P1ayer-1 commited on
Commit
7125b09
1 Parent(s): e73885d

Upload 9 files

Browse files
Files changed (8) hide show
  1. adapt_tokenizer.py +41 -0
  2. attention.py +276 -0
  3. blocks.py +41 -0
  4. hf_prefixlm_converter.py +415 -0
  5. meta_init_context.py +94 -0
  6. modeling_mpt.py +290 -0
  7. norm.py +56 -0
  8. param_init_fns.py +181 -0
adapt_tokenizer.py ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Union
2
+ from transformers import AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast
3
+ Tokenizer = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
4
+ NUM_SENTINEL_TOKENS: int = 100
5
+
6
+ def adapt_tokenizer_for_denoising(tokenizer: Tokenizer):
7
+ """Adds sentinel tokens and padding token (if missing).
8
+
9
+ Expands the tokenizer vocabulary to include sentinel tokens
10
+ used in mixture-of-denoiser tasks as well as a padding token.
11
+
12
+ All added tokens are added as special tokens. No tokens are
13
+ added if sentinel tokens and padding token already exist.
14
+ """
15
+ sentinels_to_add = [f'<extra_id_{i}>' for i in range(NUM_SENTINEL_TOKENS)]
16
+ tokenizer.add_tokens(sentinels_to_add, special_tokens=True)
17
+ if tokenizer.pad_token is None:
18
+ tokenizer.add_tokens('<pad>', special_tokens=True)
19
+ tokenizer.pad_token = '<pad>'
20
+ assert tokenizer.pad_token_id is not None
21
+ sentinels = ''.join([f'<extra_id_{i}>' for i in range(NUM_SENTINEL_TOKENS)])
22
+ _sentinel_token_ids = tokenizer(sentinels, add_special_tokens=False).input_ids
23
+ tokenizer.sentinel_token_ids = _sentinel_token_ids
24
+
25
+ class AutoTokenizerForMOD(AutoTokenizer):
26
+ """AutoTokenizer + Adaptation for MOD.
27
+
28
+ A simple wrapper around AutoTokenizer to make instantiating
29
+ an MOD-adapted tokenizer a bit easier.
30
+
31
+ MOD-adapted tokenizers have sentinel tokens (e.g., <extra_id_0>),
32
+ a padding token, and a property to get the token ids of the
33
+ sentinel tokens.
34
+ """
35
+
36
+ @classmethod
37
+ def from_pretrained(cls, *args, **kwargs):
38
+ """See `AutoTokenizer.from_pretrained` docstring."""
39
+ tokenizer = super().from_pretrained(*args, **kwargs)
40
+ adapt_tokenizer_for_denoising(tokenizer)
41
+ return tokenizer
attention.py ADDED
@@ -0,0 +1,276 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Attention layers."""
2
+ import math
3
+ import warnings
4
+ from typing import Optional
5
+ import torch
6
+ import torch.nn as nn
7
+ from einops import rearrange
8
+ from torch import nn
9
+ from .norm import LPLayerNorm
10
+
11
+ def _reset_is_causal(num_query_tokens: int, num_key_tokens: int, original_is_causal: bool):
12
+ if original_is_causal and num_query_tokens != num_key_tokens:
13
+ if num_query_tokens != 1:
14
+ raise NotImplementedError('MPT does not support query and key with different number of tokens, unless number of query tokens is 1.')
15
+ else:
16
+ return False
17
+ return original_is_causal
18
+
19
+ def scaled_multihead_dot_product_attention(query, key, value, n_heads, softmax_scale=None, attn_bias=None, key_padding_mask=None, is_causal=False, dropout_p=0.0, training=False, needs_weights=False, multiquery=False):
20
+ q = rearrange(query, 'b s (h d) -> b h s d', h=n_heads)
21
+ k = rearrange(key, 'b s (h d) -> b h d s', h=1 if multiquery else n_heads)
22
+ v = rearrange(value, 'b s (h d) -> b h s d', h=1 if multiquery else n_heads)
23
+ min_val = torch.finfo(q.dtype).min
24
+ (b, _, s_q, d) = q.shape
25
+ s_k = k.size(-1)
26
+ if softmax_scale is None:
27
+ softmax_scale = 1 / math.sqrt(d)
28
+ attn_weight = q.matmul(k) * softmax_scale
29
+ if attn_bias is not None:
30
+ if attn_bias.size(-1) != 1 and attn_bias.size(-1) != s_k or (attn_bias.size(-2) != 1 and attn_bias.size(-2) != s_q):
31
+ raise RuntimeError(f'attn_bias (shape: {attn_bias.shape}) is expected to broadcast to shape: {attn_weight.shape}.')
32
+ attn_weight = attn_weight + attn_bias
33
+ if key_padding_mask is not None:
34
+ if attn_bias is not None:
35
+ warnings.warn('Propogating key_padding_mask to the attention module ' + 'and applying it within the attention module can cause ' + 'unneccessary computation/memory usage. Consider integrating ' + 'into attn_bias once and passing that to each attention ' + 'module instead.')
36
+ attn_weight = attn_weight.masked_fill(~key_padding_mask.view((b, 1, 1, s_k)), min_val)
37
+ if is_causal:
38
+ s = max(s_q, s_k)
39
+ causal_mask = attn_weight.new_ones(s, s, dtype=torch.float16)
40
+ causal_mask = causal_mask.tril()
41
+ causal_mask = causal_mask.to(torch.bool)
42
+ causal_mask = ~causal_mask
43
+ causal_mask = causal_mask[-s_q:, -s_k:]
44
+ attn_weight = attn_weight.masked_fill(causal_mask.view(1, 1, s_q, s_k), min_val)
45
+ attn_weight = torch.softmax(attn_weight, dim=-1)
46
+ if dropout_p:
47
+ attn_weight = torch.nn.functional.dropout(attn_weight, p=dropout_p, training=training, inplace=True)
48
+ out = attn_weight.matmul(v)
49
+ out = rearrange(out, 'b h s d -> b s (h d)')
50
+ if needs_weights:
51
+ return (out, attn_weight)
52
+ return (out, None)
53
+
54
+ def check_valid_inputs(*tensors, valid_dtypes=[torch.float16, torch.bfloat16]):
55
+ for tensor in tensors:
56
+ if tensor.dtype not in valid_dtypes:
57
+ raise TypeError(f'tensor.dtype={tensor.dtype!r} must be in valid_dtypes={valid_dtypes!r}.')
58
+ if not tensor.is_cuda:
59
+ raise TypeError(f'Inputs must be cuda tensors (tensor.is_cuda={tensor.is_cuda!r}).')
60
+
61
+ def flash_attn_fn(query, key, value, n_heads, softmax_scale=None, attn_bias=None, key_padding_mask=None, is_causal=False, dropout_p=0.0, training=False, needs_weights=False, multiquery=False):
62
+ try:
63
+ from flash_attn import bert_padding, flash_attn_interface
64
+ except:
65
+ raise RuntimeError('Please install flash-attn==1.0.3.post0')
66
+ check_valid_inputs(query, key, value)
67
+ if attn_bias is not None:
68
+ raise NotImplementedError(f'attn_bias not implemented for flash attn.')
69
+ (batch_size, seqlen) = query.shape[:2]
70
+ if key_padding_mask is None:
71
+ key_padding_mask = torch.ones_like(key[:, :, 0], dtype=torch.bool)
72
+ query_padding_mask = key_padding_mask[:, -query.size(1):]
73
+ (query_unpad, indices_q, cu_seqlens_q, max_seqlen_q) = bert_padding.unpad_input(query, query_padding_mask)
74
+ query_unpad = rearrange(query_unpad, 'nnz (h d) -> nnz h d', h=n_heads)
75
+ (key_unpad, _, cu_seqlens_k, max_seqlen_k) = bert_padding.unpad_input(key, key_padding_mask)
76
+ key_unpad = rearrange(key_unpad, 'nnz (h d) -> nnz h d', h=1 if multiquery else n_heads)
77
+ (value_unpad, _, _, _) = bert_padding.unpad_input(value, key_padding_mask)
78
+ value_unpad = rearrange(value_unpad, 'nnz (h d) -> nnz h d', h=1 if multiquery else n_heads)
79
+ if multiquery:
80
+ key_unpad = key_unpad.expand(key_unpad.size(0), n_heads, key_unpad.size(-1))
81
+ value_unpad = value_unpad.expand(value_unpad.size(0), n_heads, value_unpad.size(-1))
82
+ dropout_p = dropout_p if training else 0.0
83
+ reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
84
+ output_unpad = flash_attn_interface.flash_attn_unpadded_func(query_unpad, key_unpad, value_unpad, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, dropout_p, softmax_scale=softmax_scale, causal=reset_is_causal, return_attn_probs=needs_weights)
85
+ output = bert_padding.pad_input(rearrange(output_unpad, 'nnz h d -> nnz (h d)'), indices_q, batch_size, seqlen)
86
+ return (output, None)
87
+
88
+ def triton_flash_attn_fn(query, key, value, n_heads, softmax_scale=None, attn_bias=None, key_padding_mask=None, is_causal=False, dropout_p=0.0, training=False, needs_weights=False, multiquery=False):
89
+ try:
90
+ from flash_attn import flash_attn_triton
91
+ except:
92
+ raise RuntimeError('Please install flash-attn==1.0.3.post0 and triton==2.0.0.dev20221202')
93
+ check_valid_inputs(query, key, value)
94
+ if dropout_p:
95
+ raise NotImplementedError(f'Dropout not implemented for attn_impl: triton.')
96
+ if needs_weights:
97
+ raise NotImplementedError(f'attn_impl: triton cannot return attn weights.')
98
+ if key_padding_mask is not None:
99
+ warnings.warn('Propagating key_padding_mask to the attention module ' + 'and applying it within the attention module can cause ' + 'unnecessary computation/memory usage. Consider integrating ' + 'into attn_bias once and passing that to each attention ' + 'module instead.')
100
+ (b_size, s_k) = key_padding_mask.shape[:2]
101
+ if attn_bias is None:
102
+ attn_bias = query.new_zeros(b_size, 1, 1, s_k)
103
+ attn_bias = attn_bias.masked_fill(~key_padding_mask.view((b_size, 1, 1, s_k)), torch.finfo(query.dtype).min)
104
+ query = rearrange(query, 'b s (h d) -> b s h d', h=n_heads)
105
+ key = rearrange(key, 'b s (h d) -> b s h d', h=1 if multiquery else n_heads)
106
+ value = rearrange(value, 'b s (h d) -> b s h d', h=1 if multiquery else n_heads)
107
+ if multiquery:
108
+ key = key.expand(*key.shape[:2], n_heads, key.size(-1))
109
+ value = value.expand(*value.shape[:2], n_heads, value.size(-1))
110
+ reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
111
+ attn_output = flash_attn_triton.flash_attn_func(query, key, value, attn_bias, reset_is_causal, softmax_scale)
112
+ output = attn_output.view(*attn_output.shape[:2], -1)
113
+ return (output, None)
114
+
115
+ class MultiheadAttention(nn.Module):
116
+ """Multi-head self attention.
117
+
118
+ Using torch or triton attention implemetation enables user to also use
119
+ additive bias.
