Commit
•
a2f1254
1
Parent(s):
67e342a
Delete modeling_mixformer_sequential.py
Browse files- modeling_mixformer_sequential.py +0 -935
modeling_mixformer_sequential.py
DELETED
@@ -1,935 +0,0 @@
|
|
1 |
-
# Copyright (c) Microsoft Corporation.
|
2 |
-
# Licensed under the MIT license.
|
3 |
-
#
|
4 |
-
# BSD 3-Clause License
|
5 |
-
#
|
6 |
-
# Copyright (c) 2022, Tri Dao, trid@cs.stanford.edu.
|
7 |
-
# All rights reserved.
|
8 |
-
#
|
9 |
-
# Redistribution and use in source and binary forms, with or without
|
10 |
-
# modification, are permitted provided that the following conditions are met:
|
11 |
-
#
|
12 |
-
# * Redistributions of source code must retain the above copyright notice, this
|
13 |
-
# list of conditions and the following disclaimer.
|
14 |
-
#
|
15 |
-
# * Redistributions in binary form must reproduce the above copyright notice,
|
16 |
-
# this list of conditions and the following disclaimer in the documentation
|
17 |
-
# and/or other materials provided with the distribution.
|
18 |
-
#
|
19 |
-
# * Neither the name of the copyright holder nor the names of its
|
20 |
-
# contributors may be used to endorse or promote products derived from
|
21 |
-
# this software without specific prior written permission.
|
22 |
-
#
|
23 |
-
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
24 |
-
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
25 |
-
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
26 |
-
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
27 |
-
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
28 |
-
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
29 |
-
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
30 |
-
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
31 |
-
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
32 |
-
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
33 |
-
|
34 |
-
from __future__ import annotations
|
35 |
-
|
36 |
-
import math
|
37 |
-
from typing import Any, Dict, Optional, Tuple, Union
|
38 |
-
from dataclasses import dataclass, field
|
39 |
-
|
40 |
-
import torch
|
41 |
-
import torch.nn as nn
|
42 |
-
|
43 |
-
from einops import rearrange, repeat
|
44 |
-
from transformers.activations import ACT2FN
|
45 |
-
from transformers import PretrainedConfig, PreTrainedModel
|
46 |
-
from transformers.modeling_outputs import CausalLMOutputWithPast
|
47 |
-
|
48 |
-
from .configuration_mixformer_sequential import MixFormerSequentialConfig
|
49 |
-
|
50 |
-
|
51 |
-
try:
|
52 |
-
from flash_attn.bert_padding import pad_input, unpad_input
|
53 |
-
from flash_attn.layers.rotary import RotaryEmbedding as FlashRotaryEmbedding
|
54 |
-
from flash_attn.modules.mha import FlashCrossAttention, FlashSelfAttention
|
55 |
-
from flash_attn.ops.fused_dense import FusedDense
|
56 |
-
except:
|
57 |
-
pad_input, unpad_input = None, None
|
58 |
-
FlashRotaryEmbedding = None
|
59 |
-
FlashSelfAttention, FlashCrossAttention = None, None
|
60 |
-
FusedDense = None
|
61 |
-
|
62 |
-
|
63 |
-
@dataclass
|
64 |
-
class InferenceParams:
|
65 |
-
"""Inference parameters passed to model to efficiently calculate
|
66 |
-
and store context during inference.
|
67 |
-
|
68 |
-
Reference:
|
69 |
-
https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/utils/generation.py.
|
70 |
-
|
71 |
-
Args:
|
72 |
-
max_seqlen: Maximum sequence length.
|
73 |
-
max_batch_size: Maximum batch size.
|
74 |
-
seqlen_offset: Sequence length offset.
|
75 |
-
batch_size_offset: Batch size offset.
|
76 |
-
key_value_memory_dict: Key value memory dictionary.
|
77 |
-
lengths_per_sample: Lengths per sample.
|
78 |
-
|
79 |
-
"""
|
80 |
-
|
81 |
-
max_seqlen: int = field(metadata={"help": "Maximum sequence length."})
|
82 |
-
|
83 |
-
max_batch_size: int = field(metadata={"help": "Maximum batch size."})
|
84 |
-
|
85 |
-
seqlen_offset: int = field(default=0, metadata={"help": "Sequence length offset."})
|
86 |
-
|
87 |
-
batch_size_offset: int = field(default=0, metadata={"help": "Batch size offset."})
|
88 |
-
|
89 |
-
key_value_memory_dict: Dict[str, Any] = field(
|
90 |
-
default_factory=dict, metadata={"help": "Key value memory dictionary."}
|
91 |
-
)
|
92 |
-
|
93 |
-
lengths_per_sample: torch.Tensor = field(default=None, metadata={"help": "Lengths per sample."})
|
94 |
-
|
95 |
-
|
96 |
-
class Embedding(nn.Module):
|
97 |
-
"""Token embedding with dropout."""
