Text Generation
Transformers
PyTorch
mosaic_gpt
custom_code
anas-awadalla commited on
Commit
0b68fcb
1 Parent(s): fe9ffd9
.DS_Store ADDED
Binary file (6.15 kB). View file
 
README.md ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ datasets:
4
+ - togethercomputer/RedPajama-Data-1T
5
+ ---
6
+
7
+ # MPT-1b-RedPajama-200b
8
+
9
+ MPT-1b-RedPajama-200b is a 1.3 billion parameter decoder-only transformer trained on the [RedPajama dataset](https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T).
10
+ The model was trained for 200B tokens by sampling from the subsets of the RedPajama dataset in the same proportions as were used by the [Llama series of models](https://arxiv.org/abs/2302.13971).
11
+ This model was trained by [MosaicML](https://www.mosaicml.com) and follows a modified decoder-only transformer architecture.
12
+
13
+ ## Model Date
14
+
15
+ April 20, 2023
16
+
17
+ ## How to Use
18
+
19
+ Note: This model requires that `trust_remote_code=True` be passed to the `from_pretrained` method.
20
+ This is because we use a custom model architecture `MosaicGPT` that is not yet part of the `transformers` package.
21
+ `MosaicGPT` includes options for many training efficiency features such as [FlashAttention (Dao et al. 2022)](https://arxiv.org/pdf/2205.14135.pdf), [ALIBI](https://arxiv.org/abs/2108.12409), QK LayerNorm, and more.
22
+
23
+ ```python
24
+ import transformers
25
+ model = transformers.AutoModelForCausalLM.from_pretrained('mosaicml/mpt-1b-redpajama-200b', trust_remote_code=True)
26
+ ```
27
+
28
+ To use the optimized triton implementation of FlashAttention, you can load with `attn_impl='triton'` and move the model to `bfloat16` like so:
29
+ ```python
30
+ model = transformers.AutoModelForCausalLM.from_pretrained('mosaicml/mpt-1b-redpajama-200b', trust_remote_code=True, attn_impl='triton')
31
+ model.to(device='cuda:0', dtype=torch.bfloat16)
32
+ ```
33
+
34
+ ## Model Description
35
+
36
+ This model uses the MosaicML LLM codebase, which can be found in the [MosaicML Examples Repository](https://github.com/mosaicml/examples/tree/v0.0.4/examples/llm).
37
+ The architecture is a modification of a standard decoder-only transformer.
38
+ The transformer has 24 layers, 16 attention heads, and width 2048.
39
+ The model has been modified from a standard transformer in the following ways:
40
+ * It uses ALiBi and does not use positional embeddings.
41
+ * It uses QK LayerNorm.
42
+ * It does not use biases.
43
+
44
+ ## Training Data
45
+
46
+ The model was trained for 200B tokens (batch size 2200, sequence length 2048). It was trained on the following data mix:
47
+ * 67% RedPajama Common Crawl
48
+ * 15% [C4](https://huggingface.co/datasets/c4)
49
+ * 4.5% RedPajama GitHub
50
+ * 4.5% RedPajama Wikipedia
51
+ * 4.5% RedPajama Books
52
+ * 2.5% RedPajama Arxiv
53
+ * 2% RedPajama StackExchange
54
+
55
+ This is the same mix of data as was used in the Llama series of models](https://arxiv.org/abs/2302.13971).
56
+
57
+ Each sample was chosen from one of the datasets, with the dataset selected with the probability specified above.
58
+ The examples were shuffled within each dataset.
59
+ Each example was constructed from as many sequences from that dataset as were necessary to fill the 2048 sequence length.
60
+
61
+ The data was tokenized using the [EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b) tokenizer.
62
+
63
+ ## Training Configuration
64
+
65
+ This model was trained on 440 A100-40GBs for about half a day using the [MosaicML Platform](https://www.mosaicml.com/platform). The model was trained with sharded data parallelism using FSDP.
66
+
67
+ ## Acknowledgements
68
+
69
+ This model builds on the work of [Together](https://www.together.xyz), which created the RedPajama dataset with the goal of mimicking the training data used to create the Llama series of models.
70
+ We gratefully acknowledge the hard work of the team that put together this dataset, and we hope this model serves as a useful companion to that work.
71
+
72
+ We also gratefully acknowledge the work of the researchers who created the Llama series of models, which was the impetus for our efforts and those who worked on the RedPajama project.
attention.py ADDED
@@ -0,0 +1,408 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2022 MosaicML Examples authors
2
+ # SPDX-License-Identifier: Apache-2.0
3
+
4
+ """Attention layers."""
5
+
6
+ import math
7
+ import warnings
8
+ from typing import Optional
9
+
10
+ import torch
11
+ import torch.nn as nn
12
+ from einops import rearrange
13
+ from torch import nn
14
+
15
+ from .low_precision_layernorm import LPLayerNorm
16
+
17
+
18
+ def _reset_is_causal(num_query_tokens: int, num_key_tokens: int,
19
+ original_is_causal: bool):
20
+ if original_is_causal and num_query_tokens != num_key_tokens:
21
+ if num_query_tokens != 1:
22
+ raise NotImplementedError(
23
+ 'MosaicGPT does not support query and key with different number of tokens, unless number of query tokens is 1.'
24
+ )
25
+ else:
26
+ return False
27
+ return original_is_causal
28
+
29
+
30
+ def scaled_multihead_dot_product_attention(
31
+ query,
32
+ key,
33
+ value,
34
+ n_heads,
35
+ softmax_scale=None,
36
+ attn_bias=None,
37
+ key_padding_mask=None,
38
+ is_causal=False,
39
+ dropout_p=0.0,
40
+ training=False,
41
+ needs_weights=False,
42
+ ):
43
+
44
+ q = rearrange(query, 'b s (h d) -> b h s d', h=n_heads)
45
+ k = rearrange(key, 'b s (h d) -> b h d s', h=n_heads) # includes key.t()
46
+ v = rearrange(value, 'b s (h d) -> b h s d', h=n_heads)
47
+
48
+ min_val = torch.finfo(q.dtype).min
49
+
50
+ b, _, s_q, d = q.shape
51
+ s_k = k.size(-1)
52
+
53
+ if softmax_scale is None:
54
+ softmax_scale = 1 / math.sqrt(d)
55
+
56
+ attn_weight = q.matmul(k) * softmax_scale
57
+
58
+ if attn_bias is not None:
59
+ if (attn_bias.size(-1) != 1 and
60
+ attn_bias.size(-1) != s_k) or (attn_bias.size(-2) != 1 and
61
+ attn_bias.size(-2) != s_q):
62
+ raise RuntimeError(
63
+ f'attn_bias (shape: {attn_bias.shape}) is expected to broadcast to shape: {attn_weight.shape}.'
64
+ )
65
+ attn_weight = attn_weight + attn_bias
66
+
67
+ if key_padding_mask is not None:
68
+ if attn_bias is not None:
69
+ warnings.warn(
70
+ 'Propogating key_padding_mask to the attention module ' +\
71
+ 'and applying it within the attention module can cause ' +\
72
+ 'unneccessary computation/memory usage. Consider integrating ' +\
73
+ 'into attn_bias once and passing that to each attention ' +\
74
+ 'module instead.'
75
+ )
76
+ attn_weight = attn_weight.masked_fill(
77
+ ~key_padding_mask.view((b, 1, 1, s_k)), min_val)
78
+
79
+ if is_causal:
80
+ s = max(s_q, s_k)
81
+ causal_mask = attn_weight.new_ones(s, s, dtype=torch.float16)
82
+ causal_mask = causal_mask.tril()
83
+ causal_mask = causal_mask.to(torch.bool)
84
+ causal_mask = ~causal_mask
85
+ causal_mask = causal_mask[-s_q:, -s_k:]
86
+ attn_weight = attn_weight.masked_fill(causal_mask.view(1, 1, s_q, s_k),
87
+ min_val)
88
+
89
+ attn_weight = torch.softmax(attn_weight, dim=-1)
90
+
91
+ if dropout_p:
92
+ attn_weight = torch.nn.functional.dropout(attn_weight,
93
+ p=dropout_p,
94
+ training=training,
95
+ inplace=True)
96
+
97
+ out = attn_weight.matmul(v)
98
+ out = rearrange(out, 'b h s d -> b s (h d)')
99
+
100
+ if needs_weights:
101
+ return out, attn_weight
102
+ return out, None
103
+
104
+
105
+ def check_valid_inputs(*tensors, valid_dtypes=[torch.float16, torch.bfloat16]):
106
+ for tensor in tensors:
107
+ if tensor.dtype not in valid_dtypes:
108
+ raise TypeError(f'{tensor.dtype=} must be in {valid_dtypes=}.')
109
+ if not tensor.is_cuda:
110
+ raise TypeError(f'Inputs must be cuda tensors ({tensor.is_cuda=}).')
111
+
112
+
113
+ def flash_attn_fn(
114
+ query,
115
+ key,
116
+ value,
117
+ n_heads,
118
+ softmax_scale=None,
119
+ attn_bias=None,
120
+ key_padding_mask=None,
121
+ is_causal=False,
122
+ dropout_p=0.0,
123
+ training=False,
124
+ needs_weights=False,
125
+ ):
126
+ try:
127
+ from flash_attn import bert_padding, flash_attn_interface
128
+ except:
129
+ raise RuntimeError('Please install flash_attn==0.2.8')
130
+
131
+ check_valid_inputs(query, key, value)
132
+
133
+ if attn_bias is not None:
134
+ raise NotImplementedError(f'attn_bias not implemented for flash attn.')