120
+ """
121
+
122
+ def __init__(self, d_model: int, n_heads: int, attn_impl: str='triton', clip_qkv: Optional[float]=None, qk_ln: bool=False, softmax_scale: Optional[float]=None, attn_pdrop: float=0.0, low_precision_layernorm: bool=False, device: Optional[str]=None):
123
+ super().__init__()
124
+ self.attn_impl = attn_impl
125
+ self.clip_qkv = clip_qkv
126
+ self.qk_ln = qk_ln
127
+ self.d_model = d_model
128
+ self.n_heads = n_heads
129
+ self.softmax_scale = softmax_scale
130
+ if self.softmax_scale is None:
131
+ self.softmax_scale = 1 / math.sqrt(self.d_model / self.n_heads)
132
+ self.attn_dropout_p = attn_pdrop
133
+ self.Wqkv = nn.Linear(self.d_model, 3 * self.d_model, device=device)
134
+ fuse_splits = (d_model, 2 * d_model)
135
+ self.Wqkv._fused = (0, fuse_splits)
136
+ if self.qk_ln:
137
+ layernorm_class = LPLayerNorm if low_precision_layernorm else nn.LayerNorm
138
+ self.q_ln = layernorm_class(self.d_model, device=device)
139
+ self.k_ln = layernorm_class(self.d_model, device=device)
140
+ if self.attn_impl == 'flash':
141
+ self.attn_fn = flash_attn_fn
142
+ elif self.attn_impl == 'triton':
143
+ self.attn_fn = triton_flash_attn_fn
144
+ warnings.warn('While `attn_impl: triton` can be faster than `attn_impl: flash` ' + 'it uses more memory. When training larger models this can trigger ' + 'alloc retries which hurts performance. If encountered, we recommend ' + 'using `attn_impl: flash` if your model does not use `alibi` or `prefix_lm`.')
145
+ elif self.attn_impl == 'torch':
146
+ self.attn_fn = scaled_multihead_dot_product_attention
147
+ if torch.cuda.is_available():
148
+ warnings.warn('Using `attn_impl: torch`. If your model does not use `alibi` or ' + '`prefix_lm` we recommend using `attn_impl: flash` otherwise ' + 'we recommend using `attn_impl: triton`.')
149
+ else:
150
+ raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
151
+ self.out_proj = nn.Linear(self.d_model, self.d_model, device=device)
152
+ self.out_proj._is_residual = True
153
+
154
+ def forward(self, x, past_key_value=None, attn_bias=None, attention_mask=None, is_causal=True, needs_weights=False):
155
+ qkv = self.Wqkv(x)
156
+ if self.clip_qkv:
157
+ qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
158
+ (query, key, value) = qkv.chunk(3, dim=2)
159
+ key_padding_mask = attention_mask
160
+ if self.qk_ln:
161
+ dtype = query.dtype
162
+ query = self.q_ln(query).to(dtype)
163
+ key = self.k_ln(key).to(dtype)
164
+ if past_key_value is not None:
165
+ if len(past_key_value) != 0:
166
+ key = torch.cat([past_key_value[0], key], dim=1)
167
+ value = torch.cat([past_key_value[1], value], dim=1)
168
+ past_key_value = (key, value)
169
+ if attn_bias is not None:
170
+ attn_bias = attn_bias[:, :, -query.size(1):, -key.size(1):]
171
+ (context, attn_weights) = self.attn_fn(query, key, value, self.n_heads, softmax_scale=self.softmax_scale, attn_bias=attn_bias, key_padding_mask=key_padding_mask, is_causal=is_causal, dropout_p=self.attn_dropout_p, training=self.training, needs_weights=needs_weights)
172
+ return (self.out_proj(context), attn_weights, past_key_value)
173
+
174
+ class MultiQueryAttention(nn.Module):
175
+ """Multi-Query self attention.
176
+
177
+ Using torch or triton attention implemetation enables user to also use
178
+ additive bias.
179
+ """
180
+
181
+ def __init__(self, d_model: int, n_heads: int, attn_impl: str='triton', clip_qkv: Optional[float]=None, qk_ln: bool=False, softmax_scale: Optional[float]=None, attn_pdrop: float=0.0, low_precision_layernorm: bool=False, device: Optional[str]=None):
182
+ super().__init__()
183
+ self.attn_impl = attn_impl
184
+ self.clip_qkv = clip_qkv
185
+ self.qk_ln = qk_ln
186
+ self.d_model = d_model
187
+ self.n_heads = n_heads
188
+ self.head_dim = d_model // n_heads
189
+ self.softmax_scale = softmax_scale
190
+ if self.softmax_scale is None:
191
+ self.softmax_scale = 1 / math.sqrt(self.head_dim)
192
+ self.attn_dropout_p = attn_pdrop
193
+ self.Wqkv = nn.Linear(d_model, d_model + 2 * self.head_dim, device=device)
194
+ fuse_splits = (d_model, d_model + self.head_dim)
195
+ self.Wqkv._fused = (0, fuse_splits)
196
+ if self.qk_ln:
197
+ layernorm_class = LPLayerNorm if low_precision_layernorm else nn.LayerNorm
198
+ self.q_ln = layernorm_class(d_model, device=device)
199
+ self.k_ln = layernorm_class(self.head_dim, device=device)
200
+ if self.attn_impl == 'flash':
201
+ self.attn_fn = flash_attn_fn
202
+ elif self.attn_impl == 'triton':
203
+ self.attn_fn = triton_flash_attn_fn
204
+ warnings.warn('While `attn_impl: triton` can be faster than `attn_impl: flash` ' + 'it uses more memory. When training larger models this can trigger ' + 'alloc retries which hurts performance. If encountered, we recommend ' + 'using `attn_impl: flash` if your model does not use `alibi` or `prefix_lm`.')
205
+ elif self.attn_impl == 'torch':
206
+ self.attn_fn = scaled_multihead_dot_product_attention
207
+ if torch.cuda.is_available():
208
+ warnings.warn('Using `attn_impl: torch`. If your model does not use `alibi` or ' + '`prefix_lm` we recommend using `attn_impl: flash` otherwise ' + 'we recommend using `attn_impl: triton`.')
209
+ else:
210
+ raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
211
+ self.out_proj = nn.Linear(self.d_model, self.d_model, device=device)
212
+ self.out_proj._is_residual = True
213
+
214
+ def forward(self, x, past_key_value=None, attn_bias=None, attention_mask=None, is_causal=True, needs_weights=False):
215
+ qkv = self.Wqkv(x)
216
+ if self.clip_qkv:
217
+ qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
218
+ (query, key, value) = qkv.split([self.d_model, self.head_dim, self.head_dim], dim=2)
219
+ key_padding_mask = attention_mask
220
+ if self.qk_ln:
221
+ dtype = query.dtype
222
+ query = self.q_ln(query).to(dtype)
223
+ key = self.k_ln(key).to(dtype)
224
+ if past_key_value is not None:
225
+ if len(past_key_value) != 0:
226
+ key = torch.cat([past_key_value[0], key], dim=1)
227
+ value = torch.cat([past_key_value[1], value], dim=1)
228
+ past_key_value = (key, value)
229
+ if attn_bias is not None:
230
+ attn_bias = attn_bias[:, :, -query.size(1):, -key.size(1):]
231
+ (context, attn_weights) = self.attn_fn(query, key, value, self.n_heads, softmax_scale=self.softmax_scale, attn_bias=attn_bias, key_padding_mask=key_padding_mask, is_causal=is_causal, dropout_p=self.attn_dropout_p, training=self.training, needs_weights=needs_weights, multiquery=True)
232
+ return (self.out_proj(context), attn_weights, past_key_value)
233
+
234
+ def attn_bias_shape(attn_impl, n_heads, seq_len, alibi, prefix_lm, causal, use_sequence_id):
235
+ if attn_impl == 'flash':
236
+ return None
237
+ elif attn_impl in ['torch', 'triton']:
238
+ if alibi:
239
+ if (prefix_lm or not causal) or use_sequence_id:
240
+ return (1, n_heads, seq_len, seq_len)
241
+ return (1, n_heads, 1, seq_len)
242
+ elif prefix_lm or use_sequence_id:
243
+ return (1, 1, seq_len, seq_len)
244
+ return None
245
+ else:
246
+ raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
247
+
248
+ def build_attn_bias(attn_impl, attn_bias, n_heads, seq_len, causal=False, alibi=False, alibi_bias_max=8):
249
+ if attn_impl == 'flash':
250
+ return None
251
+ elif attn_impl in ['torch', 'triton']:
252
+ if alibi:
253
+ (device, dtype) = (attn_bias.device, attn_bias.dtype)
254
+ attn_bias = attn_bias.add(build_alibi_bias(n_heads, seq_len, full=not causal, alibi_bias_max=alibi_bias_max, device=device, dtype=dtype))
255
+ return attn_bias
256
+ else:
257
+ raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
258
+
259
+ def gen_slopes(n_heads, alibi_bias_max=8, device=None):
260
+ _n_heads = 2 ** math.ceil(math.log2(n_heads))
261
+ m = torch.arange(1, _n_heads + 1, dtype=torch.float32, device=device)
262
+ m = m.mul(alibi_bias_max / _n_heads)
263
+ slopes = 1.0 / torch.pow(2, m)
264
+ if _n_heads != n_heads:
265
+ slopes = torch.concat([slopes[1::2], slopes[::2]])[:n_heads]
266
+ return slopes.view(1, n_heads, 1, 1)
267
+
268
+ def build_alibi_bias(n_heads, seq_len, full=False, alibi_bias_max=8, device=None, dtype=None):
269
+ alibi_bias = torch.arange(1 - seq_len, 1, dtype=torch.int32, device=device).view(1, 1, 1, seq_len)
270
+ if full:
271
+ alibi_bias = alibi_bias - torch.arange(1 - seq_len, 1, dtype=torch.int32, device=device).view(1, 1, seq_len, 1)
272
+ alibi_bias = alibi_bias.abs().mul(-1)
273
+ slopes = gen_slopes(n_heads, alibi_bias_max, device=device)
274
+ alibi_bias = alibi_bias * slopes
275
+ return alibi_bias.to(dtype=dtype)
276
+ ATTN_CLASS_REGISTRY = {'multihead_attention': MultiheadAttention, 'multiquery_attention': MultiQueryAttention}
blocks.py ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GPT Blocks used for the GPT Model."""
