TensorRT-LLMs/tensorrt_llm/layers/moe.py
2024-05-07 23:34:28 +08:00

502 lines
21 KiB
Python

# SPDX-FileCopyrightText: Copyright (c) 2022-2024 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.
from dataclasses import dataclass
from enum import IntEnum
from typing import List, Union
import numpy as np
import tensorrt as trt
from tensorrt_llm._utils import str_dtype_to_trt
from .._common import default_net, default_trtnet
from ..functional import (_create_tensor, allreduce, cast, div,
is_gated_activation, non_gated_version, softmax, sum,
topk)
from ..layers import MLP, GatedMLP
from ..module import Module
from ..parameter import Parameter
from ..plugin import TRT_LLM_PLUGIN_NAMESPACE
from ..quantization import QuantMode
from ..quantization.functional import quantize
from .linear import RowLinear
activation_str_to_int_map = {
# [WARNING] Keep the below in sync with cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_gemm_kernels.h
"gelu": 0,
"gelu_new": 0,
"relu": 1,
"silu": 2,
"swiglu": 3,
"geglu": 4,
"identity": 5,
}
@dataclass
class MoeConfig:
# [WARNING] Keep the below in sync with cpp/tensorrt_llm/kernels/mixtureOfExperts/moe_kernels.h
class ParallelismMode(IntEnum):
NONE = 0
EXPERT_PARALLEL = 1
TENSOR_PARALLEL = 2
class ExpertScaleNormalizationMode(IntEnum):
NONE = 0
RENORMALIZE = 1
num_experts: int = 0
top_k: int = 0
tp_mode: ParallelismMode = ParallelismMode.TENSOR_PARALLEL
normalization_mode: ExpertScaleNormalizationMode = ExpertScaleNormalizationMode.RENORMALIZE
def validate(self) -> "MoeConfig":
if (self.num_experts == 0) != (self.top_k == 0):
raise ValueError(
"Both or neither MoeConfig's num_experts and top_k must be set to 0"
)
return self
def has_moe(self) -> bool:
return self.num_experts > 1
def _moe_plugin(moe_config,
hidden_states,
routing,
finished,
expert_weight_1,
expert_weight_2,
expert_bias_1,
expert_bias_2,
expert_scale_1,
expert_scale_2,
expert_scale_3,
expert_scale_4,
hidden_size,
ffn_hidden_size,
act_fn,
dtype,
weight_dtype,
output_dtype,
quant_mode=QuantMode(0),
tp_size=1,
tp_rank=0):
if isinstance(dtype, str):
dtype = str_dtype_to_trt(dtype)
if isinstance(weight_dtype, str):
weight_dtype = str_dtype_to_trt(weight_dtype)
if isinstance(output_dtype, str):
output_dtype = str_dtype_to_trt(output_dtype)
def from_parameter(x):
if isinstance(x, Parameter):
return x.value
return x
expert_weight_1 = from_parameter(expert_weight_1)
expert_weight_2 = from_parameter(expert_weight_2)
expert_bias_1 = from_parameter(expert_bias_1)
expert_bias_2 = from_parameter(expert_bias_2)
expert_scale_1 = from_parameter(expert_scale_1)
expert_scale_2 = from_parameter(expert_scale_2)
expert_scale_3 = from_parameter(expert_scale_3)
expert_scale_4 = from_parameter(expert_scale_4)
# Create the plugin with our required state
num_experts = moe_config.num_experts
# We pass the full number of experts (not divided by tp_size) even for EP mode
p_num_experts = trt.PluginField("number_of_experts",
np.array(num_experts, dtype=np.int32),
trt.PluginFieldType.INT32)
p_top_k = trt.PluginField("top_k", np.array(moe_config.top_k,
dtype=np.int32),
trt.PluginFieldType.INT32)
p_expert_hidden_size = trt.PluginField(
"expert_hidden_size", np.array(hidden_size, dtype=np.int32),
trt.PluginFieldType.INT32)
p_expert_inter_size = trt.PluginField(
"expert_inter_size", np.array(ffn_hidden_size, dtype=np.int32),
trt.PluginFieldType.INT32)
p_activation_type = trt.PluginField(
"activation_type",
np.array(activation_str_to_int_map[act_fn], dtype=np.int32),
trt.PluginFieldType.INT32)
p_type_id = trt.PluginField("type_id", np.array([int(dtype)],
dtype=np.int32),
trt.PluginFieldType.INT32)
p_weight_type_id = trt.PluginField(
"weight_type_id", np.array([int(weight_dtype)], dtype=np.int32),
trt.PluginFieldType.INT32)
p_output_type_id = trt.PluginField(
"output_type_id", np.array([int(output_dtype)], dtype=np.int32),
trt.PluginFieldType.INT32)
p_quant_mode = trt.PluginField("quant_mode",
np.array([int(quant_mode)], dtype=np.int32),
trt.PluginFieldType.INT32)
p_use_finished = trt.PluginField(
"use_finished", np.array([int(finished is not None)], dtype=np.int32),
trt.PluginFieldType.INT32)
p_use_bias = trt.PluginField(
