# 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.
import math
import weakref
from collections import OrderedDict
from enum import IntEnum, IntFlag, auto
from functools import partial
from typing import List, Optional, Sequence, Tuple, Union
import numpy as np
# isort: off
import tensorrt as trt
# isort: on
from . import graph_rewriting as gw
from ._common import default_net, default_trtnet, precision
from ._utils import (bf16_array, bool_array, dim_resolve_negative,
dim_to_trt_axes, dims_array, fp16_array, fp32_array,
int32_array, int64_array, np_dtype_to_trt,
str_dtype_to_trt, trt_dtype_to_np, trt_dtype_to_str)
from .network import PluginInfo, set_np_weight, set_plugin_info
from .plugin import TRT_LLM_PLUGIN_NAMESPACE, current_all_reduce_helper
from .quantization import QuantMode
[docs]
class DimRange(object):
'''
One DimRange object stores the ranges of all the dimensions of one tensor in one optimization profile.
For example, tensor has 2 dimensions. Then the data members are:
self.min = [dim 0 min, dim 1 min]
self.opt = [dim 0 opt, dim 1 opt]
self.max = [dim 0 max, dim 1 max]
For static dimension, it has min==opt==max, thus the \p shape param in the ctor can be an integer
'''
def __init__(self, shape: List[Union[int, List[int], Tuple[int, int, int]]],
names: List[str]):
'''
Parameters:
shape: a list with length N, each element is an integer or a 3-elements tuple/list of int,
where N is the number of dimensions for a tensor.
When one element is an integer, it means that dimension is static.
Otherwise, when one element is a tuple/list, it means the dimension is dynamic.
The 3 elements in one tuple/list is ordered by (min, opt, max), and this function asserts
0 <= min <= opt <= max.
Example, for a 3 rank tensor, with 1st dimension being static and has value 16, and second dimension being dynamic with
min/opt/max values being 1/8/32, and 3rd dimension being static and has value 8.
The shape parameter could be:
[16, (1, 8, 32), 8]
It has same semantics of
[(16, 16, 16), (1, 8, 32), (8, 8, 8)]
'''
self.min = []
self.opt = []
self.max = []
self.dimension_names = names
assert len(names) == len(
shape
), "Expecting shape list and name list must have same length, got {shape=}, {name=}"
for dim in shape:
if isinstance(dim, (list, tuple)):
assert len(dim) == 3 and 0 <= dim[0] <= dim[1] <= dim[2], \
"Each dimension must specify a 3-elements tuple or list in the order of (min,opt,max), got {dim=}"
self.min.append(dim[0])
self.opt.append(dim[1])
self.max.append(dim[2])
elif isinstance(dim, int):
self.min.append(dim)
self.opt.append(dim)
self.max.append(dim)
else:
raise AttributeError(
f'Dimension should be [min, opt, max] (dynamic shape) or int (specific value). Got {type(dim)}'
)
def __eq__(self, __value: object) -> bool:
return isinstance(__value, DimRange) and \
self.dimension_names == __value.dimension_names and \
self.min == __value.min and self.opt == __value.opt and self.max == __value.max
def __repr__(self) -> str:
return str(self)
def __str__(self) -> str:
return f"{self.dimension_names=} {self.min=}, {self.opt=}, {self.max=})"
def __hash__(self) -> int:
return hash(str(self))
[docs]
class Tensor(object):
'''
The class to represent dense tensors.
A dense tensor is named, has a shape and contains typed elements. Each
dimension of a tensor can either be static or dynamic. Static dimensions
are known at engine compilation by TensorRT. Dynamic dimensions can take
values determined at runtime. The tensor can be located on the host (CPU)
or the device (GPU).
'''
def __init__(self,
name=None,
dtype=None,
shape=None,
dim_range=None,
is_network_input=True,
location=trt.TensorLocation.DEVICE,
network=None,
trt_tensor=None):
'''
Parameters:
name : str
The name of the tensor.
dtype : tensorrt.DataType
The type of the elements of the tensor. See the TensorRT
documentation for list of supported data types.
shape : tensorrt.Dims
The dimensions of the tensor. In TensorRT-LLM, tensors can have
static or dynamic dimensions (it is possible to mix static and
dynamic dimensions). A static dimension is known when the
TensorRT engine is built. A dynamic dimension can be set when
the engine is executed. Use -1 for dynamic dimensions.
dim_range : OrderedDict
An ordered dictionary (the positions of the elements matter)
that associates a name and a range of values to the dimensions.
For a static dimension, the range must be limited to a single
value. For a dynamic dimension, the range is defined by three
values [min, opt, max] where min and max are, respectively, the
smallest and largest possible values of that dimension. The
opt value is used by TensorRT to optimize the engine for the
most common case.
Assume there is N optimization profiles, each item dim_range dict is ordered by:
(dynamic dimension name : [profile 0 (min, opt, max), profile 1 (min, opt, max), ... profile N(min, opt, max)])
or it's following when the dimension is static (can think as min==opt==max):
(static dimension name : [profile 0 value, profile 1 value, ... profile N value])
For static dimension the profile 0-N value must be same, (TODO: can it be simplified to be only 1 value?)
And number of keys is equal to number of optimization profiles.
is_network_input : bool
A boolean indicating if that tensor is an input of the network.
Inputs must be provided by the user to run the engine.
location : tensorrt.TensorLocation
A flag to indicate where the tensor will be located. It can be
on the host (CPU) or the device (GPU).
network: Network
A parent Network instance, that helps to fine the users of this tensor.
trt_tensor: trt.ITensor
Construct with the ITensor instance directly, and no shape profiles are required.
'''
# Layout of self.profiles
# Opt profile 0: dim 0 (min, opt, max), dim 1 (min, opt, max) ... dim M
# Opt profile 1: dim 0 (min, opt, max), dim 1 (min, opt, max) ... dim M
# ...
# Opt profile N: dim 0 ... dim M
# So from the dim_range arg to self.profiles conversion, there is a layout transpose
# dim_range arg is: {M dimension x N profiles}, while self.profiles layout is {N profiles x M dimensions}
if isinstance(dtype, str):
dtype = str_dtype_to_trt(dtype)
self.profiles = []
self.is_tensor_wrapper = False # specially for the graph rewriter
# work as a wrapper for a trt.ITensor, this is used specially in the graph rewriter
if trt_tensor is not None:
self.is_tensor_wrapper = True
assert network is not None
self.trt_tensor = trt_tensor
self._network = weakref.ref(network)
assert not is_network_input, "is_network_input should be False when trt_tensor is not None"
return
# be cautious here, the weakref is critical to avoid circular referencing before Network and Tensor
# using strong reference will likely cause significant peak memory increase, since Network objects
# holds the weights data.
self._network = weakref.ref(default_net())
self.is_network_input = is_network_input
if is_network_input:
if dim_range is not None:
assert isinstance(dim_range, OrderedDict)
assert len(
dim_range
) >= 1, f"Each input tensor shall have at least one dimension, tensor '{name}' found {dim_range=}"
found_profiles = [
len(ranges) for _, ranges in dim_range.items()
]
assert all(
[x == found_profiles[0] for x in found_profiles]
), f"Expecting all the dimensions in the dim_range has same number of profiles, tensor '{name}' got {dim_range=}"
num_opt_profile = len(list(dim_range.items())[0][1])
assert num_opt_profile >= 1
for i in range(num_opt_profile):
range_shape = []
dimension_names = []
for dim, ranges in dim_range.items():
assert isinstance(ranges, (list, tuple))
range_shape.append(ranges[i])
dimension_names.append(dim)
self.profiles.append(DimRange(range_shape, dimension_names))
default_net()._add_input(self, name, dtype, shape, dim_range)
self.name = name
self.dtype = dtype
self.shape = shape
self.location = location
@property
def network(self):
return self._network()
@property
def name(self):
'''
The name of the tensor.
'''
return self.trt_tensor.name
@name.setter
def name(self, name):
'''
Set the name of the tensor.
'''
if name is not None:
self.trt_tensor.name = name
@property
def dtype(self):
'''
The type of the elements in the tensor.
'''
return self.trt_tensor.dtype
@dtype.setter
def dtype(self, dtype):
'''
Set the type of the elements in the tensor.
'''
if dtype is not None:
self.trt_tensor.dtype = dtype
@property
def shape(self):
'''
The shape of the tensor.
'''
return self.size()
@shape.setter
def shape(self, shape):
'''
Set the shape of the tensor. See __init__.
'''
if shape is not None:
self.trt_tensor.shape = shape
@property
def location(self):
'''
The physical location of the tensor (on the host or the device).
'''
return self.trt_tensor.location
@location.setter
def location(self, location):
'''
Set the physical location of the tensor (on the host or the device). See __init__.
'''
if location is not None:
self.trt_tensor.location = location
[docs]
def mark_output(self,
name: Optional[str] = None,
dtype: Optional[Union[str, trt.DataType]] = None):
'''
Mark a tensor as a network output.
When a tensor is marked as an output, its content can be obtained after
the execution of the TensorRT engine. The user is responsible for
allocating buffers to store the output tensors when preparing the
execution of the TensorRT engine.
'''
if name is None:
name = self.name
if isinstance(dtype, str):
dtype = str_dtype_to_trt(dtype)
assert dtype is None or isinstance(dtype, trt.DataType)
default_net()._mark_output(self, name, dtype)
def __add__(self, b):
'''
See functional.add.
'''
return add(self, b)
def __radd__(self, b):
'''
See functional.add.
'''
return add(b, self)
def __sub__(self, b):
'''
See functional.sub.
'''
return sub(self, b)
def __rsub__(self, b):
'''
See functional.sub.
'''
return sub(b, self)
def __mul__(self, b):
'''
See functional.mul.
'''
return mul(self, b)
def __rmul__(self, b):
'''
See functional.mul.
'''
return mul(b, self)
def __truediv__(self, b):
'''
See functional.div.
'''
return div(self, b)
def __floordiv__(self, b):
'''
See functional.floordiv.
'''
return floordiv(self, b)
def __mod__(self, b):
'''
See functional.floordiv.
'''
return modulo(self, b)
def __lt__(self, b):
'''
See functional.lt.
'''
return lt(self, b)
def __gt__(self, b):
'''
See functional.gt.
'''
return gt(self, b)
def __eq__(self, b):
'''
See functional.eq.
'''
if self.is_tensor_wrapper:
# for graph rewriter
return hash(self) == hash(b)
else:
# for creating the network
return eq(self, b)
def __ge__(self, b):
'''
Maps to functional.gt or functional.eq.
'''
return op_or(self.__gt__(b), self.__eq__(b))
def __le__(self, b):
'''
Maps to functional.lt or functional.eq.
'''
return op_or(self.__lt__(b), self.__eq__(b))
[docs]
def view(self, shape, zero_is_placeholder=True):
'''
See functional.view.
'''
return view(self, shape, zero_is_placeholder)
[docs]
def flatten(self, start_dim=0, end_dim=-1):
'''
See functional.flatten.
'''
return flatten(self, start_dim, end_dim)
[docs]
def permute(self, dims):
'''
See functional.permute.
'''
return permute(self, dims)
[docs]
def transpose(self, dim0, dim1):
'''
See functional.transpose.
'''
return transpose(self, dim0, dim1)
[docs]
def mean(self, dim, keepdim=False):
'''
See functional.mean.
'''
return mean(self, dim, keepdim)
[docs]
def max(self, dim, keepdim=False):
'''
See functional.max.
'''
return max(self, dim, keepdim)
[docs]
def abs(self):
'''
See functional.abs.
'''
return abs(self)
[docs]
def sqrt(self):
'''
See functional.sqrt.
'''
return sqrt(self)
[docs]
def log(self):
'''
See functional.log.
'''
return log(self)
[docs]
def cast(self, dtype):
'''
See functional.cast.
'''
return cast(self, dtype)
[docs]
def size(self, dim=None):
'''
Returns the shape of the tensor if the dim parameter is None.
Otherwise, returns a size of the dimension indicated by dim. The
behavior is undefined if dim is negative or exceeds the rank of the
tensor.
'''
if dim is None:
return self.trt_tensor.shape
return self.trt_tensor.shape[dim]
[docs]
def rank(self):
'''
Returns the rank (i.e. the number of dimensions) of the tensor.
'''
return len(self.trt_tensor.shape)
[docs]
def ndim(self):
'''
Returns the rank (i.e. the number of dimensions) of the tensor.
'''
return self.rank()
[docs]
def split(self, split_size_or_sections, dim=0):
'''
See functional.split.
'''
return split(self, split_size_or_sections, dim)
[docs]
def unbind(self, dim=0):
'''
See functional.unbind.
'''
return unbind(self, dim)
[docs]
def is_dynamic(self, dim=None):
'''
If the argument 'dim' is None, that function returns a boolean that
indicates if the tensor contains a dynamic dimension (True) or not
(False). In that case, the first dimension is excluded (as it usually
corresponds to the batch size). If the argument is an integer, that
functions returns a boolean that indicates if the dimension 'dim' is
dynamic (True) or not (False).
'''
if dim is not None:
return self.trt_tensor.shape[dim] == -1
for i, s in enumerate(self.trt_tensor.shape):
if i != 0 and s == -1:
return True
return False
# graph writer related functions
[docs]
def get_parent(self):
''' Get the layer that produces this tensor. '''
return self.network.get_tensor_parent(self)
[docs]
def get_users(self):
''' Get the layers that use this tensor as an input. '''
return self.network.get_tensor_users(self)
[docs]
def replace_all_uses_with(self, new_tensor):
'''
Replace all uses of this tensor as an input to consumer layers
'''
self.network.is_graph_altered = True
users = self.get_users()
for user in users:
inputs_changed = 0
for i in range(user.num_inputs):
if user.get_inputs(i)[0].trt_tensor is self.trt_tensor:
inputs_changed += 1
user.set_input(i, new_tensor.trt_tensor)
assert inputs_changed >= 1, "Tensor not found in layer inputs"
# update the FLayerMetadata as well
flayer = gw.FLayerInfoMemo.instance().get(user.name)
flayer and flayer.replace_input_with(self, new_tensor)
[docs]
def is_trt_wrapper(self):
'''
Check if there is a trt.ITensor member inside, which is required for
graph rewriter. In order to differentiate usages, it may be necessary
to have an inheritance hierarchy.
'''
if hasattr(self, 'trt_tensor'):
return True
else:
return False
def __hash__(self):
if self.is_trt_wrapper():
return id(self.trt_tensor)
else:
return id(None)
def __repr__(self):
return f"TensorRT-LLM Tensor: {self.name=} {self.dtype=} {self.shape=}"
def _create_tensor(trt_tensor: trt.ITensor, producer: trt.ILayer) -> Tensor:
'''
A helper function to create a TensorRT-LLM Tensor object that encapsulates
the connection between the TensorRT tensor (trt.ITensor) and the layer
(trt.ILayer) that produces it.
That function is expected to be used as:
# Insert a new layer in the network using the TensorRT API:
layer = default_trtnet().add_<some_layer>(...)
# Extract the first output of that layer and connect it to the layer.
return _create_tensor(layer.get_output(0), layer)
That function also sets the precision of the layer/producer to the default
precision of the network.
Parameters:
trt_tensor : trt.ITensor
The TensorRT tensor to connect to its producer (the layer).
producer : trt.ILayer
The producer.
Returns:
The TensorRT-LLM tensor (functional.Tensor) that encapsulates the
TensorRT tensor and the layer that produces it. The former is
accessible through the attribute 'trt_tensor' and the latter using the
attribute 'producer'.
'''
assert trt_tensor is not None
assert producer is not None
# Set the layer name since this is the only
# centralized location to pass the name from
# module space to the TRT IR
default_net()._set_layer_name(producer)
assert trt_tensor.shape.__len__(
) >= 0, f"tensor {trt_tensor.name} has an invalid shape"
tensor = Tensor(name=trt_tensor.name,
dtype=trt_tensor.dtype,
shape=trt_tensor.shape,
is_network_input=False)
tensor.trt_tensor = trt_tensor
tensor.producer = producer
# tb.print_stack(limit=10) # FOR DEBUGGING: filter producer.name if needed
if default_net().dtype is not None and not default_net().strongly_typed:
if producer.type not in [
trt.LayerType.SHAPE, trt.LayerType.CONSTANT,
trt.LayerType.GATHER, trt.LayerType.CONCATENATION
]:
producer.precision = default_net().dtype
assert tensor is not None
if gw.FLayerInfoMemo.instance().cur_flayer is not None:
gw.FLayerInfoMemo.instance().cur_flayer.layer_name = producer.name
return tensor
def _add_plugin_info(layer, plugin_creator: trt.IPluginCreator,
plugin_name: str, pfc: trt.PluginFieldCollection) -> None:
plugin_info = PluginInfo(plugin_creator, plugin_name, pfc)
set_plugin_info(default_net().trt_network, layer.name, plugin_info)
[docs]
class RotaryScalingType(IntEnum):
none = 0
linear = 1
dynamic = 2
longrope = 3
llama3 = 4
[docs]
@staticmethod
def from_string(s):
try:
return RotaryScalingType[s]
except KeyError:
raise ValueError(f'Unsupported rotary scaling type: {s}')
[docs]
class PositionEmbeddingType(IntEnum):
learned_absolute = 0
rope_gptj = 1
rope_gpt_neox = 2
long_rope = 3
alibi = 4
alibi_with_scale = 5
relative = 6
chatglm = 7
[docs]
def is_rope(self) -> bool:
return self in [self.rope_gptj, self.rope_gpt_neox, self.long_rope]
[docs]
def is_alibi(self) -> bool:
return self in [self.alibi, self.alibi_with_scale]
[docs]
@staticmethod
def choices() -> List[str]:
return [embedding.name for embedding in PositionEmbeddingType]
def __str__(self):
return self.name
[docs]
@staticmethod
def from_string(s):
try:
return PositionEmbeddingType[s]
except KeyError:
raise ValueError(f'Unsupported position embedding type: {s}')
[docs]
class AttentionMaskType(IntEnum):
padding = 0
causal = 1
sliding_window_causal = 2
bidirectional = 3
bidirectionalglm = 4 # TODO: merge this mask into bidirectional
blocksparse = 5
custom_mask = 6
[docs]
class LayerNormType(IntEnum):
LayerNorm = 0
RmsNorm = 1
GroupNorm = 2
[docs]
class LayerNormPositionType(IntEnum):
pre_layernorm = 0
post_layernorm = 1
[docs]
class MLPType(IntEnum):
MLP = 0
GatedMLP = 1
FusedGatedMLP = 2
[docs]
def activation(input: Tensor, act_type: trt.ActivationType) -> Tensor:
'''
Add an activation function.
Parameters:
input : Tensor
The input tensor on which the activation function is applied.
act_type : trt.ActivationType
The type of the activation (RELU, TANH, SIGMOID, ...).
The following closures are defined in functional.*:
relu for op=trt.ActivationType.RELU
tanh for op=trt.ActivationType.TANH
sigmoid for op=trt.ActivationType.SIGMOID
Returns:
The tensor produced by the activation layer.
'''
layer = default_trtnet().add_activation(input.trt_tensor, act_type)
return _create_tensor(layer.get_output(0), layer)
[docs]
def int_clip(input: Tensor, lower: int, upper: int) -> Tensor:
assert lower <= upper, f"Lower bound must be less than or equal to upper bound i.e. {lower} <= {upper}"
res = minimum(input, upper)
res = maximum(res, lower)
return res
[docs]
def clip(input: Tensor, alpha: float, beta: float) -> Tensor:
'''
Add a CLIP operation that sets the range to [alpha, beta].
Parameters:
input : Tensor
The input tensor on which the activation function is applied.
alpha : float
The lower bound of the CLIP function.
beta : float
The upper bound of the CLIP function.
Returns:
The tensor produced by the activation layer.
'''
layer = default_trtnet().add_activation(input.trt_tensor,
trt.ActivationType.CLIP)
layer.alpha = alpha
layer.beta = beta
return _create_tensor(layer.get_output(0), layer)
relu = partial(activation, act_type=trt.ActivationType.RELU)
tanh = partial(activation, act_type=trt.ActivationType.TANH)
sigmoid = partial(activation, act_type=trt.ActivationType.SIGMOID)
[docs]
def silu(input: Tensor) -> Tensor:
'''
Add a SiLU (`x * sigmoid(x)`) operation.
Parameters:
input : Tensor
The input tensor on which the activation function is applied.
Returns:
The tensor produced by the activation layer.
'''
return input * sigmoid(input)
[docs]
def swiglu(input: Tensor) -> Tensor:
'''
Add a SwiGLU (`x * SiLU(gate)`) operation.
That function takes a tensor, splits it into two halves along the last
dimension, applies SiLU to the second half and multiply the results. The
behavior is undefined if the last dimension is not even.
Parameters:
input : Tensor
The input tensor on which the activation function is applied.
Returns:
The tensor produced by the activation layer.
'''
x, gate = chunk(input, 2, dim=-1)
return silu(gate) * x
[docs]
def squared_relu(x: Tensor) -> Tensor:
'''
Add a Squared ReLU operation.
This function applies ReLU and squares the output.
Parameters:
input : Tensor
The input tensor on which the activation function is applied.
Returns:
The tensor produced by the activation layer.
'''
return pow(relu(x), 2.0)
[docs]
def cast(input: Tensor, dtype: Union[str, trt.DataType]):
'''
Add a cast operation.
For an input tensor of type INT8, this function sets the dynamic range of
the input to [-127, 127] for automatic dequantization. For a cast into
INT8, that function sets the dynamic range of the output to [-127, 127] for
automatic quantization.
Parameters:
input : Tensor
The input tensor on which the cast is applied.
dtype : str or trt.DataType
The data type of the output tensor after the cast. When 'dtype' is
provided as a string, it must be a name amongst the valid names.
See _str_to_trt_dtype_dict in _utils.py for a list of supported
types and type names.
Returns:
The tensor produced by the inserted layer.
'''
if isinstance(dtype, str):
cvt_dtype = str_dtype_to_trt(dtype)
elif isinstance(dtype, trt.DataType):
cvt_dtype = dtype
else:
raise TypeError("%s is not supported" % type(dtype))
if input.dtype == cvt_dtype:
# If input type and cast dtype are the same, do nothing
return input
layer = default_trtnet().add_cast(input.trt_tensor, cvt_dtype)
if not default_net().strongly_typed:
layer.set_output_type(0, cvt_dtype)
output = _create_tensor(layer.get_output(0), layer)
if not default_net().strongly_typed:
if input.dtype == str_dtype_to_trt('int8'):
layer.get_input(0).set_dynamic_range(-127, 127)
if cvt_dtype == str_dtype_to_trt('int8'):
layer.get_output(0).set_dynamic_range(-127, 127)
return output
[docs]
def flip(input: Tensor, dims: Sequence[int]) -> Tensor:
'''
Reverses the order of an n-D tensor along given axis in dims.
That flip operation maps to a TensorRT ISliceLayer. For the dimensions
listed in dims it copies the elements from the last one to the first one
(from (N-1) down to 0 with a step of -1). For the dimensions not in 'dims',
it copies the elements from the first one to the last one (from 0 to N-1
with a step of 1).
Parameters:
input : Tensor
The input tensor on which the cast is applied.
dims : list or tuple
The axes to flip. Negative indices are supported.
Returns:
The tensor produced by the inserted layer.
'''
assert not input.is_dynamic()
ndim = input.ndim()
for index, value in enumerate(dims):
assert -ndim <= value < ndim
if -ndim <= value < 0:
dims[index] += ndim
assert len(dims) == len(set(dims))
start_values = [
input.size()[i] - 1 if i in dims else 0 for i in range(ndim)
]
stride_values = [-1 if i in dims else 1 for i in range(ndim)]
layer = default_trtnet().add_slice(input.trt_tensor,
start=start_values,
shape=input.size(),
stride=stride_values)
return _create_tensor(layer.get_output(0), layer)
[docs]
def interpolate(input: Tensor,
size: Union[int, List[int]] = None,
scale_factor: Union[float, List[float]] = None,
mode: str = 'nearest',
align_corners: bool = False,
recompute_scale_factor: bool = False,
antialias: bool = False) -> Tensor:
##
## TODO: Document that function!
##
assert not input.is_dynamic()
input_ndim = input.ndim()
assert 2 < input_ndim < 6, "Only 3D, 4D and 5D input Tensors supported"
assert (size is not None) ^ (
scale_factor
is not None), "Only one of out_shape or scales should be defined"
assert mode in ('nearest', 'linear', 'bilinear', 'bicubic', 'trilinear',
'nearest-exact')
if mode == 'trilinear' and input_ndim != 5:
raise ValueError("trilinear only supports 5D tensor")
if mode == "bilinear" and input_ndim != 4:
raise ValueError("bilinear only supports 4D tensor")
if mode == "linear" and input_ndim != 3:
raise ValueError("linear only supports 3D tensor")
layer = default_trtnet().add_resize(input.trt_tensor)
input_shape = input.size()
updated_shape = []
if scale_factor:
scale_len = 1 if isinstance(scale_factor,
(float, int)) else len(scale_factor)
if scale_len == 1 and isinstance(scale_factor, (float, int)):
updated_scale = [scale_factor for _ in range(input_ndim - 2)]
else:
updated_scale = scale_factor
updated_shape = [
int(math.floor(updated_scale[i - 2] *
input_shape[i])) if i > 1 else input_shape[i]
for i in range(input_ndim)
]
else:
size_len = 1 if isinstance(size, int) else len(size)
assert size_len == input_ndim - 2
if size_len == 1 and isinstance(size, int):
updated_size = [size for _ in range(input_ndim - 2)]
else:
updated_size = size
updated_shape = [
input_shape[i] if i < 2 else updated_size[i - 2]
for i in range(input_ndim)
]
layer.shape = updated_shape
if mode in ['nearest', 'nearest-exact'] or mode is None:
layer.resize_mode = trt.InterpolationMode.NEAREST
layer.coordinate_transformation = trt.ResizeCoordinateTransformation.ASYMMETRIC
elif mode in ['linear', 'bilinear', 'trilinear']:
layer.resize_mode = trt.InterpolationMode.LINEAR
if align_corners:
layer.coordinate_transformation = trt.ResizeCoordinateTransformation.ALIGN_CORNERS
else:
layer.coordinate_transformation = trt.ResizeCoordinateTransformation.HALF_PIXEL
# TODO, need to confirm the align_corners effect on bilinear mode.
if mode == 'bilinear':
layer.coordinate_transformation = trt.ResizeCoordinateTransformation.HALF_PIXEL
elif mode in ['bicubic']:
layer.resize_mode = trt.InterpolationMode.CUBIC
layer.coordinate_transformation = trt.ResizeCoordinateTransformation.HALF_PIXEL
else:
layer.resize_mode = trt.InterpolationMode.NEAREST
layer.coordinate_transformation = trt.ResizeCoordinateTransformation.ASYMMETRIC
return _create_tensor(layer.get_output(0), layer)
[docs]
def matmul(input: Tensor,
mat2: Tensor,
transa: bool = False,
transb: bool = False,
use_fp32_acc: bool = True) -> Tensor:
'''
Add a matrix multiplication.