135
+
136
+ batch_size, seqlen = query.shape[:2]
137
+
138
+ if key_padding_mask is None:
139
+ key_padding_mask = torch.ones_like(key[:, :, 0], dtype=torch.bool)
140
+ query_padding_mask = key_padding_mask[:, -query.size(1):]
141
+
142
+ query_unpad, indices_q, cu_seqlens_q, max_seqlen_q = bert_padding.unpad_input(
143
+ query, query_padding_mask)
144
+ query_unpad = rearrange(query_unpad, 'nnz (h d) -> nnz h d', h=n_heads)
145
+
146
+ key_unpad, _, cu_seqlens_k, max_seqlen_k = bert_padding.unpad_input(
147
+ key, key_padding_mask)
148
+ key_unpad = rearrange(key_unpad, 'nnz (h d) -> nnz h d', h=n_heads)
149
+
150
+ value_unpad, _, _, _ = bert_padding.unpad_input(value, key_padding_mask)
151
+ value_unpad = rearrange(value_unpad, 'nnz (h d) -> nnz h d', h=n_heads)
152
+
153
+ dropout_p = dropout_p if training else 0.0
154
+
155
+ reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
156
+
157
+ output_unpad = flash_attn_interface.flash_attn_unpadded_func(
158
+ query_unpad,
159
+ key_unpad,
160
+ value_unpad,
161
+ cu_seqlens_q,
162
+ cu_seqlens_k,
163
+ max_seqlen_q,
164
+ max_seqlen_k,
165
+ dropout_p,
166
+ softmax_scale=softmax_scale,
167
+ causal=reset_is_causal,
168
+ return_attn_probs=needs_weights)
169
+
170
+ output = bert_padding.pad_input(
171
+ rearrange(output_unpad, 'nnz h d -> nnz (h d)'), indices_q, batch_size,
172
+ seqlen)
173
+ return output, None
174
+
175
+
176
+ def triton_flash_attn_fn(
177
+ query,
178
+ key,
179
+ value,
180
+ n_heads,
181
+ softmax_scale=None,
182
+ attn_bias=None,
183
+ key_padding_mask=None,
184
+ is_causal=False,
185
+ dropout_p=0.0,
186
+ training=False,
187
+ needs_weights=False,
188
+ ):
189
+ try:
190
+ from flash_attn import flash_attn_triton # type: ignore
191
+ except:
192
+ raise RuntimeError('Please install flash_attn==0.2.8 and triton==2.0.0.dev20221202.')
193
+
194
+ check_valid_inputs(query, key, value)
195
+
196
+ if dropout_p:
197
+ raise NotImplementedError(
198
+ f'Dropout not implemented for attn_impl: triton.')
199
+
200
+ if needs_weights:
201
+ raise NotImplementedError(
202
+ f'attn_impl: triton cannot return attn weights.')
203
+
204
+ if key_padding_mask is not None:
205
+ warnings.warn(
206
+ 'Propagating key_padding_mask to the attention module ' +\
207
+ 'and applying it within the attention module can cause ' +\
208
+ 'unnecessary computation/memory usage. Consider integrating ' +\
209
+ 'into attn_bias once and passing that to each attention ' +\
210
+ 'module instead.'
211
+ )
212
+ b_size, s_k = key_padding_mask.shape[:2]
213
+
214
+ if attn_bias is None:
215
+ attn_bias = query.new_zeros(b_size, 1, 1, s_k)
216
+
217
+ attn_bias = attn_bias.masked_fill(
218
+ ~key_padding_mask.view((b_size, 1, 1, s_k)),
219
+ torch.finfo(query.dtype).min)
220
+
221
+ query = rearrange(query, 'b s (h d) -> b s h d', h=n_heads)
222
+ key = rearrange(key, 'b s (h d) -> b s h d', h=n_heads)
223
+ value = rearrange(value, 'b s (h d) -> b s h d', h=n_heads)
224
+
225
+ reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
226
+ attn_output = flash_attn_triton.flash_attn_func(query, key, value,
227
+ attn_bias, reset_is_causal,
228
+ softmax_scale)
229
+
230
+ output = attn_output.view(*attn_output.shape[:2], -1)
231
+
232
+ return output, None
233
+
234
+
235
+ class MultiheadAttention(nn.Module):
236
+ """Multi-head self attention.
237
+
238
+ Using torch or triton attention implemetation enables user to also use
239
+ additive bias.
240
+ """
241
+
242
+ def __init__(
243
+ self,
244
+ d_model: int,
245
+ n_heads: int,
246
+ attn_impl: str = 'triton',
247
+ attn_clip_qkv: Optional[float] = None,
248
+ attn_qk_ln: bool = False,
249
+ softmax_scale: Optional[float] = None,
250
+ attn_pdrop: float = 0.0,
251
+ low_precision_layernorm: bool = False,
252
+ device: Optional[str] = None,
253
+ ):
254
+ super().__init__()
255
+
256
+ self.attn_impl = attn_impl
257
+ self.clip_qkv = attn_clip_qkv
258
+ self.attn_qk_ln = attn_qk_ln
259
+
260
+ self.d_model = d_model
261
+ self.n_heads = n_heads
262
+ self.softmax_scale = softmax_scale
263
+ if self.softmax_scale is None:
264
+ self.softmax_scale = 1 / math.sqrt(self.d_model / self.n_heads)
265
+ self.attn_dropout_p = attn_pdrop
266
+
267
+ self.Wqkv = nn.Linear(self.d_model, 3 * self.d_model, device=device)
268
+ # for param init fn; enables shape based init of fused layers
269
+ fuse_splits = (d_model, 2 * d_model)
270
+ self.Wqkv._fused = (0, fuse_splits) # type: ignore
271
+
272
+ if self.attn_qk_ln:
273
+ layernorm_class = LPLayerNorm if low_precision_layernorm else nn.LayerNorm
274
+ self.q_ln = layernorm_class(self.d_model, device=device)
275
+ self.k_ln = layernorm_class(self.d_model, device=device)
276
+
277
+ if self.attn_impl == 'flash':
278
+ self.attn_fn = flash_attn_fn
279
+ elif self.attn_impl == 'triton':
280
+ self.attn_fn = triton_flash_attn_fn
281
+ warnings.warn(
282
+ 'While `attn_impl: triton` can be faster than `attn_impl: flash` ' +\
283
+ 'it uses more memory. When training larger models this can trigger ' +\
284
+ 'alloc retries which hurts performance. If encountered, we recommend ' +\
285
+ 'using `attn_impl: flash` if your model does not use `alibi` or `prefix_lm`.')
286
+ elif self.attn_impl == 'torch':
287
+ self.attn_fn = scaled_multihead_dot_product_attention
288
+ if torch.cuda.is_available():
289
+ warnings.warn(
290
+ 'Using `attn_impl: torch`. If your model does not use `alibi` or ' +\
291
+ '`prefix_lm` we recommend using `attn_impl: flash` otherwise ' +\
292
+ 'we recommend using `attn_impl: triton`.'
293
+ )
294
+ else:
295
+ raise ValueError(f'{attn_impl=} is an invalid setting.')
296
+
297
+ self.out_proj = nn.Linear(self.d_model, self.d_model, device=device)
298
+ self.out_proj._is_residual = True # type: ignore
299
+
300
+ def forward(self,
301
+ x,
302
+ past_key_value=None,
303
+ attn_bias=None,
304
+ attention_mask=None,
305
+ is_causal=True,
306
+ needs_weights=False):
307
+ qkv = self.Wqkv(x)
308
+
309
+ if self.clip_qkv:
310
+ qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
311
+
312
+ query, key, value = qkv.chunk(3, dim=2)
313
+
314
+ key_padding_mask = attention_mask
315
+
316
+ if self.attn_qk_ln:
317
+ # Applying layernorm to qk
318
+ dtype = query.dtype
319
+ query = self.q_ln(query).to(dtype)
320
+ key = self.k_ln(key).to(dtype)
321
+
322
+ if past_key_value is not None:
323
+ if len(past_key_value) != 0:
324
+ key = torch.cat([past_key_value[0], key], dim=1)
325
+ value = torch.cat([past_key_value[1], value], dim=1)
326
+
327
+ past_key_value = (key, value)
328
+
329
+ if attn_bias is not None:
330
+ attn_bias = attn_bias[:, :, -query.size(1):, -key.size(1):]
331
+
332
+ context, attn_weights = self.attn_fn(
333
+ query,
334
+ key,
335
+ value,
336
+ self.n_heads,
337
+ softmax_scale=self.softmax_scale,
338
+ attn_bias=attn_bias,
339
+ key_padding_mask=key_padding_mask,
340
+ is_causal=is_causal,
341
+ dropout_p=self.attn_dropout_p,
342
+ training=self.training,
343
+ needs_weights=needs_weights,
344
+ )
345
+
346
+ return self.out_proj(context), attn_weights, past_key_value
347
+
348
+
349
+ def attn_bias_shape(attn_impl, n_heads, seq_len, alibi, prefix_lm, causal,
350
+ use_sequence_id):
351
+ if attn_impl == 'flash':
352
+ return None
353
+ elif attn_impl in ['torch', 'triton']:
354
+ if alibi:
355
+ if (prefix_lm or not causal) or use_sequence_id:
356
+ return (1, n_heads, seq_len, seq_len)
357
+ return (1, n_heads, 1, seq_len)
358
+ elif prefix_lm or use_sequence_id:
359
+ return (1, 1, seq_len, seq_len)
360
+ return None
361
+ else:
362
+ raise ValueError(f'{attn_impl=} is an invalid setting.')