"use_bias", np.array([int(expert_bias_1 is not None)], dtype=np.int32),
trt.PluginFieldType.INT32)
p_tp_size = trt.PluginField("tp_size", np.array(tp_size, dtype=np.int32),
trt.PluginFieldType.INT32)
p_tp_rank = trt.PluginField("tp_rank", np.array(tp_rank, dtype=np.int32),
trt.PluginFieldType.INT32)
p_parallelism_mode = trt.PluginField(
"parallelism_mode", np.array(moe_config.tp_mode, dtype=np.int32),
trt.PluginFieldType.INT32)
p_normalization_mode = trt.PluginField(
"normalization_mode",
np.array(moe_config.normalization_mode, dtype=np.int32),
trt.PluginFieldType.INT32)
pfc = trt.PluginFieldCollection([
p_num_experts, p_top_k, p_expert_hidden_size, p_expert_inter_size,
p_activation_type, p_type_id, p_weight_type_id, p_output_type_id,
p_quant_mode, p_use_finished, p_use_bias, p_tp_size, p_tp_rank,
p_parallelism_mode, p_normalization_mode
])
# Create the plugin with our constant inputs to the constructor
plugin_creator = trt.get_plugin_registry().get_plugin_creator(
'MixtureOfExperts', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert plugin_creator is not None
moe_plugin = plugin_creator.create_plugin("mixture_of_experts", pfc)
# Instantiate the plugin with our specific inputs
plugin_inputs = [hidden_states, routing, expert_weight_1, expert_weight_2]
if expert_bias_1:
assert expert_bias_2
plugin_inputs += [expert_bias_1, expert_bias_2]
if finished is not None:
plugin_inputs += [finished]
# Add conditional inputs
if quant_mode.is_weight_only() or quant_mode.has_fp8_qdq():
assert expert_scale_1
assert expert_scale_2
plugin_inputs += [expert_scale_1, expert_scale_2]
# Add conditional inputs
if quant_mode.has_fp8_qdq():
assert expert_scale_3
plugin_inputs += [expert_scale_3]
if expert_scale_4 is not None:
assert quant_mode.has_fp8_qdq()
assert output_dtype == trt.fp8
plugin_inputs += [expert_scale_4]
plugin_inputs = [i.trt_tensor for i in plugin_inputs]
layer = default_trtnet().add_plugin_v2(plugin_inputs, moe_plugin)
if not default_net().strongly_typed:
for ii in range(layer.num_inputs):
if layer.get_input(ii).dtype == str_dtype_to_trt("int8"):
layer.get_input(ii).set_dynamic_range(-127, 127)
output = _create_tensor(layer.get_output(0), layer)
return output
# This exists so that MOE can have the same name format as a regular MLP, just with different shaped weight tensors
class MOEWeightWrapper(Module):
def __init__(self, in_features: int, out_features: int,
experts_per_node: int, quant_mode: QuantMode,
dtype: Union[str, trt.DataType],
weight_dtype: Union[str, trt.DataType], has_bias: bool):
super().__init__()
self.quant_mode = quant_mode
self.expert_shape = (experts_per_node, out_features, in_features)
self.dtype = dtype
self.weight_dtype = weight_dtype
self.has_bias = has_bias
if quant_mode.is_weight_only():
bytes_per_col_scale = 2 if quant_mode.is_int4_weight_only() else 1
# We use a different shape here because the quantized weights have their own layout
self.expert_shape = (experts_per_node, in_features,
out_features // bytes_per_col_scale)
self.per_channel_scale = Parameter(shape=(experts_per_node,
out_features),
dtype=dtype)
else:
self.register_parameter('per_channel_scale', None)
self.weight = Parameter(shape=self.expert_shape, dtype=weight_dtype)
if has_bias:
self.bias = Parameter(shape=(experts_per_node, out_features),
dtype=dtype)
else:
self.register_parameter('bias', None)
if quant_mode.has_fp8_qdq():
self.activation_scaling_factor = Parameter(shape=(1, ),
dtype=trt.float32)
self.weights_scaling_factor = Parameter(shape=(experts_per_node, 1),
dtype=trt.float32)
else:
self.register_parameter('activation_scaling_factor', None)
self.register_parameter('weights_scaling_factor', None)
class MixtureOfExperts(Module):
def __init__(self,
moe_config: MoeConfig,
hidden_size: int,
ffn_hidden_size: int,
hidden_act: str,
bias: bool = True,
dtype=None,
tp_group: List[int] = None,
tp_size: int = 1,
tp_rank: int = 0,
quant_mode=QuantMode(0)):
super().__init__()
self.moe_config = moe_config
self.num_experts = moe_config.num_experts
self.top_k = moe_config.top_k
self.hidden_act = hidden_act
self.hidden_size = hidden_size
self.ffn_hidden_size = ffn_hidden_size
self.dtype = dtype
self.weight_dtype = dtype
self.tp_group = tp_group
self.tp_size = tp_size