That operation maps to a tensorrt.IMatrixMultiplyLayer layer. As explained
in the TensorRT documentation, it computes the inner product between the
two inputs after applying an optional transposition on the inputs.
Parameters:
input : Tensor
The first tensor (often called A).
mat2 : Tensor
The second tensor (often called B).
transa : bool
Is the first input transposed? Set to 'True' if you want the first
input to be transposed, 'False' otherwise.
transb : bool
Is the second input transposed? Set to 'True' if you want the
second input to be transposed, 'False' otherwise.
use_fp32_acc: bool
Set to 'True' if for accuracy reason, this fp16 matmul needs to use
fp32 accumulation. This can be a per model and per matmul decision.
Returns:
The tensor produced by the inserted layer.
'''
# This option is only supported for fp16, but not bf16 or any other precisions.
use_fp32_acc = use_fp32_acc and input.dtype == trt.DataType.HALF and mat2.dtype == trt.DataType.HALF
if use_fp32_acc:
input = cast(input, 'float32')
mat2 = cast(mat2, 'float32')
input, mat2 = broadcast_helper(input, mat2)
op0 = trt.MatrixOperation.TRANSPOSE if transa \
else trt.MatrixOperation.NONE
op1 = trt.MatrixOperation.TRANSPOSE if transb \
else trt.MatrixOperation.NONE
layer = default_trtnet().add_matrix_multiply(input.trt_tensor, op0,
mat2.trt_tensor, op1)
output = _create_tensor(layer.get_output(0), layer)
if use_fp32_acc:
output = cast(output, "float16")
return output
[docs]
def gemm_swiglu(input: Tensor,
weight: Tensor,
bias: Optional[Tensor] = None,
scale_d0: float = 1.0,
scale_d1: float = 1.0,
scale_output: float = 1.0) -> Tensor:
'''
Add a matrix multiplication, followed by SwiGLU (`x * SiLU(gate)`) operation.
The second SwiGLU operation takes the preceding tensor, splits it into two halves
along the last dimension, applies SiLU to the second half and multiply the results. The
behaviour is undefined if the last dimension is not even.
Parameters:
input : Tensor
The first tensor (often called A).
weight : Tensor
The second tensor (often called B).
bias : Optional[Tensor]
The per-channel bias. The plugin with fp8 dtype does not support bias yet.
scale_d0 : float
The scale for dequantizing x, used for fp8
scale_d1 : float
The scale for dequantizing gate, used for fp8
scale_output : float
The scale for quantizing output, used for fp8
Returns:
The tensor produced by the inserted layer.
'''
plg_creator = trt.get_plugin_registry().get_plugin_creator(
'GemmSwiglu', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert plg_creator is not None
p_dtype = default_net().plugin_config.gemm_swiglu_plugin
if p_dtype == "fp8":
assert bias == None, "fp8 gemm_swiglu does not support bias yet"
pf_type = trt.PluginField(
"type_id", np.array([int(str_dtype_to_trt(p_dtype))], np.int32),
trt.PluginFieldType.INT32)
pf_has_bias = trt.PluginField(
"has_bias", np.array(np.int8(0 if bias is None else 1), np.int8),
trt.PluginFieldType.INT8)
pf_scale_d0 = trt.PluginField("scale_d0",
np.array(scale_d0, dtype=np.float32),
trt.PluginFieldType.FLOAT32)
pf_scale_d1 = trt.PluginField("scale_d1",
np.array(scale_d1, dtype=np.float32),
trt.PluginFieldType.FLOAT32)
pf_scale_output = trt.PluginField("scale_output",
np.array(scale_output, dtype=np.float32),
trt.PluginFieldType.FLOAT32)
pfc = trt.PluginFieldCollection(
[pf_type, pf_has_bias, pf_scale_d0, pf_scale_d1, pf_scale_output])
gemm_swiglu_plug = plg_creator.create_plugin("gemm_swiglu", pfc)
# TODO(anchengc) pass nullptr when no bias
if bias is None:
bias = constant(
np.zeros([weight.shape[0]], dtype=trt_dtype_to_np(input.dtype)))
plug_inputs = [input.trt_tensor, weight.trt_tensor, bias.trt_tensor]
layer = default_trtnet().add_plugin_v2(plug_inputs, gemm_swiglu_plug)
return _create_tensor(layer.get_output(0), layer)
[docs]
def constant(ndarray: np.ndarray) -> Tensor:
'''
Add a constant layer.
TensorRT graphs encapsulate constant values in the form of constant layers
(tensorrt.IConstantLayer). That function creates such a layer from a Numpy
array of values. After compilation of the network by TensorRT, those
weights are stored in the serialized TensorRT engine.
Parameters:
ndarray : numpy.ndarray
The array of values (weights) encapsulated by this constant layer.
Returns:
The tensor produced by the inserted layer.
'''
weights = trt.Weights(np_dtype_to_trt(ndarray.dtype), ndarray.ctypes.data,
ndarray.size)
# Prevent underlying numpy array from going out of scope
default_net().register_ndarray(ndarray)
layer = default_trtnet().add_constant(trt.Dims(ndarray.shape), weights)
if not default_net().strongly_typed:
layer.set_output_type(0, np_dtype_to_trt(ndarray.dtype))
tensor = _create_tensor(layer.get_output(0), layer)
# TODO: remove this WAR after https://nvbugs/4359151 fixed.
set_np_weight(default_trtnet(), layer.name, ndarray)
return tensor
# TODO: TensorRT uses sizes of the output dimensions.
# DL framework uses ends usually. Will change it to ends.
[docs]
def slice(input: Tensor,
starts: Union[Tensor, Sequence[int]],
sizes: Union[Tensor, Sequence[int]],
strides: Union[Tensor, Sequence[int]] = None,
mode: trt.SampleMode = None) -> Tensor:
'''
Add an operation to extract a slice from a tensor.
As described in the TensorRT documentation of the ISliceLayer, the slice
layer has two variants: Static and dynamic.
For static slicing, this function takes the starts and sizes values in the
different dimensions to slice at layer creation time via a sequence of
integers. For dynamic slicing, it accepts starts and sizes as
tensorrt.ITensor`s.
The slice layer selects for each dimension a start location from within the
input tensor, and copies elements to the output tensor using a stride of 1
across the input tensor. Start and size tensors must be 1-D int32 shape
tensors if not specified as a sequence of integers.
As an example, on input = [[0, 2, 4], [1, 3, 5]], the call to
slice(input, start=[1, 0], size=[1, 2])
will produce the tensor [[1, 3]] as output. The slice operator when
executed by TensorRT will copy one row (because size[0] == 1) starting from
the 2nd row (because start[0] == 1) and two columns (size[1] == 2) starting
from the 1st column (because start[1] == 0).
In pseudo-code the behavior of that operation can be described as follows
for a 2D tensor (and easily be extended to more dimensions):
output = Tensor(shape=sizes)
for ii in range(sizes[0]):
for jj in range(sizes[1]):
output[ii][jj] = input[starts[0]+ii][starts[1]+jj]
Note that it is common in deep-learning frameworks to use ranges
[start:end] for similar operations. It can be emulated by setting the sizes
argument such that in each dimension [start:start+size] == [start:end] i.e.
size = end-start.
TensorRT supports different slice modes but that function restricts that
choice to `mode == tensorrt.SampleMode.STRICT_BOUNDS`.
Parameters:
input : Tensor
The input tensor on which the slicing is performed.
starts : Union[Tensor, Sequence[int]]
The starting points, in the input tensor, and each dimension.
sizes : Union[Tensor, Sequence[int]]
The number of elements in each dimension of the sliced tensor (output).
strides : Union[Tensor, Sequence[int]]
The step be taken from start, in input tensor.
mode : trt.SampleMode
The mode that controls how the slice operation handles out of bounds coordinates.
Returns:
The tensor produced by the slice layer.
'''
input_ndim = input.ndim()
trt_starts = starts
if isinstance(starts, Tensor):
trt_starts = [0 for _ in range(input_ndim)] # unused dummy value
trt_sizes = sizes
if isinstance(sizes, Tensor):
trt_sizes = [1 for _ in range(input_ndim)] # unused dummy value
trt_strides = strides
if isinstance(strides, Tensor) or strides is None:
trt_strides = [1 for _ in range(input_ndim)]
layer = default_trtnet().add_slice(input.trt_tensor,
start=trt_starts,
shape=trt_sizes,
stride=trt_strides)
if mode is not None:
layer.mode = mode
if isinstance(starts, Tensor):
layer.set_input(1, starts.trt_tensor)
if isinstance(sizes, Tensor):
layer.set_input(2, sizes.trt_tensor)
if isinstance(strides, Tensor):
layer.set_input(3, strides.trt_tensor)
return _create_tensor(layer.get_output(0), layer)
[docs]
def rand(shape: Tensor,
low: float = 0,
high: float = 1,
dtype: Union[str, trt.DataType] = 'float32') -> Tensor:
'''
This operation adds a fill layer that generates a random (uniform) tensor with the specified shape and data type.
Parameters:
shape: Tensor
The shape of the tensor needed to be generated.
low: float
The minimum value (inclusive) of the range used for random.
high: float
The maximum value (inclusive) of the range used for random.
dtype: Union[str, trt.DataType]
The desired data type for the output tensor.
Returns:
The generated random tensor produced by the fill layer.
'''
# NOTE: DISABLED FOR NOW UNTIL THE FILL LAYER (RANDOM_UNIFORM) in TRT IS FIXED
assert False, "The rand() op is temporarily disabled."
low = constant(fp32_array(low))
high = constant(fp32_array(high))
trt_dtype = dtype if isinstance(dtype,
trt.DataType) else str_dtype_to_trt(dtype)
layer = default_trtnet().add_fill([0], trt.FillOperation.RANDOM_UNIFORM,
trt_dtype)
layer.set_input(0, shape.trt_tensor)
layer.set_input(1, low.trt_tensor)
layer.set_input(2, high.trt_tensor)
return _create_tensor(layer.get_output(0), layer)
[docs]
def categorical_sample(probs: Tensor, rand_data: Tensor = None) -> Tensor:
'''
This is a sampling operation and an equivalent of torch.distributions.Categorical.sample()
i.e. given a probability distribution tensor, it samples an index of that tensor.
See: https://pytorch.org/docs/stable/distributions.html#torch.distributions.categorical.Categorical.sample
NOTE: This assumes that the given probabilities are **not** normalized.
Parameters:
probs: Tensor
A 1-D floating point tensor representing the probability distributions.
rand_data: Tensor (optional)
A random tensor of same shape as `probs` tensor.
If not provided, this function will add a rand() op to generate it and use for sampling.
Returns:
A tensor containing a single index of the `probs` tensor representing the sample.
'''
probs = probs / sum(probs, dim=-1, keepdim=True)
rand_shape = []
assert probs.ndim() > 0
for i in range(probs.ndim() - 1):
rand_shape.append(shape(probs, i))
rand_shape = concat(rand_shape)
if rand_data is None:
rand_data = rand(rand_shape, low=0, high=1, dtype=probs.dtype)
assert rand_shape == shape(rand_data)
rand_data = expand(unsqueeze(rand_data, -1), shape(probs))
cum_probs = cumsum(probs, dim=-1)
cmp = cast(cum_probs >= rand_data, probs.dtype)
samples = argmax(cmp, dim=-1)
return samples
[docs]
class Conditional:
'''
Add an operation to conditionally execute two code paths/subgraphs.
Usage:
1. conditional = Conditional(condition)
2. input_1_ = conditional.add_input(input_1)
...
input_n_ = conditional.add_input(input_n)
3. Construct the graph to get true_output_value and false_output_value using input_1_, ..., input_n_
4. output = conditional.add_output(true_output_value, false_output_value)
'''
def __init__(self, condition: Tensor):
self.layer = default_trtnet().add_if_conditional()
if condition.ndim() > 0:
condition = view(condition, [])
self.layer.set_condition(condition.trt_tensor)
[docs]
def add_output(self, true_value: Tensor, false_value: Tensor) -> Tensor:
out_node = self.layer.add_output(true_value.trt_tensor,
false_value.trt_tensor)
return _create_tensor(out_node.get_output(0), out_node)
# TODO: support step.
[docs]
def arange(start: Union[Tensor, int], end: Union[Tensor, int],
dtype: str) -> Tensor:
'''
Add an operation to fill a 1D tensor.
The tensor is filled with the values between start and end with a step of 1
between the different elements. In pseudo-code, it corresponds to a tensor
populated with the values:
output = Tensor([dtype(ii) for ii in range(start, end, 1)])
For example, a call to arange(3, 6, 'int32') will add an operation to the
TensorRT graph that will produce [3, 4, 5] when executed. The call to
arange(2, 5, 'float32') will add a layer to generate [2.0, 3.0, 4.0].
This operation is implemented using a tensorrt.IFillLayer in
trt.FillOperation.LINSPACE mode.
Parameters:
start : Union[Tensor, int]
The starting point of the range.
end : Union[Tensor, int]
The end point of the range.
dtype : str
The type of the elements. See _str_to_trt_dtype_dict in _utils.py
for a list of supported types and type names.
Returns:
The tensor produced by the fill layer. It is a 1D tensor containing
`end-start` elements of type `dtype`.
'''
res_dtype = str_dtype_to_trt(dtype)
if isinstance(start, int):
assert isinstance(end, int)
array_func = int32_array if res_dtype == trt.int32 else int64_array
start = constant(array_func(start))
end = constant(array_func(end))
elif isinstance(start, Tensor):
assert isinstance(end, Tensor)
assert start.dtype == trt.int32 or start.dtype == trt.int64
assert end.dtype == trt.int32 or end.dtype == trt.int64
if start.dtype != end.dtype:
if start.dtype == trt.int32: # end == trt.int64
if res_dtype == trt.int32:
end = cast(end, "int32")
else:
start = cast(start, "int64")
else: # start == trt.int64 and end == trt.int32
if res_dtype == trt.int32:
start = cast(start, "int32")
else:
end = cast(end, "int64")
else:
raise TypeError("%s is not supported" % type(start))
assert start.dtype == end.dtype, f"start type ({start.dtype}) != end type ({end.dtype})"
step = constant_to_tensor_(1, dtype=start.dtype, to_array=True)
num = end - start
num = num.view([1]).cast(trt.int64)
layer = default_trtnet().add_fill([0], trt.FillOperation.LINSPACE,
start.dtype)
layer.set_input(0, num.trt_tensor) # rank = 1
layer.set_input(1, start.trt_tensor) # rank = 0
layer.set_input(2, step.trt_tensor) # rank = 1
tensor = _create_tensor(layer.get_output(0), layer)
if tensor.dtype != res_dtype:
tensor = tensor.cast(dtype)
return tensor
[docs]
def expand(input: Tensor, expand_shape: Tensor) -> Tensor:
'''
Add an operation to expand a tensor.
The operation expands the input tensor in the singleton dimensions to the
size indicated by the corresponding dimension in the `expand_shape` tensor.
In other words, given an input tensor with dimensions of size 1, those
dimensions will be expanded to the size in `expand_shape`.
For example, a tensor of shape [4, 3, 1, 3] will be expanded to a tensor of
shape [4, 3, 2, 3] by the layer created using expand(input, [4, 3, 2, 3]).
The expansion may either replicate the values or be mapped to a view with a
stride of 0 in the expanded dimensions. For example, for a tensor [[3, 2]] of
shape [1, 2],
expand([[3, 2]], [2, 2])
can be used to expand the input to [[3, 2], [3, 2]].
This operation is implemented using a tensorrt.ISliceLayer. The current
implementation does not verify that non singleton dimensions are not
shrunk. In other words, for an input of shape [4, 1, 2],
expand(input, [3, 2, 2])
will produce a tensor of shape [3, 2, 2]. That behavior is subject to
change in the future.
Parameters:
input : Tensor
The input tensor.
expand_shape : Tensor
The new shape of the expanded tensor.
Returns:
The tensor produced by the expand layer.
'''
ndim = input.rank()
layer = default_trtnet().add_slice(
input.trt_tensor,
start=[0 for _ in range(ndim)],
shape=[1 for _ in range(ndim)], # unused dummy value
stride=[1 for _ in range(ndim)] # unused dummy value
)
# The stride is either:
# 0 for dimensions of size 1 (i.e. shape(input, i) - 1 == 1 - 1 == 0) or,
# 1 for dimensions of size > 1 since minimum(value >= 1, 1) == 1.
stride_tensor = concat(
[minimum((shape(input, i) - 1), 1) for i in range(ndim)])
layer.set_input(2, expand_shape.trt_tensor)
layer.set_input(3, stride_tensor.trt_tensor)
return _create_tensor(layer.get_output(0), layer)
[docs]
def einsum(einsum_eq: str, inputs: Sequence[Tensor]) -> Tensor:
'''
Add an Einsum operation.
That operation maps to tensorrt.IEinsumLayer. As explained in the TensorRT
documentation, this layer implements a summation over the elements of the
inputs along dimensions specified by the equation parameter, based on the
Einstein summation convention. The layer can have one or more inputs of
rank >= 0. All the inputs must be of same data type. This layer supports
all TensorRT data types except bool. There is one output tensor of the same
type as the input tensors. The shape of output tensor is determined by the
equation.
The equation specifies ASCII lower-case letters for each dimension in the
inputs in the same order as the dimensions, separated by comma for each
input. The dimensions labeled with the same subscript must match or be
able to be broadcasted. Repeated subscript labels in one input take the diagonal.
Repeating a label across multiple inputs means that those axes will be
multiplied. Omitting a label from the output means values along those axes
will be summed. In implicit mode, the indices which appear once in the
expression will be part of the output in increasing alphabetical order. In
explicit mode, the output can be controlled by specifying output subscript
labels by adding an arrow (‘->’) followed by subscripts for the output. For
example, “ij,jk->ik” is equivalent to “ij,jk”. Ellipsis (‘…’) can be used
in place of subscripts to broadcast the dimensions. See the TensorRT
Developer Guide for more details on equation syntax.
Many common operations can be expressed using the Einsum equation. For
example:
Matrix Transpose: ij->ji
Sum: ij-> Matrix-Matrix
Multiplication: ik,kj->ij
Dot Product: i,i->
Matrix-Vector Multiplication: ik,k->i
Batch Matrix Multiplication: ijk,ikl->ijl
Batch Diagonal: …ii->…i
Note that TensorRT does not support ellipsis or diagonal operations so,
neither, does TensorRT-LLM.
Parameters:
einsum_eq : str
The Einsum equation.
inputs: Sequence[Tensor]
The sequence of inputs consumed by the Einsum operation.
Returns:
The tensor produced by the Einsum operation.
'''
layer = default_trtnet().add_einsum([i.trt_tensor for i in inputs],
einsum_eq)
return _create_tensor(layer.get_output(0), layer)
[docs]
def permute(input: Tensor, dims: Sequence[int]) -> Tensor:
'''
Add an operation to permute the dimensions of a tensor.
The dimensions of the input tensor are permuted according to the sequence
of dimensions in 'dims'. That operation maps to tensorrt.IShuffleLayer where
the second transposition is described by the indices in 'dims'.
Given a tensor of rank N, the result of the permutation is a tensor of rank
N in which the i-th input dimension maps to the dims[i]-th dimension.
For example, permute(input, [1, 0]) will transpose a 2D tensor by permuting
the rows and columns.
Parameters:
input : Tensor
The input tensor to permute.
dims : Sequence[int]
The description of the permutation.
Returns:
The tensor produced by the permutation layer.
'''
dims = dim_resolve_negative(tuple(dims), input.ndim())
layer = default_trtnet().add_shuffle(input.trt_tensor)
layer.second_transpose = dims
return _create_tensor(layer.get_output(0), layer)
[docs]
def transpose(input: Tensor, dim0: int, dim1: int) -> Tensor:
'''
Add an operation to transpose two dimensions of a tensor.
That operation produces a tensor in which the dimensions 'dim0' and 'dim1'
are permuted. The other dimensions, if the rank of the tensor is greater
than 2, remain untouched.
That function is a helper built on the 'functional.permute' function.
Parameters:
input : Tensor
The input tensor to transpose.
dim0 : int
The first dimension to transpose.
dim1 : int
The second dimension to transpose.
Returns:
The tensor produced by the permutation layer.
'''
permutation = list(range(input.ndim()))
permutation[dim0] = dim1
permutation[dim1] = dim0
return permute(input, permutation)
[docs]
def view(input: Tensor,
shape: Union[Tensor, Sequence[int]],
zero_is_placeholder: bool = True) -> Tensor:
'''
Add an operation to create a view of a tensor.
That operation adds a tensorrt.IShuffleLayer to the network. If the 'shape'
parameter is a Tensor, that view is dynamic. Otherwise, it is a static
view.
Note that TensorRT limits the number of inferred dimensions to 1. It means
that the shape sequence or tensor cannot contain more than one -1. This
function enforces that constraint and will assert if it is not respected.
Parameters:
input : Tensor
The input tensor to transpose.
shape : Union[Tensor, Sequence[int]]
The shape of the new tensor.
zero_is_placeholder : bool
When that parameter is True, the 0s in 'shape' are replaced by the
sizes of the corresponding dimensions from the 'input'. Otherwise,
the dimensions corresponding to 0s are shrunk.
Returns:
The tensor produced by the view/shuffle layer.
'''
# TensorRT demands that at most one dimension is permitted to be specified as -1
def assert_no_more_than_one_inferred_dim(list):
inferred_dim_list = [i for i in list if i == -1]
assert len(inferred_dim_list) <= 1
layer = default_trtnet().add_shuffle(input.trt_tensor)
layer.zero_is_placeholder = zero_is_placeholder
if isinstance(shape, Tensor):
assert_no_more_than_one_inferred_dim(shape.shape)
layer.set_input(1, shape.trt_tensor)
elif isinstance(shape, (list, tuple)):
assert_no_more_than_one_inferred_dim(shape)
layer.reshape_dims = tuple(shape)
else:
raise TypeError("%s is not supported" % type(shape))
return _create_tensor(layer.get_output(0), layer)
[docs]
def flatten(input: Tensor, start_dim: int = 0, end_dim: int = -1):
'''
Flattens input by reshaping it into a one-dimensional tensor.
If start_dim or end_dim are passed, only dimensions starting with start_dim and
ending with end_dim are flattened. The order of elements in input is unchanged.
Parameters:
input : Tensor
The input tensor to flatten.
start_dim : int
The first dim to flatten.
end_dim : int
The last dim to flatten.
Returns:
The tensor produced by the flatten layer.
'''
shape = input.shape
ndim = input.ndim()
if start_dim < 0: start_dim += ndim
if end_dim < 0: end_dim += ndim
new_shape = list()
for i in range(start_dim):
new_shape.append(shape[i])
if end_dim - start_dim >= 0:
flat_dim = 1
for i in range(start_dim, end_dim + 1):
flat_dim *= shape[i]
new_shape.append(flat_dim)
for i in range(end_dim + 1, ndim):
new_shape.append(shape[i])
return view(input, new_shape)
[docs]
def expand_dims(input: Tensor,
dim: Union[int, Sequence[int]],
shape_cast_dtype=None) -> Tensor:
'''
Add an operation to expand the tensor shape with singleton dimensions.
That function adds a tensorrt.IShuffleLayer to the network. Given an 'input'
of rank N and a sequence of M dimensions, the output tensor produced by
this operation (when executed by TensorRT) will have a rank of N+M. Singleton
dimensions will be inserted at the different positions in 'dim'.
The pseudo-code for that operation is:
new_shape, ii = [], 0
for jj in range(input.rank() + len(dim)):
new_shape.append(1 if jj in dims else input.shape[ii++])
For example, for a tensor of shape [3, 4, 1, 5]
expand_dims(input, [0, 2])
will produce a tensor of shape [1, 3, 1, 4, 1, 5].
Parameters:
input : Tensor
The input tensor to expand.
dim : Union[int, Sequence[int]]
The positions in the output tensor where to insert singleton
dimensions.
Returns:
The tensor produced by the shuffle layer.
'''
if isinstance(dim, int):
dim = (dim, )
out_ndim = len(dim) + input.ndim()
input_shape = shape(input, cast_to_dtype=shape_cast_dtype)
out_shapes = []
j = 0
for i in range(out_ndim):
if i in dim:
out_shapes.append(1)
else:
out_shapes.append(gather(input_shape, 0, j))
j = j + 1
out_shape = concat(out_shapes)
return view(input, out_shape, zero_is_placeholder=False)
# NOTE: Jointly added with Apple
[docs]
def squeeze(input: Tensor,
dim: Optional[Union[int, Sequence[int]]] = None,
zero_is_placeholder: bool = False):
'''
Add an operation to remove singleton dimensions of a tensor.
This functions creates an operation that removes singleton dimension
(dimension of size 1) at positions 'dim' in the input tensor. It works with
negative values for the 'dim'.
For example, for a tensor 'input' of shape [1, 4, 1, 4]:
squeeze(input, 0) will produce an output of shape [4, 1, 4],
squeeze(input, 2) will produce an output of shape [1, 4, 4],
squeeze(input, [0, 2]) will produce an output of shape [4, 4],
squeeze(input, [-2]) will produce an output of shape [1, 4, 4],
Parameters:
input : Tensor
The input tensor for which the singleton dimensions will be removed.
dim : Union[int, Sequence[int]]
The index of the singleton dimensions in the input tensor.
