2025-08-06 10:17:44 +08:00
|
|
|
from unittest.mock import Mock, patch
|
|
|
|
|
|
2025-07-30 14:57:14 +08:00
|
|
|
import torch
|
|
|
|
|
|
|
|
|
|
from tests.ut.base import TestBase
|
[Refactor] Quantization Module Refactor (#5738)
### Summary
This PR refactors the `vllm_ascend/quantization` module to improve code
organization, maintainability, and extensibility. The refactoring
introduces a clear separation of concerns with a registry-based scheme
discovery pattern, abstract base classes for quantization schemes, and
dedicated wrapper classes.
### Key Changes
#### 1. **Modular Directory Structure**
| Before | After |
|--------|-------|
| Flat file structure with mixed responsibilities | Organized into
`methods/` subpackage for schemes |
| Single `quant_config.py` (600+ lines) | Separate config files:
`modelslim_config.py`, `compressed_tensors_config.py` |
| `utils.py` with scheme lookup logic | `methods/registry.py` with
decorator-based registration |
#### 2. **Registry-Based Scheme Discovery**
Replaced hardcoded `ASCEND_QUANTIZATION_METHOD_MAP` dictionary with a
decorator-based registry pattern:
```python
# Before: Manual dictionary mapping
ASCEND_QUANTIZATION_METHOD_MAP = {
"W8A8_DYNAMIC": {"linear": AscendW8A8DynamicLinearMethod, ...},
...
}
# After: Decorator-based registration
@register_scheme("W8A8_DYNAMIC", "linear")
class AscendW8A8DynamicLinearMethod(AscendLinearScheme):
...
```
#### 3. **Abstract Base Classes**
Introduced three abstract base classes in `methods/base.py`:
- `AscendLinearScheme` - Base for linear layer quantization
- `AscendMoEScheme` - Base for MoE layer quantization
- `AscendAttentionScheme` - Base for attention layer quantization
#### 4. **Separated Config and Wrapper Classes**
- **Config classes** (`AscendModelSlimConfig`,
`AscendCompressedTensorsConfig`): Handle config parsing and scheme
selection
- **Wrapper classes** (`AscendLinearMethod`, `AscendFusedMoEMethod`,
etc.): Implement vLLM interfaces and delegate to schemes
#### 5. **Cleaner Public API**
```python
# New clean module interface
from vllm_ascend.quantization import (
AscendModelSlimConfig,
AscendCompressedTensorsConfig,
)
from vllm_ascend.quantization.methods import get_scheme_class
```
### Architecture Diagram
```mermaid
classDiagram
direction TB
class QuantizationConfig {
<<vLLM Interface>>
+get_quant_method()
}
class AscendModelSlimConfig {
+quant_description
+get_quant_method()
-create_scheme_for_layer()
}
class AscendCompressedTensorsConfig {
+target_scheme_map
+get_quant_method()
-_get_scheme_from_parts()
}
class AscendLinearMethod {
<<Wrapper>>
+quant_method: AscendLinearScheme
+create_weights()
+apply()
}
class AscendFusedMoEMethod {
<<Wrapper>>
+quant_method: AscendMoEScheme
+create_weights()
+apply()
}
class AscendLinearScheme {
<<Abstract>>
+get_weight()*
+apply()*
+get_pertensor_param()
+get_perchannel_param()
}
class AscendMoEScheme {
<<Abstract>>
+get_weight()*
+get_dynamic_quant_param()*
+apply()*
}
class W8A8DynamicLinear {
+get_weight()
+apply()
}
class W8A8DynamicMoE {
+get_weight()
+apply()
}
QuantizationConfig <|-- AscendModelSlimConfig
QuantizationConfig <|-- AscendCompressedTensorsConfig
AscendModelSlimConfig ..> AscendLinearMethod : creates
AscendModelSlimConfig ..> AscendFusedMoEMethod : creates
AscendCompressedTensorsConfig ..> AscendLinearMethod : creates
AscendCompressedTensorsConfig ..> AscendFusedMoEMethod : creates
AscendLinearMethod o-- AscendLinearScheme : delegates to
AscendFusedMoEMethod o-- AscendMoEScheme : delegates to
AscendLinearScheme <|-- W8A8DynamicLinear
AscendMoEScheme <|-- W8A8DynamicMoE
```
### Scheme Registration Flow
```mermaid
sequenceDiagram
participant Module as Scheme Module
participant Registry as _SCHEME_REGISTRY
participant Config as QuantConfig
participant Wrapper as Wrapper Class
Note over Module: At import time
Module->>Registry: @register_scheme("W8A8_DYNAMIC", "linear")
Registry->>Registry: Store (quant_type, layer_type) -> Class
Note over Config: At runtime
Config->>Config: Determine quant_type from description
Config->>Registry: get_scheme_class(quant_type, layer_type)
Registry-->>Config: Return scheme class
Config->>Config: scheme = scheme_cls()
Config->>Wrapper: Create wrapper with scheme
Wrapper-->>Config: Return wrapper instance
```
### File Changes Summary
| Original Files | Refactored Files |
|----------------|------------------|
| `__init__.py` (empty) | `__init__.py` (exports public API) |