2
+ from typing import Dict, Optional, Tuple
3
+ import torch
4
+ import torch.nn as nn
5
+ from .attention import ATTN_CLASS_REGISTRY
6
+ from .norm import NORM_CLASS_REGISTRY
7
+
8
+ class MPTMLP(nn.Module):
9
+
10
+ def __init__(self, d_model: int, expansion_ratio: int, device: Optional[str]=None):
11
+ super().__init__()
12
+ self.up_proj = nn.Linear(d_model, expansion_ratio * d_model, device=device)
13
+ self.act = nn.GELU(approximate='none')
14
+ self.down_proj = nn.Linear(expansion_ratio * d_model, d_model, device=device)
15
+ self.down_proj._is_residual = True
16
+
17
+ def forward(self, x):
18
+ return self.down_proj(self.act(self.up_proj(x)))
19
+
20
+ class MPTBlock(nn.Module):
21
+
22
+ def __init__(self, d_model: int, n_heads: int, expansion_ratio: int, attn_config: Dict={'attn_type': 'multihead_attention', 'attn_pdrop': 0.0, 'attn_impl': 'triton', 'qk_ln': False, 'clip_qkv': None, 'softmax_scale': None, 'prefix_lm': False, 'attn_uses_sequence_id': False, 'alibi': False, 'alibi_bias_max': 8}, resid_pdrop: float=0.0, norm_type: str='low_precision_layernorm', device: Optional[str]=None, **kwargs):
23
+ del kwargs
24
+ super().__init__()
25
+ norm_class = NORM_CLASS_REGISTRY[norm_type.lower()]
26
+ attn_class = ATTN_CLASS_REGISTRY[attn_config['attn_type']]
27
+ self.norm_1 = norm_class(d_model, device=device)
28
+ self.attn = attn_class(attn_impl=attn_config['attn_impl'], clip_qkv=attn_config['clip_qkv'], qk_ln=attn_config['qk_ln'], softmax_scale=attn_config['softmax_scale'], attn_pdrop=attn_config['attn_pdrop'], d_model=d_model, n_heads=n_heads, device=device)
29
+ self.norm_2 = norm_class(d_model, device=device)
30
+ self.ffn = MPTMLP(d_model=d_model, expansion_ratio=expansion_ratio, device=device)
31
+ self.resid_attn_dropout = nn.Dropout(resid_pdrop)
32
+ self.resid_ffn_dropout = nn.Dropout(resid_pdrop)
33
+
34
+ def forward(self, x: torch.Tensor, past_key_value: Optional[Tuple[torch.Tensor]]=None, attn_bias: Optional[torch.Tensor]=None, attention_mask: Optional[torch.ByteTensor]=None, is_causal: bool=True) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor]]]:
35
+ a = self.norm_1(x)
36
+ (b, _, past_key_value) = self.attn(a, past_key_value=past_key_value, attn_bias=attn_bias, attention_mask=attention_mask, is_causal=is_causal)
37
+ x = x + self.resid_attn_dropout(b)
38
+ m = self.norm_2(x)
39
+ n = self.ffn(m)
40
+ x = x + self.resid_ffn_dropout(n)
41
+ return (x, past_key_value)
hf_prefixlm_converter.py ADDED
@@ -0,0 +1,415 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Converts Huggingface Causal LM to Prefix LM.
2
+
3
+ Conversion does lightweight surgery on a HuggingFace
4
+ Causal LM to convert it to a Prefix LM.
5
+
6
+ Prefix LMs accepts a `bidirectional_mask` input in `forward`
7
+ and treat the input prompt as the prefix in `generate`.
8
+ """
9
+ import math
10
+ import warnings
11
+ from types import MethodType
12
+ from typing import Any, Dict, List, Optional, Tuple, Union
13
+ import torch
14
+ from transformers.models.bloom.modeling_bloom import BaseModelOutputWithPastAndCrossAttentions, BloomForCausalLM, BloomModel, CausalLMOutputWithCrossAttentions, CrossEntropyLoss
15
+ from transformers.models.bloom.modeling_bloom import _expand_mask as _expand_mask_bloom
16
+ from transformers.models.bloom.modeling_bloom import _make_causal_mask as _make_causal_mask_bloom
17
+ from transformers.models.bloom.modeling_bloom import logging
18
+ from transformers.models.gpt2.modeling_gpt2 import GPT2LMHeadModel
19
+ from transformers.models.gpt_neo.modeling_gpt_neo import GPTNeoForCausalLM
20
+ from transformers.models.gpt_neox.modeling_gpt_neox import GPTNeoXForCausalLM
21
+ from transformers.models.gptj.modeling_gptj import GPTJForCausalLM
22
+ from transformers.models.opt.modeling_opt import OPTForCausalLM
23
+ from transformers.models.opt.modeling_opt import _expand_mask as _expand_mask_opt
24
+ from transformers.models.opt.modeling_opt import _make_causal_mask as _make_causal_mask_opt
25
+ logger = logging.get_logger(__name__)
26
+ _SUPPORTED_GPT_MODELS = (GPT2LMHeadModel, GPTJForCausalLM, GPTNeoForCausalLM, GPTNeoXForCausalLM)
27
+ CAUSAL_GPT_TYPES = Union[GPT2LMHeadModel, GPTJForCausalLM, GPTNeoForCausalLM, GPTNeoXForCausalLM]
28
+
29
+ def _convert_gpt_causal_lm_to_prefix_lm(model: CAUSAL_GPT_TYPES) -> CAUSAL_GPT_TYPES:
30
+ """Converts a GPT-style Causal LM to a Prefix LM.
31
+
32
+ Supported HuggingFace model classes:
33
+ - `GPT2LMHeadModel`
34
+ - `GPTNeoForCausalLM`
35
+ - `GPTNeoXForCausalLM`
36
+ - `GPTJForCausalLM`
37
+
38
+ See `convert_hf_causal_lm_to_prefix_lm` for more details.
39
+ """
40
+ if hasattr(model, '_prefix_lm_converted'):
41
+ return model
42
+ assert isinstance(model, _SUPPORTED_GPT_MODELS)
43
+ assert model.config.add_cross_attention == False, 'Only supports GPT-style decoder-only models'
44
+
45
+ def _get_attn_modules(model: CAUSAL_GPT_TYPES) -> List[torch.nn.Module]:
46
+ """Helper that gets a list of the model's attention modules.
47
+
48
+ Each module has a `bias` buffer used for causal masking. The Prefix LM
49
+ conversion adds logic to dynamically manipulate these biases to support
50
+ Prefix LM attention masking.
51
+ """
52
+ attn_modules = []
53
+ if isinstance(model, GPTNeoXForCausalLM):
54
+ blocks = model.gpt_neox.layers
55
+ else:
56
+ blocks = model.transformer.h
57
+ for block in blocks:
58
+ if isinstance(model, GPTNeoForCausalLM):
59
+ if block.attn.attention_type != 'global':
60
+ continue
61
+ attn_module = block.attn.attention
62
+ elif isinstance(model, GPTNeoXForCausalLM):
63
+ attn_module = block.attention
64
+ else:
65
+ attn_module = block.attn
66
+ attn_modules.append(attn_module)
67
+ return attn_modules
68
+ setattr(model, '_original_forward', getattr(model, 'forward'))
69
+ setattr(model, '_original_generate', getattr(model, 'generate'))
70
+
71
+ def forward(self: CAUSAL_GPT_TYPES, input_ids: Optional[torch.LongTensor]=None, past_key_values: Optional[Tuple[Tuple[torch.Tensor]]]=None, attention_mask: Optional[torch.FloatTensor]=None, bidirectional_mask: Optional[torch.Tensor]=None, token_type_ids: Optional[torch.LongTensor]=None, position_ids: Optional[torch.LongTensor]=None, head_mask: Optional[torch.FloatTensor]=None, inputs_embeds: Optional[torch.FloatTensor]=None, labels: Optional[torch.LongTensor]=None, use_cache: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, return_dict: Optional[bool]=None):
72
+ """Wraps original forward to enable PrefixLM attention."""
73
+
74
+ def call_og_forward():
75
+ if isinstance(self, GPTNeoXForCausalLM):
76
+ return self._original_forward(input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, head_mask=head_mask, inputs_embeds=inputs_embeds, labels=labels, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict)
77
+ else:
78
+ return self._original_forward(input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, token_type_ids=token_type_ids, position_ids=position_ids, head_mask=head_mask, inputs_embeds=inputs_embeds, labels=labels, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict)
79
+ if bidirectional_mask is None:
80
+ return call_og_forward()
81
+ assert isinstance(bidirectional_mask, torch.Tensor)
82
+ attn_modules = _get_attn_modules(model)
83
+ (b, s) = bidirectional_mask.shape
84
+ max_length = attn_modules[0].bias.shape[-1]
85
+ if s > max_length:
86
+ raise ValueError(f'bidirectional_mask sequence length (={s}) exceeds the ' + f'max length allowed by the model ({max_length}).')
87
+ assert s <= max_length
88
+ if s < max_length:
89
+ pad = torch.zeros((int(b), int(max_length - s)), dtype=bidirectional_mask.dtype, device=bidirectional_mask.device)
90
+ bidirectional_mask = torch.cat([bidirectional_mask, pad], dim=1)
91
+ bidirectional = bidirectional_mask.unsqueeze(1).unsqueeze(1)
92
+ for attn_module in attn_modules:
93
+ attn_module.bias.data = torch.logical_or(attn_module.bias.data, bidirectional)
94
+ output = call_og_forward()
95
+ for attn_module in attn_modules:
96
+ attn_module.bias.data = torch.tril(attn_module.bias.data[0, 0])[None, None]
97
+ return output
98
+
99
+ def generate(self: CAUSAL_GPT_TYPES, *args: tuple, **kwargs: Dict[str, Any]):
100
+ """Wraps original generate to enable PrefixLM attention."""
101
+ attn_modules = _get_attn_modules(model)
102
+ for attn_module in attn_modules:
103
+ attn_module.bias.data[:] = 1
104
+ output = self._original_generate(*args, **kwargs)
105
+ for attn_module in attn_modules:
106
+ attn_module.bias.data = torch.tril(attn_module.bias.data[0, 0])[None, None]
107
+ return output
108
+ setattr(model, 'forward', MethodType(forward, model))
109
+ setattr(model, 'generate', MethodType(generate, model))
110
+ setattr(model, '_prefix_lm_converted', True)
111
+ return model
112
+
113
+ def _convert_bloom_causal_lm_to_prefix_lm(model: BloomForCausalLM) -> BloomForCausalLM:
114
+ """Converts a BLOOM Causal LM to a Prefix LM.
115
+
116
+ Supported HuggingFace model classes:
117
+ - `BloomForCausalLM`
118
+
119
+ See `convert_hf_causal_lm_to_prefix_lm` for more details.