Returns:
The tensor produced by the layer.
'''
if dim is None:
dim = list(range(input.ndim()))
if isinstance(dim, int):
dim = (dim, )
dim = dim_resolve_negative(dim, input.ndim())
new_shape = []
for i, s in enumerate(input.shape):
if s == 1 and i in dim:
continue
new_shape.append(shape(input, i))
new_shape = concat(new_shape) if len(new_shape) > 0 else []
input = input.view(new_shape, zero_is_placeholder=zero_is_placeholder)
return input
[docs]
def unsqueeze(input: Tensor, axis: int):
'''
Add an operation to insert a singleton dimension to a tensor.
That functions creates an operation that insert a singleton dimension
(dimension of size 1) at position 'axis' in the output tensor. It works with
negative values for the 'axis'.
For example, for a tensor 'input' of shape [4, 4]:
unsqueeze(input, 0) will produce an output of shape [1, 4, 4],
unsqueeze(input, 1) will produce an output of shape [4, 1, 4],
unsqueeze(input, -1) will produce an output of shape [4, 4, 1],
unsqueeze(input, -2) will produce an output of shape [4, 1, 4],
Parameters:
input : Tensor
The input tensor to expand with a singleton dimension.
axis : int
The index of the singleton dimension in the output tensor.
Returns:
The tensor produced by the layer.
'''
if axis < 0:
axis = axis + input.ndim() + 1
return expand_dims(input, axis)
[docs]
def stack(inputs: Sequence[Tensor], dim: int = 0) -> Tensor:
'''
Add an operation to contact input tensors along a new dimension.
The function creates an operation that creates a new dim for all the
input tensors and then concatenates them along that new dim.
.
All the tensors in 'inputs' must have the same shape.
for ii in range(inputs[0].rank()):
assert all(inp.shape[ii] == inputs[0].shape[ii] for inp in inputs)
The shape of the output tensor is defined as:
output.rank() = inputs[0].rank() + 1
output.shape[dim] = len(inputs)
for ii in range(inputs[0].rank()):
if ii < dim:
output.shape[ii] = inputs[0].shape[ii]
else:
output.shape[ii+1] = inputs[0].shape[ii]
For example, given a sequence of two 2D tensors [[0, 1], [2, 3]] and
[[4, 5], [6, 7]] both of shape [2, 2],
stack(inputs, 0)
will produce [[[0, 1], [2, 3]], [[4, 5], [6, 7]]] of shape [2, 2, 2] and
stack(inputs, 1)
will produce [[[0, 1], [4, 5]], [[2, 3], [6, 7]]] of shape [2, 2, 2].
Parameters:
inputs : Sequence[Tensor]
The sequence of tensors to stack.
dim : int
The dimension in which the stack is performed.
Returns:
A tensor that contains the input tensors stacked along a new dimension.
'''
return concat([unsqueeze(inp, axis=dim) for inp in inputs], dim=dim)
[docs]
def expand_dims_like(left: Union[Tensor, int, float], right: Tensor) -> Tensor:
'''
Add an operation to expand the first tensor to the same rank as the second
tensor.
That function takes a first tensor. It also accepts an integer or a float,
in which case it creates a constant tensor from it. In both cases, the rank
of that first tensor is compared to the rank of the second tensor. If they
are of the same rank, the first tensor is returned. Otherwise, the first
tensor is expanded on the left to match the rank of the second tensor.
Note that the shapes do not have to match, only the rank is considered in
that function.
For example, for a pair of tensors of shapes [3, 4] and [4, 3, 2], the
first tensor will be expanded to a tensor of rank 3 and shape [1, 3, 4].
Parameters:
left : Union[Tensor, int, float]
The first tensor to expand. When a scalar value is provided as a
parameter, that function first creates a tensor before expanding it
(if needed).
right : Tensor
The reference tensor to match.
Returns:
The tensor produced by the shuffle layer.
'''
if isinstance(left, int):
left = constant(dims_array([left]))
elif isinstance(left, float):
if isinstance(right, Tensor) and right.dtype == trt.DataType.HALF:
left = constant(fp16_array([left]))
else:
left = constant(fp32_array([left]))
left_ndim = left.ndim()
right_ndim = right.ndim()
if right_ndim > left_ndim:
new_ndim = list(range(right_ndim - left_ndim))
return expand_dims(left, new_ndim)
return left
# If dim is None, return a 1-D TensorRT-LLM tensor of the size
# If dim is not None, return a 0-D TensorRT-LLM tensor of the dimension size
[docs]
def shape(input: Tensor,
dim: Optional[int] = None,
cast_to_dtype: Optional[Union[str, trt.DataType]] = None,
clip_before_cast: Sequence[int] = None) -> Tensor:
'''
Add an operation to create a shape tensor.
The shape tensor can either be the shape of the input tensor when the
parameter dim is None or a scalar (tensor of rank 0) that corresponds to
the size of dim-th dimension.
Parameters:
input : Tensor
The input tensor from which we want to extract the shape or the
size in one dimension.
dim : Optional[int]
The dimension from which to extract the size. If it is None, the
entire shape of the input tensor is returned.
Returns:
A tensor that contains the shape of the input tensor (if 'dim' is None)
or the size in the dimension 'dim' of the input tensor. If 'dim' is
'None', that tensor has the same rank as the input tensor, otherwise
its rank is 0.
'''
layer = default_trtnet().add_shape(input.trt_tensor)
res = _create_tensor(layer.get_output(0), layer)
if cast_to_dtype is not None:
if clip_before_cast is not None and (cast_to_dtype == 'int32'
or cast_to_dtype == trt.int32):
assert len(
clip_before_cast
) == 2, f"This parameter only expects a tuple of 2 integers (lower, upper) but got {clip_before_cast}"
res = int_clip(res, clip_before_cast[0], clip_before_cast[1])
res = cast(res, cast_to_dtype)
if dim is None:
return res
return gather(res, dim=0, indices=dim).view([])
[docs]
def gather(input: Tensor, dim: int, indices: Union[Tensor, int]) -> Tensor:
'''
Add an operation to gather elements from a tensor.
That function implements the GatherElements operator from the ONNX
specification as described in
https://github.com/onnx/onnx/blob/main/docs/Operators.md#GatherElements
The input and indices arguments must have the same rank >= 1. The operation
will produce a tensor with the same shape as the indices tensor. The axis
is the dimension to gather on.
As shown in the ONNX description, for a 3D tensor, the output is:
out[i][j][k] = input[indices[i][j][k]][j][k] if axis = 0,
out[i][j][k] = input[i][indices[i][j][k]][k] if axis = 1,
out[i][j][k] = input[i][j][indices[i][j][k]] if axis = 2.
For example,
gather([[4, 2], [5, 3]], 0, [[1, 0], [0, 1]])
will produce [[5, 2], [4, 3]].
gather([[1, 2, 3], [4, 5, 6], 1, [[1], [0]])
will produce [[2], [4]]. See the ONNX documentation for more examples.
That operation maps to the TensorRT IGatherLayer.
Parameters:
input : Tensor
The input tensor to gather elements from.
dim : int
The dimension to gather on.
indices : Union[Tensor, int]
The positions in the 'dim' dimension to gather from.
Returns:
The tensor containing the gathered elements. It has the same shape as
the indices tensor.
'''
if isinstance(indices, int):
indices = constant(int32_array([indices]))
# The input and indices tensors must have the same rank.
assert input.rank() == indices.rank()
layer = default_trtnet().add_gather_v2(input.trt_tensor,
indices.trt_tensor,
mode=trt.GatherMode.ELEMENT)
if dim < 0:
dim = input.ndim() + dim
layer.axis = dim
return _create_tensor(layer.get_output(0), layer)
[docs]
def select(input: Tensor, dim: int, index: Union[Tensor, int]) -> Tensor:
'''
Add an operation to select a slice of elements from a tensor.
Given an input tensor, that function creates an operation that selects the
index-th slice of elements in the dimension 'dim' to create a new tensor.
The output tensor has a shape in which the input dimension 'dim' is
removed.
The 'index' can either be an integer or a 1D tensor containing a single
element.
For example, on input=[[4, 2, 5], [2, 1, 2], [4, 7, 1]], which has a shape
[3, 3],
select(input, 0, 1)
will create a tensor of shape [3] that contains the [2, 1, 2].
Regarding the shape of the output tensor, the dimension 'dim' is removed.
It means that for a tensor of shape [4, 2, 6, 3],
select(input, 2, 4)
will select the 5th slice (index == 4) from the 3rd dimension (dim == 2)
and return a tensor of shape [4, 2, 3] (i.e. the 3rd dimension is removed).
That operation maps to the TensorRT IGatherLayer.
Parameters:
input : Tensor
The input tensor to select from.
dim : int
The dimension to select from.
index : Union[Tensor, int]
The index of the slice in the 'dim' dimension to select.
Returns:
The tensor containing the selected slice.
'''
if isinstance(index, int):
index = constant(int32_array([index]))
assert index.rank() == 1 and index.size(
0) == 1, f"index should have rank 1, got {index.rank()}"
new_shape = []
for i in range(input.rank()):
if i != dim:
new_shape.append(shape(input, i))
layer = default_trtnet().add_gather(input.trt_tensor, index.trt_tensor, dim)
return _create_tensor(layer.get_output(0), layer).view(concat(new_shape))
[docs]
def index_select(input: Tensor, dim: int, index: Tensor) -> Tensor:
'''
Add an operation to select slices of elements from a tensor.
Given an input tensor, that function creates an operation that selects the
slices of elements in the dimension 'dim' at the indices listed in 'index'
to create a new tensor. The output tensor has the same rank as the input
tensor.
The 'index' is a tensor of rank 1.
For example, on input=[[4, 2, 5], [2, 1, 2], [4, 7, 1]], which has a shape
[3, 3],
index_select(input, 0, [0, 1])
will create a tensor of shape [2, 3] that contains the [[4, 2, 5], [2, 1, 2]].
Regarding the shape of the output tensor, the dimension 'dim' has the same
size as the 'index' tensor. It means that for a input tensor of shape [4, 2, 6, 3],
index_select(input, 2, [1, 4])
will select the 2nd and 5th slices (index == 1 or 4) from the 3rd dimension
(dim == 2) and return a tensor of shape [4, 2, 2, 3] (i.e. the 3rd
dimension is shrunk to 2).
Note that this operation can also be used to expand a tensor in the 'dim'
dimension, for example, on input [[0, 1], [2, 3]],
index_select(input, 1, [0, 0, 0])
will produce a tensor of shape [2, 3] containing [[0, 0, 0], [2, 2, 2]].
That operation maps to the TensorRT IGatherLayer.
Parameters:
input : Tensor
The input tensor to select from.
dim : int
The dimension to select from.
index : Tensor
The indices of the slices in the 'dim' dimension to select.
Returns:
The tensor containing the selected slices.
'''
assert index.rank() == 1, f"index should have rank 1, got {index.rank()}"
new_shape = []
for i in range(input.rank()):
if i != dim:
new_shape.append(shape(input, i))
else:
new_shape.append(shape(index, 0))
layer = default_trtnet().add_gather(input.trt_tensor, index.trt_tensor, dim)
return _create_tensor(layer.get_output(0), layer).view(concat(new_shape))
# NOTE: Jointly added with Apple
[docs]
def scatter(input: Tensor, dim: int, indices: Tensor,
updates: Tensor) -> Tensor:
'''
This operation adds a layer that creates an output tensor by element-wise
copying values from the input tensor and then updating values by the given
`indices` and `updates` tensors.
For a 2D input tensor, it first copies the input to output,
then updates the output tensor like the following for each entry in `updates`:
output[indices[i][j]][j] = updates[i][j] if dim=0
output[i][indices[i][j]] = updates[i][j] if dim=1
If the `input` tensor is [[1, 2, 3], [4, 5, 6]],
the indices tensor is [[1, 2], [0, 1]],
the updates tensor is [[-1, -2], [-3, -4]], and dim=1
the output tensor will be [[1, -1, -2], [-3, -4, 6]].
Parameters:
input: Tensor
The input data that needs to be updated.
dim: int
The axis on which the scatter is to be performed.
indices: Tensor
An integer tensor of the same rank as input that indicates the positions to be updated.
updates: Tensor
A data tensor of same shape as the `indices` tensor that contains the update values.
Returns:
A tensor created by the element-wise scatter layer.
'''
layer = default_trtnet().add_scatter(input.trt_tensor,
indices.trt_tensor,
updates.trt_tensor,
mode=trt.ScatterMode.ELEMENT)
layer.axis = dim
return _create_tensor(layer.get_output(0), layer)
[docs]
def gather_nd(input: Tensor, indices: Tensor, batch_dims: int = 1) -> Tensor:
'''
Adds a layer that performs a gather with some element-wise dimensions.
See: https://onnx.ai/onnx/operators/onnx__GatherND.html
The gather is performed on dim=batch_dims.
Parameters:
input: Tensor
The tensor on which the gather operation is performed.
indices: Tensor
The tensor that indicates which entries to be gathered.
batch_dims: int
The number of first dimensions that should be skipped before gather starts.
Returns:
A tensor created by the gather layer with GatherMode.ND.
'''
gather_layer = default_trtnet().add_gather_v2(input.trt_tensor,
indices.trt_tensor,
mode=trt.GatherMode.ND)
gather_layer.num_elementwise_dims = batch_dims
return _create_tensor(gather_layer.get_output(0), gather_layer)
[docs]
def nonzero(input: Tensor) -> Tensor:
'''
Adds a layer that finds the indices of non-zero values of the input tensor.
Parameters:
input: Tensor
The input tensor for which we need to find the indices of non-zero values.
Returns:
A tensor of shape [D, C] where D is the number of dimensions of `input` and
C is the number of non-zero values in it.
Each column of this 2D tensor represents the index tuple for each non-zero value.
'''
non_zero_layer = default_trtnet().add_non_zero(input.trt_tensor)
return _create_tensor(non_zero_layer.get_output(0), non_zero_layer)
[docs]
def masked_select(input: Tensor, mask: Tensor) -> Tensor:
'''
Add an operation to select elements from a tensor according to a boolean
mask tensor.
Given an input tensor, that function creates an operation that selects
elements at the indices indicated by the boolean mask tensor to create
a new tensor. The output tensor is a 1-D tensor.
The input tensor must have rank >= 1. The shapes of the input tensor and
the mask tensor don’t need to match, but they must be able to be broadcasted.
For example, on input=[[4, 2, 5], [2, 1, 2], [4, 7, 1]], which has a shape
[3, 3],
masked_select(input, [[True, False, True], [False, True, False], [True, False, True]])
will create a tensor of shape [5] that contains the [4, 5, 1, 4, 1].
masked_select(input, [[True], [False], [True]])
will create a tensor of shape [6] that contains the [4, 2, 5, 4, 7, 1].
masked_select(input, [[False, False, True]])
will create a tensor of shape [3] that contains the [5, 2, 1].
masked_select(input, [False])
will create a tensor of shape [0] which is empty.
That operation is implemented by NonZero, Shuffle and GatherV2 layers
in TensorRT.
Parameters:
input : Tensor
The input tensor to select from.
mask : Tensor
The boolean mask tensor that indicates elements to select.
Returns:
The 1-D tensor containing the selected elements.
'''
assert input.rank() >= 1, "input should have rank >= 1"
input, mask = broadcast_helper(input, mask)
expanded_mask = expand(mask, shape(input))
non_zero_layer = default_trtnet().add_non_zero(expanded_mask.trt_tensor)
shuffle_layer = default_trtnet().add_shuffle(non_zero_layer.get_output(0))
shuffle_layer.second_transpose = (1, 0)
gather_layer = default_trtnet().add_gather_v2(input.trt_tensor,
shuffle_layer.get_output(0),
mode=trt.GatherMode.ND)
return _create_tensor(gather_layer.get_output(0), gather_layer)
[docs]
def cumsum(input: Tensor, dim: int, prefer_plugin: bool = True) -> Tensor:
'''
Add an operation to calculate inclusive cumulative sum of elements of
a tensor in a given dimension.
Given an input tensor, that function creates an operation that calculates
inclusive cumulative sum of elements in the dimension 'dim' to create
a new tensor. The output tensor has the same shape as the input tensor.
The input tensor must have rank >= 1. The 'dim' must be valid, and negative
value is supported.
For example, on input=[[4, 2, 5], [2, 1, 2], [4, 7, 1]], which has a shape
[3, 3],
cumsum(input, 0)
will produce [[4, 2, 5], [6, 3, 7], [10, 10, 8]].
cumsum(input, 1)
will produce [[4, 6, 11], [2, 3, 5], [4, 11, 12]].
That operation is implemented by TensorRT ILoopLayer.
Parameters:
input : Tensor
The input tensor to calculate the inclusive cumulative sum.
dim : int
The dimension to calculate the inclusive cumulative sum. Negative
value is supported.
prefer_plugin : bool
Whether to use the cumsumLastDim plugin if dim is last dim.
Returns:
The tensor containing the inclusive cumulative sum of input.
'''
assert input.rank() >= 1, "input should have rank >= 1"
assert dim < input.rank() and dim >= -input.rank(
), f"dim should be in [{-input.rank()}, {input.rank()}) when input have rank {input.rank()}"
dim = dim_resolve_negative(dim, input.ndim())[0]
if dim == input.ndim() - 1:
if prefer_plugin:
last_dim = input.size(-1)
if last_dim == -1: # dynamic?
last_dim = shape(input, -1)
old_shape = shape(input)
if input.ndim() == 1:
input_2d = unsqueeze(
input, 0) # special handling of rank-1 dynamic tensor
elif input.ndim() != 2:
input_2d = input.view(concat([-1, last_dim]),
zero_is_placeholder=False)
else:
input_2d = input
cumsum_last_dim_plg_creator = trt.get_plugin_registry(
).get_plugin_creator('CumsumLastDim', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert cumsum_last_dim_plg_creator is not None
input_length = trt.PluginField(
"input_length", np.array(input_2d.size(-1), dtype=np.int32),
trt.PluginFieldType.INT32)
pf_type = trt.PluginField("type_id",
np.array([int(input_2d.dtype)], np.int32),
trt.PluginFieldType.INT32)
pfc = trt.PluginFieldCollection([input_length, pf_type])
cumsum_last_dim_plug = cumsum_last_dim_plg_creator.create_plugin(
"cumsum_last_dim", pfc)
plug_inputs = [input_2d]
plug_inputs = [i.trt_tensor for i in plug_inputs]
layer = default_trtnet().add_plugin_v2(plug_inputs,
cumsum_last_dim_plug)
_add_plugin_info(layer, cumsum_last_dim_plg_creator,
"cumsum_last_dim", pfc)
output = _create_tensor(layer.get_output(0), layer)
output = output.view(old_shape, zero_is_placeholder=False)
return output
else:
# credit to Apple
reduction_length = shape(input, -1)
reduction_range = arange(constant_to_tensor_(0,
dtype='int64',
to_array=False),
reduction_length,
dtype='int64')
lower_triangle = cast(
unsqueeze(reduction_range, 0) <= unsqueeze(reduction_range, 1),
dtype=input.dtype)
output = sum(unsqueeze(input, -2) * lower_triangle, dim=-1)
return output
else:
slice_shape = []
for i in range(input.ndim()):
if i != dim:
slice_shape.append(shape(input, i))
zero_tensor = constant_to_tensor_(0, input.dtype, False)
if len(slice_shape) > 0:
zero_tensor = expand_dims(zero_tensor,
[i for i in range(len(slice_shape))])
slice_shape = concat(slice_shape)
zero_tensor = expand(zero_tensor, slice_shape)
loop_layer = default_trtnet().add_loop()
trip_limit = shape(input, dim).trt_tensor
loop_layer.add_trip_limit(trip_limit, trt.TripLimit.COUNT)
iterator_layer = loop_layer.add_iterator(input.trt_tensor, dim)
cur_slice = iterator_layer.get_output(0)
running_sum_layer = loop_layer.add_recurrence(zero_tensor.trt_tensor)
running_sum = running_sum_layer.get_output(0)
cur_sum_layer = default_trtnet().add_elementwise(
cur_slice, running_sum, trt.ElementWiseOperation.SUM)
cur_sum = cur_sum_layer.get_output(0)
running_sum_layer.set_input(1, cur_sum)
loop_output_layer = loop_layer.add_loop_output(
cur_sum, trt.LoopOutput.CONCATENATE, dim)
loop_output_layer.set_input(1, trip_limit)
return _create_tensor(loop_output_layer.get_output(0),
loop_output_layer)
[docs]
def masked_scatter(input: Tensor, mask: Tensor, source: Tensor) -> Tensor:
'''
Add the masked_scatter base on PyTorch definition.
See https://pytorch.org/docs/stable/generated/torch.Tensor.masked_scatter_.html#torch.Tensor.masked_scatter_ for a
description of that function.
Parameters:
input : Tensor
The input tensor.
mask : Tensor
The boolean mask tensor that indicates elements to select.
source: Tensor
The tensor to copy from
Returns:
The tensor containing the source tensor selected by mask.
'''
assert input.rank() >= 1, "input should have rank >= 1"
input, mask = broadcast_helper(input, mask)
expanded_mask = expand(mask, shape(input))
non_zero_layer = default_trtnet().add_non_zero(expanded_mask.trt_tensor)
shuffle_layer = default_trtnet().add_shuffle(non_zero_layer.get_output(0))
shuffle_layer.second_transpose = (1, 0)
source = source.view([-1])
scatter_layer = default_trtnet().add_scatter(input.trt_tensor,
shuffle_layer.get_output(0),
source.trt_tensor,
mode=trt.ScatterMode.ND)
return _create_tensor(scatter_layer.get_output(0), scatter_layer)
[docs]
def concat(inputs: Sequence[Union[Tensor, int]], dim: int = 0) -> Tensor:
'''
Add an operation to concatenate tensors.
The function creates an operation that concatenates the tensors from the
sequence 'inputs'. The concatenation is done along the dimension 'dim'.
All the tensors in 'inputs' must have the same shape expect for the
dimension 'dim'.
for ii in range(inputs[0].rank()):
assert (ii == dim) or all(inp.shape[ii] == inputs[0].shape[ii] for inp in inputs)
The shape of the output tensor is defined as:
for ii in range(inputs[0].rank()):
# Same size as all the inputs in dimension ii != dim.
output.shape[ii] = inputs[0].shape[ii]
# Sum of the sizes in the different inputs in dimension 'dim'.
if ii == dim:
for jj in range(1, len(inputs)):
output.shape[ii] += inputs[jj].shape[ii]
For example, given a sequence of two 2D tensors [[0, 1], [2, 3]] and
[[4, 5], [6, 7]] both of shape [2, 2],
concat(inputs, 0)
will produce [[0, 1], [2, 3], [4, 5], [6, 7]] of shape [4, 2] and
concat(inputs, 1)
will produce [[0, 1, 4, 5], [2, 3, 6, 7]] of shape [2, 4].
Parameters:
inputs : Sequence[Union[Tensor, int]]
The sequence of tensors to concatenate. For integers, that function
creates constant tensors.
dim : int
The dimension in which the concatenation is performed.
Returns:
A tensor that contains the concatenation of the tensors.
'''
assert len(
inputs
) > 0, f"Number of inputs ({len(inputs)}) to the concatenation layer must be > 0."
tmp = []
inputs = constants_to_tensors_(*inputs)
for i in inputs:
if i.rank() == 0:
tmp.append(i.view([1]))
else:
tmp.append(i)
layer = default_trtnet().add_concatenation([i.trt_tensor for i in tmp])
layer.axis = dim_resolve_negative(dim, tmp[0].ndim())[0]
return _create_tensor(layer.get_output(0), layer)
[docs]
def softmax(input: Tensor, dim: Optional[int] = None) -> Tensor:
'''
Add an operation to compute softmax on a tensor.
That operation computes the softmax on the input tensor in the dimension
'dim' if specified. Otherwise, it is applied on the last dimension.
It inserts a ISoftmaxLayer to the TensorRT graph.
Parameters:
input : Tensor
The input tensor on which to apply softmax.
dim : Optional[int]
The dimension used to apply softmax.
Returns:
The output tensor of the softmax layer.
'''
if dim is None:
dim = input.ndim() - 1
if dim < 0:
dim = input.ndim() + dim
axes = dim_to_trt_axes(dim)
layer = default_trtnet().add_softmax(input.trt_tensor)
layer.axes = axes
return _create_tensor(layer.get_output(0), layer)
def _lookup_plugin(input: Tensor, weight: Tensor, rank: int,
per_token_scale: Tensor) -> Tensor:
'''
Add an operation to perform lookup in a tensor.
That operation performs the lookup needed by embedding layers. Given a
'weight' tensor of shape [rows, cols], it produces a tensor of shape
[inputs.size(0), cols] where the ith row corresponds to the input[i] row in
the weight tensor.
It inserts a IPluginV2Layer.
Parameters:
input : Tensor
The input tensor contains the indices to perform the lookup.
weight : Tensor
The table to gather from.
rank : int
The mpi rank.
Returns:
The output tensor of the lookup layer.
'''
plg_creator = trt.get_plugin_registry().get_plugin_creator(
'Lookup', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert plg_creator is not None
p_dtype = default_net().plugin_config.lookup_plugin
pf_type = trt.PluginField(
"type_id", np.array([int(str_dtype_to_trt(p_dtype))], np.int32),
trt.PluginFieldType.INT32)
rank = trt.PluginField("rank", np.array([int(rank)], np.int32),
trt.PluginFieldType.INT32)
pfc = trt.PluginFieldCollection([pf_type, rank])
lookup_plug = plg_creator.create_plugin("lookup", pfc)
plug_inputs = [input.trt_tensor, weight.trt_tensor]
if per_token_scale is not None:
plug_inputs.append(per_token_scale.trt_tensor)
weight.trt_tensor.set_dynamic_range(-127, 127)
layer = default_trtnet().add_plugin_v2(plug_inputs, lookup_plug)
_add_plugin_info(layer, plg_creator, "lookup", pfc)
return _create_tensor(layer.get_output(0), layer)
[docs]
def embedding(input: Tensor,
weight: Tensor,
tp_size=1,
tp_group=None,
sharding_dim=0,
tp_rank=None,
per_token_scale=None) -> Tensor:
'''
Add an operation to perform embedding lookup.