| `quant_config.py` | `modelslim_config.py` + `wrappers.py` |
| `compressed_tensors/` | `compressed_tensors_config.py` |
| `utils.py` | `methods/registry.py` |
| `w8a8_dynamic.py` | `methods/w8a8_dynamic.py` |
| `w8a8.py` | `methods/w8a8_static.py` |
| `w4a4_flatquant_dynamic.py` | `methods/w4a4_flatquant.py` |
| ... | `methods/base.py` (new) |
### Benefits
1. **Extensibility**: Adding new quantization schemes only requires
implementing the base class and adding `@register_scheme` decorator
2. **Maintainability**: Clear separation between config parsing, wrapper
logic, and scheme implementation
3. **Testability**: Abstract base classes enable easier unit testing and
mocking
4. **Discoverability**: Registry pattern makes it easy to list all
supported schemes
5. **Reduced Coupling**: Config classes no longer need to know about all
scheme implementations
___
- vLLM version: v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/2f4e6548efec402b913ffddc8726230d9311948d
---------
Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
2026-01-23 14:13:47 +08:00
|
|
|
from vllm_ascend.quantization.methods.w4a8 import (
|
2025-08-06 10:17:44 +08:00
|
|
|
AscendW4A8DynamicFusedMoEMethod, AscendW4A8DynamicLinearMethod)
|
2025-07-30 14:57:14 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
class TestAscendW4A8DynamicLinearMethod(TestBase):
|
|
|
|
|
|
2025-10-21 20:18:39 +08:00
|
|
|
@patch('vllm.distributed.get_tensor_model_parallel_world_size')
|
[Refactor] Quantization Module Refactor (#5738)
### Summary
This PR refactors the `vllm_ascend/quantization` module to improve code
organization, maintainability, and extensibility. The refactoring
introduces a clear separation of concerns with a registry-based scheme
discovery pattern, abstract base classes for quantization schemes, and
dedicated wrapper classes.
### Key Changes
#### 1. **Modular Directory Structure**
| Before | After |
|--------|-------|
| Flat file structure with mixed responsibilities | Organized into
`methods/` subpackage for schemes |
| Single `quant_config.py` (600+ lines) | Separate config files:
`modelslim_config.py`, `compressed_tensors_config.py` |
| `utils.py` with scheme lookup logic | `methods/registry.py` with
decorator-based registration |
#### 2. **Registry-Based Scheme Discovery**
Replaced hardcoded `ASCEND_QUANTIZATION_METHOD_MAP` dictionary with a
decorator-based registry pattern:
```python
# Before: Manual dictionary mapping
ASCEND_QUANTIZATION_METHOD_MAP = {
"W8A8_DYNAMIC": {"linear": AscendW8A8DynamicLinearMethod, ...},
...
}
# After: Decorator-based registration
@register_scheme("W8A8_DYNAMIC", "linear")
class AscendW8A8DynamicLinearMethod(AscendLinearScheme):
...
```
#### 3. **Abstract Base Classes**
Introduced three abstract base classes in `methods/base.py`:
- `AscendLinearScheme` - Base for linear layer quantization
- `AscendMoEScheme` - Base for MoE layer quantization
- `AscendAttentionScheme` - Base for attention layer quantization
#### 4. **Separated Config and Wrapper Classes**
- **Config classes** (`AscendModelSlimConfig`,
`AscendCompressedTensorsConfig`): Handle config parsing and scheme
selection
- **Wrapper classes** (`AscendLinearMethod`, `AscendFusedMoEMethod`,
etc.): Implement vLLM interfaces and delegate to schemes
#### 5. **Cleaner Public API**
```python
# New clean module interface
from vllm_ascend.quantization import (
AscendModelSlimConfig,
AscendCompressedTensorsConfig,
)
from vllm_ascend.quantization.methods import get_scheme_class
```
### Architecture Diagram
```mermaid
classDiagram
direction TB
class QuantizationConfig {
<<vLLM Interface>>
+get_quant_method()
}
class AscendModelSlimConfig {
+quant_description
+get_quant_method()
-create_scheme_for_layer()
}
class AscendCompressedTensorsConfig {
+target_scheme_map
+get_quant_method()
-_get_scheme_from_parts()
}
class AscendLinearMethod {
<<Wrapper>>
+quant_method: AscendLinearScheme
+create_weights()
+apply()
}
class AscendFusedMoEMethod {
<<Wrapper>>
+quant_method: AscendMoEScheme
+create_weights()
+apply()
}
class AscendLinearScheme {
<<Abstract>>
+get_weight()*
+apply()*
+get_pertensor_param()
+get_perchannel_param()
}
class AscendMoEScheme {
<<Abstract>>
+get_weight()*
+get_dynamic_quant_param()*
+apply()*
}
class W8A8DynamicLinear {
+get_weight()
+apply()
}
class W8A8DynamicMoE {
+get_weight()
+apply()
}
QuantizationConfig <|-- AscendModelSlimConfig
QuantizationConfig <|-- AscendCompressedTensorsConfig
