File size: 13,474 Bytes
5000658 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 |
# 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
from typing import Optional
from .._common import default_net
from ..functional import (ACT2FN, Tensor, concat, conv2d, gather, mamba_conv1d,
permute, selective_scan, shape, split, view)
from ..module import Module
from ..parameter import Parameter
from .linear import Linear
from .normalization import RmsNorm
class MambaConv1d(Module):
def __init__(self,
d_inner,
d_conv=4,
pre_stride=0,
post_stride=0,
dtype=None,
apply_silu=True):
super().__init__()
self.d_inner = d_inner
self.d_conv = d_conv
self.pre_stride = pre_stride
self.post_stride = post_stride
self.dtype = dtype
self.weight = Parameter(shape=(self.d_inner, 1, self.d_conv, 1),
dtype=dtype)
self.bias = Parameter(shape=(self.d_inner, ), dtype=dtype)
self.apply_silu = apply_silu
def forward(self,
x: Tensor,
conv_state: Tensor,
host_request_types: Tensor,
last_token_ids: Tensor,
host_context_lengths: Optional[Tensor] = None,
slot_mapping: Optional[Tensor] = None,
conv_indices: Optional[Tensor] = None):
'''
Parameters:
x: [B, L, D] or [T, D]
conv_state: [B, W, D] or [1] of type int64 for paged state
host_request_types: [B]
last_token_ids: [B]
host_context_lengths: [B]
slot_mapping: [B]
conv_indices: [B]
'''
if default_net().plugin_config.mamba_conv1d_plugin:
transposed_weight = permute(
view(self.weight.value, shape=[self.d_inner, 1, self.d_conv]),
(1, 2, 0))
x_conv, conv_state = mamba_conv1d(
x, conv_state, transposed_weight, self.bias.value,
host_request_types, last_token_ids, self.d_inner, self.d_conv,
self.dtype, self.pre_stride, self.post_stride,
host_context_lengths, slot_mapping, self.apply_silu)
else:
assert not default_net().plugin_config.paged_state
assert len(
x.shape
) == 3, "remove_input_padding is not supported by OOTB for Mamba."
if self.pre_stride > 0:
_, x = split(x,
[self.pre_stride, self.d_inner + self.post_stride],
dim=-1)
if self.post_stride > 0:
x, _ = split(x, [self.d_inner, self.post_stride], dim=-1)
x = x.permute([0, 2, 1])
# In context phase, conv_state is a zero tensor, and it is used for padding
# In generation phase, conv_state is a tensor of the past x
x_pad = concat([conv_state, x], dim=2)
# Update conv_state
conv_state = gather(x_pad, 2, conv_indices)
# Convolution
x_pad = x_pad.view(
concat([shape(x_pad, 0),
shape(x_pad, 1),
shape(x_pad, 2), 1]))
x_conv = conv2d(x_pad,
self.weight.value,
self.bias.value,
groups=self.d_inner)
if self.apply_silu:
x_conv = ACT2FN['silu'](x_conv)
x_conv = x_conv.view(
concat([shape(x_conv, 0),
shape(x_conv, 1),
shape(x_conv, 2)]))
# Get dt, B and C
x_conv = x_conv.permute([0, 2, 1])
return x_conv, conv_state
class Mamba(Module):
def __init__(self,
d_model,
d_inner,
d_state=16,
d_conv=4,
dt_rank="auto",
bias=False,
dtype=None):
super().__init__()
self.d_model = d_model
self.d_state = d_state
self.d_conv = d_conv
self.d_inner = d_inner
self.dt_rank = math.ceil(self.d_model /
16) if dt_rank == "auto" else dt_rank
self.dtype = dtype
self.A = Parameter(shape=(self.d_state, self.d_inner), dtype="float32")
self.D = Parameter(shape=(self.d_inner, ), dtype="float32")
self.dt_bias = Parameter(shape=(self.d_inner, ), dtype="float32")
self.in_proj_x = Linear(self.d_model,
self.d_inner,
bias=bias,
dtype=dtype,
gather_output=False)
self.in_proj_z = Linear(self.d_model,
self.d_inner,
bias=bias,
dtype=dtype,
gather_output=False)
self.conv1d = MambaConv1d(self.d_inner, self.d_conv, dtype=self.dtype)
self.x_proj = Linear(self.d_inner,
self.dt_rank + self.d_state * 2,
bias=False,
dtype=dtype,
gather_output=False)
self.dt_proj = Linear(self.dt_rank,
self.d_inner,
bias=False,
dtype=dtype,
gather_output=False,
pad_lda=self.d_state * 2)
self.out_proj = Linear(self.d_inner,
self.d_model,
bias=bias,
dtype=dtype,
