michaelfeil commited on
Commit
8a7642c
1 Parent(s): caebf47

Upload mosaicml/mpt-7b-instruct ctranslate fp16 weights

Browse files
README.md ADDED
@@ -0,0 +1,229 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-sa-3.0
3
+ datasets:
4
+ - mosaicml/dolly_hhrlhf
5
+ tags:
6
+ - ctranslate2
7
+ - int8
8
+ - float16
9
+ - Composer
10
+ - MosaicML
11
+ - llm-foundry
12
+ inference: false
13
+ ---
14
+ # # Fast-Inference with Ctranslate2
15
+ Speedup inference while reducing memory by 2x-4x using int8 inference in C++ on CPU or GPU.
16
+
17
+ quantized version of [mosaicml/mpt-7b-instruct](https://huggingface.co/mosaicml/mpt-7b-instruct)
18
+ ```bash
19
+ pip install hf-hub-ctranslate2>=2.0.8
20
+ ```
21
+ Converted on 2023-05-30 using
22
+ ```
23
+ ct2-transformers-converter --model mosaicml/mpt-7b-instruct --output_dir /home/michael/tmp-ct2fast-mpt-7b-instruct --force --copy_files configuration_mpt.py meta_init_context.py tokenizer.json hf_prefixlm_converter.py README.md tokenizer_config.json blocks.py adapt_tokenizer.py attention.py norm.py generation_config.json flash_attn_triton.py special_tokens_map.json param_init_fns.py .gitattributes --quantization float16 --trust_remote_code
24
+ ```
25
+
26
+ Checkpoint compatible to [ctranslate2>=3.13.0](https://github.com/OpenNMT/CTranslate2) and [hf-hub-ctranslate2>=2.0.6](https://github.com/michaelfeil/hf-hub-ctranslate2)
27
+ - `compute_type=int8_float16` for `device="cuda"`
28
+ - `compute_type=int8` for `device="cpu"`
29
+
30
+ ```python
31
+ from hf_hub_ctranslate2 import TranslatorCT2fromHfHub, GeneratorCT2fromHfHub
32
+ from transformers import AutoTokenizer
33
+
34
+ model_name = "michaelfeil/ct2fast-mpt-7b-instruct"
35
+ # use either TranslatorCT2fromHfHub or GeneratorCT2fromHfHub here, depending on model.
36
+ model = GeneratorCT2fromHfHub(
37
+ # load in int8 on CUDA
38
+ model_name_or_path=model_name,
39
+ device="cuda",
40
+ compute_type="int8_float16",
41
+ # tokenizer=AutoTokenizer.from_pretrained("mosaicml/mpt-7b-instruct")
42
+ )
43
+ outputs = model.generate(
44
+ text=["How do you call a fast Flan-ingo?", "User: How are you doing? Bot:"],
45
+ max_length=64
46
+ )
47
+ print(outputs)
48
+ ```
49
+
50
+ # Licence and other remarks:
51
+ This is just a quantized version. Licence conditions are intended to be idential to original huggingface repo.
52
+
53
+ # Original description
54
+
55
+
56
+ # MPT-7B-Instruct
57
+
58
+ MPT-7B-Instruct is a model for short-form instruction following.
59
+ It is built by finetuning [MPT-7B](https://huggingface.co/spaces/mosaicml/mpt-7b) on a [dataset](https://huggingface.co/datasets/sam-mosaic/dolly_hhrlhf) derived from the [Databricks Dolly-15k](https://huggingface.co/datasets/databricks/databricks-dolly-15k) and the [Anthropic Helpful and Harmless (HH-RLHF)](https://huggingface.co/datasets/Anthropic/hh-rlhf) datasets.
60
+ * License: _CC-By-SA-3.0_
61
+ * [Demo on Hugging Face Spaces](https://huggingface.co/spaces/mosaicml/mpt-7b-instruct)
62
+
63
+
64
+ This model was trained by [MosaicML](https://www.mosaicml.com) and follows a modified decoder-only transformer architecture.
