fbae41697e
### What this PR does / why we need it?
* Refactor the LayerNorm and activation operator classes to decouple the
310P device implementation from the main branch.
* Refactor `mm_encoder_attention` on 310P to use the
`torch_npu._npu_flash_attention_unpad` operator.
* Refactor the QKV inputs in the prefill stage of `attention_v1` on 310P
so they are no longer padded to 16× alignment.
* Refactor `model_runner` on 310P to align the KV-cache initialization
logic with the mainline implementation.
### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?
use the e2e tests.
- vLLM version: v0.13.0
- vLLM main:
d68209402d
---------
Signed-off-by: Tflowers-0129 <2906339855@qq.com>
83 lines
2.9 KiB
Python
83 lines
2.9 KiB
Python
from unittest.mock import MagicMock, patch
|
|
|
|
import pytest
|
|
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 is_310p as is_310p_hw
|
|
|
|
enable_custom_op()
|
|
|
|
|
|
@pytest.fixture
|
|
def dummy_tensor():
|
|
return torch.randn(4, 8, dtype=torch.float16)
|
|
|
|
|
|
def mock_rms_norm(x, weight, eps):
|
|
return x + 1, None
|
|
|
|
|
|
def mock_add_rms_norm(x, residual, weight, eps):
|
|
return 2 * x, None, 2 * residual
|
|
|
|
|
|
def mock_add_rms_norm_bias(x, residual, weight, bias, eps):
|
|
if bias is None:
|
|
return 2 * x, None, 2 * residual
|
|
else:
|
|
return 2 * x + bias, None, 2 * residual
|
|
|
|
|
|
@pytest.fixture(autouse=True)
|
|
def default_vllm_config():
|
|
mock_config = MagicMock()
|
|
mock_config.compilation_config.custom_ops = ["all"]
|
|
|
|
with set_current_vllm_config(mock_config):
|
|
yield mock_config
|
|
|
|
|
|
@pytest.mark.skipif(is_310p_hw(), reason="310P operator classes have already been refactored.")
|
|
@pytest.mark.parametrize("is_310p", [True, False])
|
|
@pytest.mark.parametrize("residual", [None, torch.randn(4, 8, dtype=torch.float32)])
|
|
@patch("torch_npu.npu_rms_norm", side_effect=mock_rms_norm)
|
|
@patch("torch_npu.npu_add_rms_norm", side_effect=mock_add_rms_norm)
|
|
@patch("torch.ops._C_ascend.npu_add_rms_norm_bias", side_effect=mock_add_rms_norm_bias)
|
|
def test_RMSNorm_forward(
|
|
mock_add_rms_norm_bias, mock_add_rmsnorm, mock_rmsnorm, is_310p, residual, dummy_tensor, default_vllm_config
|
|
):
|
|
if is_310p and (not is_310p_hw()):
|
|
pytest.skip("Pseudo-310P branch is invalid on non-310P CI after refactor.")
|
|
|
|
with patch(
|
|
"vllm_ascend.utils.get_ascend_device_type",
|
|
return_value=AscendDeviceType._310P if is_310p else AscendDeviceType.A3,
|
|
):
|
|
layer = RMSNorm(hidden_size=8, eps=1e-05)
|
|
if residual is not None:
|
|
out_x, out_residual = layer.forward_oot(dummy_tensor, residual)
|
|
|
|
if is_310p:
|
|
expected_arg_x = dummy_tensor + residual.to(dummy_tensor.dtype)
|
|
expected_out_x = expected_arg_x + 1
|
|
expected_out_residual = expected_arg_x.to(residual.dtype)
|
|
|
|
mock_rmsnorm.assert_called_once()
|
|
assert torch.allclose(out_x, expected_out_x)
|
|
assert torch.allclose(out_residual, expected_out_residual)
|
|
else:
|
|
expected_out_x = 2 * dummy_tensor
|
|
expected_out_residual = 2 * residual
|
|
mock_add_rms_norm_bias.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)
|