|
98 |
-
|
99 |
-
def __init__(self, config: PretrainedConfig) -> None:
|
100 |
-
super().__init__()
|
101 |
-
|
102 |
-
self.wte = nn.Embedding(config.vocab_size, config.n_embd)
|
103 |
-
self.drop = nn.Dropout(config.embd_pdrop)
|
104 |
-
|
105 |
-
def forward(self, input_ids: torch.LongTensor) -> torch.FloatTensor:
|
106 |
-
input_shape = input_ids.size()
|
107 |
-
input_ids = input_ids.view(-1, input_shape[-1])
|
108 |
-
|
109 |
-
hidden_states = self.wte(input_ids)
|
110 |
-
hidden_states = self.drop(hidden_states)
|
111 |
-
|
112 |
-
return hidden_states
|
113 |
-
|
114 |
-
|
115 |
-
def _apply_rotary_emb(
|
116 |
-
x: torch.FloatTensor,
|
117 |
-
cos: torch.FloatTensor,
|
118 |
-
sin: torch.FloatTensor,
|
119 |
-
) -> torch.FloatTensor:
|
120 |
-
_, seqlen, _, head_dim = x.shape
|
121 |
-
rotary_seqlen, rotary_dim = cos.shape
|
122 |
-
rotary_dim *= 2
|
123 |
-
|
124 |
-
assert rotary_dim <= head_dim
|
125 |
-
assert seqlen <= rotary_seqlen
|
126 |
-
assert cos.shape == sin.shape == (rotary_seqlen, rotary_dim // 2)
|
127 |
-
|
128 |
-
x_rot = x[:, :, :, :rotary_dim]
|
129 |
-
x_pass = x[:, :, :, rotary_dim:]
|
130 |
-
|
131 |
-
x1, x2 = x_rot.chunk(2, dim=-1)
|
132 |
-
c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
|
133 |
-
x1, x2, c, s = [t.to(dtype=torch.float32) for t in [x1, x2, c, s]]
|
134 |
-
|
135 |
-
x_rot = torch.cat([x1 * c - x2 * s, x1 * s + x2 * c], axis=-1).to(x.dtype)
|
136 |
-
|
137 |
-
return torch.cat([x_rot, x_pass], axis=-1)
|
138 |
-
|
139 |
-
|
140 |
-
def _apply_rotary_emb_kv(
|
141 |
-
kv: torch.FloatTensor,
|
142 |
-
cos: torch.FloatTensor,
|
143 |
-
sin: torch.FloatTensor,
|
144 |
-
cos_k: Optional[torch.FloatTensor] = None,
|
145 |
-
sin_k: Optional[torch.FloatTensor] = None,
|
146 |
-
) -> torch.FloatTensor:
|
147 |
-
_, seqlen, two, _, head_dim = kv.shape
|
148 |
-
assert two == 2
|
149 |
-
|
150 |
-
rotary_seqlen, rotary_dim = cos.shape
|
151 |
-
rotary_dim *= 2
|
152 |
-
assert rotary_dim <= head_dim
|
153 |
-
assert seqlen <= rotary_seqlen
|
154 |
-
assert cos.shape == sin.shape == (rotary_seqlen, rotary_dim // 2)
|
155 |
-
|
156 |
-
k_rot = kv[:, :, 0, :, :rotary_dim]
|
157 |
-
k_pass = kv[:, :, 0, :, rotary_dim:]
|
158 |
-
|
159 |
-
k1, k2 = k_rot.chunk(2, dim=-1)
|
160 |
-
c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
|
161 |
-
k1, k2, c, s = [t.to(dtype=torch.float32) for t in [k1, k2, c, s]]
|
162 |
-
|
163 |
-
k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(kv.dtype)
|
164 |
-
|
165 |
-
return torch.cat(
|
166 |
-
[
|
167 |
-
torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
|
168 |
-
kv[:, :, 1:2, :, :],
|
169 |
-
],
|
170 |
-
axis=2,
|
171 |
-
)
|
172 |
-
|
173 |
-
|
174 |
-
def _apply_rotary_emb_qkv(
|
175 |
-
qkv: torch.FloatTensor,
|
176 |
-
cos: torch.FloatTensor,
|
177 |
-
sin: torch.FloatTensor,
|
178 |
-
cos_k: Optional[torch.FloatTensor] = None,
|
179 |
-
sin_k: Optional[torch.FloatTensor] = None,
|
180 |
-
) -> torch.FloatTensor:
|
181 |
-
_, seqlen, three, _, head_dim = qkv.shape
|
182 |
-
assert three == 3
|
183 |
-
|
184 |
-
rotary_seqlen, rotary_dim = cos.shape
|
185 |
-
rotary_dim *= 2
|
186 |
-
assert rotary_dim <= head_dim
|
187 |
-
assert seqlen <= rotary_seqlen
|
188 |
-
assert cos.shape == sin.shape == (rotary_seqlen, rotary_dim // 2)
|
189 |
-
|
190 |
-
q_rot = qkv[:, :, 0, :, :rotary_dim]
|
191 |
-
q_pass = qkv[:, :, 0, :, rotary_dim:]
|
192 |
-
|
193 |
-
k_rot = qkv[:, :, 1, :, :rotary_dim]
|
194 |
-
k_pass = qkv[:, :, 1, :, rotary_dim:]
|
195 |
-
|
196 |
-
q1, q2 = q_rot.chunk(2, dim=-1)
|
197 |
-
k1, k2 = k_rot.chunk(2, dim=-1)
|
198 |
-
c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
|
199 |
-
q1, q2, k1, k2, c, s = [t.to(dtype=torch.float32) for t in [q1, q2, k1, k2, c, s]]
|
200 |
-
|
201 |
-
q_rot = torch.cat([q1 * c - q2 * s, q1 * s + q2 * c], axis=-1).to(qkv.dtype)
|
202 |
-
k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(qkv.dtype)
|
203 |
-
|
204 |
-
return torch.cat(
|
205 |
-
[
|
206 |
-
torch.cat([q_rot, q_pass], axis=-1).unsqueeze(2),
|
207 |
-
torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
|
208 |
-
qkv[:, :, 2:3, :, :],
|
209 |
-
],
|
210 |
-
axis=2,
|
211 |
-
)
|
212 |
-
|
213 |
-
|
214 |
-
class RotaryEmbedding(nn.Module):
|
215 |
-
"""Rotary positional embedding (RoPE).
|
216 |
-
|
217 |
-
Reference:
|
218 |
-
RoFormer: Enhanced Transformer with Rotary Position Embedding.
|
219 |
-
https://arxiv.org/pdf/2104.09864.pdf.
