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

[3/N][Refactor][Quantization]remove packed_modules_mapping from models (#3021)

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

Some custom models in vllm-ascend define packed_modules_mapping, which
prevent keeping same model class with vllm community. So move these
custom packed_modules_mapping to quant utils.py. After this pr, some
custom models can be removed.

### Does this PR introduce _any_ user-facing change?

tested by CI

### How was this patch tested?

tested by CI

- vLLM version: v0.10.2
- vLLM main:
5089fd749c

Signed-off-by: 22dimensions <waitingwind@foxmail.com>
This commit is contained in:
22dimensions
2025-09-19 20:50:14 +08:00
committed by GitHub
parent 4ba56716f9
commit 0942d9aaab
8 changed files with 76 additions and 80 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
tests/ut/quantization/test_quant_config.py
View File

@@ -73,9 +73,12 @@ class TestAscendQuantConfig(TestBase):
self.assertIsNone(result)
def test_get_quant_method_for_linear(self):
mock_config = MagicMock()
mock_config.model_config.hf_config.model_type = None
linear_layer = MagicMock(spec=LinearBase)
# Test skipped layer
with patch.object(self.ascend_config,
with patch("vllm_ascend.quantization.quant_config.get_current_vllm_config", return_value=mock_config), \
patch.object(self.ascend_config, \
'is_layer_skipped_ascend',
return_value=True):
method = self.ascend_config.get_quant_method(linear_layer, ".attn")
@@ -83,6 +86,7 @@ class TestAscendQuantConfig(TestBase):
# Test quantized layer
with patch.object(self.ascend_config, 'is_layer_skipped_ascend', return_value=False), \
patch("vllm_ascend.quantization.quant_config.get_current_vllm_config", return_value=mock_config), \
patch('vllm_ascend.quantization.quant_config.AscendLinearMethod', return_value=MagicMock()) as mock_ascend_linear:
method = self.ascend_config.get_quant_method(linear_layer, ".attn")
@@ -93,14 +97,18 @@ class TestAscendQuantConfig(TestBase):
def test_get_quant_method_for_attention(self):
attention_layer = MagicMock(spec=Attention)
with patch('vllm_ascend.quantization.quant_config.AscendKVCacheMethod',
mock_config = MagicMock()
mock_config.model_config.hf_config.model_type = None
with patch("vllm_ascend.quantization.quant_config.get_current_vllm_config", return_value=mock_config), \
patch('vllm_ascend.quantization.quant_config.AscendKVCacheMethod', \
return_value=MagicMock()) as mock_ascend_kvcache:
# Test with fa_quant_type
method = self.ascend_config.get_quant_method(
attention_layer, ".attn")
self.assertIs(method, mock_ascend_kvcache.return_value)
with patch('vllm_ascend.quantization.quant_config.AscendKVCacheMethod',
with patch("vllm_ascend.quantization.quant_config.get_current_vllm_config", return_value=mock_config), \
patch('vllm_ascend.quantization.quant_config.AscendKVCacheMethod', \
return_value=MagicMock()) as mock_ascend_kvcache:
# Test with kv_quant_type
modified_config = {"kv_quant_type": "C8"}
@@ -113,9 +121,12 @@ class TestAscendQuantConfig(TestBase):
fused_moe_layer = MagicMock(spec=FusedMoE)
fused_moe_layer.moe = MagicMock(spec=FusedMoEConfig)
fused_moe_layer.moe_config = MagicMock(spec=FusedMoEConfig)
mock_config = MagicMock()
mock_config.model_config.hf_config.model_type = None
# Test skipped layer
with patch.object(self.ascend_config, 'is_layer_skipped_ascend', return_value=True), \
patch("vllm_ascend.quantization.quant_config.get_current_vllm_config", return_value=mock_config), \
patch('vllm_ascend.quantization.quant_config.AscendUnquantizedFusedMoEMethod', return_value=MagicMock()) as mock_ascend_moe:
method = self.ascend_config.get_quant_method(
fused_moe_layer, "moe_layer")
@@ -123,6 +134,7 @@ class TestAscendQuantConfig(TestBase):
# Test quantized layer
with patch.object(self.ascend_config, 'is_layer_skipped_ascend', return_value=False), \
patch("vllm_ascend.quantization.quant_config.get_current_vllm_config", return_value=mock_config), \
patch('vllm_ascend.quantization.quant_config.AscendFusedMoEMethod', return_value=MagicMock()) as mock_ascend_moe:
method = self.ascend_config.get_quant_method(
fused_moe_layer, "moe_layer")