363
+
364
+
365
+ def attn_bias(attn_impl,
366
+ attn_bias,
367
+ n_heads,
368
+ seq_len,
369
+ causal=False,
370
+ alibi=False,
371
+ alibi_bias_max=8):
372
+ if attn_impl == 'flash':
373
+ return None
374
+ elif attn_impl in ['torch', 'triton']:
375
+ if alibi:
376
+ # in place add alibi to attn bias
377
+ device, dtype = attn_bias.device, attn_bias.dtype
378
+ attn_bias = attn_bias.add(
379
+ alibi_bias(n_heads,
380
+ seq_len,
381
+ full=not causal,
382
+ alibi_bias_max=alibi_bias_max,
383
+ device=device,
384
+ dtype=dtype))
385
+ return attn_bias
386
+ else:
387
+ raise ValueError(f'{attn_impl=} is an invalid setting.')
388
+
389
+
390
+ def alibi_bias(n_heads,
391
+ seq_len,
392
+ full=False,
393
+ alibi_bias_max=8,
394
+ device=None,
395
+ dtype=None):
396
+ alibi_bias = torch.arange(1 - seq_len, 1, dtype=dtype,
397
+ device=device).view(1, 1, 1, seq_len)
398
+ if full:
399
+ # generate 1 x Heads x SeqLen x SeqLen alibi bias mask
400
+ # otherwise the mask is 1 x Heads x 1 x SeqLen (which is broadcast to the appropriate size)
401
+ alibi_bias = alibi_bias - torch.arange(
402
+ 1 - seq_len, 1, dtype=dtype, device=device).view(1, 1, seq_len, 1)
403
+ alibi_bias = alibi_bias.abs().mul(-1)
404
+
405
+ m = torch.arange(1, n_heads + 1, dtype=dtype, device=device)
406
+ m = m.mul(alibi_bias_max / n_heads)
407
+ alibi_bias = alibi_bias * (1. / (2**m.view(1, n_heads, 1, 1)))
408
+ return alibi_bias
config.json ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "mosaicml/mpt-1b-redpajama-200b",
3
+ "alibi": true,
4
+ "alibi_bias_max": 8,
5
+ "architectures": [
6
+ "MosaicGPT"
7
+ ],
8
+ "attn_clip_qkv": null,
9
+ "attn_impl": "torch",
10
+ "attn_pdrop": 0,
11
+ "attn_qk_ln": true,
12
+ "attn_uses_sequence_id": false,
13
+ "auto_map": {
14
+ "AutoConfig": "configuration_mosaic_gpt.MosaicGPTConfig",
15
+ "AutoModelForCausalLM": "mosaic_gpt.MosaicGPT"
16
+ },
17
+ "d_model": 2048,
18
+ "emb_init_std": null,
19
+ "emb_init_uniform_lim": null,
20
+ "emb_pdrop": 0,
21
+ "embedding_fraction": 1.0,
22
+ "fan_mode": "fan_in",
23
+ "init_device": "cpu",
24
+ "init_div_is_residual": true,
25
+ "init_gain": 0,
26
+ "init_nonlinearity": "relu",
27
+ "init_std": 0.02,
28
+ "logit_scale": null,
29
+ "low_precision_layernorm": true,
30
+ "max_seq_len": 2048,
31
+ "mlp_ratio": 4,
32
+ "model_type": "mosaic_gpt",
33
+ "n_heads": 16,
34
+ "n_layers": 24,
35
+ "no_bias": true,
36
+ "param_init_fn": "kaiming_normal_",
37
+ "prefix_lm": false,
38
+ "resid_pdrop": 0,
39
+ "softmax_scale": null,
40
+ "tokenizer_name": "EleutherAI/gpt-neox-20b",
41
+ "torch_dtype": "float32",
42
+ "transformers_version": "4.27.4",
43
+ "use_cache": false,
44
+ "verbose": 0,
45
+ "vocab_size": 50432
46
+ }
configuration_mosaic_gpt.py ADDED
@@ -0,0 +1,168 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2022 MosaicML Examples authors
2
+ # SPDX-License-Identifier: Apache-2.0
3
+
4
+ """A HuggingFace-style model configuration."""
5
+
6
+ from typing import Optional, Tuple, Union
7
+
8
+ from transformers import PretrainedConfig
9
+
10
+
11
+ class MosaicGPTConfig(PretrainedConfig):
12
+ model_type = 'mosaic_gpt'
13
+
14
+ def __init__(
15
+ self,
16
+ d_model: int = 2048,
17
+ n_heads: int = 16,
18
+ n_layers: int = 24,
19
+ mlp_ratio: int = 4,
20
+ max_seq_len: int = 2048,
21
+ vocab_size: int = 50368,
22
+ attn_pdrop: float = 0.0,
23
+ resid_pdrop: float = 0.0,
24
+ emb_pdrop: float = 0.0,
25
+ attn_impl: str = 'triton',
26
+ attn_qk_ln: bool = False,
27
+ attn_clip_qkv: Optional[float] = None,
28
+ softmax_scale: Optional[float] = None,
29
+ prefix_lm: Optional[bool] = False,
30
+ attn_uses_sequence_id: Optional[bool] = False,
31
+ alibi: bool = False,
32
+ alibi_bias_max: int = 8,
33
+ init_device: str = 'cpu',
34
+ logit_scale: Optional[Union[float, str]] = None,
35
+ no_bias: bool = False,
36
+ verbose: int = 0,
37
+ param_init_fn: str = 'kaiming_normal_',
38
+ init_div_is_residual: Union[int, float, str, bool] = True,
39
+ init_std: float = 0.02,
40
+ emb_init_std: Optional[float] = None,
41
+ emb_init_uniform_lim: Optional[Union[Tuple[float, float],
42
+ float]] = None,
43
+ init_gain: float = 0,
44
+ fan_mode: str = 'fan_in',
45
+ init_nonlinearity: str = 'relu',
46
+ embedding_fraction: float = 1.0,
47
+ low_precision_layernorm: bool = True,
48
+ use_cache: bool = False,
49
+ **kwargs,
50
+ ):
51
+ """The MosaicGPT configuration class.
52
+
53
+ Args:
54
+ d_model (int): The size of the embedding dimension of the model.
55
+ n_heads (int): The number of attention heads.
56
+ n_layers (int): The number of layers in the model.
57
+ mlp_ratio (int): The ratio of the up/down scale in the MLP.
58
+ max_seq_len (int): The maximum sequence length of the model.
59
+ vocab_size (int): The size of the vocabulary.
60
+ attn_pdrop (float): The dropout probability for the attention layers.
61
+ resid_pdrop (float): The dropout probability applied to the attention output before combining with residual.
62
+ emb_pdrop (float): The dropout probability for the embedding layer.
63
+ attn_impl (str): The attention implementation to use. One of 'torch', 'flash', or 'triton'.
64
+ attn_qk_ln (bool): Whether to apply layer normalization to the queries and keys in the attention layer.
65
+ attn_clip_qkv (Optional[float]): If not None, clip the queries, keys, and values in the attention layer to
66
+ this value.
67
+ softmax_scale (Optional[float]): If not None, scale the softmax in the attention layer by this value. If None,
68
+ use the default scale of ``1/sqrt(d_keys)``.
69
+ prefix_lm (Optional[bool]): Whether the model should operate as a Prefix LM. This requires passing an
70
+ extra `prefix_mask` argument which indicates which tokens belong to the prefix. Tokens in the prefix
71
+ can attend to one another bi-directionally. Tokens outside the prefix use causal attention.
72
+ attn_uses_sequence_id (Optional[bool]): Whether to restrict attention to tokens that have the same sequence_id.
73
+ When the model is in `train` mode, this requires passing an extra `sequence_id` argument which indicates
74
+ which sub-sequence each token belongs to.
75
+ Defaults to ``False`` meaning any provided `sequence_id` will be ignored.
76
+ alibi (bool): Whether to use the alibi bias instead of position embeddings.
77
+ alibi_bias_max (int): The maximum value of the alibi bias.
78
+ init_device (str): The device to use for parameter initialization.
79
+ logit_scale (Optional[Union[float, str]]): If not None, scale the logits by this value.
80
+ no_bias (bool): Whether to use bias in all layers.
81
+ verbose (int): The verbosity level. 0 is silent.
82
+ param_init_fn (str): The parameter initialization scheme to use. One of 'default_', 'baseline_', 'kaiming_uniform_',
83
+ 'kaiming_normal_', 'neox_init_', 'small_init_', 'xavier_uniform_', or 'xavier_normal_'.
84
+ init_div_is_residual (Union[int, float, str, bool]): Value to divide initial weights by if ``module._is_residual`` is True.
85
+ init_std (float): The standard deviation of the normal distribution used to initialize the model,
86
+ if using the baseline_ parameter initialization scheme.
87
+ emb_init_std (Optional[float]): The standard deviation of the normal distribution used to initialize the embedding layer.
88
+ emb_init_uniform_lim (Optional[Union[Tuple[float, float], float]]): The lower and upper limits of the uniform distribution
89
+ used to initialize the embedding layer. Mutually exclusive with ``emb_init_std``.
90
+ init_gain (float): The gain to use for parameter initialization with kaiming or xavier initialization schemes.
91
+ fan_mode (str): The fan mode to use for parameter initialization with kaiming initialization schemes.
92
+ init_nonlinearity (str): The nonlinearity to use for parameter initialization with kaiming initialization schemes.
93
+ embedding_fraction (float): The fraction to scale the gradients of the embedding layer by.
94
+ low_precision_layernorm (bool): Whether to use low precision layer normalization.