120
+ """
121
+ if hasattr(model, '_prefix_lm_converted'):
122
+ return model
123
+ assert isinstance(model, BloomForCausalLM)
124
+ assert model.config.add_cross_attention == False, 'Only supports BLOOM decoder-only models'
125
+
126
+ def _prepare_attn_mask(self: BloomModel, attention_mask: torch.Tensor, bidirectional_mask: Optional[torch.Tensor], input_shape: Tuple[int, int], past_key_values_length: int) -> torch.BoolTensor:
127
+ combined_attention_mask = None
128
+ device = attention_mask.device
129
+ (_, src_length) = input_shape
130
+ if src_length > 1:
131
+ combined_attention_mask = _make_causal_mask_bloom(input_shape, device=device, past_key_values_length=past_key_values_length)
132
+ if bidirectional_mask is not None:
133
+ assert attention_mask.shape == bidirectional_mask.shape
134
+ expanded_bidirectional_mask = _expand_mask_bloom(bidirectional_mask, tgt_length=src_length)
135
+ combined_attention_mask = torch.logical_and(combined_attention_mask, expanded_bidirectional_mask)
136
+ expanded_attn_mask = _expand_mask_bloom(attention_mask, tgt_length=src_length)
137
+ combined_attention_mask = expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask | combined_attention_mask
138
+ return combined_attention_mask
139
+
140
+ def _build_alibi_tensor(self: BloomModel, batch_size: int, query_length: int, key_length: int, dtype: torch.dtype, device: torch.device) -> torch.Tensor:
141
+ num_heads = self.config.n_head
142
+ closest_power_of_2 = 2 ** math.floor(math.log2(num_heads))
143
+ base = torch.tensor(2 ** (-2 ** (-(math.log2(closest_power_of_2) - 3))), device=device, dtype=torch.float32)
144
+ powers = torch.arange(1, 1 + closest_power_of_2, device=device, dtype=torch.int32)
145
+ slopes = torch.pow(base, powers)
146
+ if closest_power_of_2 != num_heads:
147
+ extra_base = torch.tensor(2 ** (-2 ** (-(math.log2(2 * closest_power_of_2) - 3))), device=device, dtype=torch.float32)
148
+ num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2)
149
+ extra_powers = torch.arange(1, 1 + 2 * num_remaining_heads, 2, device=device, dtype=torch.int32)
150
+ slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)
151
+ qa = torch.arange(query_length, device=device, dtype=torch.int32).view(-1, 1)
152
+ ka = torch.arange(key_length, device=device, dtype=torch.int32).view(1, -1)
153
+ diffs = qa - ka + key_length - query_length
154
+ diffs = -diffs.abs()
155
+ alibi = slopes.view(1, num_heads, 1, 1) * diffs.view(1, 1, query_length, key_length)
156
+ alibi = alibi.expand(batch_size, -1, -1, -1).reshape(-1, query_length, key_length)
157
+ return alibi.to(dtype)
158
+ KeyValueT = Tuple[torch.Tensor, torch.Tensor]
159
+
160
+ def forward(self: BloomModel, input_ids: Optional[torch.LongTensor]=None, past_key_values: Optional[Tuple[KeyValueT, ...]]=None, attention_mask: Optional[torch.Tensor]=None, bidirectional_mask: Optional[torch.Tensor]=None, head_mask: Optional[torch.LongTensor]=None, inputs_embeds: Optional[torch.LongTensor]=None, use_cache: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, return_dict: Optional[bool]=None, **deprecated_arguments) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPastAndCrossAttentions]:
161
+ if deprecated_arguments.pop('position_ids', False) is not False:
162
+ warnings.warn('`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. ' + 'You can safely ignore passing `position_ids`.', FutureWarning)
163
+ if len(deprecated_arguments) > 0:
164
+ raise ValueError(f'Got unexpected arguments: {deprecated_arguments}')
165
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
166
+ output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
167
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
168
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
169
+ if input_ids is not None and inputs_embeds is not None:
170
+ raise ValueError('You cannot specify both input_ids and inputs_embeds at the same time')
171
+ elif input_ids is not None:
172
+ (batch_size, seq_length) = input_ids.shape
173
+ elif inputs_embeds is not None:
174
+ (batch_size, seq_length, _) = inputs_embeds.shape
175
+ else:
176
+ raise ValueError('You have to specify either input_ids or inputs_embeds')
177
+ if past_key_values is None:
178
+ past_key_values = tuple([None] * len(self.h))
179
+ head_mask = self.get_head_mask(head_mask, self.config.n_layer)
180
+ if inputs_embeds is None:
181
+ inputs_embeds = self.word_embeddings(input_ids)
182
+ hidden_states = self.word_embeddings_layernorm(inputs_embeds)
183
+ presents = () if use_cache else None
184
+ all_self_attentions = () if output_attentions else None
185
+ all_hidden_states = () if output_hidden_states else None
186
+ seq_length_with_past = seq_length
187
+ past_key_values_length = 0
188
+ if past_key_values[0] is not None:
189
+ tmp = past_key_values[0][0]
190
+ past_key_values_length = tmp.shape[2]
191
+ seq_length_with_past = seq_length_with_past + past_key_values_length
192
+ if attention_mask is None:
193
+ attention_mask = torch.ones((batch_size, seq_length_with_past), device=hidden_states.device)
194
+ else:
195
+ attention_mask = attention_mask.to(hidden_states.device)
196
+ alibi = self._build_alibi_tensor(batch_size=batch_size, query_length=seq_length, key_length=seq_length_with_past, dtype=hidden_states.dtype, device=hidden_states.device)
197
+ causal_mask = self._prepare_attn_mask(attention_mask, bidirectional_mask, input_shape=(batch_size, seq_length), past_key_values_length=past_key_values_length)
198
+ for (i, (block, layer_past)) in enumerate(zip(self.h, past_key_values)):
199
+ if output_hidden_states:
200
+ hst = (hidden_states,)
201
+ all_hidden_states = all_hidden_states + hst
202
+ if self.gradient_checkpointing and self.training:
203
+ if use_cache:
204
+ logger.warning('`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...')
205
+ use_cache = False
206
+
207
+ def create_custom_forward(module):
208
+
209
+ def custom_forward(*inputs):
210
+ return module(*inputs, use_cache=use_cache, output_attentions=output_attentions)
211
+ return custom_forward
212
+ outputs = torch.utils.checkpoint.checkpoint(create_custom_forward(block), hidden_states, alibi, causal_mask, head_mask[i])
213
+ else:
214
+ outputs = block(hidden_states, layer_past=layer_past, attention_mask=causal_mask, head_mask=head_mask[i], use_cache=use_cache, output_attentions=output_attentions, alibi=alibi)
215
+ hidden_states = outputs[0]
216
+ if use_cache is True:
217
+ presents = presents + (outputs[1],)
218
+ if output_attentions:
219
+ oa = (outputs[2 if use_cache else 1],)
220
+ all_self_attentions = all_self_attentions + oa
221
+ hidden_states = self.ln_f(hidden_states)
222
+ if output_hidden_states:
223
+ hst = (hidden_states,)
224
+ all_hidden_states = all_hidden_states + hst
225
+ if not return_dict:
226
+ return tuple((v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None))
227
+ return BaseModelOutputWithPastAndCrossAttentions(last_hidden_state=hidden_states, past_key_values=presents, hidden_states=all_hidden_states, attentions=all_self_attentions)
228
+ setattr(model.transformer, '_prepare_attn_mask', MethodType(_prepare_attn_mask, model.transformer))
229
+ setattr(model.transformer, '_build_alibi_tensor', MethodType(_build_alibi_tensor, model.transformer))
230
+ setattr(model.transformer, 'forward', MethodType(forward, model.transformer))
231
+ KeyValueT = Tuple[torch.Tensor, torch.Tensor]
232
+
233
+ def forward(self: BloomForCausalLM, input_ids: Optional[torch.LongTensor]=None, past_key_values: Optional[Tuple[KeyValueT, ...]]=None, attention_mask: Optional[torch.Tensor]=None, bidirectional_mask: Optional[torch.Tensor]=None, head_mask: Optional[torch.Tensor]=None, inputs_embeds: Optional[torch.Tensor]=None, labels: Optional[torch.Tensor]=None, use_cache: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, return_dict: Optional[bool]=None, **deprecated_arguments) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
234
+ """Replacement forward method for BloomCausalLM."""
235
+ if deprecated_arguments.pop('position_ids', False) is not False:
236
+ warnings.warn('`position_ids` have no functionality in BLOOM and will be removed ' + 'in v5.0.0. You can safely ignore passing `position_ids`.', FutureWarning)
237
+ if len(deprecated_arguments) > 0:
238
+ raise ValueError(f'Got unexpected arguments: {deprecated_arguments}')
239
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
240
+ transformer_outputs = self.transformer(input_ids, past_key_values=past_key_values, attention_mask=attention_mask, bidirectional_mask=bidirectional_mask, head_mask=head_mask, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict)
241
+ hidden_states = transformer_outputs[0]
242
+ lm_logits = self.lm_head(hidden_states)
243
+ loss = None
244
+ if labels is not None:
245
+ shift_logits = lm_logits[..., :-1, :].contiguous()
246
+ shift_labels = labels[..., 1:].contiguous()
247
+ (batch_size, seq_length, vocab_size) = shift_logits.shape
248
+ loss_fct = CrossEntropyLoss()
249
+ loss = loss_fct(shift_logits.view(batch_size * seq_length, vocab_size), shift_labels.view(batch_size * seq_length))
250
+ if not return_dict:
251
+ output = (lm_logits,) + transformer_outputs[1:]
252
+ return (loss,) + output if loss is not None else output
253
+ return CausalLMOutputWithCrossAttentions(loss=loss, logits=lm_logits, past_key_values=transformer_outputs.past_key_values, hidden_states=transformer_outputs.hidden_states, attentions=transformer_outputs.attentions)
254
+
255
+ def prepare_inputs_for_generation(self: BloomForCausalLM, input_ids: torch.LongTensor, past: Optional[torch.Tensor]=None, attention_mask: Optional[torch.Tensor]=None, **kwargs) -> dict:
256
+ if past:
257
+ input_ids = input_ids[:, -1].unsqueeze(-1)
258
+ bidirectional_mask = None
259
+ if past[0][0].shape[0] == input_ids.shape[0]:
260
+ past = self._convert_to_bloom_cache(past)
261
+ else:
262
+ bidirectional_mask = torch.ones_like(input_ids)
263
+ return {'input_ids': input_ids, 'past_key_values': past, 'use_cache': True, 'attention_mask': attention_mask, 'bidirectional_mask': bidirectional_mask}
264
+ setattr(model, 'forward', MethodType(forward, model))
265
+ setattr(model, 'prepare_inputs_for_generation', MethodType(prepare_inputs_for_generation, model))
266
+ setattr(model, '_prefix_lm_converted', True)
267
+ return model
268
+
269
+ def _convert_opt_causal_lm_to_prefix_lm(model: OPTForCausalLM) -> OPTForCausalLM:
270
+ """Converts an OPT Causal LM to a Prefix LM.
271
+
272
+ Supported HuggingFace model classes:
273
+ - `OPTForCausalLM`
274
+
275
+ See `convert_hf_causal_lm_to_prefix_lm` for more details.