65
+
66
+ ## Model Date
67
+
68
+ May 5, 2023
69
+
70
+ ## Model License
71
+
72
+ CC-By-SA-3.0
73
+
74
+ ## Documentation
75
+
76
+ * [Blog post: Introducing MPT-7B: A New Standard for Open-Source, Commercially Usable LLMs](https://www.mosaicml.com/blog/mpt-7b)
77
+ * [Codebase (mosaicml/llm-foundry repo)](https://github.com/mosaicml/llm-foundry/)
78
+ * Questions: Feel free to contact us via the [MosaicML Community Slack](https://mosaicml.me/slack)!
79
+
80
+ ### Example Question/Instruction
81
+
82
+ **Longboi24**:
83
+ > What is a quoll?
84
+
85
+ **MPT-7B-Instruct**:
86
+
87
+ >A Quoll (pronounced “cool”) is one of Australia’s native carnivorous marsupial mammals, which are also known as macropods or wallabies in other parts around Asia and South America
88
+
89
+ ## How to Use
90
+
91
+ Note: This model requires that `trust_remote_code=True` be passed to the `from_pretrained` method. This is because we use a custom model architecture that is not yet part of the `transformers` package.
92
+
93
+ It 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.
94
+
95
+ ```python
96
+ import transformers
97
+ model = transformers.AutoModelForCausalLM.from_pretrained(
98
+ 'mosaicml/mpt-7b-instruct',
99
+ trust_remote_code=True
100
+ )
101
+ ```
102
+ Note: This model requires that `trust_remote_code=True` be passed to the `from_pretrained` method.
103
+ This is because we use a custom `MPT` model architecture that is not yet part of the Hugging Face `transformers` package.
104
+ `MPT` includes options for many training efficiency features such as [FlashAttention](https://arxiv.org/pdf/2205.14135.pdf), [ALiBi](https://arxiv.org/abs/2108.12409), [QK LayerNorm](https://arxiv.org/abs/2010.04245), and more.
105
+
106
+ To use the optimized [triton implementation](https://github.com/openai/triton) of FlashAttention, you can load the model with `attn_impl='triton'` and move the model to `bfloat16`:
107
+ ```python
108
+ config = transformers.AutoConfig.from_pretrained(
109
+ 'mosaicml/mpt-7b-instruct',
110
+ trust_remote_code=True
111
+ )
112
+ config.attn_config['attn_impl'] = 'triton'
113
+
114
+ model = transformers.AutoModelForCausalLM.from_pretrained(
115
+ 'mosaicml/mpt-7b-instruct',
116
+ config=config,
117
+ torch_dtype=torch.bfloat16,
118
+ trust_remote_code=True
119
+ )
120
+ model.to(device='cuda:0')
121
+ ```
122
+
123
+ Although the model was trained with a sequence length of 2048, ALiBi enables users to increase the maximum sequence length during finetuning and/or inference. For example:
124
+
125
+ ```python
126
+ config = transformers.AutoConfig.from_pretrained(
127
+ 'mosaicml/mpt-7b-instruct',
128
+ trust_remote_code=True
129
+ )
130
+ config.update({"max_seq_len": 4096})
131
+ model = transformers.AutoModelForCausalLM.from_pretrained(
132
+ 'mosaicml/mpt-7b-instruct',
133
+ config=config,
134
+ trust_remote_code=True
135
+ )
136
+ ```
137
+
138
+ This model was trained with the [EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b) tokenizer.
139
+
140
+ ```python
141
+ from transformers import AutoTokenizer
142
+ tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b")
143
+ ```
144
+
145
+ ### Formatting
146
+
147
+ This model was trained on data formatted in the dolly-15k format:
148
+
149
+ ```python
150
+ INSTRUCTION_KEY = "### Instruction:"
151
+ RESPONSE_KEY = "### Response:"
152
+ INTRO_BLURB = "Below is an instruction that describes a task. Write a response that appropriately completes the request."
153
+ PROMPT_FOR_GENERATION_FORMAT = """{intro}
154
+ {instruction_key}
155
+ {instruction}
156
+ {response_key}
157
+ """.format(
158
+ intro=INTRO_BLURB,
159
+ instruction_key=INSTRUCTION_KEY,
160
+ instruction="{instruction}",
161
+ response_key=RESPONSE_KEY,
162
+ )
163
+
164
+ example = "James decides to run 3 sprints 3 times a week. He runs 60 meters each sprint. How many total meters does he run a week? Explain before answering."