|
220 |
-
|
221 |
-
"""
|
222 |
-
|
223 |
-
def __init__(
|
224 |
-
self,
|
225 |
-
dim: int,
|
226 |
-
base: int = 10000,
|
227 |
-
scale_base: Optional[float] = None,
|
228 |
-
pos_idx_in_fp32: bool = True,
|
229 |
-
device: Optional[str] = None,
|
230 |
-
**kwargs,
|
231 |
-
) -> None:
|
232 |
-
super().__init__()
|
233 |
-
|
234 |
-
if scale_base is not None:
|
235 |
-
raise NotImplementedError
|
236 |
-
|
237 |
-
self.dim = dim
|
238 |
-
self.base = float(base)
|
239 |
-
self.scale_base = scale_base
|
240 |
-
self.pos_idx_in_fp32 = pos_idx_in_fp32
|
241 |
-
self.device = device
|
242 |
-
|
243 |
-
# Generate and save the inverse frequency buffer (non-trainable)
|
244 |
-
inv_freq = self._compute_inv_freq(device)
|
245 |
-
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
246 |
-
|
247 |
-
# Generate and save the scale buffer (non-trainable)
|
248 |
-
scale = (
|
249 |
-
(torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim) / (1.4 * dim)
|
250 |
-
if scale_base is not None
|
251 |
-
else None
|
252 |
-
)
|
253 |
-
self.register_buffer("scale", scale, persistent=False)
|
254 |
-
|
255 |
-
self._seq_len_cached = 0
|
256 |
-
self._cos_cached = None
|
257 |
-
self._sin_cached = None
|
258 |
-
self._cos_k_cached = None
|
259 |
-
self._sin_k_cached = None
|
260 |
-
|
261 |
-
def _compute_inv_freq(self, device: Optional[str] = None) -> torch.FloatTensor:
|
262 |
-
return 1.0 / (self.base ** (torch.arange(0, self.dim, 2, device=device, dtype=torch.float32) / self.dim))
|
263 |
-
|
264 |
-
def _update_cos_sin_cache(
|
265 |
-
self, seqlen: int, device: Optional[str] = None, dtype: Optional[torch.dtype] = None
|
266 |
-
) -> None:
|
267 |
-
# Reset the tables if sequence length has been chaned, if we are on a
|
268 |
-
# new device or if we are switching from inference mode to training
|
269 |
-
if (
|
270 |
-
seqlen > self._seq_len_cached
|
271 |
-
or self._cos_cached is None
|
272 |
-
or self._cos_cached.device != device
|
273 |
-
or self._cos_cached.dtype != dtype
|
274 |
-
or (self.training and self._cos_cached.is_inference())
|
275 |
-
):
|
276 |
-
self._seq_len_cached = seqlen
|
277 |
-
|
278 |
-
# fp32 is preferred since the output of `torch.arange` can be quite large
|
279 |
-
# and bf16 would lose a lot of precision
|
280 |
-
if self.pos_idx_in_fp32:
|
281 |
-
t = torch.arange(seqlen, device=device, dtype=torch.float32)
|
282 |
-
if self.inv_freq.dtype != torch.float32:
|
283 |
-
inv_freq = self._compute_inv_freq(device=device)
|
284 |
-
else:
|
285 |
-
inv_freq = self.inv_freq
|
286 |
-
else:
|
287 |
-
t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
|
288 |
-
inv_freq = self.inv_freq
|
289 |
-
|
290 |
-
# `torch.outer` is preferred since `torch.einsum` converts from fp32 to fp16 if used with AMP
|
291 |
-
freqs = torch.outer(t, inv_freq)
|
292 |
-
if self.scale is None:
|
293 |
-
self._cos_cached = torch.cos(freqs).to(dtype)
|
294 |
-
self._sin_cached = torch.sin(freqs).to(dtype)
|
295 |
-
else:
|
296 |
-
power = (
|
297 |
-
torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device) - seqlen // 2
|
298 |
-
) / self.scale_base
|
299 |
-
scale = self.scale.to(device=power.device) ** rearrange(power, "s -> s 1")
|
300 |
-
|
301 |
-
# Force the scale multiplication to happen in fp32
|
302 |
-
self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
|
303 |
-
self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
|
304 |
-
self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
|
305 |
-
self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
|
306 |
-
|
307 |
-
def forward(
|
308 |
-
self,
|
309 |
-
qkv: torch.Tensor,
|
310 |
-
kv: Optional[torch.Tensor] = None,
|
311 |
-
seqlen_offset: int = 0,
|
312 |
-
max_seqlen: Optional[int] = None,
|
313 |
-
) -> Tuple[torch.Tensor, torch.Tensor]:
|
314 |
-
seqlen = qkv.shape[1]
|
315 |
-
|
316 |
-
if max_seqlen is not None:
|
317 |
-
self._update_cos_sin_cache(max_seqlen, device=qkv.device, dtype=qkv.dtype)
|
318 |
-
else:
|
319 |
-
self._update_cos_sin_cache(seqlen + seqlen_offset, device=qkv.device, dtype=qkv.dtype)
|
320 |
-
|
321 |
-
if kv is None:
|
322 |
-
return _apply_rotary_emb_qkv(qkv, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:])
|
323 |
-
else:
|
324 |
-
q = _apply_rotary_emb(qkv, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:])
|
325 |
-
kv = _apply_rotary_emb_kv(kv, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:])
|
326 |
-
|
327 |
-
return q, kv
|
328 |
-
|
329 |
-
|
330 |
-
class MLP(nn.Module):
|
331 |
-
"""Multi-Layer Perceptron.
|
332 |
-
|
333 |
-
Reference:
|
334 |
-
Attention Is All You Need.
|
335 |
-
https://arxiv.org/pdf/1706.03762.pdf.
|
336 |
-
|
337 |
-
"""
|
338 |
-
|
339 |
-
def __init__(self, config: PretrainedConfig, n_inner: Optional[int] = None, act_fn: Optional[str] = None) -> None:
|
340 |
-
super().__init__()
|
341 |
-
|
342 |
-
act_fn = config.activation_function if act_fn is None else act_fn
|
343 |
-
assert act_fn in ACT2FN.keys(), f"`act_fn` must be one of: {ACT2FN.keys()}."
|
344 |
-
|
345 |
-
n_inner = getattr(config, "n_inner", None) if n_inner is None else n_inner
|
346 |
-
n_inner = n_inner if n_inner is not None else 4 * config.n_embd
|
347 |
-
|
348 |
-
self.fc1 = nn.Linear(config.n_embd, n_inner)
|
349 |
-
self.fc2 = nn.Linear(n_inner, config.n_embd)
|
350 |
-
self.act = ACT2FN[act_fn]
|
351 |
-
|
352 |
-
def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
|
353 |
-
hidden_states = self.fc1(hidden_states)
|
354 |
-
hidden_states = self.act(hidden_states)
|
355 |
-
hidden_states = self.fc2(hidden_states)
|
356 |
-
|
357 |
-
return hidden_states
|
358 |
-
|
359 |
-
|
360 |
-
class SelfAttention(nn.Module):
|
361 |
-
"""Self-attention layer (compatible with PyTorch).
