# SPDX-FileCopyrightText: Copyright (c) 2022-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# 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.
import numpy as np
from .._utils import trt_dtype_to_np
from ..functional import ACT2FN
from ..module import Module
from ..quantization import QuantMode
from ..quantization.layers import FP8Linear, FP8RowLinear
from .linear import ColumnLinear, RowLinear
[docs]
class MLP(Module):
def __init__(self,
hidden_size,
ffn_hidden_size,
hidden_act,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
quant_mode=QuantMode(0),
instance_id: int = 0):
super().__init__()
if hidden_act not in ACT2FN:
raise ValueError(
'unsupported activation function: {}'.format(hidden_act))
fc_output_size = 2 * ffn_hidden_size if hidden_act == 'swiglu' else ffn_hidden_size
self.use_fp8_qdq = quant_mode.has_fp8_qdq()
if self.use_fp8_qdq:
self.fc = FP8Linear(hidden_size,
fc_output_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False)
self.proj = FP8RowLinear(ffn_hidden_size,
hidden_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
instance_id=instance_id)
else:
self.fc = ColumnLinear(hidden_size,
fc_output_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False)
self.proj = RowLinear(ffn_hidden_size,
hidden_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
instance_id=instance_id)
self.hidden_act = hidden_act
self.dtype = dtype
[docs]
def forward(self, hidden_states, workspace=None):
inter = self.fc(hidden_states)
inter = ACT2FN[self.hidden_act](inter)
output = self.proj(inter, workspace)
return output
[docs]
class GatedMLP(MLP):
def __init__(self,
hidden_size,
ffn_hidden_size,
hidden_act,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
quant_mode=QuantMode(0),
instance_id: int = 0):
self.use_fp8_qdq = quant_mode.has_fp8_qdq()
super().__init__(hidden_size,
ffn_hidden_size,
hidden_act,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
quant_mode=quant_mode,
instance_id=instance_id)
self.hidden_size = hidden_size
self.ffn_hidden_size = ffn_hidden_size
self.hidden_act = hidden_act
self.bias = bias
self.dtype = dtype
self.tp_group = tp_group
self.tp_size = tp_size
self.quant_mode = quant_mode
self.instance_id = instance_id
if self.use_fp8_qdq:
self.gate = FP8Linear(hidden_size,
ffn_hidden_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False)
else:
self.gate = ColumnLinear(hidden_size,
ffn_hidden_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False)
[docs]
def forward(self, hidden_states, workspace=None):
inter = self.fc(hidden_states)
inter = ACT2FN[self.hidden_act](inter)
gate = self.gate(hidden_states)
output = self.proj(inter * gate, workspace)
return output
[docs]
class FusedGatedMLP(GatedMLP):
def __init__(self,
hidden_size,
ffn_hidden_size,
hidden_act,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
quant_mode=QuantMode(0),
instance_id: int = 0):
super().__init__(hidden_size,
ffn_hidden_size,
hidden_act,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
quant_mode=quant_mode,
instance_id=instance_id)
[docs]
def forward(self, hidden_states, workspace=None):
# Combine the following pattern
#
# SiLU(FC(x)) + Gate(x)
#
# into:
#
# SwiGLU(FusedFC(x))
#
# Upside is we don't need to modify 4 different weight loading paths just to concat weights
_np_dtype = trt_dtype_to_np(self.dtype)
concat_weight = np.concatenate(
[self.gate.weight.raw_value, self.fc.weight.raw_value],
axis=0).astype(_np_dtype)
if self.bias:
concat_bias = np.concatenate(
[self.gate.bias.raw_value, self.fc.bias.raw_value],
axis=0).astype(_np_dtype)
if self.use_fp8_qdq:
gate_weights_scaling_factor = self.gate.weights_scaling_factor.raw_value
fc_weights_scaling_factor = self.fc.weights_scaling_factor.raw_value
fc_activation_scaling_factor = self.fc.activation_scaling_factor.raw_value
gate_activation_scaling_factor = self.gate.activation_scaling_factor.raw_value
assert fc_activation_scaling_factor == gate_activation_scaling_factor, "Activation scales should be identical"
# Remove dangling TRT-LLM parameter references after the graph rewrite.
for param, _ in list(self.gate.named_parameters()):
self.gate._parameters.pop(param)
self.gate = None
if self.use_fp8_qdq:
self.fc = FP8Linear(self.hidden_size,
self.ffn_hidden_size * 2,
bias=self.bias,
dtype=self.dtype,
tp_group=self.tp_group,
tp_size=self.tp_size,
gather_output=False)
else:
self.fc = ColumnLinear(self.hidden_size,
self.ffn_hidden_size * 2,
bias=self.bias,
dtype=self.dtype,
tp_group=self.tp_group,
tp_size=self.tp_size,
gather_output=False)
self.fc.weight.value = concat_weight
if self.use_fp8_qdq:
self.fc.activation_scaling_factor.value = fc_activation_scaling_factor
# TODO: need to align with quantization toolkit; preferably put a constraint to equalize
# fc/gate weight scaling factor to allow horizontal fusion without accuracy loss
self.fc.weights_scaling_factor.value = max(
gate_weights_scaling_factor, fc_weights_scaling_factor)
if self.bias:
self.fc.bias.value = concat_bias
inter = self.fc(hidden_states)
if self.hidden_act == 'silu':
inter = ACT2FN['swiglu'](inter)
else:
raise NotImplementedError(
f"Activation {self.hidden_act} not yet implemented for FusedGatedMLP"
)
output = self.proj(inter, workspace)
return output