95
+ use_cache (bool): Whether or not the model should return the last key/values attentions
96
+ """
97
+ self.d_model = d_model
98
+ self.n_heads = n_heads
99
+ self.n_layers = n_layers
100
+ self.mlp_ratio = mlp_ratio
101
+ self.max_seq_len = max_seq_len
102
+ self.vocab_size = vocab_size
103
+ self.attn_pdrop = attn_pdrop
104
+ self.resid_pdrop = resid_pdrop
105
+ self.emb_pdrop = emb_pdrop
106
+ self.attn_impl = attn_impl
107
+ self.attn_qk_ln = attn_qk_ln
108
+ self.attn_clip_qkv = attn_clip_qkv
109
+ self.softmax_scale = softmax_scale
110
+ self.prefix_lm = prefix_lm
111
+ self.attn_uses_sequence_id = attn_uses_sequence_id
112
+ self.alibi = alibi
113
+ self.alibi_bias_max = alibi_bias_max
114
+ self.init_device = init_device
115
+ self.logit_scale = logit_scale
116
+ self.no_bias = no_bias
117
+ self.verbose = verbose
118
+ self.param_init_fn = param_init_fn
119
+ self.init_div_is_residual = init_div_is_residual
120
+ self.init_std = init_std
121
+ self.emb_init_std = emb_init_std
122
+ self.emb_init_uniform_lim = emb_init_uniform_lim
123
+ self.init_std = init_std
124
+ self.init_gain = init_gain
125
+ self.fan_mode = fan_mode
126
+ self.init_nonlinearity = init_nonlinearity
127
+ self.embedding_fraction = embedding_fraction
128
+ self.low_precision_layernorm = low_precision_layernorm
129
+ self.use_cache = use_cache
130
+ if 'name' in kwargs:
131
+ del kwargs['name']
132
+ if 'loss_fn' in kwargs:
133
+ del kwargs['loss_fn']
134
+ super().__init__(**kwargs)
135
+
136
+ self._validate_config()
137
+
138
+ def _validate_config(self):
139
+ if self.d_model % self.n_heads != 0:
140
+ raise ValueError('d_model must be divisible by n_heads')
141
+ if any(prob < 0 or prob > 1
142
+ for prob in [self.attn_pdrop, self.resid_pdrop, self.emb_pdrop]):
143
+ raise ValueError(
144
+ 'attn_pdrop, resid_pdrop, emb_pdrop are probabilities and must be between 0 and 1'
145
+ )
146
+ if self.attn_impl not in ['torch', 'flash', 'triton']:
147
+ raise ValueError(f'Unknown attn_impl={self.attn_impl}')
148
+ if self.prefix_lm and self.attn_impl not in ['torch', 'triton']:
149
+ raise NotImplementedError(
150
+ 'prefix_lm only implemented with torch and triton attention.')
151
+ if self.alibi and self.attn_impl not in ['torch', 'triton']:
152
+ raise NotImplementedError(
153
+ 'alibi only implemented with torch and triton attention.')
154
+ if self.attn_uses_sequence_id and self.attn_impl not in [
155
+ 'torch', 'triton'
156
+ ]:
157
+ raise NotImplementedError(
158
+ 'attn_uses_sequence_id only implemented with torch and triton attention.'
159
+ )
160
+ if self.embedding_fraction > 1 or self.embedding_fraction <= 0:
161
+ raise ValueError(
162
+ 'model.embedding_fraction must be between 0 (exclusive) and 1 (inclusive)!'
163
+ )
164
+ if isinstance(self.logit_scale,
165
+ str) and self.logit_scale != 'inv_sqrt_d_model':
166
+ raise ValueError(
167
+ f"{self.logit_scale=} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'."
168
+ )
generation_config.json ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "transformers_version": "4.27.4",
4
+ "use_cache": false
5
+ }
gpt_blocks.py ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2022 MosaicML Examples authors
2
+ # SPDX-License-Identifier: Apache-2.0
3
+
4
+ """GPT Blocks used for the GPT Model."""
5
+
6
+ from typing import Optional, Tuple
7
+
8
+ import torch
9
+ import torch.nn as nn
10
+
11
+ from .attention import MultiheadAttention
12
+ from .low_precision_layernorm import LPLayerNorm
13
+
14
+
15
+ class GPTMLP(nn.Module):
16
+
17
+ def __init__(self,
18
+ d_model: int,
19
+ mlp_ratio: int,
20
+ device: Optional[str] = None):
21
+ super().__init__()
22
+ self.mlp_up = nn.Linear(d_model, mlp_ratio * d_model, device=device)
23
+ self.mlp_act = nn.GELU(approximate='none')
24
+ self.mlp_down = nn.Linear(mlp_ratio * d_model, d_model, device=device)
25
+ self.mlp_down._is_residual = True # type: ignore
26
+
27
+ def forward(self, x):
28
+ return self.mlp_down(self.mlp_act(self.mlp_up(x)))
29
+
30
+
31
+ class GPTBlock(nn.Module):
32
+
33
+ def __init__(self,
34
+ attn_impl: str,
35
+ d_model: int,
36
+ n_heads: int,
37
+ mlp_ratio: int,
38
+ attn_clip_qkv: Optional[float] = None,
39
+ attn_qk_ln: bool = False,
40
+ softmax_scale: Optional[float] = None,
41
+ attn_pdrop: float = 0.0,
42
+ alibi: bool = False,
43
+ resid_pdrop: float = 0.0,
44
+ low_precision_layernorm: bool = False,
45
+ device: Optional[str] = None,
46
+ **kwargs):
47
+ del kwargs # unused, just to capture any extra args from the config
48
+ super().__init__()
49
+
50
+ layernorm_class = LPLayerNorm if low_precision_layernorm else nn.LayerNorm
51
+
52
+ self.ln_1 = layernorm_class(d_model, device=device)
53
+ self.attn = MultiheadAttention(
54
+ attn_impl=attn_impl,
55
+ attn_clip_qkv=attn_clip_qkv,
56
+ attn_qk_ln=attn_qk_ln,
57
+ softmax_scale=softmax_scale,
58
+ attn_pdrop=attn_pdrop,
59
+ d_model=d_model,
60
+ n_heads=n_heads,
61
+ device=device,
62
+ )
63
+ self.ln_2 = layernorm_class(d_model, device=device)
64
+ self.mlp = GPTMLP(
65
+ d_model=d_model,
66
+ mlp_ratio=mlp_ratio,
67
+ device=device,
68
+ )
69
+ self.resid_attn_dropout = nn.Dropout(resid_pdrop)
70
+ self.resid_mlp_dropout = nn.Dropout(resid_pdrop)
71
+
72
+ def forward(
73
+ self,
74
+ x: torch.Tensor,
75
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
76
+ attn_bias: Optional[torch.Tensor] = None,
77
+ attention_mask: Optional[torch.ByteTensor] = None,
78
+ is_causal: bool = True,
79
+ ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor]]]:
80
+ a = self.ln_1(x)
81
+ b, _, past_key_value = self.attn(a,
82
+ past_key_value=past_key_value,
83
+ attn_bias=attn_bias,
84
+ attention_mask=attention_mask,
85
+ is_causal=is_causal)
86
+ x = x + self.resid_attn_dropout(b)
87
+ m = self.ln_2(x)
88
+ n = self.mlp(m)
89
+ x = x + self.resid_mlp_dropout(n)
90
+ return x, past_key_value
low_precision_layernorm.py ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn.functional as F
3
+
4
+ class LPLayerNorm(torch.nn.LayerNorm):
5
+ def __init__(self, normalized_shape, eps=1e-05, elementwise_affine=True, device=None, dtype=None):
6
+ super().__init__(
7
+ normalized_shape=normalized_shape,
8
+ eps=eps,
9
+ elementwise_affine=elementwise_affine,
10
+ device=device,
11
+ dtype=dtype,
12
+ )
13
+
14
+ def forward(self, x):
15
+ module_device = x.device
16
+ downcast_x = _cast_if_autocast_enabled(x)
17
+ downcast_weight = _cast_if_autocast_enabled(self.weight) if self.weight is not None else self.weight
18
+ downcast_bias = _cast_if_autocast_enabled(self.bias) if self.bias is not None else self.bias
19
+ with torch.autocast(enabled=False, device_type=module_device.type):
20
+ return F.layer_norm(downcast_x, self.normalized_shape, downcast_weight, downcast_bias, self.eps)
21
+
22
+ def _cast_if_autocast_enabled(tensor):
23
+ if torch.is_autocast_enabled():
24
+ if tensor.device.type == 'cuda':
25
+ dtype = torch.get_autocast_gpu_dtype()
26
+ elif tensor.device.type == 'cpu':
27
+ dtype = torch.get_autocast_cpu_dtype()
28
+ else:
29
+ raise NotImplementedError()
30
+ return tensor.to(dtype=dtype)
31
+ return tensor
mosaic_gpt.py ADDED
@@ -0,0 +1,449 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2022 MosaicML Examples authors
2
+ # SPDX-License-Identifier: Apache-2.0
3
+
4
+ """A simple, flexible implementation of a GPT model.
5
+
6
+ Inspired by https://github.com/karpathy/minGPT/blob/master/mingpt/model.py
7
+ """
8
+
9
+ import math
10
+ import warnings
11
+ from typing import List, Optional, Tuple
12
+
13
+ import torch
14
+ import torch.nn as nn
15
+ import torch.nn.functional as F
16
+ from transformers import AutoTokenizer, PreTrainedModel
17
+ from transformers.modeling_outputs import CausalLMOutputWithPast
18
+
19
+ from .attention import attn_bias as module_attn_bias, attn_bias_shape as module_attn_bias_shape
20
+ from .gpt_blocks import GPTBlock
21
+ from .configuration_mosaic_gpt import \
22
+ MosaicGPTConfig
23
+ from .param_init_fns import MODEL_INIT_REGISTRY
24
+ from .low_precision_layernorm import LPLayerNorm
25
+
26
+
27
+ class MosaicGPT(PreTrainedModel):
28
+ config_class = MosaicGPTConfig
29
+ base_model_prefix = 'mosaic_gpt'
30
+
31
+ def __init__(self, config: MosaicGPTConfig):
32
+ super().__init__(config)
33
+
34
+ if config.attn_impl == 'flash' and config.alibi:
35
+ raise RuntimeError("ALiBi is not supported with flash attention. Please use triton or torch.")