|
362 |
-
Reference:
|
363 |
-
https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py.
|
364 |
-
"""
|
365 |
-
|
366 |
-
def __init__(
|
367 |
-
self,
|
368 |
-
causal: bool = True,
|
369 |
-
softmax_scale: Optional[float] = None,
|
370 |
-
attention_dropout: float = 0.0,
|
371 |
-
) -> None:
|
372 |
-
super().__init__()
|
373 |
-
|
374 |
-
self.causal = causal
|
375 |
-
self.softmax_scale = softmax_scale
|
376 |
-
self.drop = nn.Dropout(attention_dropout)
|
377 |
-
|
378 |
-
def forward(
|
379 |
-
self,
|
380 |
-
qkv: torch.FloatTensor,
|
381 |
-
causal: bool = None,
|
382 |
-
key_padding_mask: Optional[torch.BoolTensor] = None,
|
383 |
-
**kwargs,
|
384 |
-
) -> torch.FloatTensor:
|
385 |
-
batch_size, seqlen = qkv.shape[0], qkv.shape[1]
|
386 |
-
q, k, v = qkv.unbind(dim=2)
|
387 |
-
|
388 |
-
causal = self.causal if causal is None else causal
|
389 |
-
softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
|
390 |
-
|
391 |
-
scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)
|
392 |
-
|
393 |
-
if key_padding_mask is not None:
|
394 |
-
padding_mask = torch.full((batch_size, seqlen), -10000.0, dtype=scores.dtype, device=scores.device)
|
395 |
-
padding_mask.masked_fill_(key_padding_mask, 0.0)
|
396 |
-
|
397 |
-
scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")
|
398 |
-
|
399 |
-
if causal:
|
400 |
-
causal_mask = torch.triu(torch.full((seqlen, seqlen), -10000.0, device=scores.device), 1)
|
401 |
-
scores = scores + causal_mask.to(dtype=scores.dtype)
|
402 |
-
|
403 |
-
attention = torch.softmax(scores, dim=-1, dtype=v.dtype)
|
404 |
-
attention = self.drop(attention)
|
405 |
-
|
406 |
-
output = torch.einsum("bhts,bshd->bthd", attention, v)
|
407 |
-
|
408 |
-
return output
|
409 |
-
|
410 |
-
|
411 |
-
class CrossAttention(nn.Module):
|
412 |
-
"""Cross-attention layer (compatible with PyTorch).
|
413 |
-
Reference:
|
414 |
-
https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py.
|
415 |
-
"""
|
416 |
-
|
417 |
-
def __init__(
|
418 |
-
self,
|
419 |
-
causal: bool = True,
|
420 |
-
softmax_scale: Optional[float] = None,
|
421 |
-
attention_dropout: float = 0.0,
|
422 |
-
) -> None:
|
423 |
-
super().__init__()
|
424 |
-
|
425 |
-
self.causal = causal
|
426 |
-
self.softmax_scale = softmax_scale
|
427 |
-
self.drop = nn.Dropout(attention_dropout)
|
428 |
-
|
429 |
-
def forward(
|
430 |
-
self,
|
431 |
-
q: torch.FloatTensor,
|
432 |
-
kv: torch.FloatTensor,
|
433 |
-
causal: bool = None,
|
434 |
-
key_padding_mask: Optional[torch.BoolTensor] = None,
|
435 |
-
**kwargs,
|
436 |
-
) -> torch.FloatTensor:
|
437 |
-
batch_size, seqlen_q = q.shape[0], q.shape[1]
|
438 |
-
seqlen_k = kv.shape[1]
|
439 |
-
|
440 |
-
if kv.shape[3] != q.shape[2]:
|
441 |
-
kv = repeat(kv, "... hkv d -> ... (hkv g) d", g=q.shape[2] // kv.shape[3])
|
442 |
-
k, v = kv.unbind(dim=2)
|
443 |
-
|
444 |
-
causal = self.causal if causal is None else causal
|
445 |
-
softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
|
446 |
-
|
447 |
-
scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)
|
448 |
-
|
449 |
-
if key_padding_mask is not None:
|
450 |
-
padding_mask = torch.full(
|
451 |
-
(batch_size, seqlen_k),
|
452 |
-
-10000.0,
|
453 |
-
dtype=scores.dtype,
|
454 |
-
device=scores.device,
|
455 |
-
)
|
456 |
-
padding_mask.masked_fill_(key_padding_mask, 0.0)
|
457 |
-
|
458 |
-
scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")
|
459 |
-
|
460 |
-
if causal:
|
461 |
-
rows = rearrange(torch.arange(seqlen_q, device=q.device, dtype=torch.long), "s -> s 1")
|
462 |
-
cols = torch.arange(seqlen_k, device=k.device, dtype=torch.long)
|
463 |
-
causal_mask = cols > rows + seqlen_k - seqlen_q
|
464 |
-
|
465 |
-
scores = scores.masked_fill(causal_mask, -10000.0)
|
466 |
-
|
467 |
-
attention = torch.softmax(scores, dim=-1, dtype=v.dtype)
|
468 |
-
attention = self.drop(attention)
|
469 |
-
|
470 |
-
output = torch.einsum("bhts,bshd->bthd", attention, v)
|
471 |
-
|
472 |
-
return output
|
473 |
-
|
474 |
-
|
475 |
-
def _find_mha_dims(
|
476 |
-
config: PretrainedConfig,
|
477 |
-
n_head: Optional[int] = None,
|
478 |
-
n_head_kv: Optional[int] = None,
|
479 |
-
head_dim: Optional[int] = None,
|
480 |
-
) -> Tuple[int, int]:
|
481 |
-
assert all(
|
482 |
-
hasattr(config, attr) for attr in ["n_embd", "n_head"]
|
483 |
-
), "`config` must have `n_embd` and `n_head` attributes."
