5df450bca4
### What this PR does / why we need it?
New Quantization Method: Introduced support for the W8A8SC static linear
quantization scheme specifically for 310P hardware, enabling more
efficient model compression.
Refactored the save_sharded_state_310.py to avoid multi-process issue.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
W8A8SC quant E2E test.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: pu-zhe <zpuaa@outlook.com>
117 lines
4.5 KiB
Python
117 lines
4.5 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.
|
|
#
|
|
|
|
import math
|
|
from typing import Any
|
|
|
|
import torch
|
|
import torch_npu
|
|
from vllm.distributed import get_tensor_model_parallel_rank
|
|
|
|
from vllm_ascend.ops.linear import AscendRowParallelLinear
|
|
from vllm_ascend.quantization.methods.base import AscendLinearScheme
|
|
|
|
from .registry import register_scheme
|
|
|
|
|
|
@register_scheme("W8A8SC", "linear")
|
|
class AscendW8A8SCLinearMethod310(AscendLinearScheme):
|
|
"""310P-only W8A8SC 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]:
|
|
"""
|
|
Get the weight tensors for the W8A8SC quantization scheme.
|
|
|
|
Args:
|
|
input_size: Size of the input dimension (k)
|
|
output_size: Size of the output dimension (n)
|
|
params_dtype: Data type for parameters, default is torch.float16
|
|
|
|
Returns:
|
|
A dictionary containing:
|
|
- "weight": The compressed weight tensor with shape [c], where c is greater than 0
|
|
and not larger than k * n
|
|
- "index": Compression index generated simultaneously with compressed weights,
|
|
with shape [x], where x = k_index * n_index * 8, k_index = ceil(k1 / tilingK),
|
|
n_index = ceil(n1 / tilingN), k1 = k / 32, n1 = n / 16
|
|
- "info": Compression information with length 5, containing compression block
|
|
information tilingN, tilingK, original shape of the pre-compression x2 matrix,
|
|
and identifier for the compression block traversal direction
|
|
"""
|
|
self.input_size = input_size
|
|
index_len = math.ceil(input_size / 256) * math.ceil(output_size / 128) * 8
|
|
return {
|
|
"weight": torch.empty(input_size * output_size, dtype=torch.int8),
|
|
"index": torch.empty(index_len, dtype=torch.int8),
|
|
"info": torch.empty(5, dtype=torch.int64),
|
|
}
|
|
|
|
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]:
|
|
return {
|
|
"quant_bias": torch.empty(output_size, dtype=torch.int32),
|
|
"deq_scale": torch.empty(output_size, dtype=torch.int64),
|
|
}
|
|
|
|
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,
|
|
)
|
|
|
|
return torch_npu.npu_matmul_compress_dequant(
|
|
x,
|
|
layer.weight,
|
|
layer.index,
|
|
layer.quant_bias,
|
|
layer.deq_scale,
|
|
)
|
|
|
|
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
|
layer.aclnn_input_scale = layer.input_scale.data.repeat(self.input_size)
|
|
layer.aclnn_input_scale_reciprocal = 1.0 / layer.aclnn_input_scale.data
|
|
layer.aclnn_input_offset = layer.input_offset.data.repeat(self.input_size).to(layer.aclnn_input_scale.dtype)
|
|
layer.deq_scale.data = layer.deq_scale.data.unsqueeze(0).to(torch.uint64)
|
|
layer.quant_bias.data = layer.quant_bias.data.unsqueeze(0)
|
|
# Only apply bias on row_parallel_linear when tp_rank is 0.
|
|
# torch_npu.npu_matmul_compress_dequant's quant_bias cannot be None.
|
|
if isinstance(layer, AscendRowParallelLinear) and get_tensor_model_parallel_rank() != 0:
|
|
layer.quant_bias.data = torch.zeros_like(layer.quant_bias)
|