AscendModelSlimConfig ..> AscendLinearMethod : creates
AscendModelSlimConfig ..> AscendFusedMoEMethod : creates
AscendCompressedTensorsConfig ..> AscendLinearMethod : creates
AscendCompressedTensorsConfig ..> AscendFusedMoEMethod : creates
AscendLinearMethod o-- AscendLinearScheme : delegates to
AscendFusedMoEMethod o-- AscendMoEScheme : delegates to
AscendLinearScheme <|-- W8A8DynamicLinear
AscendMoEScheme <|-- W8A8DynamicMoE
```
### Scheme Registration Flow
```mermaid
sequenceDiagram
participant Module as Scheme Module
participant Registry as _SCHEME_REGISTRY
participant Config as QuantConfig
participant Wrapper as Wrapper Class
Note over Module: At import time
Module->>Registry: @register_scheme("W8A8_DYNAMIC", "linear")
Registry->>Registry: Store (quant_type, layer_type) -> Class
Note over Config: At runtime
Config->>Config: Determine quant_type from description
Config->>Registry: get_scheme_class(quant_type, layer_type)
Registry-->>Config: Return scheme class
Config->>Config: scheme = scheme_cls()
Config->>Wrapper: Create wrapper with scheme
Wrapper-->>Config: Return wrapper instance
```
### File Changes Summary
| Original Files | Refactored Files |
|----------------|------------------|
| `__init__.py` (empty) | `__init__.py` (exports public API) |
| `quant_config.py` | `modelslim_config.py` + `wrappers.py` |
| `compressed_tensors/` | `compressed_tensors_config.py` |
| `utils.py` | `methods/registry.py` |
| `w8a8_dynamic.py` | `methods/w8a8_dynamic.py` |
| `w8a8.py` | `methods/w8a8_static.py` |
| `w4a4_flatquant_dynamic.py` | `methods/w4a4_flatquant.py` |
| ... | `methods/base.py` (new) |
### Benefits
1. **Extensibility**: Adding new quantization schemes only requires
implementing the base class and adding `@register_scheme` decorator
2. **Maintainability**: Clear separation between config parsing, wrapper
logic, and scheme implementation
3. **Testability**: Abstract base classes enable easier unit testing and
mocking
4. **Discoverability**: Registry pattern makes it easy to list all
supported schemes
5. **Reduced Coupling**: Config classes no longer need to know about all
scheme implementations
___
- vLLM version: v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/2f4e6548efec402b913ffddc8726230d9311948d
---------
Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
2026-01-23 14:13:47 +08:00
|
|
|
@patch('vllm_ascend.quantization.methods.w4a8.get_current_vllm_config')
|
2025-10-21 20:18:39 +08:00
|
|
|
def setUp(self, mock_get_current_vllm_config, mock_get_tp_world_size):
|
|
|
|
|
mock_get_tp_world_size.return_value = 1
|
|
|
|
|
mock_vllm_config = Mock()
|
|
|
|
|
mock_vllm_config.quant_config = Mock(
|
|
|
|
|
quant_description={"group_size": 256})
|
|
|
|
|
mock_vllm_config.scheduler_config = Mock(max_num_batched_tokens=2048,
|
|
|
|
|
max_model_len=2048,
|
|
|
|
|
enable_chunked_prefill=False)
|
|
|
|
|
mock_get_current_vllm_config.return_value = mock_vllm_config
|
|
|
|
|
self.method = AscendW4A8DynamicLinearMethod()
|
|
|
|
|
self.method.group_size = 8
|
2025-07-30 14:57:14 +08:00
|
|
|
|
|
|
|
|
def test_get_weight(self):
|
|
|
|
|
weight = self.method.get_weight(8, 32, torch.bfloat16)
|
|
|
|
|
self.assertEqual(weight["weight"].dtype, torch.int8)
|
|
|
|
|
self.assertEqual(weight["weight"].shape, (32, 8))
|
2025-10-21 20:18:39 +08:00
|
|
|
# new quant version weight
|
|
|
|
|
self.method.new_quant_version = True
|
|
|
|
|
weight = self.method.get_weight(8, 32, torch.bfloat16)
|
|
|
|
|
self.assertEqual(weight["weight"].dtype, torch.int8)
|
|
|
|
|
self.assertEqual(weight["weight"].shape, (16, 8))
|
|
|
|
|
self.assertEqual(weight["_packed_dim"], 0)
|
|
|
|
|
self.assertEqual(weight["_packed_factor"], 2)
|
2025-07-30 14:57:14 +08:00
|
|
|
|
|
|
|
|
def test_get_pergroup_param(self):
|
|
|
|
|
params = self.method.get_pergroup_param(8, 32, torch.bfloat16)
|
|
|
|
|
self.assertEqual(params["weight_scale"].dtype, torch.bfloat16)
|
|
|
|
|
self.assertEqual(params["weight_scale"].shape, (32, 1))
|
|
|
|
|
self.assertEqual(params["weight_offset"].dtype, torch.bfloat16)
|
|
|
|
|
self.assertEqual(params["weight_offset"].shape, (32, 1))
|
|
|
|
|
self.assertEqual(params["weight_scale_second"].dtype, torch.bfloat16)
|
|
|
|
|
self.assertEqual(params["weight_scale_second"].shape, (32, 1))
|
|
|
|
|
self.assertEqual(params["weight_offset_second"].dtype, torch.bfloat16)
|
|
|
|
|
self.assertEqual(params["weight_offset_second"].shape, (32, 1))
|
2025-10-21 20:18:39 +08:00
|
|
|
