02886e2641
### What this PR does / why we need it?
This PR introduces support for W8A8 dynamic quantization for
Mixture-of-Experts (MoE) models on Ascend 310P devices. This is achieved
by:
- Implementing a new quantization scheme
`AscendW8A8DynamicFusedMoEMethod310`.
- Adding a unified MLP implementation (`unified_apply_mlp`) for 310P
that handles both quantized and unquantized paths.
- Refactoring the MoE and quantization configuration logic to correctly
route to the new 310P-specific implementations.
- Adding new e2e and unit tests to verify the functionality of MoE W8A8
quantization.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- Added a new e2e test `test_qwen3_moe_tp2_w8a8` to test MoE W8A8
quantization in a multi-card setup.
- Added several new unit tests for the 310P-specific MoE components,
including `experts_selector`, `fused_moe`, `moe_comm_method`, `moe_mlp`,
and the new `w8a8_dynamic` quantization method.
- vLLM version: v0.15.0
- vLLM main:
d7e17aaacd
---------
Signed-off-by: pu-zhe <zpuaa@outlook.com>
102 lines
3.5 KiB
Python
102 lines
3.5 KiB
Python
#
|
|
# Copyright (c) 2026 Huawei Technologies Co., Ltd. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
# This file is a part of the vllm-ascend project.
|
|
#
|
|
|
|
from typing import Any
|
|
|
|
import torch
|
|
import torch_npu
|
|
|
|
from vllm_ascend.quantization.methods.base import AscendLinearScheme
|
|
|
|
from .registry import register_scheme
|
|
|
|
|
|
@register_scheme("W8A8", "linear")
|
|
class AscendW8A8LinearMethod310(AscendLinearScheme):
|
|
"""310P-only W8A8 static linear scheme.
|
|
|
|
Notes:
|
|
- This scheme is discovered via 310P local registry.
|
|
"""
|
|
|
|
def get_weight(
|
|
self,
|
|
input_size: int,
|
|
output_size: int,
|
|
params_dtype: torch.dtype = torch.float16,
|
|
) -> dict[str, Any]:
|
|
return {"weight": torch.empty(output_size, input_size, dtype=torch.int8)}
|
|
|
|
def get_pertensor_param(self, params_dtype: torch.dtype) -> dict[str, Any]:
|
|
return {
|
|
"input_scale": torch.empty(1, dtype=params_dtype),
|
|
"input_offset": torch.empty(1, dtype=torch.int8),
|
|
}
|
|
|
|
def get_perchannel_param(self, output_size: int, params_dtype: torch.dtype) -> dict[str, Any]:
|
|
params: dict[str, Any] = {}
|
|
params["quant_bias"] = torch.empty(output_size, dtype=torch.int32)
|
|
params["deq_scale"] = torch.empty(output_size, dtype=torch.int64)
|
|
params["weight_scale"] = torch.empty(output_size, 1, dtype=params_dtype)
|
|
params["weight_offset"] = torch.empty(output_size, 1, dtype=params_dtype)
|
|
return params
|
|
|
|
def apply(
|
|
self,
|
|
layer: torch.nn.Module,
|
|
x: torch.Tensor,
|
|
bias: torch.Tensor | None = None,
|
|
tp_rank: int | None = 0,
|
|
) -> torch.Tensor:
|
|
if x.dtype != torch.int8:
|
|
x = torch.ops.vllm.quantize(
|
|
x,
|
|
layer.aclnn_input_scale,
|
|
layer.aclnn_input_scale_reciprocal,
|
|
layer.aclnn_input_offset,
|
|
)
|
|
|
|
quant_bias = layer.quant_bias if tp_rank == 0 else None
|
|
|
|
return torch_npu.npu_quant_matmul(
|
|
x,
|
|
layer.weight,
|
|
layer.deq_scale,
|
|
bias=quant_bias,
|
|
output_dtype=layer.params_dtype,
|
|
)
|
|
|
|
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
|
expanding_factor = layer.weight.data.shape[1]
|
|
layer.aclnn_input_scale = torch.nn.Parameter(
|
|
layer.input_scale.data.repeat(expanding_factor),
|
|
requires_grad=False,
|
|
)
|
|
layer.aclnn_input_scale_reciprocal = torch.nn.Parameter(
|
|
1.0 / layer.aclnn_input_scale.data,
|
|
requires_grad=False,
|
|
)
|
|
layer.aclnn_input_offset = torch.nn.Parameter(
|
|
layer.input_offset.data.repeat(expanding_factor),
|
|
requires_grad=False,
|
|
).to(layer.aclnn_input_scale.dtype)
|
|
|
|
layer.weight.data = layer.weight.data.transpose(0, 1).contiguous()
|
|
|
|
layer.weight_scale.data = torch.flatten(layer.weight_scale.data)
|
|
layer.weight_offset.data = torch.flatten(layer.weight_offset.data)
|