|
484 |
-
|
485 |
-
if head_dim is None:
|
486 |
-
assert (
|
487 |
-
config.n_embd % config.n_head == 0
|
488 |
-
), f"Hidden size ({config.n_embd}) must be divisible by the number of heads ({config.n_head})."
|
489 |
-
|
490 |
-
if n_head is None and head_dim is None:
|
491 |
-
head_dim = config.n_embd // config.n_head
|
492 |
-
n_head = config.n_head
|
493 |
-
elif n_head is None or head_dim is None:
|
494 |
-
raise ValueError("`n_head` and `head_dim` must be both specified or `None`.")
|
495 |
-
|
496 |
-
if n_head_kv is None:
|
497 |
-
n_head_kv = getattr(config, "n_head_kv", None) or n_head
|
498 |
-
assert n_head % n_head_kv == 0, "`n_head` must be divisible by `n_head_kv`."
|
499 |
-
|
500 |
-
return n_head, n_head_kv, head_dim
|
501 |
-
|
502 |
-
|
503 |
-
def _update_kv_cache(kv: torch.FloatTensor, inference_params: InferenceParams, layer_idx: int) -> torch.FloatTensor:
|
504 |
-
num_heads, head_dim = kv.shape[-2:]
|
505 |
-
|
506 |
-
if layer_idx not in inference_params.key_value_memory_dict:
|
507 |
-
kv_cache = torch.empty(
|
508 |
-
inference_params.max_batch_size,
|
509 |
-
inference_params.max_seqlen,
|
510 |
-
2,
|
511 |
-
num_heads,
|
512 |
-
head_dim,
|
513 |
-
dtype=kv.dtype,
|
514 |
-
device=kv.device,
|
515 |
-
)
|
516 |
-
inference_params.key_value_memory_dict[layer_idx] = kv_cache
|
517 |
-
else:
|
518 |
-
kv_cache = inference_params.key_value_memory_dict[layer_idx]
|
519 |
-
|
520 |
-
batch_start = inference_params.batch_size_offset
|
521 |
-
batch_end = batch_start + kv.shape[0]
|
522 |
-
assert batch_end <= kv_cache.shape[0]
|
523 |
-
|
524 |
-
sequence_start = inference_params.seqlen_offset
|
525 |
-
sequence_end = sequence_start + kv.shape[1]
|
526 |
-
assert sequence_end <= kv_cache.shape[1]
|
527 |
-
|
528 |
-
assert kv_cache is not None
|
529 |
-
kv_cache[batch_start:batch_end, sequence_start:sequence_end, ...] = kv
|
530 |
-
kv = kv_cache[batch_start:batch_end, :sequence_end, ...]
|
531 |
-
|
532 |
-
return kv
|
533 |
-
|
534 |
-
|
535 |
-
class MHA(nn.Module):
|
536 |
-
"""Multi-head attention layer."""
|
537 |
-
|
538 |
-
def __init__(
|
539 |
-
self,
|
540 |
-
config: PretrainedConfig,
|
541 |
-
dtype: Optional[torch.dtype] = None,
|
542 |
-
device: Optional[str] = None,
|
543 |
-
rotary_dim: Optional[int] = None,
|
544 |
-
rotary_scale_base: Optional[float] = None,
|
545 |
-
n_head: Optional[int] = None,
|
546 |
-
n_head_kv: Optional[int] = None,
|
547 |
-
head_dim: Optional[int] = None,
|
548 |
-
bias: bool = True,
|
549 |
-
causal: bool = True,
|
550 |
-
softmax_scale: Optional[float] = None,
|
551 |
-
layer_idx: Optional[int] = None,
|
552 |
-
return_residual: bool = False,
|
553 |
-
checkpointing: bool = False,
|
554 |
-
) -> None:
|
555 |
-
super().__init__()
|
556 |
-
|
557 |
-
# Rotary embedding
|
558 |
-
self.rotary_dim = rotary_dim if rotary_dim is not None else getattr(config, "rotary_dim", 0)
|
559 |
-
if self.rotary_dim > 0:
|
560 |
-
rotary_kwargs = {"device": device}
|
561 |
-
if rotary_scale_base is not None and rotary_scale_base > 0.0:
|
562 |
-
rotary_kwargs["scale_base"] = rotary_scale_base
|
563 |
-
|
564 |
-
rotary_cls = FlashRotaryEmbedding if config.flash_rotary else RotaryEmbedding
|
565 |
-
if rotary_cls is None:
|
566 |
-
rotary_cls = RotaryEmbedding
|
567 |
-
self.rotary_emb = rotary_cls(self.rotary_dim, **rotary_kwargs)
|
568 |
-
|
569 |
-
# MLP
|
570 |
-
self.n_head, self.n_head_kv, self.head_dim = _find_mha_dims(config, n_head=n_head, n_head_kv=n_head_kv, head_dim=head_dim)
|
571 |
-
op_size = self.head_dim * (self.n_head + 2 * self.n_head_kv)
|
572 |
-
hidden_size = config.n_embd
|
573 |
-
|
574 |
-
linear_cls = FusedDense if config.fused_dense else nn.Linear
|
575 |
-
if linear_cls is None:
|
576 |
-
linear_cls = nn.Linear
|
577 |
-
|
578 |
-
self.Wqkv = linear_cls(hidden_size, op_size, bias=bias, device=device, dtype=dtype)
|
579 |
-
self.out_proj = linear_cls(hidden_size, hidden_size, bias=bias, device=device, dtype=dtype)
|
580 |
-
|
581 |
-
# Attention
|
582 |
-
attn_cls = FlashSelfAttention if config.flash_attn else SelfAttention
|
583 |
-
if attn_cls is None:
|
584 |
-
attn_cls = SelfAttention
|
585 |
-
|
586 |
-
cross_attn_cls = FlashCrossAttention if config.flash_attn else CrossAttention
|
587 |
-
if cross_attn_cls is None:
|
588 |
-
cross_attn_cls = CrossAttention
|
589 |
-
|
590 |
-
self.inner_attn = attn_cls(causal=causal, softmax_scale=softmax_scale, attention_dropout=config.attn_pdrop)