# new quant version weight
|
|
|
|
|
self.method.new_quant_version = True
|
|
|
|
|
params = self.method.get_pergroup_param(8,
|
|
|
|
|
32,
|
|
|
|
|
torch.bfloat16,
|
|
|
|
|
layer_type="column")
|
|
|
|
|
self.assertEqual(params["scale_bias"].dtype, torch.float32)
|
|
|
|
|
self.assertEqual(params["scale_bias"].shape, (32, 1))
|
|
|
|
|
params = self.method.get_pergroup_param(8,
|
|
|
|
|
32,
|
|
|
|
|
torch.bfloat16,
|
|
|
|
|
layer_type="row")
|
|
|
|
|
self.assertEqual(params["scale_bias"].dtype, torch.float32)
|
|
|
|
|
self.assertEqual(params["scale_bias"].shape, (32, 16))
|
|
|
|
|
|
|
|
|
|
@patch('torch_npu.npu_convert_weight_to_int4pack')
|
|
|
|
|
@patch('torch.Tensor.npu')
|
2025-12-19 14:27:24 +08:00
|
|
|
@patch("torch_npu.npu_format_cast")
|
|
|
|
|
def test_process_weights_after_loading(self, mock_format_cast, mock_npu,
|
2025-10-21 20:18:39 +08:00
|
|
|
mock_npu_convert_weight):
|
|
|
|
|
mock_npu.side_effect = lambda: torch.zeros(
|
|
|
|
|
(1, 32), dtype=torch.float32)
|
|
|
|
|
mock_npu_convert_weight.return_value = torch.zeros((32, 4),
|
|
|
|
|
dtype=torch.int32)
|
|
|
|
|
# old quant version weight
|
|
|
|
|
layer = torch.nn.Module()
|
|
|
|
|
layer.weight = torch.nn.Parameter(torch.zeros((32, 8),
|
|
|
|
|
dtype=torch.int8),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
layer.weight_scale = torch.nn.Parameter(torch.ones(
|
|
|
|
|
(32, 1), dtype=torch.float32),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
layer.weight_offset = torch.nn.Parameter(torch.empty_like(
|
|
|
|
|
layer.weight_scale.data),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
layer.weight_scale_second = torch.nn.Parameter(torch.ones(
|
|
|
|
|
(32, 1), dtype=torch.float32),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
layer.weight_offset_second = torch.nn.Parameter(torch.empty_like(
|
|
|
|
|
layer.weight_scale_second.data),
|
|
|
|
|
requires_grad=False)
|
2025-12-19 14:27:24 +08:00
|
|
|
mock_format_cast.return_value = layer.weight.data.transpose(
|
|
|
|
|
0, 1).contiguous()
|
2025-10-21 20:18:39 +08:00
|
|
|
self.method.process_weights_after_loading(layer)
|
|
|
|
|
self.assertTrue(hasattr(layer, "weight_scale_bias"))
|
|
|
|
|
self.assertEqual(layer.weight_scale_bias.data.shape, (32, ))
|
|
|
|
|
self.assertEqual(layer.weight_scale_bias.data.dtype, torch.float32)
|
|
|
|
|
# new quant version weight
|
|
|
|
|
self.method.new_quant_version = True
|
|
|
|
|
new_layer = torch.nn.Module()
|
|
|
|
|
new_layer.weight = torch.nn.Parameter(torch.zeros((16, 8),
|
|
|
|
|
dtype=torch.int8),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
new_layer.weight_scale = torch.nn.Parameter(torch.ones(
|
|
|
|
|
(32, 1), dtype=torch.float32),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
new_layer.weight_offset = torch.nn.Parameter(torch.empty_like(
|
|
|
|
|
new_layer.weight_scale.data),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
new_layer.weight_scale_second = torch.nn.Parameter(torch.ones(
|
|
|
|
|
(32, 1), dtype=torch.float32),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
new_layer.weight_offset_second = torch.nn.Parameter(
|
|
|
|
|
torch.empty_like(new_layer.weight_scale_second.data),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
new_layer.scale_bias = torch.nn.Parameter(torch.zeros(
|
|
|
|
|
(32, 1), dtype=torch.float32),
|
|
|
|
|
requires_grad=False)
|
2025-12-19 14:27:24 +08:00
|
|
|
mock_format_cast.return_value = new_layer.weight.data.transpose(
|
|
|
|
|
0, 1).contiguous()
|
2025-10-21 20:18:39 +08:00
|
|
|
self.method.process_weights_after_loading(new_layer)
|
|
|
|
|
self.assertEqual(new_layer.scale_bias.data.shape, (32, ))
|
|
|
|
|
self.assertTrue(hasattr(new_layer, "weight_scale_second"))
|
|
|
|
|
self.assertEqual(new_layer.weight_scale_second.data.shape, (1, 32))
|
2025-08-06 10:17:44 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
class TestAscendW4A8DynamicFusedMoEMethod(TestBase):
|
2025-08-20 20:25:18 +08:00
|
|
|
experts = 8
|
|
|
|
|
input_size = 16
|
|
|
|
|
output_size = 56
|
|
|
|
|
group_size = 2
|
2025-08-06 10:17:44 +08:00
|
|
|
|
[Refactor] Quantization Module Refactor (#5738)
### Summary
This PR refactors the `vllm_ascend/quantization` module to improve code
organization, maintainability, and extensibility. The refactoring
introduces a clear separation of concerns with a registry-based scheme
discovery pattern, abstract base classes for quantization schemes, and
dedicated wrapper classes.
