Spaces:
Running
on
Zero
Running
on
Zero
File size: 6,016 Bytes
96fe5d9 7e08af8 e5d1aa9 96fe5d9 3c478e0 96fe5d9 |
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 |
# Copyright (c) 2024 Alibaba Inc
#
# 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 torch.nn as nn
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from inspiremusic.utils.mask import make_pad_mask
from inspiremusic.utils.hinter import hint_once
class QwenEncoder(nn.Module):
def __init__(
self,
input_size: int,
pretrain_path: str = "Qwen/Qwen2.0-0.5B",
trainable: bool = False,
do_fusion_emb: bool = False,
fusion_drop_rate: float = 0.0,
):
super(QwenEncoder, self).__init__()
self.input_size = input_size
self.trainable = trainable
self.model = AutoModelForCausalLM.from_pretrained(pretrain_path, device_map="cpu")
self._output_size = self.model.config.hidden_size
self.do_fusion_emb = do_fusion_emb
self.hidden_norm = torch.nn.LayerNorm(self._output_size)
self.fusion_dropout = nn.Dropout(fusion_drop_rate)
if do_fusion_emb:
self.fusion_layer = torch.nn.Linear(self._output_size * 2, self._output_size)
self.emb_norm = torch.nn.LayerNorm(self._output_size)
self.fusion_norm = torch.nn.LayerNorm(self._output_size)
from inspiremusic.transformer.activation import Swish
self.fusion_act = Swish(self)
if not self.trainable:
self.model.eval()
def output_size(self) -> int:
return self._output_size
def forward(
self,
input_ids: torch.Tensor,
ilens: torch.Tensor,
):
device = input_ids.device
input_ids = torch.clamp(input_ids, min=0, max=None)
input_masks = (~make_pad_mask(ilens)).to(device).long()
if not self.trainable:
with torch.no_grad():
model_outputs = self.model(
input_ids=input_ids,
attention_mask=input_masks,
output_hidden_states=True
)
else:
model_outputs = self.model(
input_ids=input_ids,
attention_mask=input_masks,
output_hidden_states=True
)
outs = model_outputs.hidden_states[-1]
outs = self.hidden_norm(outs)
if self.do_fusion_emb:
hint_once("fuse embedding and LM outputs", "fuse_emb")
outs = self.fusion_dropout(self.fusion_act(outs))
emb = model_outputs.hidden_states[0]
emb = self.fusion_dropout(self.fusion_act(self.emb_norm(emb)))
outs = self.fusion_layer(
torch.cat([outs, emb], dim=-1)
)
outs = self.fusion_act(self.fusion_norm(outs))
return outs, ilens
class QwenEmbeddingEncoder(nn.Module):
def __init__(
self,
input_size: int,
pretrain_path: str = "Qwen/Qwen2.0-0.5B",
):
super(QwenEmbeddingEncoder, self).__init__()
self.input_size = input_size
from transformers import Qwen2ForCausalLM
# self.model = Qwen2ForCausalLM.from_pretrained(pretrain_path, device_map="cpu", attn_implementation="flash_attention_2")
self.model = Qwen2ForCausalLM.from_pretrained(pretrain_path,
device_map="cpu")
self._output_size = self.model.config.hidden_size
def output_size(self) -> int:
return self._output_size
def forward(
self,
input_embeds: torch.Tensor,
ilens: torch.Tensor,
):
input_masks = (~make_pad_mask(ilens)).to(input_embeds.device).long()
outs = self.model(
inputs_embeds=input_embeds,
attention_mask=input_masks,
output_hidden_states=True,
return_dict=True,
)
return outs.hidden_states[-1], input_masks
def forward_one_step(self, xs, masks, cache=None):
outs = self.model(
inputs_embeds=xs,
attention_mask=masks,
output_hidden_states=True,
return_dict=True,
use_cache=True,
past_key_values=cache,
)
xs = outs.hidden_states[-1]
new_cache = outs.past_key_values
return xs, masks, new_cache
class QwenInputOnlyEncoder(nn.Module):
def __init__(
self,
input_size: int,
pretrain_path: str = "Qwen/Qwen2.0-0.5B",
):
super(QwenInputOnlyEncoder, self).__init__()
self.input_size = input_size
from transformers import Qwen2ForCausalLM
# model = Qwen2ForCausalLM.from_pretrained(pretrain_path, device_map="cpu", attn_implementation="flash_attention_2")
model = Qwen2ForCausalLM.from_pretrained(pretrain_path,
device_map="cpu")
self.embed = model.model.embed_tokens
for p in self.embed.parameters():
p.requires_grad = False
# set text embedding to non-trainable
# self.post_embed = model.model.rotary_emb
self._output_size = model.config.hidden_size
def output_size(self) -> int:
return self._output_size
def forward(
self,
input_ids: torch.Tensor,
ilens: torch.Tensor,
):
input_masks = (~make_pad_mask(ilens)).to(input_ids.device).long()
outs = self.embed(input_ids)
return outs, input_masks
|