|
591 |
-
self.inner_cross_attn = cross_attn_cls(
|
592 |
-
causal=causal, softmax_scale=softmax_scale, attention_dropout=config.attn_pdrop
|
593 |
-
)
|
594 |
-
|
595 |
-
self.flash_attn = config.flash_attn and attn_cls is FlashSelfAttention
|
596 |
-
self.layer_idx = layer_idx
|
597 |
-
self.return_residual = return_residual
|
598 |
-
self.checkpointing = checkpointing
|
599 |
-
|
600 |
-
def _forward_self_attn(
|
601 |
-
self, x: torch.FloatTensor, key_padding_mask: Optional[torch.BoolTensor]
|
602 |
-
) -> torch.FloatTensor:
|
603 |
-
qkv = self.Wqkv(x)
|
604 |
-
qkv = rearrange(qkv, "... (three h d) -> ... three h d", three=3, d=self.head_dim)
|
605 |
-
|
606 |
-
if self.rotary_dim > 0:
|
607 |
-
qkv = self.rotary_emb(qkv)
|
608 |
-
|
609 |
-
if self.flash_attn:
|
610 |
-
batch_size, seqlen = qkv.shape[0], qkv.shape[1]
|
611 |
-
|
612 |
-
cu_seqlens, max_seqlen = None, None
|
613 |
-
if key_padding_mask is not None:
|
614 |
-
# If `key_padding_mask` is supplied, we need to unpad the input and retrieve
|
615 |
-
# the `cu_seqlens` and `max_seqlen` to be used by `flash-attn`
|
616 |
-
qkv, indices, cu_seqlens, max_seqlen = unpad_input(qkv, key_padding_mask)
|
617 |
-
|
618 |
-
if self.checkpointing:
|
619 |
-
attn_output = torch.utils.checkpoint.checkpoint(
|
620 |
-
self.inner_attn, qkv, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen
|
621 |
-
)
|
622 |
-
else:
|
623 |
-
attn_output = self.inner_attn(qkv, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen).to(qkv.device)
|
624 |
-
|
625 |
-
# If `key_padding_mask` is supplied, we need to pad the output back to the original shape
|
626 |
-
return pad_input(attn_output, indices, batch_size, seqlen) if key_padding_mask is not None else attn_output
|
627 |
-
|
628 |
-
if self.checkpointing:
|
629 |
-
return torch.utils.checkpoint.checkpoint(self.inner_attn, qkv, key_padding_mask=key_padding_mask)
|
630 |
-
|
631 |
-
return self.inner_attn(qkv, key_padding_mask=key_padding_mask)
|
632 |
-
|
633 |
-
def _forward_cross_attn(
|
634 |
-
self,
|
635 |
-
x: torch.FloatTensor,
|
636 |
-
past_key_values: Optional[InferenceParams],
|
637 |
-
key_padding_mask: Optional[torch.BoolTensor],
|
638 |
-
) -> torch.FloatTensor:
|
639 |
-
batch_size = x.shape[0]
|
640 |
-
|
641 |
-
qkv = self.Wqkv(x)
|
642 |
-
|
643 |
-
q = qkv[..., : self.n_head * self.head_dim]
|
644 |
-
q = rearrange(q, "... (h d) -> ... h d", d=self.head_dim)
|
645 |
-
|
646 |
-
kv = qkv[..., self.n_head * self.head_dim :]
|
647 |
-
kv = rearrange(kv, "... (two hkv d) -> ... two hkv d", two=2, d=self.head_dim)
|
648 |
-
|
649 |
-
seqlen_offset = past_key_values.seqlen_offset if past_key_values is not None else 0
|
650 |
-
causal = None if seqlen_offset == 0 else False
|
651 |
-
if self.rotary_dim > 0:
|
652 |
-
q, kv = self.rotary_emb(q, kv=kv, seqlen_offset=seqlen_offset)
|
653 |
-
|
654 |
-
if past_key_values is not None:
|
655 |
-
kv = _update_kv_cache(kv, past_key_values, self.layer_idx)
|
656 |
-
|
657 |
-
if self.flash_attn:
|
658 |
-
batch_size, seqlen_q = q.shape[0], q.shape[1]
|
659 |
-
seqlen_k = kv.shape[1]
|
660 |
-
|
661 |
-
cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k = None, None, None, None
|
662 |
-
if key_padding_mask is not None:
|
663 |
-
kv, _, cu_seqlens_k, max_seqlen_k = unpad_input(kv, key_padding_mask)
|
664 |
-
|
665 |
-
if seqlen_q == 1:
|
666 |
-
key_padding_mask = torch.ones(batch_size, 1, device=q.device)
|
667 |
-
elif seqlen_q != seqlen_k:
|
668 |
-
key_padding_mask = key_padding_mask[:, -seqlen_q:]
|
669 |
-
|
670 |
-
q, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(q, key_padding_mask)
|
671 |
-
|
672 |
-
if self.checkpointing:
|
673 |
-
attn_output = torch.utils.checkpoint.checkpoint(
|
674 |
-
self.inner_cross_attn,
|
675 |
-
q,
|
676 |
-
kv,
|
677 |
-
causal=causal,
|
678 |
-
cu_seqlens=cu_seqlens_q,
|
679 |
-
max_seqlen=max_seqlen_q,
|
680 |
-
cu_seqlens_k=cu_seqlens_k,
|
681 |
-
max_seqlen_k=max_seqlen_k,
|
682 |
-
)
|
683 |
-
else:
|
684 |
-
attn_output = self.inner_cross_attn(
|
685 |
-
q,
|
686 |
-
kv,
|
687 |
-
causal=causal,
|
688 |
-
cu_seqlens=cu_seqlens_q,
|
689 |
-
max_seqlen=max_seqlen_q,
|
690 |
-
cu_seqlens_k=cu_seqlens_k,
|
691 |
-
max_seqlen_k=max_seqlen_k,
|
692 |
-
)
|
693 |
-
|
694 |
-
return (
|
695 |
-
pad_input(attn_output, indices_q, batch_size, max_seqlen_q)
|
696 |
-
if key_padding_mask is not None
|
697 |
-