That operation performs the embedding lookup. The 'input' tensor contains
the identifiers of the rows of 'weight' to gather.
1. Distribute the embedding lookup table over multiple GPU
When 'tp_size' is greater than 1 and the 'tp_group' is defined, this
embedding lookup is distributed among multiple GPUs.
When 'sharding_dim==0', each GPU stores a subset of the rows of the embedding
table rows(that number of rows per GPU is given by weights.shape[0] and the offset to
the 1st row stored on the GPU is given by rank * weights.shape[0]). Each
parallel rank will query all the indices and set 0s for the weights that
are not stored on the associated GPU. To compute the final result, a
parallel all-reduce operation is added to the TensorRT graph. That lookup
can be performed using either the plugin or the operators TensorRT support.
When'sharding_dim==1', each GPU stores a subset of the embedding table's columns.
Each rank can obtain a portion of the embedding results.
Then the embedding is collected using the all-gather operation.
Related transposition operations are also used to obtain the final results.
2. Store embedding lookup table as a whole
When 'tp_size' is not greater than 1, the embedding lookup table will not
be divided. In this case, when the default_net().plugin_config.lookup_plugin is set,
the operation is implemented using a plugin (without the all-reduce operation).
Otherwise, this operation is implemented using the standard IGatherLayer in TensorRT.
Parameters:
input : Tensor
The input tensor the contains the indices to perform the lookup.
weight : Tensor
The table to gather from.
tp_size : int
The number of GPUs collaborating to perform that embedding.
tg_group : Optional[List[int]]
The group of world ranks participating in the all-reduce when
tp_size > 1.
sharding_dim : int
sharding_dim = 0 means that we shard the embedding table in vocab dim;
sharding_dim = 1 means that we shard the embedding table in embedding dim.
tp_rank : int
The tensor parallelism rank. Used to calculate offset in TP on vocab dim.
Returns:
The tensor produced by the embedding lookup layer.
'''
# Distribute embedding lookup table across multiple GPU
if tp_size > 1 and tp_group is not None:
if sharding_dim == 0: # TP on vocab_size dimension
if tp_rank == None:
raise ValueError(
"Rank cannot be none for tensor parallelism on vocab dim")
if default_net().plugin_config.lookup_plugin:
x = _lookup_plugin(input, weight, tp_rank, per_token_scale)
x = allreduce(x, tp_group)
else:
shape_weight = shape(weight)
vocab_size = slice(shape_weight, starts=[0], sizes=[1])
tmp_input = input - vocab_size * tp_rank
# Identify the valid indices
is_qualified = op_and(tmp_input >= 0, tmp_input < vocab_size)
is_qualified_expand = expand_dims(is_qualified,
[is_qualified.ndim()])
# Replace the invalid ones to zero
placeholder_input = where(is_qualified, tmp_input, 0)
# Get the temporal results
layer = default_trtnet().add_gather(
weight.trt_tensor, placeholder_input.trt_tensor, 0)
tmp_output = _create_tensor(layer.get_output(0), layer)
# Set zero for invalid results
placeholder_tmp = cast(is_qualified_expand, tmp_output.dtype)
placeholder = placeholder_tmp - placeholder_tmp
x = where(is_qualified_expand, tmp_output, placeholder)
# Use all reduce to collect the results
x = allreduce(x, tp_group)
elif sharding_dim == 1: # TP on hidden dimension
layer = default_trtnet().add_gather(weight.trt_tensor,
input.trt_tensor, 0)
x = _create_tensor(layer.get_output(0), layer)
# [dim0, local_dim] -> [dim0 * tp_size, local_dim] --> [dim0, local_dim * tp_size]
x = allgather(x, tp_group, gather_dim=-1)
else:
raise ValueError(
'Tensor Parallelism only support splitting Embedding lookup along hidden (sharding_dim==1) and vocab (sharding_dim==0) dimensionis'
)
# Store embedding lookup table as a whole
else:
if default_net().plugin_config.lookup_plugin:
x = _lookup_plugin(input,
weight,
rank=0,
per_token_scale=per_token_scale)
else:
layer = default_trtnet().add_gather(weight.trt_tensor,
input.trt_tensor, 0)
x = _create_tensor(layer.get_output(0), layer)
return x
[docs]
def constant_to_tensor_(input: Union[Tensor, int, float, bool],
dtype: Union[trt.DataType, str] = None,
to_array=True) -> Tensor:
if dtype is None:
# deduce the type from the given value
# NOTE: bool is a subtype of int, so bool needs to be checked first
if isinstance(input, bool):
dtype = trt.bool
elif isinstance(input, int):
dtype = trt.int32
else:
dtype = trt.float32
if not isinstance(input, Tensor):
if isinstance(dtype, str):
dtype = str_dtype_to_trt(dtype)
array_fn_dict = {
trt.int64: int64_array,
trt.int32: int32_array,
trt.float32: fp32_array,
trt.float16: fp16_array,
trt.bfloat16: bf16_array,
trt.bool: bool_array,
}
assert dtype in array_fn_dict
return constant(array_fn_dict[dtype]([input] if to_array else input))
return input
[docs]
def constants_to_tensors_(
*inputs: Union[Tensor, int, float]) -> Tuple[Tensor, ...]:
'''
Helper function to create tensors from multiple inputs.
For each inputs, that function first creates a constant tensor if the input
is an integer or a float. Then, if any input is int64, it upcasts other
integer inputs to int64.
Parameters:
inputs : Tuple[Union[Tensor, int, float], ...]
The inputs to create tensors from.
Returns:
A tuple of tensors.
'''
has_int64: bool = False
for i in inputs:
if isinstance(i, int) and (i >= 2**31 or i < -2**31)\
or isinstance(i, Tensor) and i.dtype == trt.int64:
has_int64 = True
break
if not has_int64:
return tuple(constant_to_tensor_(i) for i in inputs)
result = []
for i in inputs:
if isinstance(i, int) or isinstance(i, Tensor) and i.dtype == trt.int32:
result.append(
constant_to_tensor_(i, trt.int64 if has_int64 else trt.int32))
else:
result.append(constant_to_tensor_(i))
return tuple(result)
[docs]
def broadcast_helper(left: Union[Tensor, int, float],
right: Union[Tensor, int, float]) -> Tuple[Tensor, Tensor]:
'''
Helper function to perform a broadcast.
For each input, that function first creates a constant tensor if the input
is an integer or a float. Then, if needed, it expands the smaller tensor to
make sure its rank is the same as the larger one.
Parameters:
left : Union[Tensor, int, float]
The first input. If that input is an integer or a float, the
function creates a constant tensor.
right : Union[Tensor, int, float]
The second input. If that input is an integer or a float, the
function creates a constant tensor.
Returns:
A pair of tensors of same rank.
'''
if not default_net().strongly_typed:
left = constant_to_tensor_(left)
right = constant_to_tensor_(right)
else:
left = constant_to_tensor_(
left, right.dtype if isinstance(right, Tensor) else None)
right = constant_to_tensor_(right, left.dtype)
if left.rank() == right.rank():
return (left, right)
if left.rank() < right.rank():
left = expand_dims_like(left, right)
return (left, right)
if left.rank() > right.rank():
right = expand_dims_like(right, left)
return (left, right)
[docs]
def elementwise_binary(left: Union[Tensor, int,
float], right: Union[Tensor, int, float],
op: trt.ElementWiseOperation) -> Tensor:
'''
Add an elementwise operation with two inputs.
For each input, that function first creates a constant tensor if the input
is an integer or a float. Then, if needed, it expands the smaller tensor to
make sure its rank is the same as the larger one. Then, it performs the
elementwise operation 'op'.
The following closures are defined in functional.*:
add for op=trt.ElementWiseOperation.SUM
sub for op=trt.ElementWiseOperation.SUB
mul for op=trt.ElementWiseOperation.PROD
div for op=trt.ElementWiseOperation.DIV
floordiv for op=trt.ElementWiseOperation.FLOOR_DIV
gt for op=trt.ElementWiseOperation.GREATER
lt for op=trt.ElementWiseOperation.LESS
op_and for op=trt.ElementWiseOperation.AND
op_or for op=trt.ElementWiseOperation.OR
eq for op=trt.ElementWiseOperation.EQUAL
minimum for op=trt.ElementWiseOperation.MIN
maximum for op=trt.ElementWiseOperation.MAX
pow for op=trt.ElementWiseOperation.POW
It is implemented using the IElementWiseLayer from TensorRT.
Parameters:
left : Union[Tensor, int, float]
The first input. If that input is an integer or a float, the
function creates a constant tensor.
right : Union[Tensor, int, float]
The second input. If that input is an integer or a float, the
function creates a constant tensor.
op : trt.ElementWiseOperation
The binary operation to perform.
Returns:
The tensor produced by this elementwise operation.
'''
left, right = broadcast_helper(left, right)
if left.dtype == trt.int32 and right.dtype == trt.int64:
left = cast(left, trt.int64)
if left.dtype == trt.int64 and right.dtype == trt.int32:
right = cast(right, trt.int64)
layer = default_trtnet().add_elementwise(left.trt_tensor, right.trt_tensor,
op)
return _create_tensor(layer.get_output(0), layer)
add = partial(elementwise_binary, op=trt.ElementWiseOperation.SUM)
sub = partial(elementwise_binary, op=trt.ElementWiseOperation.SUB)
mul = partial(elementwise_binary, op=trt.ElementWiseOperation.PROD)
div = partial(elementwise_binary, op=trt.ElementWiseOperation.DIV)
floordiv = partial(elementwise_binary, op=trt.ElementWiseOperation.FLOOR_DIV)
gt = partial(elementwise_binary, op=trt.ElementWiseOperation.GREATER)
lt = partial(elementwise_binary, op=trt.ElementWiseOperation.LESS)
op_and = partial(elementwise_binary, op=trt.ElementWiseOperation.AND)
op_or = partial(elementwise_binary, op=trt.ElementWiseOperation.OR)
eq = partial(elementwise_binary, op=trt.ElementWiseOperation.EQUAL)
minimum = partial(elementwise_binary, op=trt.ElementWiseOperation.MIN)
maximum = partial(elementwise_binary, op=trt.ElementWiseOperation.MAX)
pow = partial(elementwise_binary, op=trt.ElementWiseOperation.POW)
[docs]
def modulo(x: Tensor, y: Union[Tensor, int]) -> Tensor:
'''
This function adds an element-wise modulo (x % y) operation for a given tensor.
Since there is no TensorRT layer that can directly perform this,
this function implements it using some of the basic operations.
Returns:
A tensor that represents (x % y) modulo operation.
'''
return x - (x // y) * y
[docs]
def where(condition: Union[Tensor, bool], left: Union[Tensor, int, float],
right: Union[Tensor, int, float]) -> Tensor:
'''
Add a where (aka select or if-then-else) operation.
Assuming the three input parameters have the same shape, that function creates
the operation to compute a tensor of the same shape such that:
for ii in range(mul(condition.shape)):
output[ii] = left[ii] if condition[ii] else right[ii]
For each input, that function first creates a constant tensor if the
condition is boolean or the left/right input is an integer or a float.
Then, if needed, it expands the smaller tensor to make sure its
rank is the same as the larger one. Then, it performs the selection.
It is implemented using the ISelectLayer from TensorRT.
Parameters:
condition : Union[Tensor, bool]
The condition. If that input is a boolean, the function
creates a constant tensor.
left : Union[Tensor, int, float]
The first input. If that input is an integer or a float, the
function creates a constant tensor.
right : Union[Tensor, int, float]
The second input. If that input is an integer or a float, the
function creates a constant tensor.
Returns:
The tensor produced by this where operation.
'''
# Convert to tensors.
condition = constant_to_tensor_(condition)
left, right = constants_to_tensors_(left, right)
# Find the tensor with the largest rank of the three.
largest = condition
if largest.rank() < left.rank():
largest = left
if largest.rank() < right.rank():
largest = right
# Expand the tensors to match the largest one.
if condition is not largest:
condition = expand_dims_like(condition, largest)
if left is not largest:
left = expand_dims_like(left, largest)
if right is not largest:
right = expand_dims_like(right, largest)
# Insert the operation.
layer = default_trtnet().add_select(condition.trt_tensor, left.trt_tensor,
right.trt_tensor)
return _create_tensor(layer.get_output(0), layer)
[docs]
def unary(input: Tensor, op: trt.UnaryOperation) -> Tensor:
'''
Add an elementwise operation on a single input.
The following closures are defined in functional.*:
round for op=trt.UnaryOperation.ROUND
sqrt for op=trt.UnaryOperation.SQRT
exp for op=trt.UnaryOperation.EXP
sin for op=trt.UnaryOperation.SIN
cos for op=trt.UnaryOperation.COS
abs for op=trt.UnaryOperation.ABS
log for op=trt.UnaryOperation.LOG
It is implemented using the IUnaryLayer from TensorRT.
Parameters:
input : Tensor
The input tensor.
op : trt.UnaryOperation
The unary operation to perform.
Returns:
The tensor produced by this elementwise operation.
'''
layer = default_trtnet().add_unary(input.trt_tensor, op)
return _create_tensor(layer.get_output(0), layer)
round = partial(unary, op=trt.UnaryOperation.ROUND)
sqrt = partial(unary, op=trt.UnaryOperation.SQRT)
exp = partial(unary, op=trt.UnaryOperation.EXP)
sin = partial(unary, op=trt.UnaryOperation.SIN)
cos = partial(unary, op=trt.UnaryOperation.COS)
abs = partial(unary, op=trt.UnaryOperation.ABS)
log = partial(unary, op=trt.UnaryOperation.LOG)
not_op = partial(unary, op=trt.UnaryOperation.NOT)
[docs]
def log_softmax(input: Tensor, dim: int) -> Tensor:
'''
This function is equivalent of torch.nn.functional.log_softmax() i.e.
it performs log(softmax(input)) in a safer and faster way.
Parameters:
input: Tensor
The data tensor on which log_softmax to be computed.
dim: int
The dimension of the input tensor along which log_softmax will be computed.
Returns:
A tensor of same shape as input with log_softmax computed on the specified dim.
'''
x_max = max(input, dim=dim, keepdim=True)
x = input - x_max
return x - log(sum(exp(x), dim=dim, keepdim=True))
[docs]
def reduce(input: Tensor,
op: trt.ReduceOperation,
dim: int,
keepdim: bool = False) -> Tensor:
'''
Add an reduction operation to do along a dimension.
It is implemented using the IReduceLayer from TensorRT.
Parameters:
input : Tensor
The input tensor.
op : trt.ReduceOperation
The reduction operation to perform.
Options: SUM, PROD, MAX, MIN, AVG
dim : int
The dimension along which the reduction is performed.
keepdim : bool
Is the dimension kept in the reduced tensor? When True the
dimension is kept, it is removed from the shape otherwise.
Returns:
The tensor produced by this reduction operation.
'''
dim = dim_resolve_negative(dim, input.ndim())
axes = dim_to_trt_axes(dim)
layer = default_trtnet().add_reduce(input.trt_tensor,
op,
axes,
keep_dims=keepdim)
return _create_tensor(layer.get_output(0), layer)
prod = partial(reduce, op=trt.ReduceOperation.PROD)
min = partial(reduce, op=trt.ReduceOperation.MIN)
[docs]
def mean(input: Tensor, dim: int, keepdim: bool = False) -> Tensor:
'''
Add an operation to compute the mean along a dimension.
Computes the mean along the dimension 'dim' of the input tensor.
It is implemented using the IReduceLayer from TensorRT.
Parameters:
input : Tensor
The input tensor.
dim : int
The dimension along which the mean is computed.
keepdim : bool
Is the dimension kept in the reduced tensor? When True the
dimension is kept, it is removed from the shape otherwise.
Returns:
The tensor produced by this reduction operation.
'''
return reduce(input, op=trt.ReduceOperation.AVG, dim=dim, keepdim=keepdim)
[docs]
def max(input: Tensor, dim: int, keepdim: bool = False) -> Tensor:
'''
Add an operation to compute the max along a dimension.
Computes the max along the dimension 'dim' of the input tensor.
It is implemented using the IReduceLayer from TensorRT.
Parameters:
input : Tensor
The input tensor.
dim : int
The dimension along which the mean is computed.
keepdim : bool
Is the dimension kept in the reduced tensor? When True the
dimension is kept, it is removed from the shape otherwise.
Returns:
The tensor produced by this reduction operation.
'''
return reduce(input, op=trt.ReduceOperation.MAX, dim=dim, keepdim=keepdim)
[docs]
def sum(input: Tensor, dim: int, keepdim: bool = False) -> Tensor:
'''
Add an operation to compute the sum along a dimension.
Computes the sum along the dimension 'dim' of the input tensor.
It is implemented using the IReduceLayer from TensorRT.
Parameters:
input : Tensor
The input tensor.
dim : int
The dimension along which the mean is computed.
keepdim : bool
Is the dimension kept in the reduced tensor? When True the
dimension is kept, it is removed from the shape otherwise.
Returns:
The tensor produced by this reduction operation.
'''
return reduce(input, op=trt.ReduceOperation.SUM, dim=dim, keepdim=keepdim)
[docs]
def identity(input: Tensor) -> Tensor:
'''
Add an identity operation.
TODO: Document why it can be done using a plugin!!!
Parameters:
input : Tensor
The input tensor.
Returns:
The tensor produced by this identity operation.
'''
if not default_net().plugin_config.identity_plugin:
layer = default_trtnet().add_identity(input.trt_tensor)
else:
plg_creator = trt.get_plugin_registry().get_plugin_creator(
'Identity', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert plg_creator is not None
pfc = trt.PluginFieldCollection()
id_plug = plg_creator.create_plugin("identity", pfc)
plug_inputs = [input.trt_tensor]
layer = default_trtnet().add_plugin_v2(plug_inputs, id_plug)
_add_plugin_info(layer, plg_creator, "identity", pfc)
return _create_tensor(layer.get_output(0), layer)
[docs]
def argmax(input: Tensor, dim: int, keepdim: bool = False) -> Tensor:
'''
Add an argmax operation.
As explained in the ONNX documentation,
https://github.com/onnx/onnx/blob/main/docs/Operators.md#argmax
that function creates a layer computing the indices of the max elements of
the input tensor's element along the provided dim. The resulting tensor
has the same rank as the input if keepdims is True. If keepdims is False,
then the resulting tensor has the reduced dimension pruned.
Parameters:
input : Tensor
The input tensor.
dim : int
The dimension in which to compute the argmax indices.
keepdim : bool
Do we keep the dimension along which the reduction is performed?
Yes, if set to True, no otherwise.
Returns:
The tensor produced by this argmax operation.
'''
dim = dim_resolve_negative(dim, input.ndim())
axes = dim_to_trt_axes(dim)
layer = default_trtnet().add_topk(input.trt_tensor, trt.TopKOperation.MAX,
1, axes)
output = layer.get_output(1)
if keepdim:
return _create_tensor(output, layer)
output = _create_tensor(output, layer)
a = list(range(input.ndim()))
for d in dim:
a.pop(d)
indices = constant(int32_array(a))
output_shape = shape(output)
new_shape = gather(output_shape, 0, indices)
return view(output, new_shape)
[docs]
def gelu(x: Tensor) -> Tensor:
'''
Add a GELU operation.
Parameters:
input : Tensor
The input tensor on which the activation function is applied.
Returns:
The tensor produced by the activation layer.
'''
return 0.5 * x * (
tanh(math.sqrt(2.0 / math.pi) * (x + 0.044715 * pow(x, 3.0))) + 1.0)
[docs]
def geglu(x: Tensor) -> Tensor:
'''
Add a Gated-GELU operation.
That function takes a tensor, splits it into two halves along the last
dimension, applies GELU to the second half and multiply the results. The
behavior is undefined if the last dimension is not even.
Parameters:
input : Tensor
The input tensor on which the activation function is applied.
Returns:
The tensor produced by the activation layer.
'''
a, b = chunk(x, 2, dim=-1)
return a * gelu(b)
[docs]
def quick_gelu(x: Tensor) -> Tensor:
return x * sigmoid(1.702 * x)
[docs]
def gegelu(x: Tensor, limit: Optional[float] = None) -> Tensor:
# a, b = x[..., ::2], x[..., 1::2]
ndim = x.ndim()
a_starts = [0 for i in range(ndim)]
b_starts = [1 if i == (ndim - 1) else 0 for i in range(ndim)]
shapes = concat([
shape(x, i) / 2 if i == (ndim - 1) else shape(x, i) for i in range(ndim)
])
strides = [2 if i == (ndim - 1) else 1 for i in range(ndim)]
a = slice(x, a_starts, shapes, strides)
b = slice(x, b_starts, shapes, strides)
if limit is not None:
a = clip(a, alpha=float(-1e20), beta=limit)
b = clip(b, alpha=-limit, beta=limit)
# C = B + 1
const1 = arange(constant(int32_array(1)), constant(int32_array(2)),
trt_dtype_to_str(b.dtype))
for _ in range(ndim - 1):
const1 = expand_dims(const1, 0)
b_shape = concat([shape(b, i) for i in range(ndim)])
const1_arr = expand(const1, b_shape)
return quick_gelu(a) * (b + const1_arr)
[docs]
def group_norm(input: Tensor,
num_groups: int,
weight: Optional[Tensor] = None,
bias: Optional[Tensor] = None,
eps: float = 1e-05):
##
## TODO: Document that function!
##
assert not input.is_dynamic(1)
num_channels = input.size()[1]
ndim = input.ndim()
old_shape = shape(input)
new_shape = concat([
input.size(0),
num_groups,
num_channels // num_groups,
] + [input.size(i) for i in range(2, ndim)])
x = input.view(new_shape)
reduce_dim = tuple(range(2, ndim + 1))
ux = x.mean(dim=reduce_dim, keepdim=True)
numerator = x - ux
varx = numerator * numerator
varx = varx.mean(dim=reduce_dim, keepdim=True)
denom = varx + eps
denom = denom.sqrt()
y = numerator / denom
y = y.view(old_shape)
new_shape = concat([num_channels] + [1 for _ in range(2, ndim)])
if weight is not None:
y = y * weight.view(new_shape)
if bias is not None:
y = y + bias.view(new_shape)
return y
[docs]
def softplus(input: Tensor, beta: float, threshold: float) -> Tensor:
'''
Add the softplus activation base on PyTorch definition.
See https://pytorch.org/docs/stable/generated/torch.nn.functional.softplus.html for a
description of that function.
Parameters:
input : Tensor
Input TensorRT-LLM Tensor.
beta : float
The parameter for softplus computation.
threshold : float
The threshold for reverting to the linear function when input * beta > threshold
Returns:
The output tensor created by that layer.
'''
sf_layer = default_trtnet().add_activation(input.trt_tensor,
trt.ActivationType.SOFTPLUS)
sf_layer.alpha = 1 / beta
sf_layer.beta = beta
prod_tensor = input * beta
result = prod_tensor > threshold
return where(result, input, _create_tensor(sf_layer.get_output(0),
sf_layer))
[docs]
def outer(input: Tensor, vec2: Tensor) -> Tensor:
'''
Add an operation to compute the outer product between two tensors.
That operation creates an Einsum node.
Parameters:
input : Tensor
The first input tensor.
vec2 : Tensor
The second input tensor.
Returns:
The output tensor produced by this layer.