36
+
37
+ self.attn_impl = config.attn_impl
38
+ self.prefix_lm = config.prefix_lm
39
+ self.attn_uses_sequence_id = config.attn_uses_sequence_id
40
+ self.alibi = config.alibi
41
+ self.alibi_bias_max = config.alibi_bias_max
42
+
43
+ layernorm_class = LPLayerNorm if config.low_precision_layernorm else nn.LayerNorm
44
+
45
+ # CogView (https://arxiv.org/abs/2105.13290) and GLM-130B (https://arxiv.org/abs/2210.02414)
46
+ # both report this helping with stabilizing training
47
+ self.embedding_fraction = config.embedding_fraction
48
+
49
+ self.transformer = nn.ModuleDict({
50
+ 'wte':
51
+ nn.Embedding(config.vocab_size,
52
+ config.d_model,
53
+ device=config.init_device)
54
+ })
55
+ if not self.alibi:
56
+ self.transformer.update({
57
+ 'wpe':
58
+ nn.Embedding(config.max_seq_len,
59
+ config.d_model,
60
+ device=config.init_device)
61
+ })
62
+ self.transformer.update({'emb_drop': nn.Dropout(config.emb_pdrop)})
63
+ self.transformer.update({
64
+ 'blocks':
65
+ nn.ModuleList([
66
+ GPTBlock(device=config.init_device,
67
+ **config.to_dict())
68
+ for _ in range(config.n_layers)
69
+ ])
70
+ })
71
+ self.transformer.update({
72
+ 'ln_f': layernorm_class(config.d_model, device=config.init_device)
73
+ })
74
+
75
+ # enables scaling output logits; similar to a softmax "temperature"
76
+ # PaLM paper uses scale 1/sqrt(config.d_model)
77
+ self.logit_scale = None
78
+ if config.logit_scale is not None:
79
+ logit_scale = config.logit_scale
80
+ if isinstance(logit_scale, str):
81
+ if logit_scale == 'inv_sqrt_d_model':
82
+ logit_scale = 1 / math.sqrt(config.d_model)
83
+ else:
84
+ raise ValueError(
85
+ f"{logit_scale=} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'."
86
+ )
87
+ self.logit_scale = logit_scale
88
+
89
+ if config.init_device != 'meta':
90
+ print(
91
+ f'You are using {config.init_device=}, but you can also use config.init_device="meta" with Composer + FSDP for fast initialization.'
92
+ )
93
+ self.apply(self.param_init_fn)
94
+
95
+ self.is_causal = not self.prefix_lm
96
+
97
+ # define attn mask
98
+ self._attn_bias_initialized = False
99
+ self.attn_bias = None
100
+ self.attn_bias_shape = module_attn_bias_shape(
101
+ self.attn_impl,
102
+ config.n_heads,
103
+ config.max_seq_len,
104
+ self.alibi,
105
+ prefix_lm=self.prefix_lm,
106
+ causal=self.is_causal,
107
+ use_sequence_id=self.attn_uses_sequence_id)
108
+
109
+ if config.no_bias:
110
+ for module in self.modules():
111
+ if hasattr(module, 'bias') and isinstance(
112
+ module.bias, nn.Parameter):
113
+ if config.verbose:
114
+ print(f'Removing bias ({module.bias}) from {module}.')
115
+ module.register_parameter('bias', None)
116
+
117
+ if config.verbose and config.verbose > 2:
118
+ print(self)
119
+
120
+ @torch.no_grad()
121
+ def _attn_bias(self,
122
+ device,
123
+ dtype,
124
+ attention_mask: Optional[torch.ByteTensor] = None,
125
+ prefix_mask: Optional[torch.ByteTensor] = None,
126
+ sequence_id: Optional[torch.LongTensor] = None):
127
+ if not self._attn_bias_initialized:
128
+ if self.attn_bias_shape:
129
+ self.attn_bias = torch.zeros(self.attn_bias_shape,
130
+ device=device,
131
+ dtype=dtype)
132
+ self.attn_bias = module_attn_bias(
133
+ self.attn_impl,
134
+ self.attn_bias,
135
+ self.config.n_heads,
136
+ self.config.max_seq_len,
137
+ causal=self.is_causal,
138
+ alibi=self.alibi,
139
+ alibi_bias_max=self.alibi_bias_max)
140
+ self._attn_bias_initialized = True
141
+
142
+ # flash does not support prefix_lm and will incorporate any
143
+ # attention_mask inside the attention module
144
+ if self.attn_impl == 'flash':
145
+ return self.attn_bias, attention_mask
146
+
147
+ attn_bias = self.attn_bias
148
+
149
+ # If using torch or triton, we incorporate the prefix_mask (if appropriate)
150
+ if self.prefix_lm:
151
+ assert isinstance(attn_bias, torch.Tensor) # pyright
152
+ assert isinstance(prefix_mask, torch.Tensor) # pyright
153
+ attn_bias = self._apply_prefix_mask(attn_bias, prefix_mask)
154
+
155
+ # If using torch or triton, we incorporate sequence_id (if appropriate)
156
+ if self.attn_uses_sequence_id and sequence_id is not None:
157
+ assert isinstance(attn_bias, torch.Tensor) # pyright
158
+ attn_bias = self._apply_sequence_id(attn_bias, sequence_id)
159
+
160
+ # If using torch or triton, we incorporate attention_mask. This will output
161
+ # None in place of attention_mask since it will not be further needed in the
162
+ # attention modules.
163
+ if attention_mask is not None:
164
+ s_k = attention_mask.shape[-1]
165
+ if attn_bias is None:
166
+ attn_bias = torch.zeros((1, 1, 1, s_k),
167
+ device=device,
168
+ dtype=dtype)
169
+ else:
170
+ attn_bias = attn_bias[:, :, :, -s_k:]
171
+ if prefix_mask is not None and (attention_mask.shape !=
172
+ prefix_mask.shape):
173
+ raise ValueError(
174
+ f'attention_mask shape={attention_mask.shape} ' +\
175
+ f'and prefix_mask shape={prefix_mask.shape} are not equal.'
176
+ )
177
+ min_val = torch.finfo(attn_bias.dtype).min
178
+ attn_bias = attn_bias.masked_fill(
179
+ ~attention_mask.view(-1, 1, 1, s_k), min_val)
180
+
181
+ return attn_bias, None
182
+
183
+ def _apply_prefix_mask(self, attn_bias: torch.Tensor,
184
+ prefix_mask: torch.Tensor):
185
+ s_k, s_q = attn_bias.shape[-2:]
186
+ if (s_k != self.config.max_seq_len) or (s_q != self.config.max_seq_len):
187
+ raise ValueError(
188
+ 'attn_bias does not match the expected shape. ' +\
189
+ f'The last two dimensions should both be {self.config.max_length} ' +\
190
+ f'but are {s_k} and {s_q}.'
191
+ )
192
+ seq_len = prefix_mask.shape[-1]
193
+ if seq_len > self.config.max_seq_len:
194
+ raise ValueError(
195
+ f'prefix_mask sequence length cannot exceed max_seq_len={self.config.max_seq_len}'
196
+ )
197
+
198
+ # select seq_len subset of attn mask
199
+ attn_bias = attn_bias[..., :seq_len, :seq_len]
200
+
201
+ # Mix the causal max and the bidirectional mask to get the full
202
+ # allowable attention (i.e. full = not accounting for padding yet)
203
+ causal = torch.tril(
204
+ torch.ones((seq_len, seq_len),
205
+ dtype=torch.bool,
206
+ device=prefix_mask.device)).view(1, 1, seq_len, seq_len)
207
+ prefix = prefix_mask.view(-1, 1, 1, seq_len)
208
+ cannot_attend = ~torch.logical_or(causal, prefix.bool())
209
+
210
+ min_val = torch.finfo(attn_bias.dtype).min
211
+ attn_bias = attn_bias.masked_fill(cannot_attend, min_val)
212
+
213
+ return attn_bias
214
+
215
+ def _apply_sequence_id(self, attn_bias: torch.Tensor,
216
+ sequence_id: torch.LongTensor):
217
+ seq_len = sequence_id.shape[-1]
218
+ if seq_len > self.config.max_seq_len:
219
+ raise ValueError(
220
+ f'sequence_id sequence length cannot exceed max_seq_len={self.config.max_seq_len}'
221
+ )
222
+
223
+ # select seq_len subset of attn mask
224
+ attn_bias = attn_bias[..., :seq_len, :seq_len]
225
+
226
+ # Restrict attention to tokens that share the same value
227
+ # in sequence_id
228
+ cannot_attend = torch.logical_not(
229
+ torch.eq(sequence_id.view(-1, seq_len, 1),
230
+ sequence_id.view(-1, 1, seq_len))).unsqueeze(1)
231
+ min_val = torch.finfo(attn_bias.dtype).min
232
+ attn_bias = attn_bias.masked_fill(cannot_attend, min_val)
233
+
234
+ return attn_bias
235
+
236
+ def forward(
237
+ self,
238
+ input_ids: torch.LongTensor,
239
+ past_key_values: Optional[List[Tuple[torch.FloatTensor]]] = None,
240
+ attention_mask: Optional[torch.ByteTensor] = None,
241
+ prefix_mask: Optional[torch.ByteTensor] = None,
242
+ sequence_id: Optional[torch.LongTensor] = None,
243
+ return_dict: Optional[bool] = None,
244
+ output_attentions: Optional[bool] = None,
245
+ output_hidden_states: Optional[bool] = None,
246
+ use_cache: Optional[bool] = None):
247
+ return_dict = return_dict if return_dict is not None else self.config.return_dict
248
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
249
+
250
+ # These args are passed in by keyword in huggingface's generate function
251
+ # https://github.com/huggingface/transformers/blob/68287689f2f0d8b7063c400230b3766987abf18d/src/transformers/generation/utils.py#L2201-L2206
252
+ # but have not yet been fully implemented in MosaicGPT
253
+ if not return_dict:
254
+ raise NotImplementedError(
255
+ 'return_dict False is not implemented yet for MosaicGPT')
256
+ if output_attentions:
257
+ raise NotImplementedError(
258
+ 'output_attentions is not implemented yet for MosaicGPT')
259
+
260
+ if attention_mask is not None and attention_mask[:, 0].sum(
261
+ ) != attention_mask.shape[0] and self.training:
262
+ raise NotImplementedError(
263
+ 'MosaicGPT does not support training with left padding.')