else attn_output
|
698 |
-
)
|
699 |
-
|
700 |
-
if self.checkpointing:
|
701 |
-
return torch.utils.checkpoint.checkpoint(
|
702 |
-
self.inner_cross_attn, q, kv, key_padding_mask=key_padding_mask, causal=causal
|
703 |
-
)
|
704 |
-
|
705 |
-
return self.inner_cross_attn(q, kv, key_padding_mask=key_padding_mask, causal=causal)
|
706 |
-
|
707 |
-
def forward(
|
708 |
-
self,
|
709 |
-
x: torch.FloatTensor,
|
710 |
-
past_key_values: Optional[InferenceParams] = None,
|
711 |
-
attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
|
712 |
-
**kwargs,
|
713 |
-
) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
|
714 |
-
# TODO: Need an alternative way for dynamic control flow: torch.any(~attention_mask.bool())
|
715 |
-
if attention_mask is not None:
|
716 |
-
attention_mask = attention_mask.bool()
|
717 |
-
else:
|
718 |
-
attention_mask = None
|
719 |
-
|
720 |
-
# MHA
|
721 |
-
if self.n_head == self.n_head_kv:
|
722 |
-
if past_key_values is None:
|
723 |
-
# If `past_key_values` are not supplied, we run self-attention
|
724 |
-
attn_output = self._forward_self_attn(x, attention_mask)
|
725 |
-
else:
|
726 |
-
# If `past_key_values` are supplied, it means that we might have cached values and
|
727 |
-
# could take advantage of cross-attention
|
728 |
-
attn_output = self._forward_cross_attn(x, past_key_values, attention_mask)
|
729 |
-
# MQA / GQA
|
730 |
-
else:
|
731 |
-
# Regardless of `past_key_values` being supplied or not, it always use cross-attention
|
732 |
-
# because `q` and `kv` lengths might be different
|
733 |
-
attn_output = self._forward_cross_attn(x, past_key_values, attention_mask)
|
734 |
-
|
735 |
-
output = rearrange(attn_output, "... h d -> ... (h d)")
|
736 |
-
output = self.out_proj(output)
|
737 |
-
|
738 |
-
return output if not self.return_residual else (output, x)
|
739 |
-
|
740 |
-
|
741 |
-
class ParallelBlock(nn.Module):
|
742 |
-
"""Parallel block.
|
743 |
-
|
744 |
-
This block applies parallel mixer and MLP layers to the input (used in GPT-J and CodeGen).
|
745 |
-
|
746 |
-
"""
|
747 |
-
|
748 |
-
def __init__(
|
749 |
-
self,
|
750 |
-
config: PretrainedConfig,
|
751 |
-
block_idx: Optional[int] = None,
|
752 |
-
) -> None:
|
753 |
-
super().__init__()
|
754 |
-
|
755 |
-
self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
|
756 |
-
self.resid_dropout = nn.Dropout(config.resid_pdrop)
|
757 |
-
self.block_idx = block_idx
|
758 |
-
|
759 |
-
self.mixer = MHA(config, layer_idx=block_idx)
|
760 |
-
self.mlp = MLP(config)
|
761 |
-
|
762 |
-
def forward(
|
763 |
-
self,
|
764 |
-
hidden_states: torch.FloatTensor,
|
765 |
-
past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
|
766 |
-
attention_mask: Optional[torch.BoolTensor] = None,
|
767 |
-
**kwargs,
|
768 |
-
) -> torch.FloatTensor:
|
769 |
-
residual = hidden_states
|
770 |
-
hidden_states = self.ln(hidden_states)
|
771 |
-
|
772 |
-
attn_outputs = self.mixer(hidden_states, past_key_values=past_key_values, attention_mask=attention_mask)
|
773 |
-
if isinstance(attn_outputs, tuple):
|
774 |
-
attn_outputs = attn_outputs[0]
|
775 |
-
|
776 |
-
attn_outputs = self.resid_dropout(attn_outputs)
|
777 |
-
feed_forward_hidden_states = self.resid_dropout(self.mlp(hidden_states))
|
778 |
-
|
779 |
-
hidden_states = attn_outputs + feed_forward_hidden_states + residual
|
780 |
-
|
781 |
-
return hidden_states
|
782 |
-
|
783 |
-
|
784 |
-
class CausalLMHead(nn.Module):
|
785 |
-
"""Causal Language Modeling head.
|
786 |
-
|
787 |
-
Reference:
|
788 |
-
Improving Language Understanding by Generative Pre-Training.
|
789 |
-
https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.
|
790 |
-
|
791 |
-
"""
|
792 |
-
|
793 |
-
def __init__(self, config: PretrainedConfig) -> None:
|
794 |
-
super().__init__()
|
795 |
-
|
796 |
-
self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
|
797 |
-
self.linear = nn.Linear(config.n_embd, config.vocab_size)
|
798 |
-
|
799 |
-
def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
|
800 |
-
hidden_states = self.ln(hidden_states)
|
801 |
-
logits = self.linear(hidden_states).to(torch.float32)
|
802 |
-
|
803 |
-
return logits
|
804 |
-
|
805 |
-
|
806 |
-
class CausalLMLoss(nn.Module):
|
807 |
-
"""Causal Language Modeling loss.
|
808 |
-
|
809 |
-
Reference:
|
810 |
-
Improving Language Understanding by Generative Pre-Training.
|
811 |
-
https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.