'''
return einsum('i,j->ij', [input, vec2])
[docs]
def avg_pool2d(input: Tensor,
kernel_size: Tuple[int],
stride: Optional[Tuple[int]] = None,
padding: Optional[Tuple[int]] = (0, 0),
ceil_mode: bool = False,
count_include_pad: bool = True) -> Tensor:
##
## TODO: Document that function!
##
assert not input.is_dynamic()
ndim = input.ndim()
if ndim == 3:
input = expand_dims(input, 0)
layer = default_trtnet().add_pooling_nd(input.trt_tensor,
trt.PoolingType.AVERAGE,
kernel_size)
if stride is None:
stride = kernel_size
layer.stride_nd = stride
output = _create_tensor(layer.get_output(0), layer)
if ndim == 3:
return output.view(
concat([output.size(1),
output.size(2),
output.size(3)]))
return output
[docs]
def conv1d(input: Tensor,
weight: Tensor,
bias: Optional[Tensor] = None,
stride: int = 1,
padding: int = 0,
dilation: int = 1,
groups: int = 1) -> Tensor:
noutput = weight.size()[0]
kernel_size = weight.size()[-2]
is_weight_constant = (weight.producer is not None
and weight.producer.type == trt.LayerType.CONSTANT)
weight = weight.producer.weights if is_weight_constant else trt.Weights()
if bias is not None:
is_bias_constant = (bias.producer is not None
and bias.producer.type == trt.LayerType.CONSTANT)
bias = bias.producer.weights if is_bias_constant else trt.Weights()
input_shuffled = stack([input], dim=input.ndim())
kernel_size = trt.Dims([kernel_size, 1])
layer = default_trtnet().add_convolution_nd(input_shuffled.trt_tensor,
noutput, kernel_size, weight,
bias)
layer.stride_nd = (stride, 2)
layer.padding_nd = (padding, 0)
layer.dilation_nd = (dilation, 2)
layer.num_groups = groups
if not is_weight_constant:
layer.set_input(1, weight.trt_tensor)
if bias is not None and not is_bias_constant:
layer.set_input(2, bias.trt_tensor)
output_2d = _create_tensor(layer.get_output(0), layer)
output_1d = squeeze(output_2d, dim=-1)
return output_1d
[docs]
def conv2d(input: Tensor,
weight: Tensor,
bias: Optional[Tensor] = None,
stride: Tuple[int, int] = (1, 1),
padding: Tuple[int, int] = (0, 0),
dilation: Tuple[int, int] = (1, 1),
groups: int = 1) -> Tensor:
##
## TODO: Document that function!
##
ndim = input.ndim()
if ndim == 3:
input = expand_dims(input, 0)
noutput = weight.size()[0]
kernel_size = (weight.size()[-2], weight.size()[-1])
is_weight_constant = (weight.producer is not None
and weight.producer.type == trt.LayerType.CONSTANT)
weight = weight.producer.weights if is_weight_constant else trt.Weights()
if bias is not None:
is_bias_constant = (bias.producer is not None
and bias.producer.type == trt.LayerType.CONSTANT)
bias = bias.producer.weights if is_bias_constant else trt.Weights()
layer = default_trtnet().add_convolution_nd(input.trt_tensor, noutput,
kernel_size, weight, bias)
layer.stride_nd = stride
layer.padding_nd = padding
layer.dilation_nd = dilation
layer.num_groups = groups
layer.dilation_nd = dilation
if not is_weight_constant:
layer.set_input(1, weight.trt_tensor)
if bias is not None and not is_bias_constant:
layer.set_input(2, bias.trt_tensor)
output = _create_tensor(layer.get_output(0), layer)
if ndim == 3:
return output.view(
concat([output.size(1),
output.size(2),
output.size(3)]))
return output
[docs]
def conv_transpose2d(input: Tensor,
weight: Tensor,
bias: Optional[Tensor] = None,
stride: Tuple[int, int] = (1, 1),
padding: Tuple[int, int] = (0, 0),
output_padding: Tuple[int, int] = (0, 0),
dilation: Tuple[int, int] = (1, 1),
groups: int = 1) -> Tensor:
##
## TODO: Document that function!
##
assert not input.is_dynamic()
ndim = input.ndim()
if ndim == 3:
input = expand_dims(input, 0)
noutput = weight.size()[1]
kernel_size = (weight.size()[-2], weight.size()[-1])
is_weight_constant = (weight.producer is not None
and weight.producer.type == trt.LayerType.CONSTANT)
weight = weight.producer.weights if is_weight_constant else trt.Weights()
if bias is not None:
is_bias_constant = (bias.producer is not None
and bias.producer.type == trt.LayerType.CONSTANT)
bias = bias.producer.weights if is_bias_constant else trt.Weights()
layer = default_trtnet().add_deconvolution_nd(input.trt_tensor, noutput,
kernel_size, weight, bias)
layer.stride_nd = stride
layer.padding_nd = padding
layer.num_groups = groups
if not is_weight_constant:
layer.set_input(1, weight.trt_tensor)
if bias is not None and not is_bias_constant:
layer.set_input(2, bias.trt_tensor)
output = _create_tensor(layer.get_output(0), layer)
if ndim == 3:
return output.view(
concat([output.size(1),
output.size(2),
output.size(3)]))
return output
[docs]
def split(tensor: Tensor,
split_size_or_sections: Union[int, Sequence[int]],
dim: int = 0) -> Sequence[Tensor]:
'''
Add an operation that splits a tensor into sub-tensors.
This operation creates a list of tensors that are obtained from the input
tensor by slicing it along the dimension 'dim'. If 'split_size_or_sections'
is an integer, the tensor is split into 'input.shape[dim] /
split_size_or_sections' slices. If 'split_size_or_sections' is a list of
sizes, the tensor is split into 'len(split_size_or_sections)' slices and
the size of the ith slice is given by 'split_size_or_sections[i]'.
There are several constraints with the current implementation:
- The input tensor must be static (no dynamic dimension),
- If 'split_size_or_sections' is an integer, the number of elements in
the 'dim' dimension of the input must be a multiple of
'split_size_or_sections': 'input.shape[dim] % split_size_or_sections == 0'.
- If 'split_size_or_sections' is a sequence, the sum of the elements in
'split_size_or_sections' must be equal to the size in the dimension
'dim': 'input.shape[dim] == sum(ii for ii in split_size_or_sections)'.
That operation is implemented using a 'slice' operation for each output
slice.
Parameters:
tensor : Tensor
The input tensor to slice.
split_size_or_sections : Union[int, Sequence[int]]
If it is an integer, it encodes the size of each slice. Otherwise,
if it is a sequence, it is the size of each slice.
dim : int
The dimension of the tensor to slice.
Returns:
The list of tensors produced by the different operations.
'''
assert not tensor.is_dynamic(dim)
ndim = tensor.ndim()
if dim < 0:
dim += ndim
dim_value = tensor.size()[dim]
starts = [constant(dims_array([0])) for _ in range(ndim)]
sizes = [shape(tensor, i) for i in range(ndim)]
if isinstance(split_size_or_sections, int):
# TODO: support non-divisible cases
assert dim_value % split_size_or_sections == 0
num_sections = dim_value // split_size_or_sections
sizes[dim] = constant(dims_array([split_size_or_sections]))
outputs = []
for i in range(num_sections):
starts[dim] = constant(dims_array([split_size_or_sections * i]))
outputs.append(slice(tensor, concat(starts), concat(sizes)))
return outputs
else:
total_size = 0
for i in split_size_or_sections:
total_size += i
assert dim_value == total_size
num_sections = len(split_size_or_sections)
outputs = []
for i in range(num_sections):
if i > 0:
starts[dim] = starts[dim] + sizes[dim]
sizes[dim] = constant(dims_array([split_size_or_sections[i]]))
outputs.append(slice(tensor, concat(starts), concat(sizes)))
return outputs
[docs]
def chunk(tensor: Tensor, chunks: int, dim: int = 0) -> Tensor:
'''
Add an operation that splits a tensor into sub-tensors.
This operation creates a list of tensors that are obtained from the input
tensor by chunking it along the dimension 'dim'. It produces 'chunks'
sub-tensors.
That operation is only defined for static tensors (no dynamic dimension)
and the size of the tensor in the dimension 'dim' must be a multiple of
'chunks': 'input.shape[dim] % chunks == 0'.
It maps to 'split' with 'split_size = input.shape[dim] / chunks'.
Parameters:
tensor : Tensor
The input tensor to slice.
chunks : int
The number of slices to split the input tensor into.
dim : int
The dimension of the tensor to slice.
Returns:
The list of tensors produced by the different operations.
'''
assert not tensor.is_dynamic(dim)
ndim = tensor.ndim()
if dim < 0:
dim += ndim
dim_value = tensor.size()[dim]
assert dim_value % chunks == 0
return split(tensor, dim_value // chunks, dim)
[docs]
def unbind(input: Tensor, dim: int = 0):
'''
Removes a tensor dimension.
Returns a tuple of all slices along a given dimension, already without it.
'''
ndim = input.ndim()
outputs = split(input, 1, dim)
output_shape = [input.shape[i] for i in range(ndim) if i != dim]
return [output.view(output_shape) for output in outputs]
[docs]
class AllReduceStrategy(IntEnum):
"""
Warning: actual definition is in cpp/tensorrt_llm/kernels/customAllReduceKernels.h
they must be kept in sync
"""
NCCL = 0
ONESHOT = 1
TWOSHOT = 2
AUTO = 3
[docs]
class AllReduceConfig(IntFlag):
"""
Warning: actual definition is in cpp/tensorrt_llm/kernels/customAllReduceKernels.h
they must be kept in sync
"""
USE_MEMCPY = auto()
PUSH_MODE = auto()
[docs]
class AllReduceFusionOp(IntFlag):
"""
Warning: actual definition is in cpp/tensorrt_llm/kernels/customAllReduceKernels.h
they must be kept in sync
"""
NONE = 0
RESIDUAL_RMS_NORM = 1
[docs]
class AllReduceFusionParams():
def __init__(self,
fusion_op: AllReduceFusionOp = AllReduceFusionOp.NONE,
bias: Optional[Tensor] = None,
residual: Optional[Tensor] = None,
norm_weight: Optional[Tensor] = None,
eps: float = 1e-06):
self.fusion_op = fusion_op
self.bias = bias
self.residual = residual
self.norm_weight = norm_weight
self.eps = eps
assert fusion_op == AllReduceFusionOp.NONE or (residual is not None)
[docs]
def has_affine(self):
return 1 if self.norm_weight is not None else 0
[docs]
def has_bias(self):
return 1 if self.bias is not None else 0
[docs]
def create_allreduce_plugin(
network: trt.INetworkDefinition,
tensor: trt.ITensor,
workspace: Optional[trt.ITensor],
group: np.array,
strategy: AllReduceStrategy,
dtype: trt.DataType,
config: AllReduceConfig,
reduce_fusion_params: AllReduceFusionParams,
):
allreduce_plg_creator = trt.get_plugin_registry().get_plugin_creator(
'AllReduce', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert allreduce_plg_creator is not None
pf_group = trt.PluginField("group", group, trt.PluginFieldType.INT32)
pf_dtype = trt.PluginField("type_id", np.array([int(dtype)], np.int32),
trt.PluginFieldType.INT32)
pfc = [pf_group, pf_dtype]
p_strategy = trt.PluginField("strategy", np.array([int(strategy)], np.int8),
trt.PluginFieldType.INT8)
pfc.append(p_strategy)
p_config = trt.PluginField("config", np.array([int(config)], np.int8),
trt.PluginFieldType.INT8)
pfc.append(p_config)
p_fusion_op = trt.PluginField(
"fusion_op", np.array([int(reduce_fusion_params.fusion_op)], np.int8),
trt.PluginFieldType.INT8)
pfc.append(p_fusion_op)
p_eps = trt.PluginField(
"eps", np.array([float(reduce_fusion_params.eps)], np.float32),
trt.PluginFieldType.FLOAT32)
pfc.append(p_eps)
p_affine = trt.PluginField(
"affine", np.array([int(reduce_fusion_params.has_affine())], np.int8),
trt.PluginFieldType.INT8)
pfc.append(p_affine)
p_bias = trt.PluginField(
"bias", np.array([int(reduce_fusion_params.has_bias())], np.int8),
trt.PluginFieldType.INT8)
pfc.append(p_bias)
pfc = trt.PluginFieldCollection(pfc)
ar_plug = allreduce_plg_creator.create_plugin("allreduce", pfc)
plug_inputs = [tensor]
if strategy != AllReduceStrategy.NCCL:
plug_inputs.append(workspace)
if reduce_fusion_params.fusion_op == AllReduceFusionOp.RESIDUAL_RMS_NORM:
if reduce_fusion_params.has_bias() == 1:
plug_inputs.append(reduce_fusion_params.bias.trt_tensor)
plug_inputs.append(reduce_fusion_params.residual.trt_tensor)
if reduce_fusion_params.has_affine() == 1:
plug_inputs.append(reduce_fusion_params.norm_weight.trt_tensor)
layer = network.add_plugin_v2(plug_inputs, ar_plug)
return layer, allreduce_plg_creator, pfc
[docs]
def allreduce(
tensor: Tensor,
group: List[int],
strategy: Optional[AllReduceStrategy] = AllReduceStrategy.AUTO,
config: AllReduceConfig = AllReduceConfig(0),
reduce_fusion_params: Optional[AllReduceFusionParams] = None) -> Tensor:
'''
Add an operation that performs a collective all-reduce.
Let's define 'world_size' as the length of the 'group' list. That functions
creates a layer to compute the sum of 'world_size' tensors distributed
amongst the 'world_size' participating ranks (one GPU per rank).
The list 'group' contains the identifiers of the ranks participating into
the collective operation.
The tensors in the different ranks must be 1D tensors (or views) and the output
tensor will have that same shape. The output tensor will be replicated on
the 'world_size' ranks.
That operation is implemented using a plugin that wraps the NCCL all-reduce
collective operation. See
https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/usage/collectives.html#allreduce
for details.
Parameters:
tensor : Tensor
The input tensor.
group : List[int]
The ranks participating into the all-reduce operation.
strategy: AllReduceStrategy
NCCL delegates all-reduce to NCCL while ONESHOT and TWOSHOT are custom latency-optimal algorithms.
AUTO chooses amongst the three based on a message-size heuristic.
Returns:
The tensor produced by that layer.
'''
# TODO(TRTLLM-996): remove this WAR when custom allreduce is supported
# for encoder models in C++ runtime.
if current_all_reduce_helper().workspace is None:
strategy = AllReduceStrategy.NCCL
workspace = None
if strategy != AllReduceStrategy.NCCL:
workspace = current_all_reduce_helper().workspace.trt_tensor
if reduce_fusion_params is None:
reduce_fusion_params = AllReduceFusionParams()
dtype = default_net().plugin_config.nccl_plugin
layer, allreduce_plg_creator, pfc = create_allreduce_plugin(
network=default_trtnet(),
tensor=tensor.cast(dtype).trt_tensor,
workspace=workspace,
group=np.array(group, dtype=np.int32),
strategy=strategy,
dtype=str_dtype_to_trt(dtype),
config=config,
reduce_fusion_params=reduce_fusion_params,
)
_add_plugin_info(layer, allreduce_plg_creator, "allreduce", pfc)
if reduce_fusion_params.fusion_op == AllReduceFusionOp.RESIDUAL_RMS_NORM:
final_output = _create_tensor(layer.get_output(0),
layer).cast(tensor.dtype)
inter_output = _create_tensor(layer.get_output(1),
layer).cast(tensor.dtype)
return final_output, inter_output
else:
return _create_tensor(layer.get_output(0), layer).cast(tensor.dtype)
[docs]
def allgather(tensor: Tensor, group: List[int], gather_dim: int = 0) -> Tensor:
'''
Add an operation that performs a collective all-gather.
Let's define 'group_size' as the length of the 'group' list. That functions
creates a layer to gather 'group_size' tensors distributed
amongst the 'group_size' participating ranks (one GPU per rank).
The list 'group' contains the identifiers of the ranks participating into
the collective operation.
Note that 'group' here can be either TP group or PP group, because allgather communication is not limited to a specific split pattern. Therefore 'group_size' does not need to equal MPI 'world_size'.
The tensors in the different ranks must be 1D tensors (or views) and the
output tensor will have that same shape.
Given the 'section_size = input.shape[0] / group_size', each rank
contributes a section of its input tensor that correspond to
'rank*section_size:(rank+1)*section_size'.
That operation is implemented using a plugin that wraps the NCCL all-gather
collective operation. See
https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/usage/collectives.html#allgather
for details.
Parameters:
tensor : Tensor
The input tensor.
group : List[int]
The ranks participating into the all-gather operation.
gather_dim: int = 0
Gather along given dimension. By default 0, i.e. treated as 1D tensor.
Returns:
The tensor produced by that layer.
'''
allgather_plg_creator = trt.get_plugin_registry().get_plugin_creator(
'AllGather', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert allgather_plg_creator is not None
group_size = len(group)
group = trt.PluginField("group", np.array(group, dtype=np.int32),
trt.PluginFieldType.INT32)
p_dtype = default_net().plugin_config.nccl_plugin
pf_type = trt.PluginField(
"type_id", np.array([int(str_dtype_to_trt(p_dtype))], np.int32),
trt.PluginFieldType.INT32)
pfc = trt.PluginFieldCollection([group, pf_type])
allgather = allgather_plg_creator.create_plugin("allgather", pfc)
plug_inputs = [tensor.cast(p_dtype).trt_tensor]
layer = default_trtnet().add_plugin_v2(plug_inputs, allgather)
_add_plugin_info(layer, allgather_plg_creator, "allgather", pfc)
x = _create_tensor(layer.get_output(0), layer).cast(tensor.dtype)
# gather along a given dimension other than dim0
if gather_dim != 0:
# also support -1 type of dim representation
if gather_dim < 0:
gather_dim = x.ndim() + gather_dim
# plugin above gathers as 1D flattened tensor
# 1. [dim0, ...dimi, ...dimN] -> [group_size * dim0, ...dimi, ...dimN]
# now we need to gather-by-dim via split-concat
# 2. [group_size * dim0, ...dimi, ...dimN] -> [dim0, ...group_size * dimi, ...dimN]
# 2.1 split
split_size = shape(x, dim=0) / group_size
ndim = x.ndim()
starts = [constant(dims_array([0])) for _ in range(ndim)]
sizes = [shape(x, dim=d) for d in range(ndim)]
sizes[0] = split_size
sections = []
for i in range(group_size):
starts[0] = split_size * i
sections.append(slice(x, concat(starts), concat(sizes)))
# 2.2 concat
x = concat(sections, dim=gather_dim)
return x
[docs]
def send(tensor: Tensor, tgt: int) -> Tensor:
'''
Add an operation that performs a send from a rank to another.
The send operation sends a tensor from one rank to another. If a rank 'i'
sends a tensor to a rank 'j', the rank 'j' must have a corresponding 'recv'
operation from rank 'i'. See 'recv'.
That operation is implemented using a plugin that wraps the NCCL send
point-to-point operation. See
https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/api/p2p.html#ncclsend
for details.
Parameters:
tensor : Tensor
The input tensor.
tgt : int
The rank that receives the tensor.
Returns:
The tensor produced by that layer.
'''
send_plg_creator = trt.get_plugin_registry().get_plugin_creator(
'Send', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert send_plg_creator is not None
tgt = trt.PluginField("tgt_rank", np.array(tgt, dtype=np.int32),
trt.PluginFieldType.INT32)
p_dtype = default_net().plugin_config.nccl_plugin
pf_type = trt.PluginField(
"type_id", np.array([int(str_dtype_to_trt(p_dtype))], np.int32),
trt.PluginFieldType.INT32)
pfc = trt.PluginFieldCollection([tgt, pf_type])
send_plug = send_plg_creator.create_plugin("send", pfc)
plug_inputs = [tensor.cast(p_dtype).trt_tensor]
layer = default_trtnet().add_plugin_v2(plug_inputs, send_plug)
_add_plugin_info(layer, send_plg_creator, "send", pfc)
return _create_tensor(layer.get_output(0), layer).cast(tensor.dtype)
[docs]
def recv(tensor: Tensor, src: int) -> Tensor:
'''
Add an operation that performs a recv to a rank from another.
The recv operation receives a tensor from on a rank from another. If a rank 'i'
receives a tensor from a rank 'j', the rank 'j' must have a corresponding 'send'
operation to rank 'j'. See 'send'.
That operation is implemented using a plugin that wraps the NCCL recv
point-to-point operation. See
https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/api/p2p.html#ncclrecv
for details.
Parameters:
tensor : Tensor
The input tensor.
src : int
The rank that sends the tensor to.
Returns:
The tensor produced by that layer.
'''
recv_plg_creator = trt.get_plugin_registry().get_plugin_creator(
'Recv', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert recv_plg_creator is not None
src = trt.PluginField("src_rank", np.array(src, dtype=np.int32),
trt.PluginFieldType.INT32)
p_dtype = default_net().plugin_config.nccl_plugin
pf_type = trt.PluginField(
"type_id", np.array([int(str_dtype_to_trt(p_dtype))], np.int32),
trt.PluginFieldType.INT32)
pfc = trt.PluginFieldCollection([src, pf_type])
recv_plug = recv_plg_creator.create_plugin("recv", pfc)
plug_inputs = [tensor.cast(p_dtype).trt_tensor]
layer = default_trtnet().add_plugin_v2(plug_inputs, recv_plug)
_add_plugin_info(layer, recv_plg_creator, "recv", pfc)
return _create_tensor(layer.get_output(0), layer).cast(tensor.dtype)
[docs]
def bert_attention(tensor: Tensor,
input_lengths: Tensor,
num_heads: int,
head_size: int,
q_scaling: float,
relative_attention: bool = False,
relative_attention_bias: Tensor = None,
max_distance: int = 0,
max_input_length: Tensor = None) -> Tuple[Tensor]:
'''
Add an operation that performs the multi-head attention in BERT.
The multi-head attention (MHA) is the sequence of a batched matmul, a
softmax and a batched matmul as described in
https://arxiv.org/abs/1706.03762. That function adds an operation that
performs those computations using a single GPU kernel.
The input tensor contains the Q, K and V elements. It is a 2D tensor and
its shape is '[sum_of_tokens, 3*hidden_dim]' where the 'sum_of_tokens' is
the sum of the sequence lengths in the batch.
In MHA, the output of the Q*K^T product is scaled by a constant value that
is computed as:
1.f / (q_scaling * sqrt(head_size)).
That 'q_scaling' constant is the last argument of that function.
That layer is implemented using a plugin (see bertAttentionPlugin).
Parameters:
tensor : Tensor
The QKV input tensor.
input_lengths : Tensor
The length of each sequence. It is a 1D tensor of size 'batch_size'.
num_heads : int
The number of heads.
head_size : int
The size of each head.
q_scaling : float
The factor to compute the scaling factor to scale the output of the
'Q*K^T' product.
relative_attention: bool = False
If enable relative attention.
relative_attention_bias: Tensor = None
The relative attention bias [num_heads, max_seq_len, max_seq_len], or The relative attention embedding table for implicit mode, [num_heads, num_buckets].
max_distance: int = 0
The maximum distance of relative position in attention, for implicit mode.
Default value is 0, meaning to use the regular mode of relative attention bias.
Implicit mode is only enabled when passing in non-zero positive max_distance value.
See relative attention bias in docs/source/advanced/gpt-attention.md
max_input_length: Tensor = None
The maximum input sequence length represented by Tensor shape. Requires for remove_input_padding to pre-define plugin workspace size.
Returns:
The tensor produced by that layer.
'''
attn_plg_creator = trt.get_plugin_registry().get_plugin_creator(
'BertAttention', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert attn_plg_creator is not None
nheads = trt.PluginField("num_heads", np.array(num_heads, dtype=np.int32),
trt.PluginFieldType.INT32)
head_size = trt.PluginField("head_size", np.array(head_size,
dtype=np.int32),
trt.PluginFieldType.INT32)
q_scaling = trt.PluginField("q_scaling",
np.array(q_scaling, dtype=np.float32),
trt.PluginFieldType.FLOAT32)
context_fmha_type = trt.PluginField(
"context_fmha_type",
np.array(np.int8(default_net().plugin_config.context_fmha_type),
dtype=np.int8), trt.PluginFieldType.INT8)
p_dtype = default_net().plugin_config.bert_attention_plugin
pf_type = trt.PluginField(
"type_id", np.array([int(str_dtype_to_trt(p_dtype))], np.int32),
trt.PluginFieldType.INT32)
do_relative_attention = trt.PluginField(
"do_relative_attention",
np.array(np.int8(relative_attention), dtype=np.int8),
trt.PluginFieldType.INT8)
max_distance = trt.PluginField("max_distance",
np.array(max_distance, dtype=np.int32),
trt.PluginFieldType.INT32)
remove_padding = trt.PluginField(
"remove_padding",
np.array(np.int8(default_net().plugin_config.remove_input_padding),
dtype=np.int8), trt.PluginFieldType.INT8)
pfc = trt.PluginFieldCollection([
nheads, head_size, q_scaling, context_fmha_type, pf_type,
do_relative_attention, max_distance, remove_padding
])
attn_plug = attn_plg_creator.create_plugin("padding_attn", pfc)
plug_inputs = [tensor, input_lengths]
if max_input_length is not None:
# for remove padding mode
plug_inputs += [max_input_length]
if relative_attention_bias is not None:
# for relative attention mode
plug_inputs += [relative_attention_bias]
plug_inputs = [i.trt_tensor for i in plug_inputs]
layer = default_trtnet().add_plugin_v2(plug_inputs, attn_plug)
_add_plugin_info(layer, attn_plg_creator, "padding_attn", pfc)
assert layer.num_outputs == 1, \
f"Plugin outputs number mismatch with expected, got {layer.num_outputs}, expected 1"
output = _create_tensor(layer.get_output(0), layer)
assert output is not None
return output
[docs]
class RopeEmbeddingUtils:
[docs]
@staticmethod
# ref: https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_rope_utils.py#L298
def apply_llama3_scaling(inv_freqs: np.ndarray, rope_scaling_config: dict):
scale_factor = rope_scaling_config.get("factor", 8.0)
low_freq_factor = rope_scaling_config.get("low_freq_factor", 1.0)
high_freq_factor = rope_scaling_config.get("high_freq_factor", 4.0)
old_context_len = rope_scaling_config.get(
"original_max_position_embeddings", 8192)
low_freq_wavelen = old_context_len / low_freq_factor
high_freq_wavelen = old_context_len / high_freq_factor
new_inv_freqs = []
for inv_freq in inv_freqs:
wavelen = 2 * math.pi / inv_freq
if wavelen < high_freq_wavelen:
new_inv_freqs.append(inv_freq)
elif wavelen > low_freq_wavelen:
new_inv_freqs.append(inv_freq / scale_factor)
else:
assert low_freq_wavelen != high_freq_wavelen
smooth = (old_context_len / wavelen - low_freq_factor) / (
high_freq_factor - low_freq_factor)
new_inv_freqs.append((1 - smooth) * inv_freq / scale_factor +
smooth * inv_freq)
return np.array(new_inv_freqs, dtype=inv_freqs.dtype)
[docs]
@staticmethod
def create_sinusoidal_positions(num_pos: int,
dim: int,
theta: float = 10000.0,
dtype=np.float32):
inv_freq = 1.0 / (theta**(np.arange(0, dim, 2) / dim)).astype(dtype)
sinusoid_inp = np.einsum("i , j -> i j",
np.arange(num_pos, dtype=dtype),
inv_freq,
dtype=dtype)
concat = np.concatenate((np.sin(sinusoid_inp), np.cos(sinusoid_inp)),
axis=1)
return np.expand_dims(concat, axis=0).astype(dtype)
[docs]
@staticmethod
def create_sinusoidal_positions_for_attention_plugin(
num_pos: int,
dim: int,
theta: float = 10000.0,
scale: float = 1.0,
scale_type: RotaryScalingType = RotaryScalingType.none,
# Other scaling configs that only used by certain scaling types.
rope_scaling_config: dict = None,
dtype=np.float32):
if scale_type == RotaryScalingType.linear:
scale = 1.0 / scale
if scale_type == RotaryScalingType.llama3:
assert rope_scaling_config is not None, "rotary_scaling config must be provided."