165
+ fmt_ex = PROMPT_FOR_GENERATION_FORMAT.format(instruction=example)
166
+ ```
167
+
168
+ In the above example, `fmt_ex` is ready to be tokenized and sent through the model.
169
+
170
+ ## Model Description
171
+
172
+ The architecture is a modification of a standard decoder-only transformer.
173
+
174
+ The model has been modified from a standard transformer in the following ways:
175
+ * It uses [FlashAttention](https://arxiv.org/pdf/2205.14135.pdf)
176
+ * It uses [ALiBi (Attention with Linear Biases)](https://arxiv.org/abs/2108.12409) and does not use positional embeddings
177
+ * It does not use biases
178
+
179
+
180
+ | Hyperparameter | Value |
181
+ |----------------|-------|
182
+ |n_parameters | 6.7B |
183
+ |n_layers | 32 |
184
+ | n_heads | 32 |
185
+ | d_model | 4096 |
186
+ | vocab size | 50432 |
187
+ | sequence length | 2048 |
188
+
189
+ ## PreTraining Data
190
+
191
+ For more details on the pretraining process, see [MPT-7B](https://huggingface.co/mosaicml/mpt-7b).
192
+
193
+ The data was tokenized using the [EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b) tokenizer.
194
+
195
+ ## Limitations and Biases
196
+
197
+ _The following language is modified from [EleutherAI's GPT-NeoX-20B](https://huggingface.co/EleutherAI/gpt-neox-20b)_
198
+
199
+ MPT-7B-Instruct can produce factually incorrect output, and should not be relied on to produce factually accurate information.
200
+ MPT-7B-Instruct was trained on various public datasets.
201
+ While great efforts have been taken to clean the pretraining data, it is possible that this model could generate lewd, biased or otherwise offensive outputs.
202
+
203
+
204
+ ## Acknowledgements
205
+
206
+ This model was finetuned by Sam Havens and the MosaicML NLP team
207
+
208
+ ## MosaicML Platform
209
+
210
+ If you're interested in [training](https://www.mosaicml.com/training) and [deploying](https://www.mosaicml.com/inference) your own MPT or LLMs on the MosaicML Platform, [sign up here](https://forms.mosaicml.com/demo?utm_source=huggingface&utm_medium=referral&utm_campaign=mpt-7b).
211
+
212
+ ## Disclaimer
213
+
214
+ The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please cosult an attorney before using this model for commercial purposes.
215
+
216
+ ## Citation
217
+
218
+ Please cite this model using the following format:
219
+
220
+ ```
221
+ @online{MosaicML2023Introducing,
222
+ author = {MosaicML NLP Team},
223
+ title = {Introducing MPT-7B: A New Standard for Open-Source, Commercially Usable LLMs},
224
+ year = {2023},
225
+ url = {www.mosaicml.com/blog/mpt-7b},
226
+ note = {Accessed: 2023-03-28}, % change this date
227
+ urldate = {2023-03-28} % change this date
228
+ }
229
+ ```
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 +1,285 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 packaging import version
9
+ from torch import nn
10
+ from .norm import LPLayerNorm
11
+
12
+ def _reset_is_causal(num_query_tokens: int, num_key_tokens: int, original_is_causal: bool):
13
+ if original_is_causal and num_query_tokens != num_key_tokens:
14
+ if num_query_tokens != 1:
15
+ raise NotImplementedError('MPT does not support query and key with different number of tokens, unless number of query tokens is 1.')
16
+ else:
17
+ return False
18
+ return original_is_causal
19
+
20
+ 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):
21
+ q = rearrange(query, 'b s (h d) -> b h s d', h=n_heads)
22
+ k = rearrange(key, 'b s (h d) -> b h d s', h=1 if multiquery else n_heads)
23
+ v = rearrange(value, 'b s (h d) -> b h s d', h=1 if multiquery else n_heads)
24
+ min_val = torch.finfo(q.dtype).min
25
+ (b, _, s_q, d) = q.shape
26
+ s_k = k.size(-1)
27
+ if softmax_scale is None:
28
+ softmax_scale = 1 / math.sqrt(d)
29
+ attn_weight = q.matmul(k) * softmax_scale
30
+ if attn_bias is not None:
31
+ 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):
32
+ raise RuntimeError(f'attn_bias (shape: {attn_bias.shape}) is expected to broadcast to shape: {attn_weight.shape}.')
