Markus28 commited on
Commit
0211324
1 Parent(s): 4ed66c0

Added GLUMLP, changed config accordingly, added code to convert state_dict

Browse files
Files changed (4) hide show
  1. configuration_bert.py +3 -3
  2. convert_v2_weights.py +126 -0
  3. mlp.py +40 -0
  4. modeling_bert.py +18 -5
configuration_bert.py CHANGED
@@ -75,7 +75,7 @@ class JinaBertConfig(PretrainedConfig):
75
  pad_token_id=0,
76
  window_size=(-1, -1),
77
  dense_seq_output=False,
78
- fused_mlp=False,
79
  mlp_checkpoint_lvl=0,
80
  last_layer_subset=False,
81
  fused_dropout_add_ln=False,
@@ -92,7 +92,7 @@ class JinaBertConfig(PretrainedConfig):
92
  assert 'max_position_embeddings' not in kwargs
93
  super().__init__(pad_token_id=pad_token_id, **kwargs)
94
 
95
- if fused_mlp and hidden_act not in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"]:
96
  raise ValueError('Fused MLP only supports approximate gelu')
97
 
98
  self.vocab_size = vocab_size
@@ -108,7 +108,7 @@ class JinaBertConfig(PretrainedConfig):
108
  self.layer_norm_eps = layer_norm_eps
109
  self.window_size = window_size
110
  self.dense_seq_output = dense_seq_output
111
- self.fused_mlp = fused_mlp
112
  self.mlp_checkpoint_lvl = mlp_checkpoint_lvl
113
  self.last_layer_subset = last_layer_subset
114
  self.fused_dropout_add_ln = fused_dropout_add_ln
 
75
  pad_token_id=0,
76
  window_size=(-1, -1),
77
  dense_seq_output=False,
78
+ mlp_type='mlp',
79
  mlp_checkpoint_lvl=0,
80
  last_layer_subset=False,
81
  fused_dropout_add_ln=False,
 
92
  assert 'max_position_embeddings' not in kwargs
93
  super().__init__(pad_token_id=pad_token_id, **kwargs)
94
 
95
+ if mlp_type == 'fused_mlp' and hidden_act not in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"]:
96
  raise ValueError('Fused MLP only supports approximate gelu')
97
 