276
+ """
277
+ if hasattr(model, '_prefix_lm_converted'):
278
+ return model
279
+ assert isinstance(model, OPTForCausalLM)
280
+ assert model.config.add_cross_attention == False, 'Only supports OPT decoder-only models'
281
+ setattr(model, '_original_forward', getattr(model, 'forward'))
282
+ setattr(model, '_original_generate', getattr(model, 'generate'))
283
+ model.model.decoder.bidirectional_mask = None
284
+
285
+ def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
286
+ combined_attention_mask = None
287
+ if input_shape[-1] > 1:
288
+ if self.bidirectional_mask == 'g':
289
+ (bsz, src_length) = input_shape
290
+ combined_attention_mask = torch.zeros((bsz, 1, src_length, src_length + past_key_values_length), dtype=inputs_embeds.dtype, device=inputs_embeds.device)
291
+ else:
292
+ combined_attention_mask = _make_causal_mask_opt(input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length).to(inputs_embeds.device)
293
+ if self.bidirectional_mask is not None:
294
+ assert attention_mask.shape == self.bidirectional_mask.shape
295
+ expanded_bidirectional_mask = _expand_mask_opt(self.bidirectional_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(inputs_embeds.device)
296
+ combined_attention_mask = torch.maximum(expanded_bidirectional_mask, combined_attention_mask)
297
+ if attention_mask is not None:
298
+ expanded_attn_mask = _expand_mask_opt(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(inputs_embeds.device)
299
+ combined_attention_mask = expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
300
+ return combined_attention_mask
301
+ setattr(model.model.decoder, '_prepare_decoder_attention_mask', MethodType(_prepare_decoder_attention_mask, model.model.decoder))
302
+
303
+ def forward(self: OPTForCausalLM, input_ids: Optional[torch.LongTensor]=None, attention_mask: Optional[torch.Tensor]=None, bidirectional_mask: Optional[torch.ByteTensor]=None, head_mask: Optional[torch.Tensor]=None, past_key_values: Optional[List[torch.FloatTensor]]=None, inputs_embeds: Optional[torch.FloatTensor]=None, labels: Optional[torch.LongTensor]=None, use_cache: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, return_dict: Optional[bool]=None):
304
+
305
+ def call_og_forward():
306
+ return self._original_forward(input_ids=input_ids, attention_mask=attention_mask, head_mask=head_mask, past_key_values=past_key_values, inputs_embeds=inputs_embeds, labels=labels, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict)
307
+ if bidirectional_mask is None:
308
+ return call_og_forward()
309
+ self.model.decoder.bidirectional_mask = bidirectional_mask
310
+ try:
311
+ outputs = call_og_forward()
312
+ except:
313
+ self.model.decoder.bidirectional_mask = None
314
+ raise
315
+ self.model.decoder.bidirectional_mask = None
316
+ return outputs
317
+
318
+ def generate(self: OPTForCausalLM, *args: tuple, **kwargs: Dict[str, Any]):
319
+ """Wraps original generate to enable PrefixLM-style attention."""
320
+ self.model.decoder.bidirectional_mask = 'g'
321
+ try:
322
+ output = self._original_generate(*args, **kwargs)
323
+ except:
324
+ self.model.decoder.bidirectional_mask = None
325
+ raise
326
+ self.model.decoder.bidirectional_mask = None
327
+ return output
328
+ setattr(model, 'forward', MethodType(forward, model))
329
+ setattr(model, 'generate', MethodType(generate, model))
330
+ setattr(model, '_prefix_lm_converted', True)
331
+ return model
332
+ _SUPPORTED_HF_MODELS = _SUPPORTED_GPT_MODELS + (BloomForCausalLM, OPTForCausalLM)
333
+ CAUSAL_LM_TYPES = Union[GPT2LMHeadModel, GPTJForCausalLM, GPTNeoForCausalLM, GPTNeoXForCausalLM, BloomForCausalLM, OPTForCausalLM]
334
+
335
+ def convert_hf_causal_lm_to_prefix_lm(model: CAUSAL_LM_TYPES) -> CAUSAL_LM_TYPES:
336
+ """Converts a HuggingFace Causal LM to a Prefix LM.
337
+
338
+ Supported HuggingFace model classes:
339
+ - `GPT2LMHeadModel`
340
+ - `GPTNeoForCausalLM`
341
+ - `GPTNeoXForCausalLM`
342
+ - `GPTJForCausalLM`
343
+ - `BloomForCausalLM`
344
+ - `OPTForCausalLM`
345
+
346
+ Conversion to a Prefix LM is done by modifying the `forward` method, and possibly also the
347
+ `generate` method and/or select underlying methods depending on the model class.
348
+
349
+ These changes preserve the model API, but add a new input to `forward`: "bidirectional_mask".
350
+
351
+ Notes on training:
352
+ To actually train the converted model as a Prefix LM, training batches will need to indicate
353
+ the prefix/target structure by including `bidirectional_mask` as part of the batch inputs.
354
+
355
+ **This is not a standard input and requires custom layers either within or after your dataloader.**
356
+
357
+ In addition to adding `bidirectional_mask` to the batch, this custom code should modify `labels`
358
+ such that `batch['labels'][batch['bidirectional_mask'] == 1] == -100`.
359
+ That is, the prefix portion of the sequence should not generate any loss. Loss should only be
360
+ generated by the target portion of the sequence.
361
+
362
+ Notes on `GPTNeoForCausalLM`:
363
+ To simplify the implementation, "global" and "local" attention layers are handled differently.
364
+ For "global" layers, we handle conversion as described above. For "local" layers, which use a
365
+ causal attention mask within a restricted local window, we do not alter the masking.
366
+
367
+ Notes on `forward` method conversion:
368
+ After conversion, the `forward` method will handle a new input, `bidirectional_mask`,
369
+ which should be a [batch_size, seq_length] byte tensor, where 1 indicates token positions
370
+ belonging to the prefix (prefix tokens can attend to one another bidirectionally), and
371
+ 0 indicates token positions belonging to the target.
372
+
373
+ The new `forward` method will incorporate `bidirectional_mask` (if supplied) into the existing
374
+ causal mask, call the original `forward` method, and (if the causal mask is a buffer) reset
375
+ the causal masks before returning the result.
376
+
377
+ Notes on `generate` method conversion:
378
+ After conversion, the `generate` method will have the same signature but will internally
379
+ convert all causal masks to be purely bidirectional, call the original `generate` method, and
380
+ (where appropriate) reset the causal masks before returning the result.
381
+
382
+ This works thanks to the logic of the HuggingFace `generate` API, which first encodes the token
383
+ "prompt" passed to `generate` (which is treated as the prefix) and then sequentially generates
384
+ each new token. Encodings are cached as generation happens, so all prefix tokens can attend to one
385
+ another (as expected in a Prefix LM) and generated tokens can only attend to prefix tokens and
386
+ previously-generated tokens (also as expected in a Prefix LM).
387
+
388
+ To preserve the API, the original methods are renamed to `_original_forward` and
389
+ `_original_generate`, and replaced with new `forward` and `generate` methods that wrap
390
+ them, respectively. Although implementation details vary by model class.
391
+ """
392
+ if isinstance(model, _SUPPORTED_GPT_MODELS):
393
+ return _convert_gpt_causal_lm_to_prefix_lm(model)
394
+ elif isinstance(model, BloomForCausalLM):
395
+ return _convert_bloom_causal_lm_to_prefix_lm(model)
396
+ elif isinstance(model, OPTForCausalLM):
397
+ return _convert_opt_causal_lm_to_prefix_lm(model)
398
+ else:
399
+ raise TypeError(f'Cannot convert model to Prefix LM. ' + f'Model does not belong to set of supported HF models:' + f'\n{_SUPPORTED_HF_MODELS}')
400
+
401
+ def add_bidirectional_mask_if_missing(batch: Dict[str, Any]):
402
+ """Attempts to add bidirectional_mask to batch if missing.
403
+
404
+ Raises:
405
+ KeyError if bidirectional_mask is missing and can't be inferred
406
+ """
407
+ if 'bidirectional_mask' not in batch:
408
+ if batch.get('mode', None) == 'icl_task':
409
+ batch['bidirectional_mask'] = batch['attention_mask'].clone()
410
+ for (i, continuation_indices) in enumerate(batch['continuation_indices']):
411
+ batch['bidirectional_mask'][i, continuation_indices] = 0
412
+ elif 'labels' in batch and 'attention_mask' in batch:
413
+ batch['bidirectional_mask'] = torch.logical_and(torch.eq(batch['attention_mask'], 1), torch.eq(batch['labels'], -100)).type_as(batch['attention_mask'])
414
+ else:
415
+ raise KeyError('No bidirectional_mask in batch and not sure how to construct one.')
meta_init_context.py ADDED
@@ -0,0 +1,94 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from contextlib import contextmanager
2
+ import torch
3
+ import torch.nn as nn
4
+
5
+ @contextmanager
6
+ def init_empty_weights(include_buffers: bool=False):
7
+ """Meta initialization context manager.
8
+
9
+ A context manager under which models are initialized with all parameters
10
+ on the meta device, therefore creating an empty model. Useful when just
11
+ initializing the model would blow the available RAM.
12
+
13
+ Args:
14
+ include_buffers (`bool`, *optional*, defaults to `False`): Whether or
15
+ not to also put all buffers on the meta device while initializing.
16
+
17
+ Example:
18
+ ```python
19
+ import torch.nn as nn
20
+
21
+ # Initialize a model with 100 billions parameters in no time and without using any RAM.
22
+ with init_empty_weights():
23
+ tst = nn.Sequential(*[nn.Linear(10000, 10000) for _ in range(1000)])
24
+ ```
25
+
26
+ <Tip warning={true}>
27
+
28
+ Any model created under this context manager has no weights. As such you can't do something like
29
+ `model.to(some_device)` with it. To load weights inside your empty model, see [`load_checkpoint_and_dispatch`].
30
+
31
+ </Tip>
32
+ """
33
+ with init_on_device(torch.device('meta'), include_buffers=include_buffers) as f:
34
+ yield f
35
+
36
+ @contextmanager
37
+ def init_on_device(device: torch.device, include_buffers: bool=False):
38
+ """Device initialization context manager.
39
+
40
+ A context manager under which models are initialized with all parameters
41
+ on the specified device.
42
+
43
+ Args:
44
+ device (`torch.device`): Device to initialize all parameters on.
45
+ include_buffers (`bool`, *optional*, defaults to `False`): Whether or
46
+ not to also put all buffers on the meta device while initializing.
47
+
48
+ Example:
49
+ ```python
50
+ import torch.nn as nn
51
+
52
+ with init_on_device(device=torch.device("cuda")):
53
+ tst = nn.Liner(100, 100) # on `cuda` device
54
+ ```
55
+ """
56
+ old_register_parameter = nn.Module.register_parameter
57
+ if include_buffers:
58
+ old_register_buffer = nn.Module.register_buffer
59
+
60
+ def register_empty_parameter(module, name, param):
61
+ old_register_parameter(module, name, param)
62
+ if param is not None:
63
+ param_cls = type(module._parameters[name])
64
+ kwargs = module._parameters[name].__dict__
65
+ module._parameters[name] = param_cls(module._parameters[name].to(device), **kwargs)
66
+
67
+ def register_empty_buffer(module, name, buffer):
68
+ old_register_buffer(module, name, buffer)
69
+ if buffer is not None:
70
+ module._buffers[name] = module._buffers[name].to(device)
71
+ if include_buffers:
72
+ tensor_constructors_to_patch = {torch_function_name: getattr(torch, torch_function_name) for torch_function_name in ['empty', 'zeros', 'ones', 'full']}
73
+ else:
74
+ tensor_constructors_to_patch = {}
75
+
76
+ def patch_tensor_constructor(fn):
77
+
78
+ def wrapper(*args, **kwargs):
79
+ kwargs['device'] = device
80
+ return fn(*args, **kwargs)
81
+ return wrapper
82
+ try:
83
+ nn.Module.register_parameter = register_empty_parameter
84
+ if include_buffers:
85
+ nn.Module.register_buffer = register_empty_buffer
86
+ for torch_function_name in tensor_constructors_to_patch.keys():
87
+ setattr(torch, torch_function_name, patch_tensor_constructor(getattr(torch, torch_function_name)))
88
+ yield
89
+ finally:
90
+ nn.Module.register_parameter = old_register_parameter
91
+ if include_buffers:
92
+ nn.Module.register_buffer = old_register_buffer
93
+ for (torch_function_name, old_torch_function) in tensor_constructors_to_patch.items():
94
+ setattr(torch, torch_function_name, old_torch_function)
modeling_mpt.py ADDED
@@ -0,0 +1,290 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A simple, flexible implementation of a GPT model.
