EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity
2,783 Commits 2 Branches 0 Tags
v0.18.0
Commit Graph

8 Commits

Author SHA1 Message Date
linfeng-yuan
d452d04656 [A5][bugfix] Fix fused MoE A5 MXFP8 scale normalization, load-balance routing and gating_topk ops (#7573) 
### What this PR does / why we need it?
This PR fixes A5 MXFP8 MoE scale handling in the fused MoE path.

- It normalizes MXFP8 activation scales to the packed 3D layout expected
by A5 kernels, including both precomputed dynamic_scale inputs and gmm1
output scales before they are consumed by downstream grouped matmul ops.
- It also refines the MXFP8 force load-balancing path in profiling runs.
- This PR also enables npu_gating_top_k from torch_npu instead of custom
op when running ascend950 chip.
### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
CI and E2E serving tests on Ascend950DT passed.

---------

Signed-off-by: linfeng-yuan <1102311262@qq.com>
 2026-03-25 17:20:28 +08:00
linfeng-yuan
88d03a783f [refactor] replace scattered business kwargs with typed request objects and explicit stage boundaries (#7024) 
### What this PR does / why we need it?
Refactor `vllm_ascend/ops/fused_moe` to replace scattered MoE business
`**kwargs` with typed request objects and explicit stage boundaries.

- Prepare, dispatch, MLP, and quant stages now have clearer ownership.
- Main MoE path no longer depends on business `kwargs.get(...)` lookups.
- Comm and dispatcher interfaces are request-only on the main path.
- UTs can assert stage-level fields directly instead of inferring
behavior indirectly.

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
CI passed.

---------

Signed-off-by: linfeng-yuan <1102311262@qq.com>
 2026-03-20 23:23:57 +08:00
Siyuan Kong
a16c99141b Adapt w8a8mxfp8 quantization for Qwen VL models (#7417) 
### What this PR does / why we need it?

This PR adapts the `w8a8_mxfp8` quantization method to support Qwen
Vision-Language (VL) models. Key changes include:
- Reshaping multi-dimensional input tensors to 2D before the quantized
matrix multiplication.
- Reshaping the 2D output back to its original multi-dimensional format.
- Adding specific output reshaping for the visual components of Qwen VL
models.
- Casting the bias tensor to `float32` to comply with the
`npu_quant_matmul` kernel requirements.

These changes are necessary to enable `w8a8_mxfp8` quantization for
models with multi-modal inputs like Qwen VL.

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

No, this is a backend enhancement to extend quantization support to new
model architectures. There are no user-facing API or behavior changes.

### How was this patch tested?

CI is expected to pass. Manual testing should be performed with a Qwen
VL model using `w8a8_mxfp8` quantization to verify correctness and
performance.

- vLLM version: v0.17.0
- vLLM main:
4497431df6

---------

Signed-off-by: ksiyuan <ksiyuan@umich.edu>
 2026-03-20 16:18:58 +08:00
Ronald
c980e68d40 [Feature] support aclgraph for model runner v2 (#7110) 
### What this PR does / why we need it?
This PR aims to support aclgraph for model runner v2, please see RFC
#5208. The PR contains these modifications:
- adapt to newest commit of vllm main branch.
- supply a unified interface of extra forward context for both model
runner v1 and model runner v2.
- implement graph mode for main model. 

### Does this PR introduce _any_ user-facing change?
no

### How was this patch tested?

- vLLM version: v0.16.0
- vLLM main:
4034c3d32e

---------

Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
 2026-03-13 09:11:46 +08:00
linfeng-yuan
68d8d20ca2 [misc] move mxfp_compat into device to decouple from quantization init chain (#6918) 
### What this PR does / why we need it?
`mxfp_compat` only provides dtype/symbol compatibility helpers for
different `torch_npu` versions, but it was placed under
`vllm_ascend.quantization`. Importing it from device/ops paths could
trigger `quantization/__init__.py` and pull in heavy quantization method
dependencies, increasing startup coupling and causing import-cycle risk
(especially on 310P paths).

### Does this PR introduce _any_ user-facing change?
No functional behavior change intended.

### How was this patch tested?
CI passed.

- vLLM version: v0.16.0
- vLLM main:
15d76f74e2

---------

Signed-off-by: linfeng-yuan <1102311262@qq.com>
 2026-03-02 18:17:01 +08:00
Eric-dot
3c66a970f2 add mxfp8 moe quantization (#6670) 
### What this PR does / why we need it?
support mxfp8 quantization (Qwen MOE )
Using adaptor to make the hardware-specific behavior clearer and more
maintainable
### How was this patch tested?

- vLLM version: v0.15.0
- vLLM main:
13397841ab

---------

Signed-off-by: fangrongcan <17343701736@163.com>
Signed-off-by: wangyao-i <iwangyao@outlook.com>
Signed-off-by: linfeng-yuan <1102311262@qq.com>
Signed-off-by: Eric-dot <60131170+Eric-dot@users.noreply.github.com>
Co-authored-by: fangrongcan <f00876277@china.huawei.com>
Co-authored-by: wangyao-i <iwangyao@outlook.com>
Co-authored-by: linfeng-yuan <1102311262@qq.com>
 2026-03-02 11:04:06 +08:00
SILONG ZENG
99aedaff63 [Lint]Style: Convert vllm-ascend/ to ruff format(Batch #7) (#6023) 
### What this PR does / why we need it?
**Scope of Changes**:
| File Path |
| :--- |
|` vllm_ascend/quantization/compressed_tensors/compressed_tensors.py`|
|` vllm_ascend/quantization/quant_config.py`|
|` vllm_ascend/quantization/utils.py`|
|` vllm_ascend/quantization/w4a16.py`|
|` vllm_ascend/quantization/w4a4_flatquant_dynamic.py`|
|` vllm_ascend/quantization/w4a8_dynamic.py`|
|` vllm_ascend/quantization/w8a16.py`|
|` vllm_ascend/quantization/w8a8.py`|
|` vllm_ascend/quantization/w8a8_dynamic.py`|
|` vllm_ascend/quantization/w8a8_pdmix.py`|
|` vllm_ascend/quantization/w8a8mxfp8.py`|
|` vllm_ascend/sample/rejection_sampler.py`|
|` vllm_ascend/sample/sampler.py`|
|` vllm_ascend/worker/block_table.py`|

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

### How was this patch tested?

- vLLM version: v0.13.0
- vLLM main:
2c24bc6996

Signed-off-by: MrZ20 <2609716663@qq.com>
 2026-02-06 14:56:53 +08:00
Cao Yi
a69ef10c3a [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:
2f4e6548ef

---------

Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
 2026-01-23 14:13:47 +08:00