### Key Changes
#### 1. **Modular Directory Structure**
| Before | After |
|--------|-------|
| Flat file structure with mixed responsibilities | Organized into
`methods/` subpackage for schemes |
| Single `quant_config.py` (600+ lines) | Separate config files:
`modelslim_config.py`, `compressed_tensors_config.py` |
| `utils.py` with scheme lookup logic | `methods/registry.py` with
decorator-based registration |
#### 2. **Registry-Based Scheme Discovery**
Replaced hardcoded `ASCEND_QUANTIZATION_METHOD_MAP` dictionary with a
decorator-based registry pattern:
```python
# Before: Manual dictionary mapping
ASCEND_QUANTIZATION_METHOD_MAP = {
"W8A8_DYNAMIC": {"linear": AscendW8A8DynamicLinearMethod, ...},
...
}
# After: Decorator-based registration
@register_scheme("W8A8_DYNAMIC", "linear")
class AscendW8A8DynamicLinearMethod(AscendLinearScheme):
...
```
#### 3. **Abstract Base Classes**
Introduced three abstract base classes in `methods/base.py`:
- `AscendLinearScheme` - Base for linear layer quantization
- `AscendMoEScheme` - Base for MoE layer quantization
- `AscendAttentionScheme` - Base for attention layer quantization
#### 4. **Separated Config and Wrapper Classes**
- **Config classes** (`AscendModelSlimConfig`,
`AscendCompressedTensorsConfig`): Handle config parsing and scheme
selection
- **Wrapper classes** (`AscendLinearMethod`, `AscendFusedMoEMethod`,
etc.): Implement vLLM interfaces and delegate to schemes
#### 5. **Cleaner Public API**
```python
# New clean module interface
from vllm_ascend.quantization import (
AscendModelSlimConfig,
AscendCompressedTensorsConfig,
)
from vllm_ascend.quantization.methods import get_scheme_class
```
### Architecture Diagram
```mermaid
classDiagram
direction TB
class QuantizationConfig {
<<vLLM Interface>>
+get_quant_method()
}
class AscendModelSlimConfig {
+quant_description
+get_quant_method()
-create_scheme_for_layer()
}
class AscendCompressedTensorsConfig {
+target_scheme_map
+get_quant_method()
-_get_scheme_from_parts()
}
class AscendLinearMethod {
<<Wrapper>>
+quant_method: AscendLinearScheme
+create_weights()
+apply()
}
class AscendFusedMoEMethod {
<<Wrapper>>
+quant_method: AscendMoEScheme
+create_weights()
+apply()
}
class AscendLinearScheme {
<<Abstract>>
+get_weight()*
+apply()*
+get_pertensor_param()
+get_perchannel_param()
}
class AscendMoEScheme {
<<Abstract>>
+get_weight()*
+get_dynamic_quant_param()*
+apply()*
}
class W8A8DynamicLinear {
+get_weight()
+apply()
}
class W8A8DynamicMoE {
+get_weight()
+apply()
}
QuantizationConfig <|-- AscendModelSlimConfig
QuantizationConfig <|-- AscendCompressedTensorsConfig
AscendModelSlimConfig ..> AscendLinearMethod : creates
AscendModelSlimConfig ..> AscendFusedMoEMethod : creates
AscendCompressedTensorsConfig ..> AscendLinearMethod : creates
AscendCompressedTensorsConfig ..> AscendFusedMoEMethod : creates
AscendLinearMethod o-- AscendLinearScheme : delegates to
AscendFusedMoEMethod o-- AscendMoEScheme : delegates to
AscendLinearScheme <|-- W8A8DynamicLinear
AscendMoEScheme <|-- W8A8DynamicMoE
```
### Scheme Registration Flow
```mermaid
sequenceDiagram
participant Module as Scheme Module
participant Registry as _SCHEME_REGISTRY
participant Config as QuantConfig
participant Wrapper as Wrapper Class
Note over Module: At import time
Module->>Registry: @register_scheme("W8A8_DYNAMIC", "linear")
Registry->>Registry: Store (quant_type, layer_type) -> Class
Note over Config: At runtime
Config->>Config: Determine quant_type from description
Config->>Registry: get_scheme_class(quant_type, layer_type)
Registry-->>Config: Return scheme class
Config->>Config: scheme = scheme_cls()
Config->>Wrapper: Create wrapper with scheme
Wrapper-->>Config: Return wrapper instance
```
### File Changes Summary
| Original Files | Refactored Files |