264
+
265
+ if self.prefix_lm and prefix_mask is None:
266
+ raise ValueError(
267
+ 'prefix_mask is a required argument when MosaicGPT is configured with prefix_lm=True.'
268
+ )
269
+
270
+ if self.training:
271
+ if self.attn_uses_sequence_id and sequence_id is None:
272
+ raise ValueError(
273
+ 'sequence_id is a required argument when MosaicGPT is configured with attn_uses_sequence_id=True ' +\
274
+ 'and the model is in train mode.'
275
+ )
276
+ elif (self.attn_uses_sequence_id is False) and (sequence_id
277
+ is not None):
278
+ warnings.warn(
279
+ 'MosaicGPT received non-None input for `sequence_id` but is configured with attn_uses_sequence_id=False. ' +\
280
+ 'This input will be ignored. If you want the model to use `sequence_id`, set attn_uses_sequence_id to True.'
281
+ )
282
+
283
+ S = input_ids.size(1)
284
+
285
+ assert (
286
+ S <= self.config.max_seq_len
287
+ ), f'Cannot forward input with seq_len={S}, this model only supports seq_len<={self.config.max_seq_len}'
288
+
289
+ tok_emb = self.transformer.wte(input_ids) # type: ignore
290
+ if self.alibi:
291
+ x = tok_emb
292
+ else:
293
+ past_position = 0
294
+ if past_key_values is not None:
295
+ if len(past_key_values) != self.config.n_layers:
296
+ raise ValueError(
297
+ f'past_key_values must provide a past_key_value for each attention ' +\
298
+ f'layer in the network ({len(past_key_values)=}; {self.config.n_layers=}).'
299
+ )
300
+ # get the key tensor whose spec should be (batch, seq, dim), and
301
+ # collect the `seq`, so that the position embedding is shifted
302
+ past_position = past_key_values[0][0].size(1)
303
+
304
+ if S + past_position > self.config.max_seq_len:
305
+ raise ValueError(
306
+ f'Cannot forward input with past sequence length {past_position} and current sequence length '
307
+ f'{S + 1}, this model only supports total sequence length <= {self.config.max_seq_len}.'
308
+ )
309
+ pos = torch.arange(past_position,
310
+ S + past_position,
311
+ dtype=torch.long,
312
+ device=input_ids.device).unsqueeze(0)
313
+ if attention_mask is not None:
314
+ # adjust the position indices to account for padding tokens
315
+ pos = torch.clamp(pos - torch.cumsum(
316
+ (~attention_mask).to(torch.int32), dim=1)[:,
317
+ past_position:],
318
+ min=0)
319
+
320
+ pos_emb = self.transformer.wpe(pos) # type: ignore
321
+ x = tok_emb + pos_emb
322
+
323
+ if self.embedding_fraction == 1:
324
+ x = self.transformer.emb_drop(x) # type: ignore
325
+ else:
326
+ # this implementation is proposed on page 7 of the GLM-130B paper https://arxiv.org/abs/2210.02414
327
+ x_shrunk = (x * self.embedding_fraction) + (
328
+ x.detach() * (1 - self.embedding_fraction))
329
+ assert isinstance(self.transformer.emb_drop, nn.Module) # pyright
330
+ x = self.transformer.emb_drop(x_shrunk)
331
+
332
+ attn_bias, attention_mask = self._attn_bias(
333
+ device=x.device,
334
+ dtype=x.dtype,
335
+ attention_mask=attention_mask,
336
+ prefix_mask=prefix_mask,
337
+ sequence_id=sequence_id)
338
+
339
+ # initialize the past key values cache if it should be used
340
+ if use_cache and past_key_values is None:
341
+ past_key_values = [() for _ in range(self.config.n_layers)
342
+ ] # type: ignore
343
+
344
+ all_hidden_states = () if output_hidden_states else None
345
+ for b_idx, block in enumerate(self.transformer.blocks): # type: ignore
346
+ if output_hidden_states:
347
+ assert all_hidden_states is not None # pyright
348
+ all_hidden_states = all_hidden_states + (x,)
349
+ past_key_value = past_key_values[
350
+ b_idx] if past_key_values is not None else None
351
+ x, past_key_value = block(x,
352
+ past_key_value=past_key_value,
353
+ attn_bias=attn_bias,
354
+ attention_mask=attention_mask,
355
+ is_causal=self.is_causal)
356
+ if past_key_values is not None:
357
+ past_key_values[b_idx] = past_key_value
358
+
359
+ x = self.transformer.ln_f(x) # type: ignore
360
+
361
+ # output embedding weight tied to input embedding
362
+ assert isinstance(self.transformer.wte, nn.Module) # pyright
363
+ assert isinstance(self.transformer.wte.weight, torch.Tensor) # pyright
364
+ logits = F.linear(x, self.transformer.wte.weight, None)
365
+
366
+ if self.logit_scale is not None:
367
+ if self.logit_scale == 0:
368
+ warnings.warn(
369
+ f'Multiplying logits by {self.logit_scale=}. This will produce uniform (uninformative) outputs.'
370
+ )
371
+ logits *= self.logit_scale
372
+
373
+ return CausalLMOutputWithPast(logits=logits,
374
+ past_key_values=past_key_values,
375
+ hidden_states=all_hidden_states)
376
+
377
+ # Param Initialization, needed for device='meta' fast initialization
378
+ def param_init_fn(self, module):
379
+ init_fn_name = self.config.param_init_fn
380
+ if self.config.verbose > 1:
381
+ warnings.warn(f'Using {init_fn_name} initialization.')
382
+ MODEL_INIT_REGISTRY[init_fn_name](module=module,
383
+ **self.config.to_dict())
384
+
385
+ # FSDP Wrap function
386
+ def fsdp_wrap_fn(self, module):
387
+ return isinstance(module, GPTBlock)
388
+
389
+ # Activation Checkpointing
390
+ def activation_checkpointing_fn(self, module):
391
+ return isinstance(module, GPTBlock)
392
+
393
+ def prepare_inputs_for_generation(self,
394
+ input_ids,
395
+ past_key_values=None,
396
+ inputs_embeds=None,
397
+ **kwargs):
398
+ if inputs_embeds is not None:
399
+ raise NotImplementedError(
400
+ 'inputs_embeds is not implemented for MosaicGPT yet')
401
+
402
+ attention_mask = kwargs['attention_mask'].bool()
403
+ if attention_mask[:, -1].sum() != attention_mask.shape[0]:
404
+ raise NotImplementedError(
405
+ 'MosaicGPT does not support generation with right padding.')
406
+
407
+ if self.attn_uses_sequence_id and self.training:
408
+ sequence_id = torch.zeros_like(input_ids[:1])
409
+ else:
410
+ sequence_id = None
411
+
412
+ if past_key_values is not None:
413
+ input_ids = input_ids[:, -1].unsqueeze(-1)
414
+
415
+ if self.prefix_lm:
416
+ # Leverage a convenience of sequential generation!
417
+ prefix_mask = torch.ones_like(attention_mask)
418
+ # This requires that we're using the cache
419
+ if kwargs.get('use_cache') == False:
420
+ raise NotImplementedError(
421
+ 'MosaicGPT with prefix_lm=True does not support use_cache=False.'
422
+ )
423
+ else:
424
+ prefix_mask = None
425
+
426
+ return {
427
+ 'input_ids': input_ids,
428
+ 'attention_mask': attention_mask,
429
+ 'prefix_mask': prefix_mask,
430
+ 'sequence_id': sequence_id,
431
+ 'past_key_values': past_key_values,
432
+ 'use_cache': kwargs.get('use_cache', True),
433
+ }
434
+
435
+ @staticmethod
436
+ def _reorder_cache(past_key_values, beam_idx):
437
+ """Used by HuggingFace generate when using beam search with kv-caching.
438
+
439
+ See https://github.com/huggingface/transformers/blob/3ec7a47664ebe40c40f4b722f6bb1cd30c3821ec/src/transformers/models/gpt2/modeling_gpt2.py#L1122-L1133
440
+ for an example in transformers.