self.tp_rank = tp_rank
self.quant_mode = quant_mode
self.has_bias = bias
self.experts_per_node = self.num_experts
self.tp_mode = moe_config.tp_mode
if moe_config.tp_mode == MoeConfig.ParallelismMode.EXPERT_PARALLEL:
if self.num_experts % self.tp_size != 0:
raise ValueError(
f"MixtureOfExperts - Number of experts {self.num_experts} is not a multiple of EP size {self.tp_size}"
)
self.experts_per_node = self.experts_per_node // tp_size
elif moe_config.tp_mode == MoeConfig.ParallelismMode.TENSOR_PARALLEL:
if self.ffn_hidden_size % self.tp_size != 0:
raise ValueError(
f"MixtureOfExperts - FFN Hidden Size {self.ffn_hidden_size} is not a multiple of TP size {self.tp_size}"
)
self.ffn_hidden_size = self.ffn_hidden_size // tp_size
if quant_mode.has_fp8_qdq() and self.has_bias:
# TODO (dastokes) We will need to revisit this if we have a use case for it
raise ValueError(
f"MixtureOfExperts - Bias is not supported with FP8")
if quant_mode.is_weight_only():
self.weight_dtype = trt.int8
elif quant_mode.has_fp8_qdq():
self.weight_dtype = trt.fp8
# Since output dimension is usually low (in the order of 10s), no TP at
# all is more efficient as no allreduce required in the end.
# Note that if we see models that have large number of experts, we may
# need to consider add TP back here.
# TODO: Arctic has large # experts, we may need to add TP back here.
self.router = RowLinear(
hidden_size,
self.num_experts,
bias=False,
dtype=trt.
float32, # Routing is sensitive since it conditions what experts are used
tp_group=None,
tp_size=1,
strict_dtype=True)
# Note we use horizontal fusion for gated activation to do the operation in one GEMM invocation
# The left matrix is a linear projection (no activation applied)
# The right matrix is the gating value (activation applied)
# The naming convention is the inverse of GatedMLP, but the same as `tensorrt_llm/functional.py`
expert_1_out_size = self.ffn_hidden_size * 2 if is_gated_activation(
hidden_act) else self.ffn_hidden_size
self.fc = MOEWeightWrapper(hidden_size, expert_1_out_size,
self.experts_per_node, self.quant_mode,
self.dtype, self.weight_dtype, self.has_bias)
self.proj = MOEWeightWrapper(self.ffn_hidden_size, hidden_size,
self.experts_per_node, self.quant_mode,
self.dtype, self.weight_dtype,
self.has_bias)
ClsMLP = GatedMLP if is_gated_activation(self.hidden_act) else MLP
# In OOTB mode, when ParallelismMode mode is TENSOR_PARALLEL, using MLP class to do TP settings
# pass self.ffn_hidden_size to original size,
# self.experts only inference in OOTB mode.
if moe_config.tp_mode == MoeConfig.ParallelismMode.TENSOR_PARALLEL:
ffn_hidden_size = self.ffn_hidden_size * self.tp_size
else:
tp_size = 1
tp_group = None
ffn_hidden_size = self.ffn_hidden_size
self.experts = [
ClsMLP(self.hidden_size, ffn_hidden_size,
non_gated_version(self.hidden_act), bias, dtype, tp_group,
tp_size, quant_mode) for _ in range(self.experts_per_node)
]
def set_ootb_weight(self):
for i, expert in enumerate(self.experts):
is_gated_act = is_gated_activation(self.hidden_act)
# Gated weight pack in expert1 weights
# expert_weight_1
experts_weight_1_raw = self.fc.weight.raw_value
expert.fc.weight.value = experts_weight_1_raw[
i, -self.ffn_hidden_size:, :]
if is_gated_act:
expert.gate.weight.value = experts_weight_1_raw[
i, :self.ffn_hidden_size, :]
# expert_weight_2
experts_weight_2_raw = self.proj.weight.raw_value
expert.proj.weight.value = experts_weight_2_raw[i, :, :]
has_bias = self.has_bias
if has_bias:
experts_bias_1_raw = self.fc.bias.raw_value
expert.fc.bias.value = experts_bias_1_raw[
i, -self.ffn_hidden_size:]
experts_bias_2_raw = self.proj.bias.raw_value
expert.proj.bias.value = experts_bias_2_raw[i, :]
if is_gated_act:
expert.gate.bias.value = experts_bias_1_raw[
i, :self.ffn_hidden_size]
def forward(self, hidden_states, finished=None, lora_layer_params=None):
assert lora_layer_params is None, "LoRA + MoE is not supported for the moment"
routing_input = cast(hidden_states, trt.float32)
routing = self.router(routing_input)
if not default_net().plugin_config.moe_plugin:
# Depending on the value of plugin_config.moe_plugin, weights must be assigned differently. Hence the need to do that in .forward().