|----------------|------------------|
| `__init__.py` (empty) | `__init__.py` (exports public API) |
| `quant_config.py` | `modelslim_config.py` + `wrappers.py` |
| `compressed_tensors/` | `compressed_tensors_config.py` |
| `utils.py` | `methods/registry.py` |
| `w8a8_dynamic.py` | `methods/w8a8_dynamic.py` |
| `w8a8.py` | `methods/w8a8_static.py` |
| `w4a4_flatquant_dynamic.py` | `methods/w4a4_flatquant.py` |
| ... | `methods/base.py` (new) |
### Benefits
1. **Extensibility**: Adding new quantization schemes only requires
implementing the base class and adding `@register_scheme` decorator
2. **Maintainability**: Clear separation between config parsing, wrapper
logic, and scheme implementation
3. **Testability**: Abstract base classes enable easier unit testing and
mocking
4. **Discoverability**: Registry pattern makes it easy to list all
supported schemes
5. **Reduced Coupling**: Config classes no longer need to know about all
scheme implementations
___
- vLLM version: v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/2f4e6548efec402b913ffddc8726230d9311948d
---------
Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
2026-01-23 14:13:47 +08:00
|
|
|
@patch('vllm_ascend.quantization.methods.w4a8.get_ascend_config')
|
|
|
|
|
@patch('vllm_ascend.quantization.methods.w4a8.get_current_vllm_config')
|
|
|
|
|
@patch('vllm_ascend.quantization.methods.w4a8.get_ep_group')
|
|
|
|
|
@patch('vllm_ascend.quantization.methods.w4a8.get_mc2_group')
|
2025-08-06 10:17:44 +08:00
|
|
|
@patch('torch.distributed.get_rank', return_value=0)
|
2025-08-28 09:10:03 +08:00
|
|
|
def setUp(self, mock_get_rank, mock_get_mc2_group, mock_get_ep_group,
|
2025-09-22 19:12:58 +08:00
|
|
|
get_current_vllm_config, mock_get_ascend_config):
|
|
|
|
|
# Mock ascend config
|
|
|
|
|
mock_ascend_config = Mock()
|
2026-01-15 10:26:44 +08:00
|
|
|
mock_ascend_config.eplb_config.dynamic_eplb = False
|
2025-09-22 19:12:58 +08:00
|
|
|
mock_get_ascend_config.return_value = mock_ascend_config
|
|
|
|
|
|
2025-08-20 20:25:18 +08:00
|
|
|
mock_vllm_config = Mock()
|
|
|
|
|
mock_vllm_config.quant_config = Mock(quant_description={
|
|
|
|
|
"group_size": self.group_size,
|
|
|
|
|
"version": "0.0.0"
|
|
|
|
|
})
|
|
|
|
|
mock_vllm_config.parallel_config = Mock(enable_expert_parallel=True)
|
2025-09-22 19:12:58 +08:00
|
|
|
mock_vllm_config.scheduler_config = Mock(max_num_batched_tokens=2048,
|
|
|
|
|
max_model_len=2048,
|
|
|
|
|
enable_chunked_prefill=False)
|
2025-08-20 20:25:18 +08:00
|
|
|
get_current_vllm_config.return_value = mock_vllm_config
|
2025-08-06 10:17:44 +08:00
|
|
|
self.quant_method = AscendW4A8DynamicFusedMoEMethod()
|
|
|
|
|
|
|
|
|
|
def test_get_weight(self):
|
2025-08-20 20:25:18 +08:00
|
|
|
# old quant version w4a8 weight
|
|
|
|
|
param_dict = self.quant_method.get_weight(self.experts,
|
|
|
|
|
self.input_size,
|
|
|
|
|
self.output_size,
|
|
|
|
|
torch.bfloat16)
|
|
|
|
|
self.assertEqual(param_dict["w13_weight"].dtype, torch.int8)
|
|
|
|
|
self.assertEqual(param_dict["w13_weight"].shape,
|
|
|
|
|
(self.experts, 2 * self.input_size, self.output_size))
|
|
|
|
|
# new quant version weight
|
|
|
|
|
self.quant_method.new_quant_version = True
|
|
|
|
|
param_dict = self.quant_method.get_weight(self.experts,
|
|
|
|
|
self.input_size,
|
|
|
|
|
self.output_size,
|
|
|
|
|
torch.bfloat16)
|
2025-08-06 10:17:44 +08:00
|
|
|
self.assertEqual(param_dict["w13_weight"].dtype, torch.int8)
|
2025-08-20 20:25:18 +08:00
|
|
|
self.assertEqual(param_dict["w13_weight"].shape,
|
|
|
|
|
(self.experts, self.input_size, self.output_size))
|
2025-08-06 10:17:44 +08:00
|
|
|
|
2025-08-20 20:25:18 +08:00
|
|
|
def test_get_dynamic_quant_param(self):
|
|
|
|
|
# old quant version weight
|
2025-08-06 10:17:44 +08:00
|
|
|
param_dict = self.quant_method.get_dynamic_quant_param(