2
+
3
+ Inspired by https://github.com/karpathy/minGPT/blob/master/mingpt/model.py
4
+ """
5
+ import math
6
+ import warnings
7
+ from typing import List, Optional, Tuple, Union
8
+ import torch
9
+ import torch.nn as nn
10
+ import torch.nn.functional as F
11
+ from transformers import PreTrainedModel, PreTrainedTokenizer, PreTrainedTokenizerFast
12
+ from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
13
+ from .attention import attn_bias_shape, build_attn_bias
14
+ from .blocks import MPTBlock
15
+ from .norm import NORM_CLASS_REGISTRY
16
+ from .configuration_mpt import MPTConfig
17
+ from .adapt_tokenizer import AutoTokenizerForMOD, adapt_tokenizer_for_denoising
18
+ from .hf_prefixlm_converter import add_bidirectional_mask_if_missing, convert_hf_causal_lm_to_prefix_lm
19
+ from .meta_init_context import init_empty_weights
20
+ from .param_init_fns import MODEL_INIT_REGISTRY, generic_param_init_fn_
21
+ Tokenizer = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
22
+
23
+ class MPTPreTrainedModel(PreTrainedModel):
24
+ config_class = MPTConfig
25
+ base_model_prefix = 'model'
26
+
27
+ class MPTModel(MPTPreTrainedModel):
28
+
29
+ def __init__(self, config: MPTConfig):
30
+ config._validate_config()
31
+ super().__init__(config)
32
+ self.attn_impl = config.attn_config['attn_impl']
33
+ self.prefix_lm = config.attn_config['prefix_lm']
34
+ self.attn_uses_sequence_id = config.attn_config['attn_uses_sequence_id']
35
+ self.alibi = config.attn_config['alibi']
36
+ self.alibi_bias_max = config.attn_config['alibi_bias_max']
37
+ if config.norm_type.lower() not in NORM_CLASS_REGISTRY.keys():
38
+ norm_options = ' | '.join(NORM_CLASS_REGISTRY.keys())
39
+ raise NotImplementedError(f'Requested norm type ({config.norm_type}) is not implemented within this repo (Options: {norm_options}).')
40
+ norm_class = NORM_CLASS_REGISTRY[config.norm_type.lower()]
41
+ self.embedding_fraction = config.embedding_fraction
42
+ self.wte = nn.Embedding(config.vocab_size, config.d_model, device=config.init_device)
43
+ if not self.alibi:
44
+ self.wpe = nn.Embedding(config.max_seq_len, config.d_model, device=config.init_device)
45
+ self.emb_drop = nn.Dropout(config.emb_pdrop)
46
+ self.blocks = nn.ModuleList([MPTBlock(device=config.init_device, **config.to_dict()) for _ in range(config.n_layers)])
47
+ self.norm_f = norm_class(config.d_model, device=config.init_device)
48
+ if config.init_device != 'meta':
49
+ self.apply(self.param_init_fn)
50
+ self.is_causal = not self.prefix_lm
51
+ self._attn_bias_initialized = False
52
+ self.attn_bias = None
53
+ self.attn_bias_shape = attn_bias_shape(self.attn_impl, config.n_heads, config.max_seq_len, self.alibi, prefix_lm=self.prefix_lm, causal=self.is_causal, use_sequence_id=self.attn_uses_sequence_id)
54
+ if config.no_bias:
55
+ for module in self.modules():
56
+ if hasattr(module, 'bias') and isinstance(module.bias, nn.Parameter):
57
+ if config.verbose:
58
+ warnings.warn(f'Removing bias ({module.bias}) from {module}.')
59
+ module.register_parameter('bias', None)
60
+ if config.verbose and config.verbose > 2:
61
+ print(self)
62
+ if 'verbose' not in self.config.init_config:
63
+ self.config.init_config['verbose'] = self.config.verbose
64
+ if self.config.init_config['verbose'] > 1:
65
+ init_fn_name = self.config.init_config['name']
66
+ warnings.warn(f'Using {init_fn_name} initialization.')
67
+
68
+ def get_input_embeddings(self):
69
+ return self.wte
70
+
71
+ def set_input_embeddings(self, value):
72
+ self.wte = value
73
+
74
+ @torch.no_grad()
75
+ def _attn_bias(self, device, dtype, attention_mask: Optional[torch.ByteTensor]=None, prefix_mask: Optional[torch.ByteTensor]=None, sequence_id: Optional[torch.LongTensor]=None):
76
+ if not self._attn_bias_initialized:
77
+ if self.attn_bias_shape:
78
+ self.attn_bias = torch.zeros(self.attn_bias_shape, device=device, dtype=dtype)
79
+ self.attn_bias = build_attn_bias(self.attn_impl, self.attn_bias, self.config.n_heads, self.config.max_seq_len, causal=self.is_causal, alibi=self.alibi, alibi_bias_max=self.alibi_bias_max)
80
+ self._attn_bias_initialized = True
81
+ if self.attn_impl == 'flash':
82
+ return (self.attn_bias, attention_mask)
83
+ if self.attn_bias is not None:
84
+ self.attn_bias = self.attn_bias.to(dtype=dtype, device=device)
85
+ attn_bias = self.attn_bias
86
+ if self.prefix_lm:
87
+ assert isinstance(attn_bias, torch.Tensor)
88
+ assert isinstance(prefix_mask, torch.Tensor)
89
+ attn_bias = self._apply_prefix_mask(attn_bias, prefix_mask)
90
+ if self.attn_uses_sequence_id and sequence_id is not None:
91
+ assert isinstance(attn_bias, torch.Tensor)
92
+ attn_bias = self._apply_sequence_id(attn_bias, sequence_id)
93
+ if attention_mask is not None:
94
+ s_k = attention_mask.shape[-1]
95
+ if attn_bias is None:
96
+ attn_bias = torch.zeros((1, 1, 1, s_k), device=device, dtype=dtype)
97
+ else:
98
+ attn_bias = attn_bias[:, :, :, -s_k:]
99
+ if prefix_mask is not None and attention_mask.shape != prefix_mask.shape:
100
+ raise ValueError(f'attention_mask shape={attention_mask.shape} ' + f'and prefix_mask shape={prefix_mask.shape} are not equal.')
101
+ min_val = torch.finfo(attn_bias.dtype).min
102
+ attn_bias = attn_bias.masked_fill(~attention_mask.view(-1, 1, 1, s_k), min_val)
103
+ return (attn_bias, None)
104
+
105
+ def _apply_prefix_mask(self, attn_bias: torch.Tensor, prefix_mask: torch.Tensor):
106
+ (s_k, s_q) = attn_bias.shape[-2:]
107
+ if s_k != self.config.max_seq_len or s_q != self.config.max_seq_len:
108
+ raise ValueError('attn_bias does not match the expected shape. ' + f'The last two dimensions should both be {self.config.max_length} ' + f'but are {s_k} and {s_q}.')
109
+ seq_len = prefix_mask.shape[-1]
110
+ if seq_len > self.config.max_seq_len:
111
+ raise ValueError(f'prefix_mask sequence length cannot exceed max_seq_len={self.config.max_seq_len}')
112
+ attn_bias = attn_bias[..., :seq_len, :seq_len]
113
+ causal = torch.tril(torch.ones((seq_len, seq_len), dtype=torch.bool, device=prefix_mask.device)).view(1, 1, seq_len, seq_len)
114
+ prefix = prefix_mask.view(-1, 1, 1, seq_len)
115
+ cannot_attend = ~torch.logical_or(causal, prefix.bool())
116
+ min_val = torch.finfo(attn_bias.dtype).min
117
+ attn_bias = attn_bias.masked_fill(cannot_attend, min_val)
118
+ return attn_bias
119
+
120
+ def _apply_sequence_id(self, attn_bias: torch.Tensor, sequence_id: torch.LongTensor):
121
+ seq_len = sequence_id.shape[-1]
122
+ if seq_len > self.config.max_seq_len:
123
+ raise ValueError(f'sequence_id sequence length cannot exceed max_seq_len={self.config.max_seq_len}')
124
+ attn_bias = attn_bias[..., :seq_len, :seq_len]
125
+ cannot_attend = torch.logical_not(torch.eq(sequence_id.view(-1, seq_len, 1), sequence_id.view(-1, 1, seq_len))).unsqueeze(1)
126
+ min_val = torch.finfo(attn_bias.dtype).min
127
+ attn_bias = attn_bias.masked_fill(cannot_attend, min_val)
128
+ return attn_bias
129
+
130
+ def forward(self, input_ids: torch.LongTensor, past_key_values: Optional[List[Tuple[torch.FloatTensor]]]=None, attention_mask: Optional[torch.ByteTensor]=None, prefix_mask: Optional[torch.ByteTensor]=None, sequence_id: Optional[torch.LongTensor]=None, return_dict: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, use_cache: Optional[bool]=None):
131
+ return_dict = return_dict if return_dict is not None else self.config.return_dict
132
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
133
+ if attention_mask is not None:
134
+ attention_mask = attention_mask.bool()
135
+ if prefix_mask is not None:
136
+ prefix_mask = prefix_mask.bool()
137
+ if not return_dict:
138
+ raise NotImplementedError('return_dict False is not implemented yet for MPT')
139
+ if output_attentions:
140
+ raise NotImplementedError('output_attentions is not implemented yet for MPT')
141
+ if attention_mask is not None and attention_mask[:, 0].sum() != attention_mask.shape[0] and self.training:
142
+ raise NotImplementedError('MPT does not support training with left padding.')
143
+ if self.prefix_lm and prefix_mask is None:
144
+ raise ValueError('prefix_mask is a required argument when MPT is configured with prefix_lm=True.')
145
+ if self.training:
146
+ if self.attn_uses_sequence_id and sequence_id is None:
147
+ raise ValueError('sequence_id is a required argument when MPT is configured with attn_uses_sequence_id=True ' + 'and the model is in train mode.')
148
+ elif self.attn_uses_sequence_id is False and sequence_id is not None:
149
+ warnings.warn('MPT received non-None input for `sequence_id` but is configured with attn_uses_sequence_id=False. ' + 'This input will be ignored. If you want the model to use `sequence_id`, set attn_uses_sequence_id to True.')
150
+ S = input_ids.size(1)
151
+ assert S <= self.config.max_seq_len, f'Cannot forward input with seq_len={S}, this model only supports seq_len<={self.config.max_seq_len}'
152
+ tok_emb = self.wte(input_ids)
153
+ if self.alibi:
154
+ x = tok_emb
155
+ else:
156
+ past_position = 0
157
+ if past_key_values is not None:
158
+ if len(past_key_values) != self.config.n_layers:
159
+ raise ValueError(f'past_key_values must provide a past_key_value for each attention ' + f'layer in the network (len(past_key_values)={len(past_key_values)!r}; self.config.n_layers={self.config.n_layers!r}).')