98
  self.vocab_size = vocab_size
 
108
  self.layer_norm_eps = layer_norm_eps
109
  self.window_size = window_size
110
  self.dense_seq_output = dense_seq_output
111
+ self.mlp_type= mlp_type
112
  self.mlp_checkpoint_lvl = mlp_checkpoint_lvl
113
  self.last_layer_subset = last_layer_subset
114
  self.fused_dropout_add_ln = fused_dropout_add_ln
convert_v2_weights.py ADDED
@@ -0,0 +1,126 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ from collections import OrderedDict
3
+ from transformers import AutoModel
4
+ from .configuration_bert import JinaBertConfig
5
+ import torch
6
+ from .modeling_bert import BertModel
7
+
8
+ def remap_state_dict(state_dict, config: JinaBertConfig):
9
+ """
10
+ Map the state_dict of a Huggingface BERT model to be flash_attn compatible.
11
+ """
12
+
13
+ # LayerNorm
14
+ def key_mapping_ln_gamma_beta(key):
15
+ key = re.sub(r"LayerNorm.gamma$", "LayerNorm.weight", key)
16
+ key = re.sub(r"LayerNorm.beta$", "LayerNorm.bias", key)
17
+ return key
18
+
19
+ state_dict = OrderedDict((key_mapping_ln_gamma_beta(k), v) for k, v in state_dict.items())
20
+
21
+ # Layers
22
+ def key_mapping_layers(key):
23
+ return re.sub(r"^encoder.layer.", "encoder.layers.", key)
24
+
25
+ state_dict = OrderedDict((key_mapping_layers(k), v) for k, v in state_dict.items())
26
+
27
+ # LayerNorm
28
+ def key_mapping_ln(key):
29
+ key = re.sub(r"^embeddings.LayerNorm.", "emb_ln.", key)
30
+ key = re.sub(
31
+ r"^encoder.layers.(\d+).attention.output.LayerNorm.(weight|bias)",
32
+ r"encoder.layers.\1.norm1.\2",
33
+ key,
34
+ )
35
+ key = re.sub(
36
+ r"^encoder.layers.(\d+).output.LayerNorm.(weight|bias)",
37
+ r"encoder.layers.\1.norm2.\2",
38
+ key,
39
+ )
40
+ key = re.sub(
41
+ r"^cls.predictions.transform.LayerNorm.(weight|bias)",
42
+ r"cls.predictions.transform.layer_norm.\1",
43
+ key,
44
+ )
45
+ return key
46
+
47
+ state_dict = OrderedDict((key_mapping_ln(k), v) for k, v in state_dict.items())
48
+
49
+ # MLP
50
+ def key_mapping_mlp(key):
51
+ key = re.sub(
52
+ r"^encoder.layers.(\d+).intermediate.dense.(weight|bias)",
53
+ r"encoder.layers.\1.mlp.fc1.\2",
54
+ key,
55
+ )
56
+ key = re.sub(
57
+ r"^encoder.layers.(\d+).output.dense.(weight|bias)",
58
+ r"encoder.layers.\1.mlp.fc2.\2",
59
+ key,
60
+ )
61
+ return key
62
+
63
+ state_dict = OrderedDict((key_mapping_mlp(k), v) for k, v in state_dict.items())
64
+
65
+ # Attention
66
+ last_layer_subset = getattr(config, "last_layer_subset", False)
67
+ for d in range(config.num_hidden_layers):
68
+ Wq = state_dict.pop(f"encoder.layers.{d}.attention.self.query.weight")
69
+ Wk = state_dict.pop(f"encoder.layers.{d}.attention.self.key.weight")
70
+ Wv = state_dict.pop(f"encoder.layers.{d}.attention.self.value.weight")
71
+ bq = state_dict.pop(f"encoder.layers.{d}.attention.self.query.bias")
72
+ bk = state_dict.pop(f"encoder.layers.{d}.attention.self.key.bias")
73
+ bv = state_dict.pop(f"encoder.layers.{d}.attention.self.value.bias")
74
+ if not (last_layer_subset and d == config.num_hidden_layers - 1):
75
+ state_dict[f"encoder.layers.{d}.mixer.Wqkv.weight"] = torch.cat(
76
+ [Wq, Wk, Wv], dim=0
77
+ )
78
+ state_dict[f"encoder.layers.{d}.mixer.Wqkv.bias"] = torch.cat([bq, bk, bv], dim=0)
79
+ else:
80
+ state_dict[f"encoder.layers.{d}.mixer.Wq.weight"] = Wq
81
+ state_dict[f"encoder.layers.{d}.mixer.Wkv.weight"] = torch.cat([Wk, Wv], dim=0)
82
+ state_dict[f"encoder.layers.{d}.mixer.Wq.bias"] = bq
83
+ state_dict[f"encoder.layers.{d}.mixer.Wkv.bias"] = torch.cat([bk, bv], dim=0)
84
+
85
+ def key_mapping_attn(key):
86
+ return re.sub(
87
+ r"^encoder.layers.(\d+).attention.output.dense.(weight|bias)",
88
+ r"encoder.layers.\1.mixer.out_proj.\2",
89
+ key,
90
+ )
91
+
92
+ state_dict = OrderedDict((key_mapping_attn(k), v) for k, v in state_dict.items())
93
+
94
+ def key_mapping_decoder_bias(key):
95
+ return re.sub(r"^cls.predictions.bias", "cls.predictions.decoder.bias", key)
96
+
97
+ state_dict = OrderedDict((key_mapping_decoder_bias(k), v) for k, v in state_dict.items())
98
+
99
+ # Word embedding
100
+ pad_vocab_size_multiple = getattr(config, "pad_vocab_size_multiple", 1)
101
+ if pad_vocab_size_multiple > 1:
102
+ word_embeddings = state_dict["embeddings.word_embeddings.weight"]
103
+ state_dict["embeddings.word_embeddings.weight"] = F.pad(
104
+ word_embeddings, (0, 0, 0, config.vocab_size - word_embeddings.shape[0])
105
+ )
106
+ decoder_weight = state_dict["cls.predictions.decoder.weight"]
107
+ state_dict["cls.predictions.decoder.weight"] = F.pad(
108
+ decoder_weight, (0, 0, 0, config.vocab_size - decoder_weight.shape[0])
109
+ )
110
+ # If the vocab was padded, we want to set the decoder bias for those padded indices to be
111
+ # strongly negative (i.e. the decoder shouldn't predict those indices).
112
+ # TD [2022-05-09]: I don't think it affects the MLPerf training.
113
+ decoder_bias = state_dict["cls.predictions.decoder.bias"]
114
+ state_dict["cls.predictions.decoder.bias"] = F.pad(
115
+ decoder_bias, (0, config.vocab_size - decoder_bias.shape[0]), value=-100.0
116
+ )
117
+
118
+ return state_dict
119
+
120
+
121
+ v2_model = AutoModel.from_pretrained('jinaai/jina-embeddings-v2-base-en', trust_remote_code=True)
122
+ config = JinaBertConfig(vocab_size=30528, use_qk_norm=False, mlp_type='glu', hidden_act='gelu')
123
+ state_dict = v2_model.state_dict()
124
+ new_state_dict = remap_state_dict(state_dict, config)
125
+ flash_model = BertModel(config)
126
+ flash_model.load_state_dict(new_state_dict)
mlp.py CHANGED
@@ -27,6 +27,46 @@ except ImportError:
27
  FusedMLP, ParallelFusedMLP = None, None
28
 
29
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
30
  class Mlp(nn.Module):
31
  def __init__(
32
  self,
 