inv_freq = 1.0 / (theta**(np.arange(0, dim, 2) / dim)).astype(dtype)
inv_freq = RopeEmbeddingUtils.apply_llama3_scaling(
inv_freq, rope_scaling_config)
else:
inv_freq = scale / (theta
**(np.arange(0, dim, 2) / dim)).astype(dtype)
sinusoid_inp = np.expand_dims(np.einsum("i , j -> i j",
np.arange(num_pos, dtype=dtype),
inv_freq,
dtype=dtype),
axis=-1)
# fuse cos/sin into float2 (cos, sin).
concat = np.concatenate((np.cos(sinusoid_inp), np.sin(sinusoid_inp)),
axis=-1)
return inv_freq, concat.reshape(1, -1).astype(dtype)
[docs]
@staticmethod
def create_sinusoidal_positions_for_cogvlm_attention_plugin(
num_pos: int,
dim: int,
theta: float = 10000.0,
scale: float = 1.0,
scale_type: RotaryScalingType = RotaryScalingType.none,
vision_start: int = 1,
vision_length: int = 1225,
dtype=np.float32):
if scale_type == RotaryScalingType.linear:
scale = 1.0 / scale
inv_freq = scale / (theta**(np.arange(0, dim, 2) / dim)).astype(dtype)
position_id = np.hstack([
np.arange(0, vision_start + 1, dtype=dtype),
np.full(vision_length, vision_start + 1, dtype=dtype),
np.arange(vision_start + 2,
num_pos - (vision_length - 1),
dtype=dtype)
])
sinusoid_inp = np.expand_dims(np.einsum("i , j -> i j",
position_id,
inv_freq,
dtype=dtype),
axis=-1)
# fuse cos/sin into float2 (cos, sin).
concat = np.concatenate((np.cos(sinusoid_inp), np.sin(sinusoid_inp)),
axis=-1)
return inv_freq, concat.reshape(1, -1).astype(dtype)
[docs]
def create_sinusoidal_positions_long_rope(
num_pos: int,
num_orig_pos: int,
dim: int,
theta: float = 10000.0,
scaling_short_factors: Tensor = 1.0,
scaling_long_factors: Tensor = 1.0,
short_mscale=None,
long_mscale=None,
dtype=np.float32):
def _calc_mscale(scale):
if scale <= 1.0:
return 1.0
return math.sqrt(1 + math.log(scale) / math.log(num_orig_pos))
if short_mscale is None:
short_mscale = _calc_mscale(num_pos / num_orig_pos)
long_mscale = short_mscale
def _compute_sinusoidal_positions(scale_factors, is_short,
for_attention_plugin):
inv_freq = 1 / (scale_factors *
(theta**(np.arange(0, dim, 2) / dim)).astype(dtype))
sinusoid_inp = np.einsum("i , j -> i j",
np.arange(num_pos, dtype=dtype),
inv_freq,
dtype=dtype)
if for_attention_plugin:
sinusoid_inp = np.expand_dims(sinusoid_inp, axis=-1)
concat = np.concatenate(
(np.cos(sinusoid_inp), np.sin(sinusoid_inp)), axis=-1)
else:
concat = np.concatenate(
(np.sin(sinusoid_inp), np.cos(sinusoid_inp)), axis=1)
concat = np.expand_dims(concat, axis=0)
mscale = short_mscale if is_short else long_mscale
# gpt attention plugins also need inv_freq.
if for_attention_plugin:
return inv_freq, concat.astype(dtype) * mscale
else:
return concat.astype(dtype) * mscale
return _compute_sinusoidal_positions(
scaling_short_factors, True, False), _compute_sinusoidal_positions(
scaling_long_factors,
False, False), _compute_sinusoidal_positions(
scaling_short_factors, True,
True), _compute_sinusoidal_positions(
scaling_long_factors, False, True), short_mscale
[docs]
@staticmethod
def rotate_every_two(tensor: Tensor) -> Tensor:
assert tensor.ndim() == 4
shape_tensor = concat([
shape(tensor, i) / 2 if i == (tensor.ndim() -
1) else shape(tensor, i)
for i in range(tensor.ndim())
])
x1 = slice(tensor, [0, 0, 0, 0], shape_tensor, [1, 1, 1, 2])
x2 = slice(tensor, [0, 0, 0, 1], shape_tensor, [1, 1, 1, 2])
x1 = expand_dims(x1, 4)
x2 = expand_dims(x2, 4)
zero = constant(
np.ascontiguousarray(
np.zeros([1], dtype=trt_dtype_to_np(tensor.dtype))))
x2 = zero - x2
x = concat([x2, x1], 4)
return view(
x, concat([shape(x, 0),
shape(x, 1),
shape(x, 2),
shape(x, 3) * 2]))
[docs]
@staticmethod
def rotate_half(tensor: Tensor) -> Tensor:
# [bs, num_attention_kv_heads, seqlen, attention_head_size]
assert tensor.ndim() == 4
shape_tensor = concat([
shape(tensor, i) / 2 if i == (tensor.ndim() -
1) else shape(tensor, i)
for i in range(tensor.ndim())
])
last_dim = shape(tensor, tensor.ndim() - 1) / 2
x1 = slice(tensor, [0, 0, 0, 0], shape_tensor, [1, 1, 1, 1])
x2 = slice(tensor, concat([0, 0, 0, last_dim]), shape_tensor,
[1, 1, 1, 1])
zero = constant(
np.ascontiguousarray(
np.zeros([1], dtype=trt_dtype_to_np(tensor.dtype))))
x2 = zero - x2
x = concat([x2, x1], 3)
return x
[docs]
@staticmethod
def apply_rotary_pos_emb(
tensor: Tensor,
position_embedding: List[Tensor] = None,
pos_emb_type: PositionEmbeddingType = PositionEmbeddingType.rope_gptj
) -> Tensor:
rotate_func = None
if pos_emb_type == PositionEmbeddingType.rope_gpt_neox or pos_emb_type == PositionEmbeddingType.long_rope:
assert len(position_embedding) == 2
cos, sin = position_embedding
sin = expand_dims(sin, 2)
cos = expand_dims(cos, 2)
sin = concat([sin, sin], 3)
cos = concat([cos, cos], 3)
rotate_func = RopeEmbeddingUtils.rotate_half
elif pos_emb_type == PositionEmbeddingType.rope_gptj:
assert len(position_embedding) == 2
cos, sin = position_embedding
sin = expand_dims(sin, 2)
cos = expand_dims(cos, 2)
sin = repeat_interleave(sin, 2, 3)
cos = repeat_interleave(cos, 2, 3)
rotate_func = RopeEmbeddingUtils.rotate_every_two
elif pos_emb_type == PositionEmbeddingType.chatglm:
assert len(position_embedding) == 4
cos0, cos1, sin0, sin1 = position_embedding
shape_tensor = concat([
shape(tensor, i) / 2 if i == (tensor.ndim() -
1) else shape(tensor, i)
for i in range(tensor.ndim())
])
last_dim = shape(tensor, tensor.ndim() - 1) / 2
x_part0 = slice(tensor, [0, 0, 0, 0], shape_tensor, [1, 1, 1, 1])
x_part1 = slice(tensor, concat([0, 0, 0, last_dim]), shape_tensor,
[1, 1, 1, 1])
y_part0 = (x_part0 *
cos0) + (RopeEmbeddingUtils.rotate_half(x_part0) * sin0)
y_part1 = (x_part1 *
cos1) + (RopeEmbeddingUtils.rotate_half(x_part1) * sin1)
result = concat([y_part0, y_part1], dim=3)
return result.view(shape(tensor))
else:
raise ValueError('The PositionEmbeddingType is not RoPE')
return (tensor * cos) + (rotate_func(tensor) * sin)
[docs]
@staticmethod
def apply_rotary_pos_emb_chatglm(qkv, position_embedding,
num_attention_heads, attention_head_size,
max_position_embeddings,
rotary_embedding_scale,
remove_input_padding) -> Tensor:
half_head_size = attention_head_size // 2
input = qkv[0] if isinstance(qkv, list) else qkv
input_shape = shape(input)
batch_size = 1 if remove_input_padding else shape(input, 0)
seqlen = shape(input, 0 if remove_input_padding else 1)
if isinstance(qkv, list):
query, key, value = qkv
else:
qkv = qkv.view(
concat([
batch_size,
seqlen,
num_attention_heads,
3,
attention_head_size,
]))
query, key, value = split(qkv, 1, dim=3)
q_shape = concat([
batch_size,
seqlen,
num_attention_heads,
attention_head_size,
])
query = query.view(q_shape)
key = key.view(q_shape)
value = value.view(q_shape)
embedding_weight = RopeEmbeddingUtils.create_sinusoidal_positions(
max_position_embeddings, half_head_size)
embedding_weight /= rotary_embedding_scale
embedding_weight = np.split(embedding_weight.squeeze(0), 2, axis=1)
embedding_weight = np.concatenate(
[
embedding_weight[0],
embedding_weight[0],
embedding_weight[1],
embedding_weight[1],
],
axis=1,
)
if remove_input_padding:
position_embedding = unsqueeze(position_embedding, 0)
embedding_weight = embedding_weight.astype(trt_dtype_to_np(query.dtype))
embedding_weight = constant(embedding_weight)
position_embedding = embedding(position_embedding, embedding_weight)
position_embedding, block_embedding = split(
position_embedding,
1,
dim=1,
)
sin0, cos0 = split(position_embedding, half_head_size, dim=3)
sin1, cos1 = split(block_embedding, half_head_size, dim=3)
new_shape = concat([
batch_size,
seqlen,
1,
half_head_size,
])
position_embedding = [
tensor.view(new_shape) for tensor in [cos0, cos1, sin0, sin1]
]
query = RopeEmbeddingUtils.apply_rotary_pos_emb(
tensor=query,
position_embedding=position_embedding,
pos_emb_type=PositionEmbeddingType.chatglm)
key = RopeEmbeddingUtils.apply_rotary_pos_emb(
tensor=key,
position_embedding=position_embedding,
pos_emb_type=PositionEmbeddingType.chatglm)
if isinstance(qkv, list):
qkv = [
query.view(input_shape),
key.view(input_shape),
value.view(input_shape),
]
else:
qkv = concat([query, key, value], dim=2)
qkv = qkv.view(input_shape)
return qkv
[docs]
@staticmethod
def apply_rotary_pos_emb_cogvlm(qkv, position_embedding,
num_attention_heads, attention_head_size,
max_position_embeddings,
rotary_embedding_scale,
remove_input_padding) -> Tensor:
input = qkv[0] if isinstance(qkv, list) else qkv
input_shape = shape(input)
batch_size = 1 if remove_input_padding else shape(input, 0)
seqlen = shape(input, 0 if remove_input_padding else 1)
if isinstance(qkv, list):
query, key, value = qkv
else:
qkv = qkv.view(
concat([
batch_size,
seqlen,
3,
num_attention_heads,
attention_head_size,
]))
query, key, value = split(qkv, 1, dim=2)
q_shape = concat([
batch_size,
seqlen,
num_attention_heads,
attention_head_size,
])
query = query.view(q_shape)
key = key.view(q_shape)
value = value.view(q_shape)
embedding_weight = RopeEmbeddingUtils.create_sinusoidal_positions(
max_position_embeddings, attention_head_size).squeeze(0)
embedding_weight /= rotary_embedding_scale # [max_position_embeddings, attention_head_size]
if remove_input_padding:
position_embedding = unsqueeze(position_embedding, 0) # [1, seqlen]
embedding_weight = constant(embedding_weight) # float32
position_embedding = embedding(
position_embedding,
embedding_weight) # [1, seqlen, attention_head_size]
sin, cos = split(position_embedding, attention_head_size // 2,
dim=-1) # [1, seqlen, attention_head_size//2]
input_dtype = query.dtype
fp32_query = cast(query, "float32")
fp32_key = cast(key, "float32")
fp32_query = RopeEmbeddingUtils.apply_rotary_pos_emb(
tensor=fp32_query,
position_embedding=[cos, sin],
pos_emb_type=PositionEmbeddingType.rope_gpt_neox)
fp32_key = RopeEmbeddingUtils.apply_rotary_pos_emb(
tensor=fp32_key,
position_embedding=[cos, sin],
pos_emb_type=PositionEmbeddingType.rope_gpt_neox)
query = cast(fp32_query, input_dtype)
key = cast(fp32_key, input_dtype)
if isinstance(qkv, list):
qkv = [
query.view(input_shape),
key.view(input_shape),
value.view(input_shape),
]
else:
qkv = concat([query, key, value], dim=2)
qkv = qkv.view(input_shape)
return qkv
[docs]
@gw.record_signature
def gpt_attention(
*,
qkv: Tensor,
past_key_value: Tensor,
context_fmha_custom_mask: Optional[Tensor] = None,
sequence_length: Tensor,
host_past_key_value_lengths: Optional[Tensor],
host_max_attention_window_sizes: Tensor,
host_sink_token_length: Tensor,
context_lengths: Optional[Tensor],
cache_indirection: Optional[Tensor],
host_request_types: Tensor,
layer_idx: int,
num_heads: int,
num_kv_heads: int,
hidden_size_per_head: int,
q_scaling: float,
qk_tanh_scale: float = 0.0,
rotary_embedding_dim: int = 0,
rotary_embedding_base: float = 10000.0,
rotary_embedding_scale_type: RotaryScalingType = RotaryScalingType.none,
rotary_embedding_short_m_scale: float = 1.0,
rotary_embedding_long_m_scale: float = 1.0,
rotary_embedding_scale: float = 1.0,
rotary_embedding_max_positions: int = 1024,
rotary_embedding_original_max_positions: int = 1024,
position_embedding_type: PositionEmbeddingType = PositionEmbeddingType.
learned_absolute,
rotary_inv_freq: Optional[Tensor] = None,
rotary_cos_sin: Optional[Tensor] = None,
kv_orig_quant_scale: Optional[Tensor] = None,
kv_quant_orig_scale: Optional[Tensor] = None,
attention_output_orig_quant_scale: Optional[Tensor] = None,
kv_cache_quant_mode: QuantMode = QuantMode(0),
max_context_length: Optional[int] = None,
mask_type: AttentionMaskType = AttentionMaskType.causal,
block_sparse_block_size: int = 64,
block_sparse_homo_head_pattern: bool = False,
block_sparse_num_local_blocks: int = 16,
block_sparse_vertical_stride: int = 8,
alibi_slopes: Optional[Tensor] = None,
tp_size: int = 1,
tp_rank: int = 0,
vision_start: int = -1,
vision_length: int = -1,
kv_cache_block_offsets: Optional[Tensor] = None,
host_kv_cache_block_offsets: Tensor = None,
host_kv_cache_pool_pointers: Tensor = None,
host_kv_cache_pool_mapping: Tensor = None,
do_cross_attention: bool = False,
cross_qkv: Optional[Tensor] = None, # for cross attention
cross_qkv_length: Optional[Tensor] = None, # for cross attention
encoder_input_lengths: Optional[Tensor] = None, # for cross attention
relative_attention_bias: Optional[Tensor] = None, # for relative attention
max_distance: int = 0, # for relative attention
host_context_lengths: Optional[Tensor] = None, # for pad-free input mode
qkv_bias: Optional[Tensor] = None,
use_cache: bool = True,
spec_decoding_is_generation_length_variable: bool = False,
spec_decoding_max_generation_length: int = 0,
spec_decoding_generation_lengths: Tensor = None,
spec_decoding_position_offsets: Tensor = None,
spec_decoding_packed_mask: Tensor = None,
host_runtime_perf_knobs: Optional[Tensor] = None,
layer_idx_in_cache_pool: Optional[int] = None,
) -> Tuple[Tensor, Optional[Tensor]]:
'''
Add an operation that performs the multi-head attention in GPT-like models.
The signature of the function will change in the future release - we are in
the process of simplifying the API. The current version is still
work-in-progress! The following API is provided with hints regarding the
arguments that are likely to be removed or merged with others in the future
release.
See docs/source/advanced/gpt-attention.md for the documentation of that function.
Parameters:
qkv: Tensor (On GPU)
The input QKV tensor. Its shape is [batch_beam_size, max_seqlen, qkv_dim] in padded mode and [1, num_tokens, qkv_dim] in
packed mode. Where qkv_dim depends on using MQA, GQA, or MHA. See QKV Input in docs/source/advanced/gpt-attention.md,
past_key_value: Tensor (On GPU)
The tensor that stores KV cache data. Its shape is
[max_batch_size * max_beam_width, 2, num_kv_heads, max_seqlen, hidden_dim_per_head]
in contiguous mode and
[max_blocks, 2, num_kv_heads, num_tokens_per_block, hidden_dim_per_head]
in paged mode. See KV Cache in docs/source/advanced/gpt-attention.md,
context_fmha_custom_mask: Tensor (On GPU)
The tensor that stores the packed custom mask for fmha.
Its shape is [num_tokens, max_kv_seqlen / 32].
sequence_lengths: Tensor (On GPU)
The tensor that stores the length of each sequence. Its shape is
[batch_size]. See QKV Input in docs/source/advanced/gpt-attention.md,
host_past_key_value_lengths: Tensor (On CPU)
An INT32 tensor of shape [batch_size],
host_max_attention_window_sizes: Tensor (On CPU)
An INT32 tensor of shape [1].
by default, the max_attention_window_size is determined by the shape of cache_indir_table.
And we support independent max_attention_window_size for each layer.
This controls the sliding-window-attention/cyclic-kv-cache features.
context_lengths: Tensor (On GPU)
The tensor that stores the context-phase sequence length of each request. Its shape
is [batch_size]. See QKV Input in doc/functional.py,
cache_indirection: Tensor (On GPU)
The tensor to reconstruct the paths when using beam-search. Its
shape is [batch_size, beam_width, max_seqlen]. See Beam-Search in
docs/source/advanced/gpt-attention.md,
host_request_types: Tensor = None (On CPU)
The tensor on the host that indicates if a request is in context or
generation phase. Its shape is [batch_size]. See Inflight Batching
in docs/source/advanced/gpt-attention.md,
layer_idx: int
The index of this attention layer, used to access kv_cache_block_offsets,
num_heads: int
The number of heads,
num_kv_heads: int
The number of KV heads, generic to handle MHA/MQA/GQA,
hidden_size_per_head: int
The hidden size per head,
q_scaling: float
The value used to compute the scaling factor applied to the output
of the Q*K^T product. See Scaling Factors in docs/source/advanced/gpt-attention.md,
qk_tanh_scale: float
The scale * tanh(value / scale) used to compute the scaling factor applied to the output
of the Q*K^T product. Note this is only used by grok models.
rotary_embedding_dim: int
The dimension to compute RoPE. Use 0 when position_embedding_type is not RoPE.
rotary_embedding_base: float
The theta value to use for RoPE. Ignored when position_embedding_type is not RoPE.
rotary_embedding_scale_type: RotaryScalingType
The scaling type of RoPE. Ignored when position_embedding_type is not RoPE.
Possible rotary scaling type:
* RotaryScalingType.none
* RotaryScalingType.linear
* RotaryScalingType.dynamic
* RotaryScalingType.longrope
* RotaryScalingType.llama3
rotary_embedding_scale: float
The scale value to use for linear/dynamic scaling in RoPE.
Ignored when position_embedding_type is not RoPE.
Must be set to 1 (default) if rotary_embedding_scale_type is `none`.
rotary_inv_freq: float Tensor
The rotary inv freq with shape [head_size / 2].
rotary_cos_sin: float2(cos/sin) Tensor
The rotary cos/sin cache, which will be reused among different requests.
It is taken as constant tensor.
rotary_embedding_max_positions: int
Needed only for `dynamic` RoPE scaling. Ignored otherwise.
position_embedding_type: PositionEmbeddingType
The position embedding type:
* PositionEmbeddingType.learned_absolute
* PositionEmbeddingType.relative
* PositionEmbeddingType.rope_gptj
* PositionEmbeddingType.rope_gpt_neox
* PositionEmbeddingType.alibi
* PositionEmbeddingType.alibi_with_scale
kv_orig_quant_scale: Tensor
The tensor to store the scaling factor for quantization to INT8/FP8
in the KV cache. Its shape is [1]. See INT8/FP8 KV Cache in
docs/source/advanced/gpt-attention.md,
kv_quant_orig_scale: Tensor
The tensor to store the scaling factor for dequantization from
INT8/FP8 in the KV cache. Its shape is [1]. See INT8/FP8 KV Cache
in docs/source/advanced/gpt-attention.md,
attention_output_orig_quant_scale: Tensor
The tensor to store the scaling factor for quantization to FP8
in the KV cache. Its shape is [1].
kv_cache_quant_mode: QuantMode (int flags)
Do we enable the INT8 or FP8 KV cache?
max_context_length: int32_t
The length of the longest input sequence. See QKV Input in
docs/source/advanced/gpt-attention.md,
mask_type: int = 1
The type of mask:
* tensorrt_llm.layers.AttentionMaskType.padding for BERT,
* tensorrt_llm.layers.AttentionMaskType.causal for GPT,
* tensorrt_llm.layers.AttentionMaskType.sliding_window_causal for GPT,
* tensorrt_llm.layers.AttentionMaskType.bidirectional for ChatGLM-6B,
* tensorrt_llm.layers.AttentionMaskType.bidirectionalglm for GLM-10B,
* tensorrt_llm.layers.AttentionMaskType.blocksparse for Phi-3-small,
* tensorrt_llm.layers.AttentionMaskType.custom_mask for any models.
block_sparse_block_size: int
Block size in block sparse attention
block_sparse_homo_head_pattern: bool
Do all attention heads share same vertical stride pattern?
block_sparse_num_local_blocks: int
Number of active blocks near diagonal
block_sparse_vertical_stride: int
Stride of active blocks in vertical dimension
alibi_slopes: Tensor
The ALiBi slopes. The ALiBi bias is computed on-the-fly in the kernel
when possible,
tp_size: int
The number of processes/GPUs when tensor parallelism is activated,
tp_rank: int
The rank of that process (when running tensor parallelism),
kv_cache_block_offsets:
The tensor of block offsets for the KV cache. Its shape is
[num_layers, max_batch_size, max_beam_width, 2, max_blocks_per_sequence * 2],
See KV cache section in docs/source/advanced/gpt-attention.md, on gpu,
host_kv_cache_block_offsets:
The same as kv_cache_block_offsets, but on cpu,
host_kv_cache_pool_pointers:
The tensor of pool pointers for the KV cache. Its shape is [num_layers, 2],
See KV cache section in docs/source/advanced/gpt-attention.md, on gpu,
host_kv_cache_pool_mapping:
The tensor of pool mapping for the different memory pools. Its shape is [num_layers,],
do_cross_attention: bool = False
Do we use this as cross attention instead of self attention,
cross_qkv: Tensor = None
The QKV tensor of encoder output hidden states. Its shape is [batch_size, max_seqlen, 3
* hidden_dim] in padded mode and [1, num_tokens, 3 * hidden_dim] in
packed mode,
cross_qkv_length: Tensor = None
The length of the longest encoder output sequence,
encoder_input_lengths: Tensor
The tensor that stores the length of each encoder input sequence. Its shape is [batch_size],
relative_attention_bias: Tensor = None
The relative attention bias [num_heads, max_seq_len, max_seq_len], or The relative attention embedding table for implicit mode, [num_heads, num_buckets].
max_distance: int = 0
The maximum distance of relative position in attention, for implicit mode.
Default value is 0, meaning to use the regular mode of relative attention bias.
Implicit mode is only enabled when passing in non-zero positive max_distance value.
See relative attention bias in docs/source/advanced/gpt-attention.md
host_context_lengths: Tensor = None (On CPU)
A host tensor that contains the lengths of the different inputs,
qkv_bias: Tensor = None,
The qkv bias tensor.
use_cache: bool = False
Do we need to store kv cache ? not needed if there is no generation phase.
spec_decoding_is_generation_length_variable: bool = False,
Whether the generation lengths can be different for each sequence in a batch.
For Medusa, this should be set False.
For Redrafter, this should be set to True.
spec_decoding_max_generation_length: int = 1,
The maximum number of tokens possible in the generation phase per sequence.
spec_decoding_generation_lengths: Tensor = None,
The generation phase tokens' lengths for each sequence.
Shape: [batch_size]
spec_decoding_position_offsets: Tensor = None,
The speculative decoding tokens's position offsets (shared by all sequences).
Shape: [batch_size, num_draft_tokens + 1].
spec_decoding_packed_mask: Tensor = None,
The speculative decoding tokens's attention mask (packed into uint32_t bits).
remove_input_padding is False:
Shape: [batch_size, num_draft_tokens + 1, divUp(num_draft_tokens + 1, 32)].
remove_input_padding is True:
Shape: [sum(spec_decoding_generation_lengths), divUp(num_draft_tokens + 1, 32)].
host_runtime_perf_knobs: Tensor = None,
The runtime perf knobs bit mask, controls whether to use certain perf knob in the runtime.
Returns:
The tensor produced by that layer.