441
+ """
442
+ reordered_past = []
443
+ for layer_past in past_key_values:
444
+ reordered_past += [
445
+ tuple(
446
+ past_state.index_select(0, beam_idx)
447
+ for past_state in layer_past)
448
+ ]
449
+ return reordered_past
param_init_fns.py ADDED
@@ -0,0 +1,464 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2022 MosaicML Examples authors
2
+ # SPDX-License-Identifier: Apache-2.0
3
+ import math
4
+ import warnings
5
+ from collections.abc import Sequence
6
+ from functools import partial
7
+ from typing import Optional, Tuple, Union
8
+
9
+ import torch
10
+ from torch import nn
11
+
12
+
13
+ def torch_default_param_init_fn_(
14
+ module: nn.Module,
15
+ verbose: int = 0,
16
+ **kwargs,
17
+ ):
18
+ del kwargs # unused, just to capture any extra args from the config
19
+ if verbose > 1:
20
+ warnings.warn(
21
+ f"Initializing network using module's reset_parameters attribute")
22
+
23
+ if hasattr(module, 'reset_parameters'):
24
+ module.reset_parameters() # type: ignore
25
+
26
+
27
+ def fused_init_helper_(module: nn.Module, init_fn_):
28
+ # parameter initialization is often based on the parameters shape.
29
+ # If a layer is fused, initialization should be based on the shapes
30
+ # of the original tensor instead of the shape of the fused tensor.
31
+ # Layers which are fused should have the _fused attibute defined.
32
+ # The first element of _fused is the dimension along which the tensor is fused.
33
+ # This is followed by an iterable of split indices."
34
+
35
+ _fused = getattr(module, '_fused', None)
36
+
37
+ if _fused is None:
38
+ raise RuntimeError(f'Internal logic error')
39
+
40
+ dim, splits = _fused
41
+ splits = (0, *splits, module.weight.size(dim)) # type: ignore
42
+ for s, e in zip(splits[:-1], splits[1:]):
43
+ slice_indices = [slice(None)] * module.weight.ndim # type: ignore
44
+ slice_indices[dim] = slice(s, e)
45
+ init_fn_(module.weight[slice_indices]) # type: ignore
46
+
47
+
48
+ def generic_param_init_fn_(
49
+ module: nn.Module,
50
+ init_fn_,
51
+ n_layers: int,
52
+ d_model: Optional[int] = None,
53
+ init_div_is_residual: Union[int, float, str, bool] = True,
54
+ emb_init_std: Optional[float] = None,
55
+ emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
56
+ verbose: int = 0,
57
+ **kwargs,
58
+ ):
59
+ del kwargs # unused, just to capture any extra args from the config
60
+ if verbose > 1:
61
+ warnings.warn(
62
+ f'If model has bias parameters they are initialized to 0.')
63
+
64
+ # enable user to divide _is_residual weights by
65
+ # a value which defaults to math.sqrt(2 * cfg.n_layers)
66
+ init_div_is_residual = init_div_is_residual
67
+
68
+ if init_div_is_residual is False:
69
+ # not used, for pyright
70
+ div_is_residual = 1.0
71
+ elif init_div_is_residual is True:
72
+ div_is_residual = math.sqrt(2 * n_layers)
73
+ elif isinstance(init_div_is_residual, float) or isinstance(
74
+ init_div_is_residual, int):
75
+ div_is_residual = init_div_is_residual
76
+ elif isinstance(init_div_is_residual,
77
+ str) and init_div_is_residual.isnumeric():
78
+ # do not trust YAML parsing to always convert numbers to numbers
79
+ div_is_residual = float(init_div_is_residual)
80
+ else:
81
+ # not used, for pyright
82
+ div_is_residual = 1.0
83
+ raise ValueError(
84
+ f'Expected init_div_is_residual to be boolean or numeric, got {init_div_is_residual}'
85
+ )
86
+
87
+ if init_div_is_residual is not False:
88
+ if verbose > 1:
89
+ warnings.warn(
90
+ f'Initializing _is_residual layers then dividing them by {div_is_residual}.' +\
91
+ f'set `init_div_is_residual: false` in model config to disable this.'
92
+ )
93
+
94
+ if isinstance(module, nn.Linear):
95
+ # Linear
96
+ if hasattr(module, '_fused'):
97
+ fused_init_helper_(module, init_fn_)
98
+ else:
99
+ init_fn_(module.weight)
100
+ if module.bias is not None:
101
+ torch.nn.init.zeros_(module.bias)
102
+
103
+ if init_div_is_residual is not False and getattr(
104
+ module, '_is_residual', False):
105
+ with torch.no_grad():
106
+ module.weight.div_(div_is_residual)
107
+
108
+ elif isinstance(module, nn.Embedding):
109
+ # Embedding
110
+ if emb_init_std is not None:
111
+ std = emb_init_std
112
+ if std == 0:
113
+ warnings.warn(f'Embedding layer initialized to 0.')
114
+ emb_init_fn_ = partial(torch.nn.init.normal_, mean=0.0, std=std)
115
+ if verbose > 1:
116
+ warnings.warn(
117
+ f'Embedding layer initialized using normal distribution with mean=0 and {std=}.'
118
+ )
119
+ elif emb_init_uniform_lim is not None:
120
+ lim = emb_init_uniform_lim
121
+ if isinstance(lim, Sequence):
122
+ if len(lim) > 2:
123
+ raise ValueError(
124
+ f'Uniform init requires a min and a max limit. User input: {lim}.'
125
+ )
126
+ if lim[0] == lim[1]:
127
+ warnings.warn(f'Embedding layer initialized to {lim[0]}.')
128
+ else:
129
+ if lim == 0:
130
+ warnings.warn(f'Embedding layer initialized to 0.')
131
+ lim = [-lim, lim]
132
+ a, b = lim
133
+ emb_init_fn_ = partial(torch.nn.init.uniform_, a=a, b=b)
134
+ if verbose > 1:
135
+ warnings.warn(
136
+ f'Embedding layer initialized using uniform distribution in range {lim}.'
137
+ )
138
+ else:
139
+ emb_init_fn_ = init_fn_
140
+
141
+ emb_init_fn_(module.weight)
142
+
143
+ elif isinstance(module, nn.LayerNorm):
144
+ # LayerNorm
145
+ if verbose > 1:
146
+ warnings.warn(
147
+ f'LayerNorm gamma weights are set to 1. If the layer has a bias it is initialized to 0.'
148
+ )
149
+ torch.nn.init.ones_(module.weight)
150
+ if module.bias is not None:
151
+ torch.nn.init.zeros_(module.bias)
152
+
153
+ elif isinstance(module, nn.MultiheadAttention):
154
+ # torch's MultiheadAttention
155
+ if module._qkv_same_embed_dim:
156
+ assert module.in_proj_weight is not None
157
+ assert module.q_proj_weight is None and module.k_proj_weight is None and module.v_proj_weight is None
158
+ assert d_model is not None
159
+ # in_proj_weight is actually 3 layers and should be split up for width based init
160
+ _d = d_model
161
+ splits = (0, _d, 2 * _d, 3 * _d)
162
+ for s, e in zip(splits[:-1], splits[1:]):
163
+ init_fn_(module.in_proj_weight[s:e])
164
+ else:
165
+ assert module.q_proj_weight is not None and module.k_proj_weight is not None and module.v_proj_weight is not None
166
+ assert module.in_proj_weight is None
167
+ init_fn_(module.q_proj_weight)
168
+ init_fn_(module.k_proj_weight)
169
+ init_fn_(module.v_proj_weight)
170
+
171
+ # bias
172
+ if module.in_proj_bias is not None:
173
+ torch.nn.init.zeros_(module.in_proj_bias)
174
+ if module.bias_k is not None:
175
+ torch.nn.init.zeros_(module.bias_k)
176
+ if module.bias_v is not None:
177
+ torch.nn.init.zeros_(module.bias_v)
178
+
179
+ # out proj
180
+ init_fn_(module.out_proj.weight)
181
+ if init_div_is_residual is not False and getattr(
182
+ module.out_proj, '_is_residual', False):
183
+ with torch.no_grad():
184
+ module.out_proj.weight.div_(div_is_residual)
185
+ if module.out_proj.bias is not None:
186
+ torch.nn.init.zeros_(module.out_proj.bias)
187
+
188
+ else:
189
+ for _ in module.parameters(recurse=False):
190
+ # raise error if uninitialized module has any parameters
191
+ raise NotImplementedError(
192
+ f'{module.__class__.__name__} parameters are not initialized by param_init_fn.'
193
+ )
194
+
195
+
196
+ def _normal_init_(std, mean=0.0):
197
+ return partial(torch.nn.init.normal_, mean=mean, std=std)
198
+
199
+
200
+ def _normal_param_init_fn_(
201
+ module: nn.Module,
202
+ std: float,
203
+ n_layers: int,
204
+ d_model: Optional[int] = None,
205
+ init_div_is_residual: Union[int, float, str, bool] = True,
206
+ emb_init_std: Optional[float] = None,
207
+ emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
208
+ verbose: int = 0,
209
+ **kwargs,
210
+ ):
211
+ del kwargs # unused, just to capture any extra args from the config
212
+ init_fn_ = _normal_init_(std=std)
213
+
214
+ if verbose > 1:
215
+ warnings.warn(
216
+ f'Using torch.nn.init.normal_ init fn mean=0.0, std={std}')
217
+
218
+ generic_param_init_fn_(
219
+ module=module,
220
+ init_fn_=init_fn_,
221
+ d_model=d_model,
222
+ n_layers=n_layers,
223
+ init_div_is_residual=init_div_is_residual,
224
+ emb_init_std=emb_init_std,
225
+ emb_init_uniform_lim=emb_init_uniform_lim,
226
+ verbose=verbose,
227
+ )
228
+
229
+
230
+ def baseline_param_init_fn_(
231
+ module: nn.Module,
232
+ init_std: float,
233
+ n_layers: int,
234
+ d_model: Optional[int] = None,
235
+ init_div_is_residual: Union[int, float, str, bool] = True,
236
+ emb_init_std: Optional[float] = None,
237
+ emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
238
+ verbose: int = 0,
239
+ **kwargs,
240
+ ):
241
+ del kwargs # unused, just to capture any extra args from the config
242
+ if init_std is None:
243
+ raise ValueError(
244
+ 'You must set model.init_std to a float value to use the default initialization scheme.'
