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

[refactor] refactor weight trans nz and transpose (#4878)

Browse Source 
### What this PR does / why we need it?

Now `VLLM_ASCEND_ENABLE_NZ` will have three options:
0: disable nz;
1: only quant case enable nz;
2: enable nz as long as possible;

And `VLLM_ASCEND_ENABLE_NZ`=1 by default.

All cases are shown in the table below:

|  | W4A4 | W4A8 | W8A8 | fp16/bf16 | fp32 |
|---|---|---|---|---|---|
| trans nz | can't support nz | trans nz by default | trans nz by
default | trans nz when VLLM_ASCEND_ENABLE_NZ is 2 | can't support nz |
| transpose | only support not transpose case | only support transpose
case | only support transpose case | linear: only support not transpose
case<br>gmm: only support transpose case | same to fp16/bf16 |

Some exceptional cases:
1. MLAPO op need to do some additional processing on the weights,
including trans nz. If use MLAPO op, some weight will be transformed to
nz forcely;
2. MLA/SFA's weight `W_UV` will be used by op
`torch.ops._C_ascend.batch_matmul_transpose`, and this op can't support
nz currently;

### Does this PR introduce _any_ user-facing change?
Now fp16/bf16 weight will not trans nz by default.

### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: zzzzwwjj <1183291235@qq.com>
This commit is contained in:
zzzzwwjj
2025-12-19 14:27:24 +08:00
committed by GitHub
parent ea8f544ce7
commit cc23067f1e
19 changed files with 156 additions and 255 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
tests/ut/quantization/test_w8a8.py
View File

@@ -1,3 +1,4 @@
import os
from unittest.mock import MagicMock, patch
import torch
@@ -132,20 +133,21 @@ class TestAscendW8A8LinearMethod(TestBase):
expected_y_output += bias
self.assertTrue(torch.equal(output, expected_y_output))
@patch("vllm_ascend.quantization.w8a8.is_enable_nz")
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_NZ": "0"})
@patch('torch_npu.npu_format_cast')
def test_process_weights_after_loading_not_nz(self, mock_npu_format_cast,
mock_is_nz):
def test_process_weights_after_loading_with_nz0(self,
mock_npu_format_cast):
layer = MagicMock()
layer.weight.data = torch.randn(128, 256)
layer.weight.data = torch.randint(-127,
128, (128, 256),
dtype=torch.int8)
layer.input_scale.data = torch.tensor([0.1])
layer.input_offset.data = torch.tensor([0])
layer.deq_scale = torch.tensor([0.5])
layer.weight_scale.data = torch.randn(128, 1)
layer.weight_offset.data = torch.randn(128, 1)
mock_is_nz.return_value = 0
mock_npu_format_cast.return_value = MagicMock
self.method.process_weights_after_loading(layer)
@@ -160,20 +162,50 @@ class TestAscendW8A8LinearMethod(TestBase):
self.assertEqual(layer.weight_offset.data.shape, (128, ))
mock_npu_format_cast.assert_not_called()
@patch("vllm_ascend.quantization.w8a8.is_enable_nz")
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_NZ": "1"})
@patch('torch_npu.npu_format_cast')
def test_process_weights_after_loading_nz(self, mock_npu_format_cast,
mock_is_nz):
def test_process_weights_after_loading_with_nz1(self,
mock_npu_format_cast):
layer = MagicMock()
layer.weight.data = torch.randn(128, 256)
layer.weight.data = torch.randint(-127,
128, (128, 256),
dtype=torch.int8)
layer.input_scale.data = torch.tensor([0.1])
layer.input_offset.data = torch.tensor([0])
layer.deq_scale = torch.tensor([0.5])
layer.weight_scale.data = torch.randn(128, 1)
layer.weight_offset.data = torch.randn(128, 1)
mock_npu_format_cast.return_value = MagicMock
self.method.process_weights_after_loading(layer)
expected_offset = torch.tensor([0]).repeat(256).to(torch.int8)
self.assertTrue(
torch.equal(layer.aclnn_input_offset.data, expected_offset))
self.assertFalse(layer.aclnn_input_offset.requires_grad)
self.assertFalse(layer.deq_scale.requires_grad)
self.assertEqual(layer.weight_scale.data.shape, (128, ))
self.assertEqual(layer.weight_offset.data.shape, (128, ))
mock_npu_format_cast.assert_called_once()
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_NZ": "2"})
@patch('torch_npu.npu_format_cast')
def test_process_weights_after_loading_with_nz2(self,
mock_npu_format_cast):
layer = MagicMock()
layer.weight.data = torch.randint(-127,
128, (128, 256),
dtype=torch.int8)
layer.input_scale.data = torch.tensor([0.1])
layer.input_offset.data = torch.tensor([0])
layer.deq_scale = torch.tensor([0.5])
layer.weight_scale.data = torch.randn(128, 1)
layer.weight_offset.data = torch.randn(128, 1)
mock_is_nz.return_value = 1
mock_npu_format_cast.return_value = MagicMock
self.method.process_weights_after_loading(layer)