self.set_ootb_weight()
if self.moe_config.normalization_mode == MoeConfig.ExpertScaleNormalizationMode.RENORMALIZE:
topk_values, topk_indices = topk(routing, self.top_k, dim=-1)
topk_values = softmax(topk_values, -1)
else:
router_probs = softmax(routing, -1)
topk_values, topk_indices = topk(router_probs,
self.top_k,
dim=-1)
output = hidden_states * 0.0 # Create output space
# Experts inference
for i, expert in enumerate(self.experts):
if self.tp_mode == MoeConfig.ParallelismMode.EXPERT_PARALLEL:
index = i + self.experts_per_node * self.tp_rank
else:
index = i
# inference expert
out = expert(hidden_states)
expert_mask = topk_indices == index
expert_weights = cast(
sum(topk_values * cast(expert_mask, topk_values.dtype),
dim=-1,
keepdim=True), self.dtype)
output += out * expert_weights
if self.tp_size > 1 and self.tp_group is not None and self.moe_config.tp_mode == MoeConfig.ParallelismMode.EXPERT_PARALLEL:
output = allreduce(output, self.tp_group)
else:
if self.quant_mode.has_fp8_qdq():
assert self.fc.weight.value.dtype == trt.fp8, (
"mlp fc weight dtype should be fp8 in the fp8 quantization mode."
)
assert self.proj.weight.value.dtype == trt.fp8, (
"mlp proj weight dtype should be fp8 in the fp8 quantization mode."
)
hidden_states_quant = hidden_states
if hidden_states_quant.dtype != trt.fp8:
hidden_states_quant = quantize(
hidden_states, self.fc.activation_scaling_factor.value,
'fp8')
dtype_quant = trt.fp8
weight_dtype_quant = trt.fp8
fc1_dequant = self.fc.weights_scaling_factor.value * self.fc.activation_scaling_factor.value
fc2_quant = div(1.0, self.proj.activation_scaling_factor.value)
fc2_dequant = self.proj.weights_scaling_factor.value * self.proj.activation_scaling_factor.value
scale_1 = fc1_dequant
scale_2 = fc2_quant
scale_3 = fc2_dequant
scale_4 = None
output_dtype_quant = self.dtype
if output_dtype_quant == trt.fp8 and scale_4 is None:
raise RuntimeError(
"Cannot output FP8 value without knowing quantization parameter"
)
else:
hidden_states_quant = hidden_states
dtype_quant = self.dtype
weight_dtype_quant = self.weight_dtype
output_dtype_quant = self.dtype
scale_1 = self.fc.per_channel_scale
scale_2 = self.proj.per_channel_scale
scale_3 = None
scale_4 = None
output = _moe_plugin(self.moe_config,
hidden_states_quant,
routing,
expert_weight_1=self.fc.weight.value,
expert_weight_2=self.proj.weight.value,
expert_bias_1=self.fc.bias,
expert_bias_2=self.proj.bias,
expert_scale_1=scale_1,
expert_scale_2=scale_2,
expert_scale_3=scale_3,
expert_scale_4=scale_4,
finished=finished,
hidden_size=self.hidden_size,
ffn_hidden_size=self.ffn_hidden_size,
act_fn=self.hidden_act,
dtype=dtype_quant,
weight_dtype=weight_dtype_quant,
output_dtype=output_dtype_quant,
quant_mode=self.quant_mode,
tp_size=self.tp_size,
tp_rank=self.tp_rank)
if self.tp_size > 1 and self.tp_group is not None and self.moe_config.tp_mode != MoeConfig.ParallelismMode.NONE:
output = allreduce(output, self.tp_group)
return output
MOE = MixtureOfExperts