160
+ past_position = past_key_values[0][0].size(1)
161
+ if S + past_position > self.config.max_seq_len:
162
+ raise ValueError(f'Cannot forward input with past sequence length {past_position} and current sequence length {S + 1}, this model only supports total sequence length <= {self.config.max_seq_len}.')
163
+ pos = torch.arange(past_position, S + past_position, dtype=torch.long, device=input_ids.device).unsqueeze(0)
164
+ if attention_mask is not None:
165
+ pos = torch.clamp(pos - torch.cumsum((~attention_mask).to(torch.int32), dim=1)[:, past_position:], min=0)
166
+ pos_emb = self.wpe(pos)
167
+ x = tok_emb + pos_emb
168
+ if self.embedding_fraction == 1:
169
+ x = self.emb_drop(x)
170
+ else:
171
+ x_shrunk = x * self.embedding_fraction + x.detach() * (1 - self.embedding_fraction)
172
+ assert isinstance(self.emb_drop, nn.Module)
173
+ x = self.emb_drop(x_shrunk)
174
+ (attn_bias, attention_mask) = self._attn_bias(device=x.device, dtype=x.dtype, attention_mask=attention_mask, prefix_mask=prefix_mask, sequence_id=sequence_id)
175
+ if use_cache and past_key_values is None:
176
+ past_key_values = [() for _ in range(self.config.n_layers)]
177
+ all_hidden_states = () if output_hidden_states else None
178
+ for (b_idx, block) in enumerate(self.blocks):
179
+ if output_hidden_states:
180
+ assert all_hidden_states is not None
181
+ all_hidden_states = all_hidden_states + (x,)
182
+ past_key_value = past_key_values[b_idx] if past_key_values is not None else None
183
+ (x, past_key_value) = block(x, past_key_value=past_key_value, attn_bias=attn_bias, attention_mask=attention_mask, is_causal=self.is_causal)
184
+ if past_key_values is not None:
185
+ past_key_values[b_idx] = past_key_value
186
+ x = self.norm_f(x)
187
+ return BaseModelOutputWithPast(last_hidden_state=x, past_key_values=past_key_values, hidden_states=all_hidden_states)
188
+
189
+ def param_init_fn(self, module):
190
+ init_fn_name = self.config.init_config['name']
191
+ MODEL_INIT_REGISTRY[init_fn_name](module=module, n_layers=self.config.n_layers, d_model=self.config.d_model, **self.config.init_config)
192
+
193
+ def fsdp_wrap_fn(self, module):
194
+ return isinstance(module, MPTBlock)
195
+
196
+ def activation_checkpointing_fn(self, module):
197
+ return isinstance(module, MPTBlock)
198
+
199
+ class MPTForCausalLM(MPTPreTrainedModel):
200
+
201
+ def __init__(self, config: MPTConfig):
202
+ super().__init__(config)
203
+ if not config.tie_word_embeddings:
204
+ raise ValueError('MPTForCausalLM only supports tied word embeddings')
205
+ self.transformer = MPTModel(config)
206
+ self.logit_scale = None
207
+ if config.logit_scale is not None:
208
+ logit_scale = config.logit_scale
209
+ if isinstance(logit_scale, str):
210
+ if logit_scale == 'inv_sqrt_d_model':
211
+ logit_scale = 1 / math.sqrt(config.d_model)
212
+ else:
213
+ raise ValueError(f"logit_scale={logit_scale!r} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'.")
214
+ self.logit_scale = logit_scale
215
+
216
+ def get_input_embeddings(self):
217
+ return self.transformer.wte
218
+
219
+ def set_input_embeddings(self, value):
220
+ self.transformer.wte = value
221
+
222
+ def get_output_embeddings(self):
223
+ return self.transformer.wte
224
+
225
+ def set_output_embeddings(self, new_embeddings):
226
+ self.transformer.wte = new_embeddings
227
+
228
+ def set_decoder(self, decoder):
229
+ self.transformer = decoder
230
+
231
+ def get_decoder(self):
232
+ return self.transformer
233
+
234
+ def forward(self, input_ids: torch.LongTensor, past_key_values: Optional[List[Tuple[torch.FloatTensor]]]=None, attention_mask: Optional[torch.ByteTensor]=None, prefix_mask: Optional[torch.ByteTensor]=None, sequence_id: Optional[torch.LongTensor]=None, labels: Optional[torch.LongTensor]=None, return_dict: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, use_cache: Optional[bool]=None):
235
+ return_dict = return_dict if return_dict is not None else self.config.return_dict
236
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
237
+ outputs = self.transformer(input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, prefix_mask=prefix_mask, sequence_id=sequence_id, return_dict=return_dict, output_attentions=output_attentions, output_hidden_states=output_hidden_states, use_cache=use_cache)
238
+ logits = F.linear(outputs.last_hidden_state, self.transformer.wte.weight)
239
+ if self.logit_scale is not None:
240
+ if self.logit_scale == 0:
241
+ warnings.warn(f'Multiplying logits by self.logit_scale={self.logit_scale!r}. This will produce uniform (uninformative) outputs.')
242
+ logits *= self.logit_scale
243
+ loss = None
244
+ if labels is not None:
245
+ labels = torch.roll(labels, shifts=-1)
246
+ labels[:, -1] = -100
247
+ loss = F.cross_entropy(logits.view(-1, logits.size(-1)), labels.to(logits.device).view(-1))
248
+ return CausalLMOutputWithPast(loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states)
249
+
250
+ def param_init_fn(self, module):
251
+ init_fn_name = self.config.init_config['name']
252
+ MODEL_INIT_REGISTRY[init_fn_name](module=module, n_layers=self.config.n_layers, d_model=self.config.d_model, **self.config.init_config)
253
+
254
+ def fsdp_wrap_fn(self, module):
255
+ return isinstance(module, MPTBlock)
256
+
257
+ def activation_checkpointing_fn(self, module):
258
+ return isinstance(module, MPTBlock)
259
+
260
+ def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
261
+ if inputs_embeds is not None:
262
+ raise NotImplementedError('inputs_embeds is not implemented for MPT yet')
263
+ attention_mask = kwargs['attention_mask'].bool()
264
+ if attention_mask[:, -1].sum() != attention_mask.shape[0]:
265
+ raise NotImplementedError('MPT does not support generation with right padding.')
266
+ if self.transformer.attn_uses_sequence_id and self.training:
267
+ sequence_id = torch.zeros_like(input_ids[:1])
268
+ else:
269
+ sequence_id = None
270
+ if past_key_values is not None:
271
+ input_ids = input_ids[:, -1].unsqueeze(-1)
272
+ if self.transformer.prefix_lm:
273
+ prefix_mask = torch.ones_like(attention_mask)
274
+ if kwargs.get('use_cache') == False:
275
+ raise NotImplementedError('MPT with prefix_lm=True does not support use_cache=False.')
276
+ else:
277
+ prefix_mask = None
278
+ return {'input_ids': input_ids, 'attention_mask': attention_mask, 'prefix_mask': prefix_mask, 'sequence_id': sequence_id, 'past_key_values': past_key_values, 'use_cache': kwargs.get('use_cache', True)}
279
+
280
+ @staticmethod
281
+ def _reorder_cache(past_key_values, beam_idx):
282
+ """Used by HuggingFace generate when using beam search with kv-caching.
283
+
284
+ See https://github.com/huggingface/transformers/blob/3ec7a47664ebe40c40f4b722f6bb1cd30c3821ec/src/transformers/models/gpt2/modeling_gpt2.py#L1122-L1133
285
+ for an example in transformers.
286
+ """
287
+ reordered_past = []
288
+ for layer_past in past_key_values:
289
+ reordered_past += [tuple((past_state.index_select(0, beam_idx) for past_state in layer_past))]
290
+ return reordered_past
norm.py ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+
3
+ def _cast_if_autocast_enabled(tensor):
4
+ if torch.is_autocast_enabled():
5
+ if tensor.device.type == 'cuda':
6
+ dtype = torch.get_autocast_gpu_dtype()
7
+ elif tensor.device.type == 'cpu':
8
+ dtype = torch.get_autocast_cpu_dtype()
9
+ else:
10
+ raise NotImplementedError()
11
+ return tensor.to(dtype=dtype)
12
+ return tensor
13
+
14
+ class LPLayerNorm(torch.nn.LayerNorm):
15
+
16
+ def __init__(self, normalized_shape, eps=1e-05, elementwise_affine=True, device=None, dtype=None):
17
+ super().__init__(normalized_shape=normalized_shape, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
18
+
19
+ def forward(self, x):
20
+ module_device = x.device
21
+ downcast_x = _cast_if_autocast_enabled(x)
22
+ downcast_weight = _cast_if_autocast_enabled(self.weight) if self.weight is not None else self.weight
23
+ downcast_bias = _cast_if_autocast_enabled(self.bias) if self.bias is not None else self.bias
24
+ with torch.autocast(enabled=False, device_type=module_device.type):
25
+ return torch.nn.functional.layer_norm(downcast_x, self.normalized_shape, downcast_weight, downcast_bias, self.eps)
26
+
27
+ def rms_norm(x, weight=None, eps=1e-05):
28
+ output = x / torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + eps)
29
+ if weight is not None:
30
+ return output * weight
31
+ return output
32
+
33
+ class RMSNorm(torch.nn.Module):
34
+
35
+ def __init__(self, normalized_shape, eps=1e-05, weight=True, dtype=None, device=None):
36
+ super().__init__()
37
+ self.eps = eps
38
+ if weight:
39
+ self.weight = torch.nn.Parameter(torch.ones(normalized_shape, dtype=dtype, device=device))
40
+ else:
41
+ self.register_parameter('weight', None)
42
+
43
+ def forward(self, x):
44
+ return rms_norm(x.float(), self.weight, self.eps).to(dtype=x.dtype)
45
+
46
+ class LPRMSNorm(RMSNorm):
47
+
48
+ def __init__(self, normalized_shape, eps=1e-05, weight=True, dtype=None, device=None):
49
+ super().__init__(normalized_shape=normalized_shape, eps=eps, weight=weight, dtype=dtype, device=device)
50
+
51
+ def forward(self, x):
52
+ downcast_x = _cast_if_autocast_enabled(x)
53
+ downcast_weight = _cast_if_autocast_enabled(self.weight) if self.weight is not None else self.weight
54
+ with torch.autocast(enabled=False, device_type=x.device.type):
55
+ return rms_norm(downcast_x, downcast_weight, self.eps).to(dtype=x.dtype)
56
+ NORM_CLASS_REGISTRY = {'layernorm': torch.nn.LayerNorm, 'low_precision_layernorm': LPLayerNorm, 'rmsnorm': RMSNorm, 'low_precision_rmsnorm': LPRMSNorm}
param_init_fns.py ADDED
@@ -0,0 +1,181 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import warnings
3
+ from collections.abc import Sequence
4
+ from functools import partial
5
+ from typing import Optional, Tuple, Union
6
+ import torch
7
+ from torch import nn
8
+ from .norm import NORM_CLASS_REGISTRY
9
+
10
+ def torch_default_param_init_fn_(module: nn.Module, verbose: int=0, **kwargs):
11
+ del kwargs
12
+ if verbose > 1:
13
+ warnings.warn(f"Initializing network using module's reset_parameters attribute")
14
+ if hasattr(module, 'reset_parameters'):
15
+ module.reset_parameters()
16
+
17
+ def fused_init_helper_(module: nn.Module, init_fn_):
18
+ _fused = getattr(module, '_fused', None)
19
+ if _fused is None:
20
+ raise RuntimeError(f'Internal logic error')
21
+ (dim, splits) = _fused
22
+ splits = (0, *splits, module.weight.size(dim))
23
+ for (s, e) in zip(splits[:-1], splits[1:]):
24
+ slice_indices = [slice(None)] * module.weight.ndim
25
+ slice_indices[dim] = slice(s, e)
26
+ init_fn_(module.weight[slice_indices])
27
+
28
+ def generic_param_init_fn_(module: nn.Module, init_fn_, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, verbose: int=0, **kwargs):
29
+ del kwargs
30
+ if verbose > 1:
31
+ warnings.warn(f'If model has bias parameters they are initialized to 0.')