|
812 |
-
|
813 |
-
"""
|
814 |
-
|
815 |
-
def __init__(self, shift_labels: bool = True) -> None:
|
816 |
-
super().__init__()
|
817 |
-
|
818 |
-
self.shift_labels = shift_labels
|
819 |
-
self.loss_fct = nn.CrossEntropyLoss()
|
820 |
-
|
821 |
-
def forward(self, logits: torch.FloatTensor, labels: torch.LongTensor) -> torch.FloatTensor:
|
822 |
-
if self.shift_labels:
|
823 |
-
logits = logits[..., :-1, :].contiguous()
|
824 |
-
labels = labels[..., 1:].contiguous()
|
825 |
-
|
826 |
-
loss = self.loss_fct(logits.view(-1, logits.size(-1)), labels.view(-1))
|
827 |
-
|
828 |
-
return loss
|
829 |
-
|
830 |
-
|
831 |
-
class MixFormerSequentialPreTrainedModel(PreTrainedModel):
|
832 |
-
"""MixFormer (sequential for DeepSpeed) pre-trained model."""
|
833 |
-
|
834 |
-
config_class = MixFormerSequentialConfig
|
835 |
-
base_model_prefix = "transformer"
|
836 |
-
supports_gradient_checkpointing = True
|
837 |
-
|
838 |
-
def __init__(self, *inputs, **kwargs) -> None:
|
839 |
-
super().__init__(*inputs, **kwargs)
|
840 |
-
|
841 |
-
def _init_weights(self, module: nn.Module) -> None:
|
842 |
-
if isinstance(module, (nn.Linear,)):
|
843 |
-
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
844 |
-
if module.bias is not None:
|
845 |
-
module.bias.data.zero_()
|
846 |
-
elif isinstance(module, nn.Embedding):
|
847 |
-
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
848 |
-
if module.padding_idx is not None:
|
849 |
-
module.weight.data[module.padding_idx].zero_()
|
850 |
-
elif isinstance(module, nn.LayerNorm):
|
851 |
-
if module.bias is not None:
|
852 |
-
module.bias.data.zero_()
|
853 |
-
module.weight.data.fill_(1.0)
|
854 |
-
|
855 |
-
def prepare_inputs_for_generation(
|
856 |
-
self,
|
857 |
-
input_ids: torch.LongTensor,
|
858 |
-
past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
|
859 |
-
attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
|
860 |
-
**kwargs,
|
861 |
-
) -> Dict[str, Any]:
|
862 |
-
if past_key_values is None or not (isinstance(past_key_values, InferenceParams)):
|
863 |
-
past_key_values = InferenceParams(
|
864 |
-
max_seqlen=self.config.n_positions,
|
865 |
-
max_batch_size=input_ids.shape[0],
|
866 |
-
seqlen_offset=0,
|
867 |
-
batch_size_offset=0,
|
868 |
-
key_value_memory_dict={},
|
869 |
-
lengths_per_sample=None,
|
870 |
-
)
|
871 |
-
else:
|
872 |
-
# Assume that `past_key_values` has cached all tokens up to the last token in `input_ids`
|
873 |
-
past_key_values.seqlen_offset = len(input_ids[0]) - 1
|
874 |
-
input_ids = input_ids[:, -1].unsqueeze(-1)
|
875 |
-
|
876 |
-
return {
|
877 |
-
"input_ids": input_ids,
|
878 |
-
"past_key_values": past_key_values,
|
879 |
-
"attention_mask": attention_mask,
|
880 |
-
}
|
881 |
-
|
882 |
-
def _set_gradient_checkpointing(self, module: nn.Module, value: bool = False) -> None:
|
883 |
-
if isinstance(module, MixFormerSequentialPreTrainedModel):
|
884 |
-
module.gradient_checkpointing = value
|
885 |
-
|
886 |
-
|
887 |
-
class MixFormerSequentialForCausalLM(MixFormerSequentialPreTrainedModel):
|
888 |
-
"""MixFormer (sequential for DeepSpeed) for Causal Language Modeling."""
|
889 |
-
|
890 |
-
_keys_to_ignore_on_load_missing = [""]
|
891 |
-
_keys_to_ignore_on_load_unexpected = [r"layers\.\d+\.mlp.(fc_in|fc_out)\.(weight|bias)"]
|
892 |
-
_no_split_modules = ["ParallelBlock"]
|
893 |
-
|
894 |
-
def __init__(self, config: MixFormerSequentialConfig) -> None:
|
895 |
-
super().__init__(config)
|
896 |
-
|
897 |
-
modules = [Embedding(config)]
|
898 |
-
modules += [ParallelBlock(config, block_idx=i) for i in range(config.n_layer)]
|
899 |
-
modules.append(CausalLMHead(config))
|
900 |
-
|
901 |
-
self.layers = nn.Sequential(*modules)
|
902 |
-
self.loss = CausalLMLoss()
|
903 |
-
|
904 |
-
self.post_init()
|
905 |
-
|
906 |
-
def get_input_embeddings(self) -> nn.Embedding:
|
907 |
-
return self.layers[0].wte
|
908 |
-
|
909 |
-
def set_input_embeddings(self, new_embeddings: nn.Embedding) -> None:
|
910 |
-
self.layers[0].wte = new_embeddings
|
911 |
-
|
912 |
-
def get_output_embeddings(self) -> nn.Linear:
|
913 |
-
return self.layers[-1].linear
|
914 |
-
|
915 |
-
def set_output_embeddings(self, new_embeddings: nn.Linear) -> None:
|
916 |
-
self.layers[-1].linear = new_embeddings
|
917 |
-
|
918 |
-
def forward(
|
919 |
-
self,
|
920 |
-
input_ids: torch.LongTensor,
|
921 |
-
past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
|
922 |
-
attention_mask: Optional[torch.BoolTensor] = None,
|
923 |
-
labels: Optional[torch.LongTensor] = None,
|
924 |
-
**kwargs,
|
925 |
-
) -> CausalLMOutputWithPast:
|
926 |
-
hidden_layer = self.layers[0](input_ids)
|
927 |
-
for module in self.layers[1:-1]:
|
928 |
-
hidden_layer = module(hidden_layer, past_key_values=past_key_values, attention_mask=attention_mask)
|
929 |
-
lm_logits = self.layers[-1](hidden_layer)
|
930 |
-
|
931 |
-
loss = None
|
932 |
-
if labels is not None:
|
933 |
-
loss = self.loss(lm_logits, labels)
|
934 |
-
|
935 |
-
return CausalLMOutputWithPast(loss=loss, logits=lm_logits, past_key_values=past_key_values)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|