103 lines
3.8 KiB
Python
103 lines
3.8 KiB
Python
#
|
|
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
|
|
# This file is a part of the vllm-ascend project.
|
|
#
|
|
# 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.
|
|
#
|
|
|
|
from typing import Any
|
|
|
|
import torch
|
|
|
|
from .base import QuantType
|
|
from .registry import register_scheme
|
|
from .w8a8_dynamic import AscendW8A8DynamicFusedMoEMethod, AscendW8A8DynamicLinearMethod
|
|
|
|
|
|
@register_scheme("W8A8FP8_DYNAMIC", "linear")
|
|
class AscendW8A8FP8DynamicLinearMethod(AscendW8A8DynamicLinearMethod):
|
|
"""Linear method for Ascend W8A8FP8_DYNAMIC.
|
|
|
|
This scheme uses FP8 dynamic per-token quantization for activations
|
|
and FP8 per-channel quantization for weights.
|
|
"""
|
|
|
|
act_quant_type: torch.dtype = torch.float8_e4m3fn
|
|
|
|
def __init__(self):
|
|
pass
|
|
|
|
def get_weight(self, input_size: int, output_size: int, params_dtype: torch.dtype) -> dict[str, Any]:
|
|
params_dict = {"weight": torch.empty(output_size, input_size, dtype=torch.float8_e4m3fn)}
|
|
return params_dict
|
|
|
|
def get_perchannel_param(
|
|
self,
|
|
output_size: int,
|
|
params_dtype: torch.dtype,
|
|
) -> dict[str, Any]:
|
|
params_dict = {}
|
|
params_dict["weight_scale"] = torch.empty(output_size, 1, dtype=torch.float32)
|
|
params_dict["weight_offset"] = torch.empty(output_size, 1, dtype=params_dtype)
|
|
return params_dict
|
|
|
|
def apply(
|
|
self,
|
|
layer: torch.nn.Module,
|
|
x: torch.Tensor,
|
|
bias: torch.Tensor | None = None,
|
|
tp_rank: int | None = 0,
|
|
) -> torch.Tensor:
|
|
output = super().apply(layer, x, bias, tp_rank)
|
|
# TODO: there is a bug in npu_quant_matmul for fp8 with bias
|
|
# after the bug is fixed, the whole apply method can be removed.
|
|
if bias is not None:
|
|
output = (output + bias).to(x.dtype)
|
|
return output
|
|
|
|
|
|
@register_scheme("W8A8FP8_DYNAMIC", "moe")
|
|
class AscendW8A8FP8DynamicFusedMoEMethod(AscendW8A8DynamicFusedMoEMethod):
|
|
"""FusedMoE method for Ascend W8A8FP8_DYNAMIC."""
|
|
|
|
quant_type: QuantType = QuantType.W8A8FP8
|
|
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
def get_weight(
|
|
self, num_experts: int, intermediate_size_per_partition: int, hidden_sizes: int, params_dtype: torch.dtype
|
|
) -> dict[str, Any]:
|
|
param_dict = {}
|
|
param_dict["w13_weight"] = torch.empty(
|
|
num_experts, 2 * intermediate_size_per_partition, hidden_sizes, dtype=torch.float8_e4m3fn
|
|
)
|
|
param_dict["w2_weight"] = torch.empty(
|
|
num_experts, hidden_sizes, intermediate_size_per_partition, dtype=torch.float8_e4m3fn
|
|
)
|
|
return param_dict
|
|
|
|
def get_dynamic_quant_param(
|
|
self, num_experts: int, intermediate_size_per_partition: int, hidden_sizes: int, params_dtype: torch.dtype
|
|
) -> dict[str, Any]:
|
|
param_dict = {}
|
|
param_dict["w13_weight_scale"] = torch.empty(
|
|
num_experts, 2 * intermediate_size_per_partition, 1, dtype=torch.float32
|
|
)
|
|
param_dict["w13_weight_offset"] = torch.empty(
|
|
num_experts, 2 * intermediate_size_per_partition, 1, dtype=params_dtype
|
|
)
|
|
param_dict["w2_weight_scale"] = torch.empty(num_experts, hidden_sizes, 1, dtype=torch.float32)
|
|
param_dict["w2_weight_offset"] = torch.empty(num_experts, hidden_sizes, 1, dtype=params_dtype)
|
|
return param_dict
|