'''
assert host_request_types is not None
assert (alibi_slopes is not None) == (position_embedding_type.is_alibi())
attn_plg_creator = trt.get_plugin_registry().get_plugin_creator(
'GPTAttention', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert attn_plg_creator is not None
assert host_context_lengths is not None or not default_net(
).plugin_config.remove_input_padding
assert isinstance(max_context_length, int)
assert host_max_attention_window_sizes is not None
assert host_sink_token_length is not None
if layer_idx_in_cache_pool is None:
layer_idx_in_cache_pool = layer_idx
paged_kv_cache_flag = default_net().plugin_config.paged_kv_cache
if isinstance(qkv, list):
is_unfuse_qkv_gemm = 1
else:
is_unfuse_qkv_gemm = 0
unfuse_qkv_gemm = trt.PluginField(
"unfuse_qkv_gemm", np.array(np.int8(is_unfuse_qkv_gemm), dtype=np.int8),
trt.PluginFieldType.INT8)
layer_idx = trt.PluginField("layer_idx", np.array(layer_idx,
dtype=np.int32),
trt.PluginFieldType.INT32)
nheads = trt.PluginField("num_heads", np.array(num_heads, dtype=np.int32),
trt.PluginFieldType.INT32)
vision_start = trt.PluginField("vision_start",
np.array(vision_start, dtype=np.int32),
trt.PluginFieldType.INT32)
vision_length = trt.PluginField("vision_length",
np.array(vision_length, dtype=np.int32),
trt.PluginFieldType.INT32)
num_kv_heads = trt.PluginField("num_kv_heads",
np.array(num_kv_heads, dtype=np.int32),
trt.PluginFieldType.INT32)
layer_idx_in_cache_pool = trt.PluginField(
"layer_idx_in_cache_pool",
np.array(layer_idx_in_cache_pool, dtype=np.int32),
trt.PluginFieldType.INT32)
head_size = trt.PluginField("head_size",
np.array(hidden_size_per_head, dtype=np.int32),
trt.PluginFieldType.INT32)
unidirectional = trt.PluginField("unidirectional",
np.array(1, dtype=np.int32),
trt.PluginFieldType.INT32)
q_scaling = trt.PluginField("q_scaling",
np.array(q_scaling, dtype=np.float32),
trt.PluginFieldType.FLOAT32)
qk_tanh_scale = trt.PluginField("qk_tanh_scale",
np.array(qk_tanh_scale, dtype=np.float32),
trt.PluginFieldType.FLOAT32)
rotary_embedding_dim = trt.PluginField(
"rotary_embedding_dim", np.array(rotary_embedding_dim, dtype=np.int32),
trt.PluginFieldType.INT32)
rotary_embedding_base = trt.PluginField(
"rotary_embedding_base",
np.array(rotary_embedding_base, dtype=np.float32),
trt.PluginFieldType.FLOAT32)
rotary_embedding_scale_type = trt.PluginField(
"rotary_embedding_scale_type",
np.array(rotary_embedding_scale_type, dtype=np.int8),
trt.PluginFieldType.INT8)
rotary_embedding_scale = trt.PluginField(
"rotary_embedding_scale",
np.array(rotary_embedding_scale, dtype=np.float32),
trt.PluginFieldType.FLOAT32)
rotary_embedding_short_m_scale = trt.PluginField(
"rotary_embedding_short_m_scale",
np.array(rotary_embedding_short_m_scale, dtype=np.float32),
trt.PluginFieldType.FLOAT32)
rotary_embedding_long_m_scale = trt.PluginField(
"rotary_embedding_long_m_scale",
np.array(rotary_embedding_long_m_scale, dtype=np.float32),
trt.PluginFieldType.FLOAT32)
rotary_embedding_max_positions = trt.PluginField(
"rotary_embedding_max_positions",
np.array(rotary_embedding_max_positions, dtype=np.int32),
trt.PluginFieldType.INT32)
rotary_embedding_original_max_positions = trt.PluginField(
"rotary_embedding_original_max_positions",
np.array(rotary_embedding_original_max_positions, dtype=np.int32),
trt.PluginFieldType.INT32)
position_embedding_type = trt.PluginField(
"position_embedding_type",
np.array(int(position_embedding_type), dtype=np.int8),
trt.PluginFieldType.INT8)
context_fmha_type = trt.PluginField(
"context_fmha_type",
np.array(np.int8(default_net().plugin_config.context_fmha_type),
dtype=np.int8), trt.PluginFieldType.INT8)
remove_input_padding = trt.PluginField(
"remove_input_padding",
np.array(np.int8(default_net().plugin_config.remove_input_padding),
dtype=np.int8), trt.PluginFieldType.INT8)
is_spec_decoding_enabled = trt.PluginField(
"is_spec_decoding_enabled",
np.array(np.int8(spec_decoding_packed_mask is not None), dtype=np.int8),
trt.PluginFieldType.INT8)
spec_decoding_is_generation_length_variable = trt.PluginField(
"spec_decoding_is_generation_length_variable",
np.array(np.int8(spec_decoding_is_generation_length_variable),
dtype=np.int8), trt.PluginFieldType.INT8)
spec_decoding_max_generation_length = trt.PluginField(
"spec_decoding_max_generation_length",
np.array(spec_decoding_max_generation_length, dtype=np.int32),
trt.PluginFieldType.INT32)
p_dtype = default_net().plugin_config.gpt_attention_plugin
pf_type = trt.PluginField(
"type_id", np.array([int(str_dtype_to_trt(p_dtype))], np.int32),
trt.PluginFieldType.INT32)
# reset mask_type to custom_mask.
if context_fmha_custom_mask is not None:
mask_type = AttentionMaskType.custom_mask
mask_type = trt.PluginField("mask_type", np.array([int(mask_type)],
np.int32),
trt.PluginFieldType.INT32)
block_sparse_block_size = trt.PluginField(
"block_sparse_block_size", np.array([block_sparse_block_size],
np.int32),
trt.PluginFieldType.INT32)
block_sparse_homo_head_pattern = trt.PluginField(
"block_sparse_homo_head_pattern",
np.array(np.int8(block_sparse_homo_head_pattern), np.int8),
trt.PluginFieldType.INT8)
block_sparse_num_local_blocks = trt.PluginField(
"block_sparse_num_local_blocks",
np.array([block_sparse_num_local_blocks], np.int32),
trt.PluginFieldType.INT32)
block_sparse_vertical_stride = trt.PluginField(
"block_sparse_vertical_stride",
np.array([block_sparse_vertical_stride], np.int32),
trt.PluginFieldType.INT32)
enable_xqa = trt.PluginField(
"enable_xqa",
np.array(np.int8(default_net().plugin_config.enable_xqa),
dtype=np.int8), trt.PluginFieldType.INT8)
tp_size = trt.PluginField("tp_size", np.array(tp_size, dtype=np.int32),
trt.PluginFieldType.INT32)
tp_rank = trt.PluginField("tp_rank", np.array(tp_rank, dtype=np.int32),
trt.PluginFieldType.INT32)
kv_cache_quant_mode_field = trt.PluginField(
"kv_cache_quant_mode", np.array(kv_cache_quant_mode, dtype=np.int32),
trt.PluginFieldType.INT32)
paged_kv_cache = trt.PluginField(
"paged_kv_cache", np.array(paged_kv_cache_flag, dtype=np.int32),
trt.PluginFieldType.INT32)
tokens_per_block = trt.PluginField(
"tokens_per_block",
np.array(default_net().plugin_config.tokens_per_block, dtype=np.int32),
trt.PluginFieldType.INT32)
max_context_length = trt.PluginField("max_context_length",
np.array(max_context_length, np.int32),
trt.PluginFieldType.INT32)
pos_shift_enabled = trt.PluginField(
"pos_shift_enabled",
np.array(np.int8(default_net().plugin_config.streamingllm),
dtype=np.int8), trt.PluginFieldType.INT8)
dense_context_fmha = trt.PluginField(
"dense_context_fmha",
np.array(np.int8(default_net().plugin_config.streamingllm),
dtype=np.int8), trt.PluginFieldType.INT8)
if qkv_bias is None:
qkv_bias_enabled = trt.PluginField("qkv_bias_enabled",
np.array(0, dtype=np.int8),
trt.PluginFieldType.INT8)
else:
qkv_bias_enabled = trt.PluginField("qkv_bias_enabled",
np.array(1, dtype=np.int8),
trt.PluginFieldType.INT8)
do_cross_attention_field = trt.PluginField(
"do_cross_attention",
np.array(np.int8(do_cross_attention), dtype=np.int8),
trt.PluginFieldType.INT8)
max_distance = trt.PluginField("max_distance",
np.array(max_distance, dtype=np.int32),
trt.PluginFieldType.INT32)
use_paged_context_fmha_field = trt.PluginField(
"use_paged_context_fmha",
np.array(np.int8(default_net().plugin_config.use_paged_context_fmha),
dtype=np.int8), trt.PluginFieldType.INT8)
use_fp8_context_fmha_field = trt.PluginField(
"use_fp8_context_fmha",
np.array(np.int8(default_net().plugin_config.use_fp8_context_fmha),
dtype=np.int8), trt.PluginFieldType.INT8)
use_cache_pf = trt.PluginField("use_cache",
np.array([use_cache], dtype=np.int32),
trt.PluginFieldType.INT32)
pfc = trt.PluginFieldCollection([
layer_idx, nheads, vision_start, vision_length, num_kv_heads,
layer_idx_in_cache_pool, head_size, unidirectional, q_scaling,
qk_tanh_scale, position_embedding_type, rotary_embedding_dim,
rotary_embedding_base, rotary_embedding_scale_type,
rotary_embedding_scale, rotary_embedding_short_m_scale,
rotary_embedding_long_m_scale, rotary_embedding_max_positions,
rotary_embedding_original_max_positions, tp_size, tp_rank,
unfuse_qkv_gemm, context_fmha_type, enable_xqa,
kv_cache_quant_mode_field, remove_input_padding, mask_type,
block_sparse_block_size, block_sparse_homo_head_pattern,
block_sparse_num_local_blocks, block_sparse_vertical_stride,
paged_kv_cache, tokens_per_block, pf_type, max_context_length,
qkv_bias_enabled, do_cross_attention_field, max_distance,
pos_shift_enabled, dense_context_fmha, use_paged_context_fmha_field,
use_fp8_context_fmha_field, use_cache_pf, is_spec_decoding_enabled,
spec_decoding_is_generation_length_variable,
spec_decoding_max_generation_length
])
attn_plug = attn_plg_creator.create_plugin("causal_attn", pfc)
assert attn_plug
plug_inputs = [*qkv] if is_unfuse_qkv_gemm else [qkv]
if context_fmha_custom_mask is not None:
plug_inputs += [context_fmha_custom_mask]
if use_cache:
plug_inputs += [
sequence_length,
host_past_key_value_lengths,
host_max_attention_window_sizes,
host_sink_token_length,
context_lengths,
cache_indirection,
host_request_types,
]
else:
plug_inputs += [
host_max_attention_window_sizes,
host_sink_token_length,
context_lengths,
host_request_types,
]
if use_cache:
if paged_kv_cache_flag:
assert kv_cache_block_offsets is not None, "Paged kv cache is enabled, the kv_cache_block_offsets tensor shall not be None"
assert host_kv_cache_block_offsets is not None, "Paged kv cache is enabled, the host_kv_cache_block_offsets tensor shall not be None"
assert host_kv_cache_pool_pointers is not None, "Paged kv cache is enabled, the host_kv_cache_pool_pointers tensor shall not be None"
assert host_kv_cache_pool_mapping is not None, "Paged kv cache is enabled, the host_kv_cache_pool_mapping tensor shall not be None"
plug_inputs += [
kv_cache_block_offsets, host_kv_cache_block_offsets,
host_kv_cache_pool_pointers, host_kv_cache_pool_mapping
]
else:
plug_inputs += [past_key_value]
if use_cache and kv_cache_quant_mode.has_kv_cache_quant():
plug_inputs += [kv_orig_quant_scale, kv_quant_orig_scale]
if attention_output_orig_quant_scale is not None:
assert default_net(
).plugin_config.use_fp8_context_fmha, "FP8 Context FMHA needs to be enabled"
plug_inputs += [attention_output_orig_quant_scale]
if rotary_inv_freq is not None:
plug_inputs += [rotary_inv_freq]
if rotary_cos_sin is not None:
plug_inputs += [rotary_cos_sin]
if alibi_slopes is not None:
plug_inputs += [alibi_slopes]
if relative_attention_bias is not None:
plug_inputs += [relative_attention_bias]
if do_cross_attention:
plug_inputs += [cross_qkv, cross_qkv_length, encoder_input_lengths]
if default_net().plugin_config.remove_input_padding:
plug_inputs += [host_context_lengths]
if qkv_bias is not None:
plug_inputs += [qkv_bias]
if spec_decoding_packed_mask is not None:
# add position_ids as well only if speculative decoding mode
assert spec_decoding_position_offsets is not None
assert spec_decoding_generation_lengths is not None
plug_inputs += [
spec_decoding_generation_lengths, spec_decoding_packed_mask,
spec_decoding_position_offsets
]
if host_runtime_perf_knobs is not None:
plug_inputs += [host_runtime_perf_knobs]
for idx, i in enumerate(plug_inputs):
assert i is not None, f"Found None input for {idx} th item in plugin inputs {plug_inputs}"
plug_inputs = [i.trt_tensor for i in plug_inputs]
layer = default_trtnet().add_plugin_v2(plug_inputs, attn_plug)
_add_plugin_info(layer, attn_plg_creator, "causal_attn", pfc)
output = _create_tensor(layer.get_output(0), layer)
present_key_value = None
if use_cache and not paged_kv_cache_flag:
present_key_value = _create_tensor(layer.get_output(1), layer)
assert present_key_value is not None
expected_outputs = 2
else:
expected_outputs = 1
assert layer.num_outputs == expected_outputs, \
f"Plugin outputs number mismatch with expected, got {layer.num_outputs}, expected {expected_outputs}"
if kv_cache_quant_mode.has_int8_kv_cache(
) and not default_net().strongly_typed:
if not paged_kv_cache_flag:
# past key value
layer.get_input(8).set_dynamic_range(-127, 127)
# present key value
layer.get_output(1).set_dynamic_range(-127, 127)
else:
layer.get_input(0).set_dynamic_range(-127, 127)
layer.get_input(1).set_dynamic_range(-127, 127)
layer.get_output(0).set_dynamic_range(-127, 127)
assert output is not None
return output, present_key_value
[docs]
def assertion(condition: Tensor, message: str = '') -> None:
default_trtnet().add_assertion(condition.trt_tensor, message)
[docs]
def layer_norm(input: Tensor,
normalized_shape: Union[int, Tuple[int]],
weight: Optional[Tensor] = None,
bias: Optional[Tensor] = None,
eps: float = 1e-05,
use_diff_of_squares: bool = True) -> Tensor:
'''
Add a layer-norm operation on a tensor.
That operation applies the layer-normalization to its input tensor. In its
simplest form, for large language models, the 'normalized_shape' should be
set to the hidden dimension of the activation tensor. Otherwise, it is the
shape of the normalized fraction of the tensor (starting from the
right-most dimension).
The 'weight' tensor corresponds to 'gamma' in the layer-norm formula and
'bias' is 'beta'. The 'eps' value is added to the variance before computing
the squared-root.
This implementation (when using the plugin) supports an additional flag to
enable/disable the use of a difference of squares ('Var = Mean(X^2) -
Mean(X)^2').
Parameters:
input : Tensor
The tensor to normalize.
normalized_shape : Union[int, Tuple[int]]
The shape of the sub-tensor that is normalized. Use 'hidden_dim' to
normalize the inner-most dimension of an activation tensor in LLMs.
weight : Optional[Tensor] = None
The 'gamma' term in layer-norm. Its shape must be
'normalized_shape'.
bias : Optional[Tensor] = None
The 'beta' term in layer-norm. Its shape must be
'normalized_shape'.
eps : float
The epsilon term to be added to the variance in the squared-root.
use_diff_of_squares : bool
Does the plugin use the difference of squares to compute the
variance?
Returns:
The output tensor of that operation.
'''
input, weight = broadcast_helper(input, weight)
input, bias = broadcast_helper(input, bias)
if isinstance(normalized_shape, int): # FIXME: better way?
axis = input.ndim() - 1
else:
axis = input.ndim() - len(normalized_shape)
axes_mask = 0
for i in range(axis, input.ndim()):
axes_mask |= 1 << i
layer = default_trtnet().add_normalization(input.trt_tensor,
weight.trt_tensor,
bias.trt_tensor, axes_mask)
layer.epsilon = eps
return _create_tensor(layer.get_output(0), layer)
[docs]
def rms_norm(input: Tensor,
normalized_shape: Union[int, Tuple[int]],
num_groups: int = 1,
weight: Optional[Tensor] = None,
eps: float = 1e-06) -> Tensor:
'''
Add a RMS norm operation on a tensor.
That operation applies the rms-normalization to its input tensor. In its
simplest form, for large language models, the 'normalized_shape' should be
set to the hidden dimension of the activation tensor. Otherwise, it is the
shape of the normalized fraction of the tensor (starting from the
right-most dimension).
The 'weight' tensor corresponds to 'gamma' in the rms-norm formula.
The 'eps' value is added to the variance before computing the squared-root.
Parameters:
input: Tensor
The tensor to normalize.
normalized_shape : Union[int, Tuple[int]]
The shape of the sub-tensor that is normalized. Use 'hidden_dim' to
normalize the inner-most dimension of an activation tensor in LLMs.
num_groups: int = 1
The group size.
weight : Optional[Tensor] = None
The 'gamma' term in layer-norm. Its shape must be
'normalized_shape'.
eps : float
The epsilon term to be added to the variance in the squared-root.weig
Returns:
The output tensor of that operation.
'''
normalized_shape = [normalized_shape] if isinstance(
normalized_shape, int) else normalized_shape
dim = tuple([-i - 1 for i in range(len(normalized_shape))])
if num_groups > 1:
assert len(normalized_shape) == 1
num_channels = input.size()[-1]
ndim = input.ndim()
old_shape = shape(input)
new_shape = concat([input.size(i) for i in range(ndim - 1)] +
[num_groups, num_channels // num_groups])
input = input.view(new_shape)
with precision("float32"):
input_dtype = input.dtype
fp32_input = cast(input, "float32")
varx = pow(fp32_input, 2.0)
varx = varx.mean(dim=dim, keepdim=True)
denom = varx + eps
denom = denom.sqrt()
fp32_y = fp32_input / denom
y = cast(fp32_y, input_dtype)
if num_groups > 1:
y = y.view(old_shape)
if weight is not None:
y = y * weight
return y
[docs]
def repeat_interleave(tensor: Tensor, repeats: int, dim: int) -> Tensor:
'''
Repeats elements of a tensor along an axis.
Parameters:
repeats : int
The number of repetitions along axis specified.
dim : int
The dimension along which repetitions are performed.
Returns:
A tensor with the same shape as input except for repeated elements along specified dim.
TODO: Allow repeats to be a list of integers and dim to be unspecified.
'''
expanded_tensor = expand_dims(tensor, dim + 1)
tile_output_size = concat([
repeats if i == (dim + 1) else shape(expanded_tensor, i)
for i in range(expanded_tensor.ndim())
])
tile = expand(expanded_tensor, tile_output_size)
tile_reshape_size = [shape(tensor, i) for i in range(tensor.ndim())]
tile_reshape_size[dim] = tile_reshape_size[dim] * repeats
tensor = tile.view(concat(tile_reshape_size))
return tensor
[docs]
def generate_alibi_slopes(num_heads: int,
tp_size: int = 1,
tp_rank: int = 0,
alibi_scale: float = 1.0,
alibi_bias_max: int = 8) -> np.ndarray:
'''
Compute the ALiBi slopes as described in https://arxiv.org/abs/2211.05100.
Parameters:
num_heads : int
The number of heads.
dtype : trt.DataType
The data type of the returned slopes
tp_size : int
The tensor parallelism size
tp_rank : int
The tensor parallelism rank
Returns:
A constant tensor that contains the ALiBi slopes.
'''
start_head_id = 0
end_head_id = num_heads
if tp_size > 1:
rank_heads = num_heads // tp_size
start_head_id = rank_heads * tp_rank
end_head_id = start_head_id + rank_heads
closest_power_of_2 = 2**np.floor(np.log2(num_heads))
# FT's implementation
# https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/kernels/gen_relative_pos_bias.cu#L248
slopes_ft = []
for h_id in range(start_head_id, end_head_id):
if h_id < closest_power_of_2:
slopes_ft.append(
np.power(
2**(-(2**-(np.log2(closest_power_of_2) -
np.log2(alibi_bias_max)))), h_id + 1))
else:
slopes_ft.append(
np.power(
2**(-(2**-(np.log2(closest_power_of_2 * 2) -
np.log2(alibi_bias_max)))),
(h_id - closest_power_of_2) * 2 + 1))
slopes = np.asarray(slopes_ft, dtype=np.float32)
slopes = alibi_scale * slopes
slopes = slopes.reshape(1, (end_head_id - start_head_id), 1, 1)
return slopes
[docs]
def generate_alibi_biases(slopes: Tensor, key_length: Tensor) -> Tensor:
'''
Compute the ALiBi biases as described in https://arxiv.org/abs/2211.05100.
The ALiBi biases are added to the result of the Q*K^T product in the
multi-head attention block.
Parameters:
slopes : Tensor
The slopes.
key_length : Tensor
The size of the K vector per head.
Returns:
A constant tensor that contains the ALiBi biases.
'''
# We don't need to care about the batch size or query length since we can just broadcast
# across the batch and query dimensions
trt_0 = constant(int32_array(0))
arange_shape = concat([1, 1, 1, key_length])
arange_tensor = arange(trt_0, key_length, "float32").view(arange_shape)
return slopes * arange_tensor
[docs]
def expand_mask(mask: Tensor, tgt_len: Optional[Tensor] = None) -> Tensor:
'''
Expand an attention mask.
That function adds the sequence of operations to expand from a tensor of
shape '[batch_size, src_seq_len]' to a tensor of shape
'[batch_size, 1, tgt_seq_len, src_seq_len]'. It can be used to create the
mask applied to the Q*K^T product before the softmax operation in the
multi-head attention block.
Parameters:
mask : Tensor
The input mask
tgt_len : Optional[Tensor]
The dimension of the 3rd dimension in the output tensor. If None,
the 2nd dimension of the input is used.
Returns:
The tensor created by that sequence of operations.
'''
bsz = shape(mask, 0)
src_len = shape(mask, 1)
tgt_len = tgt_len if tgt_len is not None else src_len
mask = mask.view(concat([bsz, 1, 1, src_len]))
mask = expand(mask, concat([bsz, 1, tgt_len, src_len]))
mask = where(mask == 0, float('-inf'), 0.0)
return mask
[docs]
def gather_last_token_logits(hidden_states: Tensor, last_token_ids: Tensor,
remove_input_padding: bool) -> Tensor:
'''
Extract the logits that correspond to the last token from the hidden states.
That function adds the operations to extract the logits of the last tokens
in a batch of sequences.
Depending on whether 'remove_input_padding' is 'True' or 'False', that
function assumes inputs of different shapes.
When 'remove_input_padding' is 'True', the 'hidden_states' tensor is
assumed to be packed. It has a shape '[num_tokens, hidden_dim]' where
'num_tokens' is the sum of the lengths of the sequences in the batch and
'hidden_dim' is the hidden dimension. The 'last_tokens_ids' is a 1D tensor
that encodes the inclusive prefix-sums of the lengths of the sequences in
the batch.
When 'remove_input_padding' is 'False', the 'hidden_states' tensor is
assumed to be padded. It has a shape '[batch_size, max_seqlen, hidden_dim]'
where 'max_seqlen' is the length of the longest sequence in the batch and
'hidden_dim' is the hidden dimension. The 'last_token_ids' is a 1D tensor
that encodes the length of each sequence in the batch.
In both cases, that function produces a tensor of shape '[batch_size,
hidden_size]' where the row at index 'i' corresponds to the logits of the
last token from the 'i'-th sequence.
Parameters:
hidden_states : Tensor
The hidden states
last_token_ids : Tensor
The inclusive prefix-sum of the lengths or the lengths of the
sequences in the batch.
remove_input_padding : bool
Indicate if the hidden_states are packed ('True') or padded
('False').
Returns:
The tensor created by that sequence of operations.
'''
if last_token_ids is None:
return hidden_states
if remove_input_padding:
hidden_states = index_select(hidden_states, 0,
last_token_ids - 1) # [seq_len, hidden]
hidden_states = hidden_states.view(
concat([shape(last_token_ids, 0),
shape(hidden_states, 1)]))
else:
ndim = last_token_ids.ndim()
if ndim == 1:
# only calculate logits for the last token
# [batch_size, seqlen, hidden_size] -> [batch_size, hidden_size]
last_token_ids = last_token_ids.view(
concat([shape(last_token_ids, 0), 1, 1]))
last_token_ids = expand(
last_token_ids,
concat([shape(last_token_ids, 0), 1,
shape(hidden_states, 2)]))
last_token_ids = last_token_ids - 1
hidden_states = gather(
hidden_states, dim=1, indices=last_token_ids).view(
concat([shape(hidden_states, 0),
shape(hidden_states, 2)]))
elif ndim == 2: # speculative decoding needs last few token's logits
# last_token_ids is of shape [batch_size, num_last_tokens]
# So [batch_size, seqlen, hidden_size] -> [batch_size, num_last_tokens, hidden_size]
last_token_ids = last_token_ids.view(
concat([shape(last_token_ids, 0),
shape(last_token_ids, 1), 1]))
last_token_ids = expand(
last_token_ids,
concat([
shape(last_token_ids, 0),
shape(last_token_ids, 1),
shape(hidden_states, 2)
]))
hidden_states = gather(hidden_states, dim=1, indices=last_token_ids)
return hidden_states
ACT2FN = {
'relu': relu,
'tanh': tanh,
'gelu': gelu,
'gelu_new': gelu,
'gelu_fast': gelu,
'gelu_pytorch_tanh': gelu,
'openai-gelu': gelu,
'geglu': geglu,
'gegelu': gegelu,
'identity': identity,
'silu': silu,
'softplus': softplus,
'squared-relu': squared_relu,
'swiglu': swiglu,
'fast-swiglu': swiglu,
}
GATED_ACT_2_ACT = {
'swiglu': 'silu',
'fast-swiglu': 'silu',
'geglu': 'gelu',
}
[docs]
def is_gated_activation(activation):
'''
Is a given activation function gated?
Parameters:
activation : str
The name of the activation function.
Returns:
True if the function is gated, False otherwise.
'''
assert activation in ACT2FN
return activation in GATED_ACT_2_ACT
[docs]
def non_gated_version(activation):
'''
Given an activation function, get the non-gated version.
If the activation function is non-gated, it returns the same activation
function name.
For example, that function returns 'silu' for 'swiglu' and 'relu' for
'relu'.
Parameters:
activation : str
The name of the activation function.
Returns:
The name of the non-gated activation function.
