internal optimization and does not change any
user-facing APIs or behaviors.

### How was this patch tested?
This patch was tested with updated unit tests
(`test_RMSNorm_forward_310p`) that mock the `npu_add_rms_norm` operation
to verify the correctness of the fused kernel integration.

---------

Signed-off-by: Tflowers-0129 <2906339855@qq.com>
This commit is contained in:
Shaoxu Cheng
2026-02-13 15:40:49 +08:00
committed by GitHub
parent 7164990904
commit f40256b697
4 changed files with 12 additions and 80 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
tests/ut/ops/test_layernorm.py
View File

@@ -5,7 +5,7 @@ import torch
from vllm.config import set_current_vllm_config
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm_ascend.utils import AscendDeviceType, enable_custom_op
from vllm_ascend.utils import enable_custom_op
from vllm_ascend.utils import is_310p as is_310p_hw
enable_custom_op()
@@ -39,8 +39,8 @@ def default_vllm_config():
with set_current_vllm_config(mock_config):
yield mock_config
@pytest.mark.skip(
"Skip as register_kernels has NPU SocName checking in CANN 8.5.0.")
@pytest.mark.skip("Skip as register_kernels has NPU SocName checking in CANN 8.5.0.")
@pytest.mark.skipif(is_310p_hw(), reason="non_310P device unittest case.")
@pytest.mark.parametrize("residual", [None, torch.randn(4, 8, dtype=torch.float32)])
@patch("torch_npu.npu_rms_norm", side_effect=mock_rms_norm)
@@ -68,19 +68,18 @@ def test_RMSNorm_forward(
@pytest.mark.skipif(not is_310p_hw(), reason="310P device unittest case.")
@pytest.mark.parametrize("residual", [None, torch.randn(4, 8, dtype=torch.float16)])
@patch("torch_npu.npu_rms_norm", side_effect=mock_rms_norm)
def test_RMSNorm_forward_310p(
mock_rmsnorm, residual, dummy_tensor, default_vllm_config
):
@patch("torch_npu.npu_add_rms_norm", side_effect=mock_add_rms_norm)
def test_RMSNorm_forward_310p(mock_add_rmsnorm, mock_rmsnorm, residual, dummy_tensor, default_vllm_config):
layer = RMSNorm(hidden_size=8, eps=1e-05)
if residual is not None:
out_x, out_residual = layer.forward_oot(dummy_tensor, residual)
expected_out_residual = dummy_tensor + residual
expected_out_x = expected_out_residual + 1
mock_rmsnorm.assert_called_once()
expected_out_x = 2 * dummy_tensor
expected_out_residual = 2 * residual
mock_add_rmsnorm.assert_called_once()
assert torch.allclose(out_x, expected_out_x)
assert torch.allclose(out_residual, expected_out_residual)
else:
out_x = layer.forward_oot(dummy_tensor, residual)
expected_out_x = dummy_tensor + 1
mock_rmsnorm.assert_called_once()
assert torch.allclose(out_x, expected_out_x)
assert torch.allclose(out_x, expected_out_x)