|
2025-08-20 20:25:18 +08:00
|
|
|
self.experts, self.input_size, self.output_size, torch.bfloat16)
|
2025-09-27 21:01:16 +08:00
|
|
|
self.assertEqual(param_dict["w13_weight_scale"].dtype, torch.float32)
|
2025-08-20 20:25:18 +08:00
|
|
|
self.assertEqual(param_dict["w13_weight_scale"].shape,
|
|
|
|
|
(self.experts, 2 * self.input_size, 1))
|
2025-08-06 10:17:44 +08:00
|
|
|
self.assertEqual(param_dict["w13_weight_scale_second"].dtype,
|
2025-09-27 21:01:16 +08:00
|
|
|
torch.float32)
|
2025-08-06 10:17:44 +08:00
|
|
|
self.assertEqual(param_dict["w13_weight_scale_second"].shape,
|
2025-08-20 20:25:18 +08:00
|
|
|
(self.experts, 2 * self.input_size,
|
|
|
|
|
self.output_size // self.group_size))
|
2025-09-27 21:01:16 +08:00
|
|
|
self.assertEqual(param_dict["w2_weight_scale"].dtype, torch.float32)
|
2025-08-20 20:25:18 +08:00
|
|
|
self.assertEqual(param_dict["w2_weight_scale"].shape,
|
|
|
|
|
(self.experts, self.output_size, 1))
|
2025-08-06 10:17:44 +08:00
|
|
|
self.assertEqual(param_dict["w2_weight_scale_second"].dtype,
|
2025-09-27 21:01:16 +08:00
|
|
|
torch.float32)
|
2025-08-06 10:17:44 +08:00
|
|
|
self.assertEqual(param_dict["w2_weight_scale_second"].shape,
|
2025-08-20 20:25:18 +08:00
|
|
|
(self.experts, self.output_size,
|
|
|
|
|
self.input_size // self.group_size))
|
|
|
|
|
# new quant version weight
|
|
|
|
|
self.quant_method.new_quant_version = True
|
|
|
|
|
param_dict = self.quant_method.get_dynamic_quant_param(
|
|
|
|
|
self.experts, self.input_size, self.output_size, torch.bfloat16)
|
|
|
|
|
self.assertEqual(param_dict["w2_scale_bias"].dtype, torch.float32)
|
|
|
|
|
self.assertEqual(
|
|
|
|
|
param_dict["w2_scale_bias"].shape,
|
|
|
|
|
(self.experts, self.output_size, 16 // self.quant_method.tp_size))
|
2025-09-27 21:01:16 +08:00
|
|
|
# per-channel weight
|
|
|
|
|
self.quant_method.is_per_channel_weight = True
|
|
|
|
|
param_dict = self.quant_method.get_dynamic_quant_param(
|
|
|
|
|
self.experts, self.input_size, self.output_size, torch.bfloat16)
|
|
|
|
|
pergroup_param = [
|
|
|
|
|
"w13_weight_scale_second", "w13_weight_offset_second",
|
|
|
|
|
"w2_weight_scale_second", "w2_weight_offset_second"
|
|
|
|
|
]
|
|
|
|
|
is_contains = any(key in param_dict for key in pergroup_param)
|
|
|
|
|
self.assertFalse(is_contains)
|
2025-08-06 10:17:44 +08:00
|
|
|
|
2025-09-27 21:01:16 +08:00
|
|
|
def build_layer(self,
|
|
|
|
|
is_new_quant_version=True,
|
|
|
|
|
is_per_channel_weight=False):
|
2025-08-06 10:17:44 +08:00
|
|
|
layer = torch.nn.Module()
|
2025-09-27 21:01:16 +08:00
|
|
|
if is_new_quant_version:
|
|
|
|
|
layer.w13_weight = torch.nn.Parameter(torch.zeros(
|
|
|
|
|
(self.experts, self.input_size, self.output_size),
|
|
|
|
|
dtype=torch.int8),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
layer.w2_weight = torch.nn.Parameter(torch.zeros(
|
|
|
|
|
(self.experts, self.output_size // 2, self.input_size),
|
|
|
|
|
dtype=torch.int8),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
w13_scale_bias = torch.zeros(
|
|
|
|
|
(self.experts, 2 * self.input_size, 1), dtype=torch.float32)
|
|
|
|
|
layer.w13_scale_bias = torch.nn.Parameter(w13_scale_bias,
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
w2_scale_bias = torch.zeros((self.experts, self.output_size,
|
|
|
|
|
16 // self.quant_method.tp_size),
|
|
|
|
|
dtype=torch.float32)
|
|
|
|
|
layer.w2_scale_bias = torch.nn.Parameter(w2_scale_bias,
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
else:
|
|
|
|
|
layer.w13_weight = torch.nn.Parameter(torch.zeros(
|
|
|
|
|
(self.experts, 2 * self.input_size, self.output_size),
|
|
|
|
|
dtype=torch.int8),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
layer.w2_weight = torch.nn.Parameter(torch.zeros(
|
|
|
|
|
(self.experts, self.output_size, self.input_size),