'''
if is_gated_activation(activation):
return GATED_ACT_2_ACT[activation]
return activation
[docs]
def lora_plugin(
input: Tensor = None,
in_hidden_size: int = 0,
out_hidden_sizes: List[int] = [0],
host_request_types: Tensor = None,
transa: bool = False,
transb: bool = False,
host_context_lengths: Tensor = None, # for pad-free input mode
max_low_rank: int = 0,
lora_ranks: List[Tensor] = None,
lora_weights_pointers: List[Tensor] = None,
weight_index: int = 0,
):
'''
Parameters:
input : Tensor (On GPU)
The input tensor. Its shape is [batch_size, seq_len, dim] or [num_tokens, dim] for remove_input_padding
in_hidden_size/out_hidden_size : int
the lora computation workflow is
[M, in_hidden_size] -> [M, low_rank] -> [M, out_hidden_size]
host_request_types : Tensor = None
The tensor on the host that indicates if a request is in context or
generation phase. Its shape is [batch_size]. See Inflight Batching
in docs/source/advanced/gpt-attention.md,
transa : bool
Is the first input transposed? Set to 'True' if you want the first
input to be transposed, 'False' otherwise.
transb : bool
Is the second input transposed? Set to 'True' if you want the
second input to be transposed, 'False' otherwise.
host_context_lengths: cpu Tensor = None
A host tensor that contains the lengths of the different inputs,
max_low_rank : int
Maximum low_rank, used to determine the workspace size.
lora_ranks : cpu Tensor with shape [batch_size]
The low_rank of each request
lora_weights_pointers : cpu int64 Tensor with shape [batch_size, 2]
The weights pointers of each request. Consist of in_pointer and out_pointer.
weight_index : int
The index of weight if the weight pointer pointing to multiple weights.
Return:
The tensor produced by that layer.
'''
assert host_context_lengths is not None or not default_net(
).plugin_config.remove_input_padding
trt.get_plugin_registry().plugin_creator_list
in_hidden_size_field = trt.PluginField(
"in_hidden_size", np.array(in_hidden_size, dtype=np.int32),
trt.PluginFieldType.INT32)
out_hidden_size_field_list = [
trt.PluginField(f"out_hidden_size_{i}", np.array(o, dtype=np.int32),
trt.PluginFieldType.INT32)
for i, o in enumerate(out_hidden_sizes)
]
transa = 1 if transa else 0
transa = trt.PluginField("transa", np.array(transa, dtype=np.int32),
trt.PluginFieldType.INT32)
transb = 1 if transb else 0
transb = trt.PluginField("transb", np.array(transb, dtype=np.int32),
trt.PluginFieldType.INT32)
plg_creator = trt.get_plugin_registry().get_plugin_creator(
'Lora', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert plg_creator is not None
p_dtype = default_net().plugin_config.lora_plugin
pf_type = trt.PluginField(
"type_id", np.array([int(str_dtype_to_trt(p_dtype))], np.int32),
trt.PluginFieldType.INT32)
remove_input_padding = trt.PluginField(
"remove_input_padding",
np.array(np.int8(default_net().plugin_config.remove_input_padding),
dtype=np.int8), trt.PluginFieldType.INT8)
max_low_rank_field = trt.PluginField("max_low_rank",
np.array(max_low_rank, dtype=np.int32),
trt.PluginFieldType.INT32)
weight_index_field = trt.PluginField("weight_index",
np.array(weight_index, dtype=np.int32),
trt.PluginFieldType.INT32)
num_lora_modules = len(out_hidden_sizes)
num_lora_modules_field = trt.PluginField(
"num_lora_modules", np.array(num_lora_modules, dtype=np.int32),
trt.PluginFieldType.INT32)
pfc = trt.PluginFieldCollection([
in_hidden_size_field, transa, transb, num_lora_modules_field, pf_type,
remove_input_padding, max_low_rank_field, weight_index_field
] + out_hidden_size_field_list)
lora_plug = plg_creator.create_plugin("lora", pfc)
plug_inputs = [input.cast(p_dtype), host_request_types
] + lora_ranks + lora_weights_pointers
if default_net().plugin_config.remove_input_padding:
plug_inputs += [host_context_lengths]
plug_inputs = [i.trt_tensor for i in plug_inputs]
layer = default_trtnet().add_plugin_v2(plug_inputs, lora_plug)
if num_lora_modules == 1:
return _create_tensor(layer.get_output(0), layer).cast(input.dtype)
else:
return [
_create_tensor(layer.get_output(i), layer).cast(input.dtype)
for i in range(num_lora_modules)
]
[docs]
def mamba_conv1d(input: Tensor,
conv_state_or_ptr: Tensor,
conv_weight: Tensor,
conv_bias: Tensor,
host_request_types: Tensor,
last_token_ids: Tensor,
dim: int,
dconv: int,
dtype: str,
pre_stride: int = 0,
post_stride: int = 0,
host_context_lengths: Optional[Tensor] = None,
slot_mapping: Optional[Tensor] = None,
apply_silu: bool = True):
'''
Parameters:
input : Tensor (On GPU)
The input tensor. Its shape is [batch_size, seq_len, dim] or [num_tokens, dim] for remove_input_padding
conv_state_or_ptr : Tensor (On GPU or CPU)
The conv state tensor. Its shape is [batch_size, dconv - 1, dim]
Or the CPU tensor of shape [1] for the pointer of paged states.
conv_weight : Tensor (On GPU)
The weight tensor. Its shape is [1, dconv, dim]
conv_bias : Tensor (On GPU)
The bias tensor. Its shape is [dim]
host_request_types : Tensor (On CPU)
The tensor on the host that indicates if a request is in context or
generation phase. Its shape is [batch_size]. See Inflight Batching
in docs/source/advanced/gpt-attention.md,
last_token_ids : Tensor (On GPU)
The inclusive prefix-sum of the lengths or the lengths of the
sequences in the batch.
dim : int
The hidden dimension of conv1d
dconv : int
The window size of conv1d
dtype: str
data type
pre_stride : int = 0
The (pre) stride size of the input tensor.
The valid values of the input tensor are input[..., pre_stride: dim-post_stride]
post_stride : int = 0
The (post) stride size of the input tensor.
The valid values of the input tensor are input[..., pre_stride: dim-post_stride]
host_context_lengths: Tensor (On CPU) (Optional)
A host tensor that contains the lengths of the different inputs,
slot_mapping: Tensor (On GPU) (Optional)
Real page index in state. Its shape is [dim], used for paged state, each page shape is [dconv, dim]
apply_silu: bool
Is there a SiLU operation after the conv1d? When True apply
SiLU activation function after the conv1d.
'''
assert host_request_types is not None
mamba_conv1d_plg_creator = trt.get_plugin_registry().get_plugin_creator(
'MambaConv1d', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert mamba_conv1d_plg_creator is not None
dim = trt.PluginField("dim", np.array(dim, dtype=np.int32),
trt.PluginFieldType.INT32)
dconv = trt.PluginField("dconv", np.array(dconv, dtype=np.int32),
trt.PluginFieldType.INT32)
pre_stride = trt.PluginField("pre_stride",
np.array(pre_stride, dtype=np.int32),
trt.PluginFieldType.INT32)
post_stride = trt.PluginField("post_stride",
np.array(post_stride, dtype=np.int32),
trt.PluginFieldType.INT32)
pf_type = trt.PluginField(
"type_id", np.array([int(str_dtype_to_trt(dtype))], np.int32),
trt.PluginFieldType.INT32)
remove_input_padding = trt.PluginField(
"remove_input_padding",
np.array(np.int8(default_net().plugin_config.remove_input_padding),
dtype=np.int8), trt.PluginFieldType.INT8)
paged_state = trt.PluginField(
"paged_state",
np.array(np.int8(default_net().plugin_config.paged_state),
dtype=np.int8), trt.PluginFieldType.INT8)
apply_silu = trt.PluginField("apply_silu",
np.array(np.int8(apply_silu), dtype=np.int8),
trt.PluginFieldType.INT8)
pfc = trt.PluginFieldCollection([
dim, dconv, pre_stride, post_stride, pf_type, remove_input_padding,
paged_state, apply_silu
])
mamba_conv1d_plug = mamba_conv1d_plg_creator.create_plugin(
"mamba_conv1d", pfc)
plug_inputs = [
input, conv_state_or_ptr, conv_weight, conv_bias, host_request_types,
last_token_ids
]
if default_net().plugin_config.remove_input_padding:
plug_inputs += [host_context_lengths]
if default_net().plugin_config.paged_state:
plug_inputs += [slot_mapping]
plug_inputs = [i.trt_tensor for i in plug_inputs]
layer = default_trtnet().add_plugin_v2(plug_inputs, mamba_conv1d_plug)
_add_plugin_info(layer, mamba_conv1d_plg_creator, "mamba_conv1d", pfc)
output = _create_tensor(layer.get_output(0), layer)
if default_net().plugin_config.paged_state:
return output, None
else:
present_state = _create_tensor(layer.get_output(1), layer)
return output, present_state
[docs]
def selective_scan(input: Tensor,
state_or_ptr: Tensor,
delta: Tensor,
delta_bias: Tensor,
A: Tensor,
BC: Tensor,
D: Tensor,
host_request_types: Tensor,
last_token_ids: Tensor,
dim: int,
dstate: int,
dt_rank: int,
delta_softplus: bool,
dtype: str,
z: Optional[Tensor] = None,
host_context_lengths: Optional[Tensor] = None,
slot_mapping: Optional[Tensor] = None,
nheads: int = 1,
ngroups: int = 1,
chunk_size: int = 256,
mamba_version: str = 'Mamba1'):
'''
Parameters:
input : Tensor (On GPU)
The input tensor. Its shape is [batch_size, seq_len, dim]
state_or_ptr : Tensor (On GPU or CPU)
The ssm state tensor. Its shape is [batch_size, dstate, dim]
Or the CPU tensor of shape [1] for the pointer of paged states.
delta : Tensor (On GPU)
The delta tensor.
mamba: Its shape is [batch_size, seq_len, dim] or [num_tokens, dim] for remove_input_padding
mamba2: Its shape is [batch_size, seq_len, nheads] or [num_tokens, nheads] for remove_input_padding
delta_bias : Tensor (On GPU)
The delta bias tensor.
mamba: Its shape is [dim]
mamba2: Its shape is [nheads]
A : Tensor (On GPU)
A matrix.
mamba: Its shape is [dstate, dim]
mamba2: Its shape is [nheads]
BC : Tensor (On GPU)
B and C matrix.
mamba: Its shape is [batch_size, seq_len, dstate * 2] or [num_tokens, dstate * 2] for remove_input_padding
mamba2: Its shape is [batch_size, seq_len, ngroups * dstate * 2] or [num_tokens, ngroups * dstate * 2] for remove_input_padding
D : Tensor (On GPU)
D matrix.
mamba: Its shape is [dim]
mamba2: Its shape is [nheads]
host_request_types : Tensor (On CPU)
The tensor on the host that indicates if a request is in context or
generation phase. Its shape is [batch_size]. See Inflight Batching
in docs/source/advanced/gpt-attention.md
last_token_ids : Tensor (On GPU)
The inclusive prefix-sum of the lengths or the lengths of the
sequences in the batch.
dim : int
The inner dimension of SSM block
dstate : int
The state dimension of SSM block
dt_rank: int
The rank dimension of dt_proj
delta_softplus : bool
Do we apply softplus to the delta.
dtype: str
data type
z : Tensor (On GPU) (Optional)
The z tensor. Its shape is [batch_size, seq_len, dim] or [num_tokens, dim] for remove_input_padding
host_context_lengths: Tensor (On CPU) (Optional)
A host tensor that contains the lengths of the different inputs,
slot_mapping: Tensor (On GPU) (Optional)
Real page index in state. Its shape is [dim], used for paged state, each page shape is [dstate, dim]
nheads: int (Optional)
The number of heads.
ngroups: int (Optional)
The number of groups.
chunk_size: int (Optional)
The chunk_size is used for the chunk_scan kernel.
mamba_version: int (Optional)
Mamba version, support Mamba1 as default.
'''
assert host_request_types is not None
selective_scan_plg_creator = trt.get_plugin_registry().get_plugin_creator(
'SelectiveScan', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert selective_scan_plg_creator is not None
dim = trt.PluginField("dim", np.array(dim, dtype=np.int32),
trt.PluginFieldType.INT32)
dstate = trt.PluginField("dstate", np.array(dstate, dtype=np.int32),
trt.PluginFieldType.INT32)
dt_rank = trt.PluginField("dt_rank", np.array(dt_rank, dtype=np.int32),
trt.PluginFieldType.INT32)
nheads = trt.PluginField("nheads", np.array(nheads, dtype=np.int32),
trt.PluginFieldType.INT32)
ngroups = trt.PluginField("ngroups", np.array(ngroups, dtype=np.int32),
trt.PluginFieldType.INT32)
chunk_size = trt.PluginField("chunk_size",
np.array(chunk_size, dtype=np.int32),
trt.PluginFieldType.INT32)
delta_softplus = trt.PluginField(
"delta_softplus", np.array(np.int8(delta_softplus), dtype=np.int8),
trt.PluginFieldType.INT8)
pf_type = trt.PluginField(
"type_id", np.array([int(str_dtype_to_trt(dtype))], np.int32),
trt.PluginFieldType.INT32)
remove_input_padding = trt.PluginField(
"remove_input_padding",
np.array(np.int8(default_net().plugin_config.remove_input_padding),
dtype=np.int8), trt.PluginFieldType.INT8)
paged_state = trt.PluginField(
"paged_state",
np.array(np.int8(default_net().plugin_config.paged_state),
dtype=np.int8), trt.PluginFieldType.INT8)
if z is None:
z_enabled = trt.PluginField("z_enabled", np.array(0, dtype=np.int8),
trt.PluginFieldType.INT8)
else:
z_enabled = trt.PluginField("z_enabled", np.array(1, dtype=np.int8),
trt.PluginFieldType.INT8)
is_mamba2 = trt.PluginField(
"is_mamba2",
np.array(1 if mamba_version == 'Mamba2' else 0, dtype=np.int8),
trt.PluginFieldType.INT8)
pfc = trt.PluginFieldCollection([
dim, dstate, dt_rank, nheads, ngroups, chunk_size, delta_softplus,
pf_type, remove_input_padding, paged_state, z_enabled, is_mamba2
])
selective_scan_plug = selective_scan_plg_creator.create_plugin(
"selective_scan", pfc)
plug_inputs = [
input, state_or_ptr, delta, delta_bias, A, BC, D, host_request_types,
last_token_ids
]
if default_net().plugin_config.remove_input_padding:
plug_inputs += [host_context_lengths]
if default_net().plugin_config.paged_state:
plug_inputs += [slot_mapping]
if z is not None:
plug_inputs += [z]
plug_inputs = [i.trt_tensor for i in plug_inputs]
layer = default_trtnet().add_plugin_v2(plug_inputs, selective_scan_plug)
_add_plugin_info(layer, selective_scan_plg_creator, "selective_scan", pfc)
output = _create_tensor(layer.get_output(0), layer)
if default_net().plugin_config.paged_state:
return output, None
else:
present_state = _create_tensor(layer.get_output(1), layer)
return output, present_state
[docs]
def rg_lru(input: Tensor,
A: Tensor,
state_or_ptr: Tensor,
host_request_types: Tensor,
last_token_ids: Tensor,
dim: int,
dtype: str,
block_size: int = 0,
y: Optional[Tensor] = None,
y_bias: Optional[Tensor] = None,
gate: Optional[Tensor] = None,
gate_bias: Optional[Tensor] = None,
gate_x: Optional[Tensor] = None,
gate_x_bias: Optional[Tensor] = None,
gate_a: Optional[Tensor] = None,
gate_a_bias: Optional[Tensor] = None,
slot_mapping: Optional[Tensor] = None):
'''
Parameters:
input : Tensor (On GPU)
The input tensor. Its shape is [batch_size, seq_len, dim]
A : Tensor (On GPU)
A matrix. Its shape is [dim]
state_or_ptr : Tensor (On GPU or CPU)
The lru state tensor. Its shape is [batch_size, dstate, dim]
Or the CPU tensor of shape [1] for the pointer of paged states.
host_request_types : Tensor (On CPU)
The tensor on the host that indicates if a request is in context or
generation phase. Its shape is [batch_size]. See Inflight Batching
in docs/source/advanced/gpt-attention.md,
last_token_ids : Tensor (On GPU)
The inclusive prefix-sum of the lengths or the lengths of the
sequences in the batch.
dim : int
The inner dimension of RG_LRU block
block_size : int
The block size of the block diagonal linear layer. It is used to
support the cases that enable fused gate.
dtype: str
data type
y : Tensor (On GPU) (Optional)
The y tensor. Its shape is [batch_size, seq_len, dim]
y_bias : Tensor (On GPU) (Optional)
The y_bias tensor. Its shape is [dim]. If y_bias is not None, we
will fuse GELU(y + y_bias) in this function.
gate : Tensor (On GPU) (Optional)
The gate tensor. Its shape is [batch_size, seq_len, 2 * dim].
If gate is not None, we will fuse the gate_x and gate_a, otherwise
use those two tensors.
gate_bias : Tensor (On GPU) (Optional)
The gate_bias tensor. Its shape is [2 * block_num, dim // block_num].
If gate_bias is not None, we will fuse the bias add in this function.
gate_x : Tensor (On GPU) (Optional)
The gate_x tensor. Its shape is [batch_size, seq_len, dim]
gate_x_bias : Tensor (On GPU) (Optional)
The gate_x_bias tensor. Its shape is [block_num, dim // block_num].
If gate_x_bias is not None, we will fuse the bias add in this function.
gate_a : Tensor (On GPU) (Optional)
The gate_a tensor. Its shape is [batch_size, seq_len, dim]
gate_a_bias : Tensor (On GPU) (Optional)
The gate_a_bias tensor. Its shape is [block_num, dim // block_num].
If gate_a_bias is not None, we will fuse the bias add in this function.
slot_mapping: Tensor (On GPU) (Optional)
Real page index in state. Its shape is [dim], used for paged state, each page shape is [dstate, dim]
'''
assert host_request_types is not None
lru_plg_creator = trt.get_plugin_registry().get_plugin_creator(
'LRU', '1', TRT_LLM_PLUGIN_NAMESPACE)
assert lru_plg_creator is not None
assert (gate_x_bias is None) == (gate_a_bias is None)
enable_fuse_gate = gate is not None
has_gate_bias = (gate_bias is not None) or (gate_x_bias is not None)
if enable_fuse_gate:
assert gate is not None
assert block_size > 0
if has_gate_bias:
assert gate_bias is not None
else:
assert gate_x is not None and gate_a is not None
if has_gate_bias:
assert gate_x_bias is not None and gate_a_bias is not None
dim = trt.PluginField("dim", np.array(dim, dtype=np.int32),
trt.PluginFieldType.INT32)
block_size = trt.PluginField("block_size",
np.array(block_size, dtype=np.int32),
trt.PluginFieldType.INT32)
pf_type = trt.PluginField(
"type_id", np.array([int(str_dtype_to_trt(dtype))], np.int32),
trt.PluginFieldType.INT32)
remove_input_padding = trt.PluginField(
"remove_input_padding",
np.array(np.int8(default_net().plugin_config.remove_input_padding),
dtype=np.int8), trt.PluginFieldType.INT8)
paged_state = trt.PluginField(
"paged_state",
np.array(np.int8(default_net().plugin_config.paged_state),
dtype=np.int8), trt.PluginFieldType.INT8)
if y is None:
y_enabled = trt.PluginField("y_enabled", np.array(0, dtype=np.int8),
trt.PluginFieldType.INT8)
else:
y_enabled = trt.PluginField("y_enabled", np.array(1, dtype=np.int8),
trt.PluginFieldType.INT8)
if y_bias is None:
y_bias_enabled = trt.PluginField("y_bias_enabled",
np.array(0, dtype=np.int8),
trt.PluginFieldType.INT8)
else:
y_bias_enabled = trt.PluginField("y_bias_enabled",
np.array(1, dtype=np.int8),
trt.PluginFieldType.INT8)
if enable_fuse_gate:
fuse_gate_enabled = trt.PluginField("fuse_gate_enabled",
np.array(1, dtype=np.int8),
trt.PluginFieldType.INT8)
else:
fuse_gate_enabled = trt.PluginField("fuse_gate_enabled",
np.array(0, dtype=np.int8),
trt.PluginFieldType.INT8)
if has_gate_bias:
gate_bias_enabled = trt.PluginField("gate_bias_enabled",
np.array(1, dtype=np.int8),
trt.PluginFieldType.INT8)
else:
gate_bias_enabled = trt.PluginField("gate_bias_enabled",
np.array(0, dtype=np.int8),
trt.PluginFieldType.INT8)
pfc = trt.PluginFieldCollection([
dim, block_size, pf_type, remove_input_padding, paged_state, y_enabled,
y_bias_enabled, fuse_gate_enabled, gate_bias_enabled
])
lru_plug = lru_plg_creator.create_plugin("rg_lru", pfc)
plug_inputs = [
input,
A,
state_or_ptr,
host_request_types,
last_token_ids,
]
if default_net().plugin_config.paged_state:
plug_inputs += [slot_mapping]
if y is not None:
plug_inputs += [y]
if y_bias is not None:
plug_inputs += [y_bias]
if enable_fuse_gate:
plug_inputs += [gate]
if has_gate_bias:
plug_inputs += [gate_bias]
else:
plug_inputs += [gate_x, gate_a]
if has_gate_bias:
plug_inputs += [gate_x_bias, gate_a_bias]
plug_inputs = [i.trt_tensor for i in plug_inputs]
layer = default_trtnet().add_plugin_v2(plug_inputs, lru_plug)
_add_plugin_info(layer, lru_plg_creator, "rg_lru", pfc)
output = _create_tensor(layer.get_output(0), layer)
if default_net().plugin_config.paged_state:
return output, None
else:
present_state = _create_tensor(layer.get_output(1), layer)
return output, present_state
[docs]
def topk(input: Tensor,
k: int,
dim: int,
largest: bool = True) -> Tuple[Tensor, Tensor]:
'''
Add an topk operation.
As explained in the ONNX documentation,
https://github.com/onnx/onnx/blob/main/docs/Operators.md#topk
NOTE: One distinction from the ONNX topk op, the output is always sorted
with TensorRT layer.
Retrieve the top-K largest elements along a specified axis.
Given an input tensor of shape [a_1, a_2, ..., a_n, r]
and integer argument k, return two outputs:
Value tensor of shape [a_1, a_2, ..., a_{axis-1}, k, a_{axis+1}, ... a_n] which contains the values of the top k elements along the specified axis
Index tensor of shape [a_1, a_2, ..., a_{axis-1}, k, a_{axis+1}, ... a_n] which contains the indices of the top k elements (original indices from the input tensor).
Parameters:
input : Tensor
The input tensor.
k : int
A single positive value corresponding to the number of top elements to retrieve
dim: int
The dimension in which to compute the topk indices.
largest: bool
Controls whether to return largest or smallest elements
Returns:
The tensors (values, indices) produced by this topk operation.
'''
dim = dim_resolve_negative(dim, input.ndim())
axes = dim_to_trt_axes(dim)
layer = default_trtnet().add_topk(
input.trt_tensor,
trt.TopKOperation.MAX if largest else trt.TopKOperation.MIN,
k=k,
axes=axes)
values = layer.get_output(0)
indices = layer.get_output(1)
return _create_tensor(values, layer), _create_tensor(indices, layer)
[docs]
def scatter_nd(input: Tensor, mask: Tensor, source: Tensor) -> Tensor:
'''
Scatter_nd is a tensor operation that writes or updates values in a tensor based on indices.
Parameters:
input: Tensor
The input tensor to be updated
mask: Tensor
A tensor of indices specifying the locations in data to be updated.
source: Tensor
A tensor of values to be written or scattered into data.
Returns:
New tensor with the same shape as the input tensor data,
where the values from the source tensor are scattered or written into the output tensor
at the locations specified by the mask tensor.
'''
scatter_layer = default_trtnet().add_scatter(input.trt_tensor,
mask.trt_tensor,
source.trt_tensor,
mode=trt.ScatterMode.ND)
return _create_tensor(scatter_layer.get_output(0), scatter_layer)
[docs]
def low_latency_gemm(input: Tensor,
mat2: Tensor,
alpha: Optional[np.ndarray] = None,
strict_dtype: Optional[trt.DataType] = None) -> Tensor:
if not default_net().plugin_config.low_latency_gemm_plugin:
raise RuntimeError("Low Latency GEMM is only support with plugin")
elif default_net().plugin_config.low_latency_gemm_plugin != "fp8":
raise RuntimeError("Low Latency GEMM plugin only support fp8")
else:
plg_creator = trt.get_plugin_registry().get_plugin_creator(
"LowLatencyGemm", "1", TRT_LLM_PLUGIN_NAMESPACE)
assert plg_creator is not None
if ((input.dtype != trt.fp8) or ((mat2.dtype) != trt.fp8)):
raise TypeError("Low Latency GEMM only support fp8 input")
if (alpha):
assert (isinstance(alpha, np.ndarray) and alpha.dtype == np.float32
and alpha.size
== 1), "`alpha` must be passed as a float32 ndarray"
alpha = alpha if alpha else np.array(1.0, dtype=np.float32)
alpha = trt.PluginField("alpha", alpha.flatten(),
trt.PluginFieldType.FLOAT32)
if strict_dtype is not None:
assert isinstance(strict_dtype, trt.DataType)
p_dtype = strict_dtype
if (p_dtype not in [trt.float32, trt.float16, trt.bfloat16]):
raise ValueError(
"strict_dtype must be float32, float16 or bfloat16 in low latency gemm plugin"
)
else:
raise RuntimeError(
"need to use strict dtype in low latency gemm plugin fp8")
pf_type = trt.PluginField("type_id", np.array([int(p_dtype)], np.int32),
trt.PluginFieldType.INT32)
pfc = trt.PluginFieldCollection([alpha, pf_type])
low_latency_gemm_plug = plg_creator.create_plugin(
"low_latency_gemm", pfc)
plug_inputs = [input.trt_tensor, mat2.trt_tensor]
layer = default_trtnet().add_plugin_v2(plug_inputs,
low_latency_gemm_plug)
_add_plugin_info(layer, plg_creator, "low_latency_gemm", pfc)
return _create_tensor(layer.get_output(0), layer)