32
+ init_div_is_residual = init_div_is_residual
33
+ if init_div_is_residual is False:
34
+ div_is_residual = 1.0
35
+ elif init_div_is_residual is True:
36
+ div_is_residual = math.sqrt(2 * n_layers)
37
+ elif isinstance(init_div_is_residual, float) or isinstance(init_div_is_residual, int):
38
+ div_is_residual = init_div_is_residual
39
+ elif isinstance(init_div_is_residual, str) and init_div_is_residual.isnumeric():
40
+ div_is_residual = float(init_div_is_residual)
41
+ else:
42
+ div_is_residual = 1.0
43
+ raise ValueError(f'Expected init_div_is_residual to be boolean or numeric, got {init_div_is_residual}')
44
+ if init_div_is_residual is not False:
45
+ if verbose > 1:
46
+ warnings.warn(f'Initializing _is_residual layers then dividing them by {div_is_residual:.3f}. ' + f'Set `init_div_is_residual: false` in init config to disable this.')
47
+ if isinstance(module, nn.Linear):
48
+ if hasattr(module, '_fused'):
49
+ fused_init_helper_(module, init_fn_)
50
+ else:
51
+ init_fn_(module.weight)
52
+ if module.bias is not None:
53
+ torch.nn.init.zeros_(module.bias)
54
+ if init_div_is_residual is not False and getattr(module, '_is_residual', False):
55
+ with torch.no_grad():
56
+ module.weight.div_(div_is_residual)
57
+ elif isinstance(module, nn.Embedding):
58
+ if emb_init_std is not None:
59
+ std = emb_init_std
60
+ if std == 0:
61
+ warnings.warn(f'Embedding layer initialized to 0.')
62
+ emb_init_fn_ = partial(torch.nn.init.normal_, mean=0.0, std=std)
63
+ if verbose > 1:
64
+ warnings.warn(f'Embedding layer initialized using normal distribution with mean=0 and std={std!r}.')
65
+ elif emb_init_uniform_lim is not None:
66
+ lim = emb_init_uniform_lim
67
+ if isinstance(lim, Sequence):
68
+ if len(lim) > 2:
69
+ raise ValueError(f'Uniform init requires a min and a max limit. User input: {lim}.')
70
+ if lim[0] == lim[1]:
71
+ warnings.warn(f'Embedding layer initialized to {lim[0]}.')
72
+ else:
73
+ if lim == 0:
74
+ warnings.warn(f'Embedding layer initialized to 0.')
75
+ lim = [-lim, lim]
76
+ (a, b) = lim
77
+ emb_init_fn_ = partial(torch.nn.init.uniform_, a=a, b=b)
78
+ if verbose > 1:
79
+ warnings.warn(f'Embedding layer initialized using uniform distribution in range {lim}.')
80
+ else:
81
+ emb_init_fn_ = init_fn_
82
+ emb_init_fn_(module.weight)
83
+ elif isinstance(module, tuple(set(NORM_CLASS_REGISTRY.values()))):
84
+ if verbose > 1:
85
+ warnings.warn(f'Norm weights are set to 1. If norm layer has a bias it is initialized to 0.')
86
+ if hasattr(module, 'weight') and module.weight is not None:
87
+ torch.nn.init.ones_(module.weight)
88
+ if hasattr(module, 'bias') and module.bias is not None:
89
+ torch.nn.init.zeros_(module.bias)
90
+ elif isinstance(module, nn.MultiheadAttention):
91
+ if module._qkv_same_embed_dim:
92
+ assert module.in_proj_weight is not None
93
+ assert module.q_proj_weight is None and module.k_proj_weight is None and (module.v_proj_weight is None)
94
+ assert d_model is not None
95
+ _d = d_model
96
+ splits = (0, _d, 2 * _d, 3 * _d)
97
+ for (s, e) in zip(splits[:-1], splits[1:]):
98
+ init_fn_(module.in_proj_weight[s:e])
99
+ else:
100
+ assert module.q_proj_weight is not None and module.k_proj_weight is not None and (module.v_proj_weight is not None)
101
+ assert module.in_proj_weight is None
102
+ init_fn_(module.q_proj_weight)
103
+ init_fn_(module.k_proj_weight)
104
+ init_fn_(module.v_proj_weight)
105
+ if module.in_proj_bias is not None:
106
+ torch.nn.init.zeros_(module.in_proj_bias)
107
+ if module.bias_k is not None:
108
+ torch.nn.init.zeros_(module.bias_k)
109
+ if module.bias_v is not None:
110
+ torch.nn.init.zeros_(module.bias_v)
111
+ init_fn_(module.out_proj.weight)
112
+ if init_div_is_residual is not False and getattr(module.out_proj, '_is_residual', False):
113
+ with torch.no_grad():
114
+ module.out_proj.weight.div_(div_is_residual)
115
+ if module.out_proj.bias is not None:
116
+ torch.nn.init.zeros_(module.out_proj.bias)
117
+ else:
118
+ for _ in module.parameters(recurse=False):
119
+ raise NotImplementedError(f'{module.__class__.__name__} parameters are not initialized by param_init_fn.')
120
+
121
+ def _normal_init_(std, mean=0.0):
122
+ return partial(torch.nn.init.normal_, mean=mean, std=std)
123
+
124
+ def _normal_param_init_fn_(module: nn.Module, std: float, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, verbose: int=0, **kwargs):
125
+ del kwargs
126
+ init_fn_ = _normal_init_(std=std)
127
+ if verbose > 1:
128
+ warnings.warn(f'Using torch.nn.init.normal_ init fn mean=0.0, std={std}')
129
+ generic_param_init_fn_(module=module, init_fn_=init_fn_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
130
+
131
+ def baseline_param_init_fn_(module: nn.Module, init_std: float, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, verbose: int=0, **kwargs):
132
+ del kwargs
133
+ if init_std is None:
134
+ raise ValueError("You must set model.init_config['init_std'] to a float value to use the default initialization scheme.")
135
+ _normal_param_init_fn_(module=module, std=init_std, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
136
+
137
+ def small_param_init_fn_(module: nn.Module, n_layers: int, d_model: int, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, verbose: int=0, **kwargs):
138
+ del kwargs
139
+ std = math.sqrt(2 / (5 * d_model))
140
+ _normal_param_init_fn_(module=module, std=std, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
141
+
142
+ def neox_param_init_fn_(module: nn.Module, n_layers: int, d_model: int, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, verbose: int=0, **kwargs):
143
+ """From section 2.3.1 of GPT-NeoX-20B:
144
+
145
+ An Open-Source AutoregressiveLanguage Model — Black et. al. (2022)
146
+ see https://github.com/EleutherAI/gpt-neox/blob/9610391ab319403cef079b438edd016a2443af54/megatron/model/init_functions.py#L151
147
+ and https://github.com/EleutherAI/gpt-neox/blob/main/megatron/model/transformer.py
148
+ """
149
+ del kwargs
150
+ residual_div = n_layers / math.sqrt(10)
151
+ if verbose > 1:
152
+ warnings.warn(f'setting init_div_is_residual to {residual_div}')
153
+ small_param_init_fn_(module=module, d_model=d_model, n_layers=n_layers, init_div_is_residual=residual_div, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
154
+
155
+ def kaiming_uniform_param_init_fn_(module: nn.Module, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, init_gain: float=0, fan_mode: str='fan_in', init_nonlinearity: str='leaky_relu', verbose: int=0, **kwargs):
156
+ del kwargs
157
+ if verbose > 1:
158
+ warnings.warn(f'Using nn.init.kaiming_uniform_ init fn with parameters: ' + f'a={init_gain}, mode={fan_mode}, nonlinearity={init_nonlinearity}')
159
+ kaiming_uniform_ = partial(nn.init.kaiming_uniform_, a=init_gain, mode=fan_mode, nonlinearity=init_nonlinearity)
160
+ generic_param_init_fn_(module=module, init_fn_=kaiming_uniform_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
161
+
162
+ def kaiming_normal_param_init_fn_(module: nn.Module, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, init_gain: float=0, fan_mode: str='fan_in', init_nonlinearity: str='leaky_relu', verbose: int=0, **kwargs):
163
+ del kwargs
164
+ if verbose > 1:
165
+ warnings.warn(f'Using nn.init.kaiming_normal_ init fn with parameters: ' + f'a={init_gain}, mode={fan_mode}, nonlinearity={init_nonlinearity}')
166
+ kaiming_normal_ = partial(torch.nn.init.kaiming_normal_, a=init_gain, mode=fan_mode, nonlinearity=init_nonlinearity)
167
+ generic_param_init_fn_(module=module, init_fn_=kaiming_normal_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
168
+
169
+ def xavier_uniform_param_init_fn_(module: nn.Module, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, init_gain: float=0, verbose: int=0, **kwargs):
170
+ del kwargs
171
+ xavier_uniform_ = partial(torch.nn.init.xavier_uniform_, gain=init_gain)
172
+ if verbose > 1:
173
+ warnings.warn(f'Using torch.nn.init.xavier_uniform_ init fn with parameters: ' + f'gain={init_gain}')
174
+ generic_param_init_fn_(module=module, init_fn_=xavier_uniform_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
175
+
176
+ def xavier_normal_param_init_fn_(module: nn.Module, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, init_gain: float=0, verbose: int=0, **kwargs):
177
+ xavier_normal_ = partial(torch.nn.init.xavier_normal_, gain=init_gain)
178
+ if verbose > 1:
179
+ warnings.warn(f'Using torch.nn.init.xavier_normal_ init fn with parameters: ' + f'gain={init_gain}')
180
+ generic_param_init_fn_(module=module, init_fn_=xavier_normal_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
181
+ MODEL_INIT_REGISTRY = {'default_': torch_default_param_init_fn_, 'baseline_': baseline_param_init_fn_, 'kaiming_uniform_': kaiming_uniform_param_init_fn_, 'kaiming_normal_': kaiming_normal_param_init_fn_, 'neox_init_': neox_param_init_fn_, 'small_init_': small_param_init_fn_, 'xavier_uniform_': xavier_uniform_param_init_fn_, 'xavier_normal_': xavier_normal_param_init_fn_}