gather_output=False)
def forward(self,
hidden_states: Tensor,
conv_state: Tensor,
ssm_state: Tensor,
host_request_types: Tensor,
last_token_ids: Tensor,
host_context_lengths: Optional[Tensor] = None,
slot_mapping: Optional[Tensor] = None,
conv_indices: Optional[Tensor] = None):
'''
Parameters:
hidden_states: [B, L, D] or [T, D]
conv_state: [B, W, D] or [1] of type int64 for paged state
ssm_state: [B, N, D] or [1] of type int64 for paged state
host_request_types: [B]
last_token_ids: [B]
host_context_lengths: [B]
slot_mapping: [B]
conv_indices: [B]
'''
# in_proj
x = self.in_proj_x(hidden_states)
z = self.in_proj_z(hidden_states)
x_conv, conv_state = self.conv1d(x, conv_state, host_request_types,
last_token_ids, host_context_lengths,
slot_mapping, conv_indices)
# Get dt, B and C
x_dbl = self.x_proj(x_conv)
if default_net().plugin_config.gemm_plugin:
dt = self.dt_proj(x_dbl)
else:
dt, _ = split(x_dbl, [self.dt_rank, self.d_state * 2], dim=-1)
dt = self.dt_proj(dt)
# selective scan
y, ssm_state = selective_scan(x_conv,
ssm_state,
dt,
self.dt_bias.value,
self.A.value,
x_dbl,
self.D.value,
host_request_types,
last_token_ids,
self.d_inner,
self.d_state,
self.dt_rank,
delta_softplus=True,
dtype=self.dtype,
z=z,
host_context_lengths=host_context_lengths,
slot_mapping=slot_mapping)
# out_proj
out = self.out_proj(y)
return out, conv_state, ssm_state
class Mamba2(Module):
def __init__(self,
d_model,
d_inner,
d_state=16,
d_conv=4,
headdim=64,
ngroups=1,
chunk_size=256,
bias=False,
rmsnorm=True,
dtype=None):
super().__init__()
self.d_model = d_model
self.d_state = d_state
self.d_conv = d_conv
self.d_inner = d_inner
self.headdim = headdim
self.ngroups = ngroups
self.chunk_size = chunk_size
self.rmsnorm = rmsnorm
self.dtype = dtype
assert self.d_inner % self.headdim == 0
self.nheads = self.d_inner // self.headdim
self.A = Parameter(shape=(self.nheads, ), dtype="float32")
self.D = Parameter(shape=(self.nheads, ), dtype="float32")
self.dt_bias = Parameter(shape=(self.nheads, ), dtype="float32")
d_in_proj = 2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads
self.in_proj = Linear(self.d_model,
d_in_proj,
bias=bias,
dtype=dtype,
gather_output=False)
self.conv_dim = self.d_inner + 2 * self.ngroups * self.d_state
self.conv1d = MambaConv1d(self.conv_dim,
self.d_conv,
pre_stride=self.d_inner,
post_stride=self.nheads,
dtype=self.dtype)
if rmsnorm:
self.norm = RmsNorm(normalized_shape=self.d_inner,
num_groups=ngroups,
eps=1e-5,
dtype=dtype)
self.out_proj = Linear(self.d_inner,
self.d_model,
bias=bias,
dtype=dtype,
gather_output=False)
def forward(self,
hidden_states: Tensor,
conv_state: Tensor,
ssm_state: Tensor,
host_request_types: Tensor,
last_token_ids: Tensor,
host_context_lengths: Optional[Tensor] = None,
slot_mapping: Optional[Tensor] = None,
conv_indices: Optional[Tensor] = None):
'''
Parameters:
hidden_states: [B, L, D] or [T, D]
conv_state: [B, W, D_conv] or [1] of type int64 for paged state
ssm_state: [B, H, N, D] or [1] of type int64 for paged state
host_request_types: [B]
last_token_ids: [B]
host_context_lengths: [B]
slot_mapping: [B]
conv_indices: [B]
'''
# in_proj
zxbcdt = self.in_proj(hidden_states)
# conv1d
xbc_conv, conv_state = self.conv1d(zxbcdt, conv_state,
host_request_types, last_token_ids,
host_context_lengths, slot_mapping,
conv_indices)
# mamba scan
y, ssm_state = selective_scan(xbc_conv,
ssm_state,
zxbcdt,
self.dt_bias.value,
self.A.value,
xbc_conv,
self.D.value,
host_request_types,
last_token_ids,
self.d_inner,
self.d_state,
dt_rank=0,
delta_softplus=True,
dtype=self.dtype,
z=zxbcdt,
host_context_lengths=host_context_lengths,
slot_mapping=slot_mapping,
nheads=self.nheads,
ngroups=self.ngroups,
chunk_size=self.chunk_size,
mamba_version='Mamba2')
# norm
if self.rmsnorm:
y = self.norm(y)
# out_proj
out = self.out_proj(y)
return out, conv_state, ssm_state
|