245
+ )
246
+ _normal_param_init_fn_(
247
+ module=module,
248
+ std=init_std,
249
+ d_model=d_model,
250
+ n_layers=n_layers,
251
+ init_div_is_residual=init_div_is_residual,
252
+ emb_init_std=emb_init_std,
253
+ emb_init_uniform_lim=emb_init_uniform_lim,
254
+ verbose=verbose,
255
+ )
256
+
257
+
258
+ def small_param_init_fn_(
259
+ module: nn.Module,
260
+ n_layers: int,
261
+ d_model: int,
262
+ init_div_is_residual: Union[int, float, str, bool] = True,
263
+ emb_init_std: Optional[float] = None,
264
+ emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
265
+ verbose: int = 0,
266
+ **kwargs,
267
+ ):
268
+ del kwargs # unused, just to capture any extra args from the config
269
+ # very close to kaiming normal
270
+ # from Transformers without Tears (2019) - Nguyen & Salazar
271
+ std = math.sqrt(2 / (5 * d_model))
272
+ _normal_param_init_fn_(
273
+ module=module,
274
+ std=std,
275
+ d_model=d_model,
276
+ n_layers=n_layers,
277
+ init_div_is_residual=init_div_is_residual,
278
+ emb_init_std=emb_init_std,
279
+ emb_init_uniform_lim=emb_init_uniform_lim,
280
+ verbose=verbose,
281
+ )
282
+
283
+
284
+ def neox_param_init_fn_(
285
+ module: nn.Module,
286
+ n_layers: int,
287
+ d_model: int,
288
+ emb_init_std: Optional[float] = None,
289
+ emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
290
+ verbose: int = 0,
291
+ **kwargs,
292
+ ):
293
+ """From section 2.3.1 of GPT-NeoX-20B:
294
+
295
+ An Open-Source AutoregressiveLanguage Model — Black et. al. (2022)
296
+ see https://github.com/EleutherAI/gpt-neox/blob/9610391ab319403cef079b438edd016a2443af54/megatron/model/init_functions.py#L151
297
+ and https://github.com/EleutherAI/gpt-neox/blob/main/megatron/model/transformer.py
298
+ """
299
+ del kwargs # unused, just to capture any extra args from the config
300
+ residual_div = n_layers / math.sqrt(10) # small std / wang std
301
+
302
+ if verbose > 1:
303
+ warnings.warn(f'setting init_div_is_residual to {residual_div}')
304
+
305
+ small_param_init_fn_(
306
+ module=module,
307
+ d_model=d_model,
308
+ n_layers=n_layers,
309
+ init_div_is_residual=residual_div,
310
+ emb_init_std=emb_init_std,
311
+ emb_init_uniform_lim=emb_init_uniform_lim,
312
+ verbose=verbose,
313
+ )
314
+
315
+
316
+ def kaiming_uniform_param_init_fn_(
317
+ module: nn.Module,
318
+ n_layers: int,
319
+ d_model: Optional[int] = None,
320
+ init_div_is_residual: Union[int, float, str, bool] = True,
321
+ emb_init_std: Optional[float] = None,
322
+ emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
323
+ init_gain: float = 0,
324
+ fan_mode: str = 'fan_in',
325
+ init_nonlinearity: str = 'leaky_relu',
326
+ verbose: int = 0,
327
+ **kwargs,
328
+ ):
329
+ del kwargs # unused, just to capture any extra args from the config
330
+
331
+ if verbose > 1:
332
+ warnings.warn(
333
+ f'Using nn.init.kaiming_uniform_ init fn with parameters: ' +\
334
+ f'a={init_gain}, mode={fan_mode}, nonlinearity={init_nonlinearity}'
335
+ )
336
+
337
+ kaiming_uniform_ = partial(nn.init.kaiming_uniform_,
338
+ a=init_gain,
339
+ mode=fan_mode,
340
+ nonlinearity=init_nonlinearity)
341
+
342
+ generic_param_init_fn_(
343
+ module=module,
344
+ init_fn_=kaiming_uniform_,
345
+ d_model=d_model,
346
+ n_layers=n_layers,
347
+ init_div_is_residual=init_div_is_residual,
348
+ emb_init_std=emb_init_std,
349
+ emb_init_uniform_lim=emb_init_uniform_lim,
350
+ verbose=verbose,
351
+ )
352
+
353
+
354
+ def kaiming_normal_param_init_fn_(
355
+ module: nn.Module,
356
+ n_layers: int,
357
+ d_model: Optional[int] = None,
358
+ init_div_is_residual: Union[int, float, str, bool] = True,
359
+ emb_init_std: Optional[float] = None,
360
+ emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
361
+ init_gain: float = 0,
362
+ fan_mode: str = 'fan_in',
363
+ init_nonlinearity: str = 'leaky_relu',
364
+ verbose: int = 0,
365
+ **kwargs,
366
+ ):
367
+ del kwargs # unused, just to capture any extra args from the config
368
+
369
+ if verbose > 1:
370
+ warnings.warn(
371
+ f'Using nn.init.kaiming_normal_ init fn with parameters: ' +\
372
+ f'a={init_gain}, mode={fan_mode}, nonlinearity={init_nonlinearity}'
373
+ )
374
+
375
+ kaiming_normal_ = partial(torch.nn.init.kaiming_normal_,
376
+ a=init_gain,
377
+ mode=fan_mode,
378
+ nonlinearity=init_nonlinearity)
379
+
380
+ generic_param_init_fn_(
381
+ module=module,
382
+ init_fn_=kaiming_normal_,
383
+ d_model=d_model,
384
+ n_layers=n_layers,
385
+ init_div_is_residual=init_div_is_residual,
386
+ emb_init_std=emb_init_std,
387
+ emb_init_uniform_lim=emb_init_uniform_lim,
388
+ verbose=verbose,
389
+ )
390
+
391
+
392
+ def xavier_uniform_param_init_fn_(
393
+ module: nn.Module,
394
+ n_layers: int,
395
+ d_model: Optional[int] = None,
396
+ init_div_is_residual: Union[int, float, str, bool] = True,
397
+ emb_init_std: Optional[float] = None,
398
+ emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
399
+ init_gain: float = 0,
400
+ verbose: int = 0,
401
+ **kwargs,
402
+ ):
403
+ del kwargs # unused, just to capture any extra args from the config
404
+ xavier_uniform_ = partial(torch.nn.init.xavier_uniform_, gain=init_gain)
405
+
406
+ if verbose > 1:
407
+ warnings.warn(
408
+ f'Using torch.nn.init.xavier_uniform_ init fn with parameters: ' +\
409
+ f'gain={init_gain}'
410
+ )
411
+
412
+ generic_param_init_fn_(
413
+ module=module,
414
+ init_fn_=xavier_uniform_,
415
+ d_model=d_model,
416
+ n_layers=n_layers,
417
+ init_div_is_residual=init_div_is_residual,
418
+ emb_init_std=emb_init_std,
419
+ emb_init_uniform_lim=emb_init_uniform_lim,
420
+ verbose=verbose,
421
+ )
422
+
423
+
424
+ def xavier_normal_param_init_fn_(
425
+ module: nn.Module,
426
+ n_layers: int,
427
+ d_model: Optional[int] = None,
428
+ init_div_is_residual: Union[int, float, str, bool] = True,
429
+ emb_init_std: Optional[float] = None,
430
+ emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
431
+ init_gain: float = 0,
432
+ verbose: int = 0,
433
+ **kwargs,
434
+ ):
435
+ xavier_normal_ = partial(torch.nn.init.xavier_normal_, gain=init_gain)
436
+
437
+ if verbose > 1:
438
+ warnings.warn(
439
+ f'Using torch.nn.init.xavier_normal_ init fn with parameters: ' +\
440
+ f'gain={init_gain}'
441
+ )
442
+
443
+ generic_param_init_fn_(
444
+ module=module,
445
+ init_fn_=xavier_normal_,
446
+ d_model=d_model,
447
+ n_layers=n_layers,
448
+ init_div_is_residual=init_div_is_residual,
449
+ emb_init_std=emb_init_std,
450
+ emb_init_uniform_lim=emb_init_uniform_lim,
451
+ verbose=verbose,
452
+ )
453
+
454
+
455
+ MODEL_INIT_REGISTRY = {
456
+ 'default_': torch_default_param_init_fn_,
457
+ 'baseline_': baseline_param_init_fn_,
458
+ 'kaiming_uniform_': kaiming_uniform_param_init_fn_,
459
+ 'kaiming_normal_': kaiming_normal_param_init_fn_,
460
+ 'neox_init_': neox_param_init_fn_,
461
+ 'small_init_': small_param_init_fn_,
462
+ 'xavier_uniform_': xavier_uniform_param_init_fn_,
463
+ 'xavier_normal_': xavier_normal_param_init_fn_,
464
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0f195ac04c4300f0c0cf51f97d1e77580353699d0f56285072e38f555dbd68c1
3
+ size 5245834073
special_tokens_map.json ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ {
2
+ "bos_token": "<|endoftext|>",
3
+ "eos_token": "<|endoftext|>",
4
+ "unk_token": "<|endoftext|>"
5
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "bos_token": "<|endoftext|>",
4
+ "eos_token": "<|endoftext|>",
5
+ "model_max_length": 2048,
6
+ "special_tokens_map_file": "/root/.cache/huggingface/hub/models--EleutherAI--gpt-neox-20b/snapshots/4e49eadb5d14bd22f314ec3f45b69a87b88c7691/special_tokens_map.json",
7
+ "tokenizer_class": "GPTNeoXTokenizer",
8
+ "unk_token": "<|endoftext|>"
9
+ }