27
  FusedMLP, ParallelFusedMLP = None, None
28
 
29
 
30
+ class GLUMLP(nn.Module):
31
+ def __init__(
32
+ self,
33
+ in_features,
34
+ hidden_features,
35
+ activation,
36
+ return_residual=False,
37
+ hidden_dropout_prob=0.1
38
+ ):
39
+ super().__init__()
40
+ self.gated_layers = nn.Linear(
41
+ in_features, hidden_features * 2, bias=False
42
+ )
43
+ if activation == 'relu':
44
+ self.act = nn.ReLU()
45
+ elif activation == 'gelu':
46
+ self.act = nn.GELU()
47
+ else:
48
+ raise ValueError(
49
+ f"activation {activation} not supported"
50
+ )
51
+ self.wo = nn.Linear(hidden_features, in_features)
52
+ self.dropout = nn.Dropout(hidden_dropout_prob)
53
+ self.return_residual = return_residual
54
+ #self.layernorm = nn.LayerNorm(in_features, eps=layer_norm_eps)
55
+
56
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
57
+ residual_connection = hidden_states
58
+ # compute the activation
59
+ hidden_states = self.gated_layers(hidden_states)
60
+ gated = hidden_states[:, :, : self.config.intermediate_size]
61
+ non_gated = hidden_states[:, :, self.config.intermediate_size :]
62
+ hidden_states = self.act(gated) * non_gated
63
+ hidden_states = self.dropout(hidden_states)
64
+ # multiply by the second matrix
65
+ hidden_states = self.wo(hidden_states)
66
+ # add the residual connection and post-LN
67
+ # hidden_states = self.layernorm(hidden_states + residual_connection)
68
+ return hidden_states if not self.return_residual else (hidden_states, residual_connection)
69
+
70
  class Mlp(nn.Module):
71
  def __init__(
72
  self,
modeling_bert.py CHANGED
@@ -39,7 +39,7 @@ from .bert_padding import (
39
  from .block import Block
40
  from .embedding import BertEmbeddings
41
  from .mha import MHA
42
- from .mlp import FusedMLP, Mlp
43
 
44
  try:
45
  from flash_attn.ops.fused_dense import FusedDense
@@ -89,12 +89,15 @@ def create_mixer_cls(config, cross_attn=False, return_residual=False):
89
 
90
  def create_mlp_cls(config, layer_idx=None, return_residual=False):
91
  inner_dim = config.intermediate_size
92
- fused_mlp = getattr(config, "fused_mlp", False)
93
- if fused_mlp:
 
94
  assert config.hidden_act in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"], (
95
  "fused_mlp only " "supports approximate gelu"
96
  )
97
- if not fused_mlp:
 
 
98
  approximate = (
99
  "tanh"
100
  if config.hidden_act in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"]
@@ -106,7 +109,15 @@ def create_mlp_cls(config, layer_idx=None, return_residual=False):
106
  activation=partial(F.gelu, approximate=approximate),
107
  return_residual=return_residual,
108
  )
109
- else:
 
 
 
 
 
 
 
 
110
  if FusedMLP is None:
111
  raise ImportError("fused_dense is not installed")
112
  mlp_checkpoint_lvl = getattr(config, "mlp_checkpoint_lvl", 0)
@@ -120,6 +131,8 @@ def create_mlp_cls(config, layer_idx=None, return_residual=False):
120
  checkpoint_lvl=mlp_checkpoint_lvl,
121
  return_residual=return_residual,
122
  )
 
 
123
  return mlp_cls
124
 
125
 
 
39
  from .block import Block
40
  from .embedding import BertEmbeddings
41
  from .mha import MHA
42
+ from .mlp import FusedMLP, Mlp, GLUMLP
43
 
44
  try:
45
  from flash_attn.ops.fused_dense import FusedDense
 
89
 
90
  def create_mlp_cls(config, layer_idx=None, return_residual=False):
91
  inner_dim = config.intermediate_size
92
+ mlp_type = config.mlp_type
93
+ assert mlp_type in ('mlp', 'fused_mlp', 'glu')
94
+ if mlp_type == 'fused_mlp':
95
  assert config.hidden_act in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"], (
96
  "fused_mlp only " "supports approximate gelu"
97
  )
98
+ if mlp_type == 'glu':
99
+ assert config.hidden_act in ('relu', 'gelu')
100
+ if mlp_type == 'mlp':
101
  approximate = (
102
  "tanh"
103
  if config.hidden_act in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"]
 
109
  activation=partial(F.gelu, approximate=approximate),
110
  return_residual=return_residual,
111
  )
112
+ elif mlp_type == 'glu':
113
+ mlp_cls = partial(
114
+ GLUMLP,
115
+ hidden_features=inner_dim,
116
+ activation=config.hidden_act,
117
+ hidden_dropout_prob=config.hidden_dropout_prob,
118
+ return_residual=return_residual,
119
+ )
120
+ elif mlp_type == 'fused_mlp':
121
  if FusedMLP is None:
122
  raise ImportError("fused_dense is not installed")
123
  mlp_checkpoint_lvl = getattr(config, "mlp_checkpoint_lvl", 0)
 
131
  checkpoint_lvl=mlp_checkpoint_lvl,
132
  return_residual=return_residual,
133
  )
134
+ else:
135
+ raise NotImplementedError
136
  return mlp_cls
137
 
138