|
|
|
|
|
dtype=torch.int8),
|
|
|
|
|
requires_grad=False)
|
2025-08-06 10:17:44 +08:00
|
|
|
layer.w13_weight_scale = torch.nn.Parameter(torch.ones(
|
2025-09-27 21:01:16 +08:00
|
|
|
(self.experts, 2 * self.input_size, 1), dtype=torch.float32),
|
2025-08-06 10:17:44 +08:00
|
|
|
requires_grad=False)
|
|
|
|
|
layer.w2_weight_scale = torch.nn.Parameter(torch.ones(
|
2025-09-27 21:01:16 +08:00
|
|
|
(self.experts, self.output_size, 1), dtype=torch.float32),
|
2025-08-06 10:17:44 +08:00
|
|
|
requires_grad=False)
|
2025-09-27 21:01:16 +08:00
|
|
|
if not is_per_channel_weight:
|
|
|
|
|
layer.w13_weight_scale_second = torch.nn.Parameter(
|
|
|
|
|
torch.ones((self.experts, 2 * self.input_size,
|
|
|
|
|
self.output_size // self.group_size),
|
|
|
|
|
dtype=torch.float32),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
layer.w13_weight_offset_second = torch.nn.Parameter(
|
|
|
|
|
torch.empty_like(layer.w13_weight_scale_second.data),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
layer.w2_weight_scale_second = torch.nn.Parameter(
|
|
|
|
|
torch.ones((self.experts, self.output_size,
|
|
|
|
|
self.input_size // self.group_size),
|
|
|
|
|
dtype=torch.float32),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
layer.w2_weight_offset_second = torch.nn.Parameter(
|
|
|
|
|
torch.empty_like(layer.w2_weight_scale_second.data),
|
|
|
|
|
requires_grad=False)
|
|
|
|
|
return layer
|
2025-08-06 10:17:44 +08:00
|
|
|
|
2025-09-27 21:01:16 +08:00
|
|
|
@patch('torch_npu.npu_format_cast')
|
|
|
|
|
@patch('torch_npu.npu_quantize')
|
|
|
|
|
@patch('torch.Tensor.npu')
|
|
|
|
|
def test_process_weights_after_loading(self, mock_npu, mock_npu_quantize,
|
|
|
|
|
mock_npu_format_cast):
|
2025-08-06 10:17:44 +08:00
|
|
|
mock_npu.return_value = torch.Tensor()
|
|
|
|
|
mock_npu_quantize.return_value = torch.Tensor()
|
2025-09-27 21:01:16 +08:00
|
|
|
|
|
|
|
|
def func_by_args(weight, num_format):
|
|
|
|
|
return weight
|
|
|
|
|
|
|
|
|
|
mock_npu_format_cast.side_effect = func_by_args
|
|
|
|
|
# old quant version weight
|
|
|
|
|
layer = self.build_layer(is_new_quant_version=False)
|
2025-08-06 10:17:44 +08:00
|
|
|
self.quant_method.process_weights_after_loading(layer)
|
|
|
|
|
self.assertTrue(hasattr(layer, "w13_scale_bias"))
|
2025-08-20 20:25:18 +08:00
|
|
|
self.assertEqual(layer.w13_scale_bias.data.shape,
|
|
|
|
|
(self.experts, 2 * self.input_size))
|
2025-08-06 10:17:44 +08:00
|
|
|
self.assertEqual(layer.w13_scale_bias.data.dtype, torch.float32)
|
|
|
|
|
self.assertTrue(hasattr(layer, "w2_scale_bias"))
|
2025-08-20 20:25:18 +08:00
|
|
|
self.assertEqual(layer.w2_scale_bias.data.shape,
|
|
|
|
|
(self.experts, self.output_size))
|
2025-08-06 10:17:44 +08:00
|
|
|
self.assertEqual(layer.w2_scale_bias.data.dtype, torch.float32)
|
2025-08-20 20:25:18 +08:00
|
|
|
# new quant version weight
|
|
|
|
|
self.quant_method.new_quant_version = True
|
2025-09-27 21:01:16 +08:00
|
|
|
new_layer = self.build_layer(is_new_quant_version=True)
|
2025-08-20 20:25:18 +08:00
|
|
|
self.quant_method.process_weights_after_loading(new_layer)
|
|
|
|
|
self.assertEqual(new_layer.w13_scale_bias.data.shape,
|
|
|
|
|
(self.experts, 2 * self.input_size))
|
|
|
|
|
self.assertEqual(new_layer.w2_scale_bias.data.shape,
|
|
|
|
|
(self.experts, self.output_size))
|
2025-09-27 21:01:16 +08:00
|
|
|
self.assertFalse(hasattr(new_layer, "w13_weight_scale_second"))
|
|
|
|
|
# per-channel weight
|
|
|
|
|
self.quant_method.is_per_channel_weight = True
|
|
|
|
|
per_channel_layer = self.build_layer(is_new_quant_version=True,
|
|
|
|
|
is_per_channel_weight=True)
|
|
|
|
|
self.quant_method.process_weights_after_loading(per_channel_layer)
|
|
|
|
|
self.assertEqual(new_layer.w13_scale_bias.data.shape,
|
|
|
|
|
(self.experts, 2 * self.input_size))
|