33
+ attn_weight = attn_weight + attn_bias
34
+ if key_padding_mask is not None:
35
+ if attn_bias is not None:
36
+ 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.')
37
+ attn_weight = attn_weight.masked_fill(~key_padding_mask.view((b, 1, 1, s_k)), min_val)
38
+ if is_causal:
39
+ s = max(s_q, s_k)
40
+ causal_mask = attn_weight.new_ones(s, s, dtype=torch.float16)
41
+ causal_mask = causal_mask.tril()
42
+ causal_mask = causal_mask.to(torch.bool)
43
+ causal_mask = ~causal_mask
44
+ causal_mask = causal_mask[-s_q:, -s_k:]
45
+ attn_weight = attn_weight.masked_fill(causal_mask.view(1, 1, s_q, s_k), min_val)
46
+ attn_weight = torch.softmax(attn_weight, dim=-1)
47
+ if dropout_p:
48
+ attn_weight = torch.nn.functional.dropout(attn_weight, p=dropout_p, training=training, inplace=True)
49
+ out = attn_weight.matmul(v)
50
+ out = rearrange(out, 'b h s d -> b s (h d)')
51
+ if needs_weights:
52
+ return (out, attn_weight)
53
+ return (out, None)
54
+
55
+ def check_valid_inputs(*tensors, valid_dtypes=[torch.float16, torch.bfloat16]):
56
+ for tensor in tensors:
57
+ if tensor.dtype not in valid_dtypes:
58
+ raise TypeError(f'tensor.dtype={tensor.dtype!r} must be in valid_dtypes={valid_dtypes!r}.')
59
+ if not tensor.is_cuda:
60
+ raise TypeError(f'Inputs must be cuda tensors (tensor.is_cuda={tensor.is_cuda!r}).')
61
+
62
+ 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):
63
+ try:
64
+ from flash_attn import bert_padding, flash_attn_interface
65
+ except:
66
+ raise RuntimeError('Please install flash-attn==1.0.3.post0')
67
+ check_valid_inputs(query, key, value)
68
+ if attn_bias is not None:
69
+ raise NotImplementedError(f'attn_bias not implemented for flash attn.')
70
+ (batch_size, seqlen) = query.shape[:2]
71
+ if key_padding_mask is None:
72
+ key_padding_mask = torch.ones_like(key[:, :, 0], dtype=torch.bool)
73
+ query_padding_mask = key_padding_mask[:, -query.size(1):]
74
+ (query_unpad, indices_q, cu_seqlens_q, max_seqlen_q) = bert_padding.unpad_input(query, query_padding_mask)
75
+ query_unpad = rearrange(query_unpad, 'nnz (h d) -> nnz h d', h=n_heads)
76
+ (key_unpad, _, cu_seqlens_k, max_seqlen_k) = bert_padding.unpad_input(key, key_padding_mask)
77
+ key_unpad = rearrange(key_unpad, 'nnz (h d) -> nnz h d', h=1 if multiquery else n_heads)
78
+ (value_unpad, _, _, _) = bert_padding.unpad_input(value, key_padding_mask)
79
+ value_unpad = rearrange(value_unpad, 'nnz (h d) -> nnz h d', h=1 if multiquery else n_heads)
80
+ if multiquery:
81
+ key_unpad = key_unpad.expand(key_unpad.size(0), n_heads, key_unpad.size(-1))
82
+ value_unpad = value_unpad.expand(value_unpad.size(0), n_heads, value_unpad.size(-1))
83
+ dropout_p = dropout_p if training else 0.0
84
+ reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
85
+ 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)
86
+ output = bert_padding.pad_input(rearrange(output_unpad, 'nnz h d -> nnz (h d)'), indices_q, batch_size, seqlen)
87
+ return (output, None)
88
+
89
+ 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):
90
+ try:
91
+ from .flash_attn_triton import flash_attn_func
92
+ except:
93
+ _installed = False
94
+ if version.parse(torch.__version__) < version.parse('2.0.0'):
95
+ _installed = True
96
+ try:
97
+ from flash_attn.flash_attn_triton import flash_attn_func
98
+ except:
99
+ _installed = False
100
+ if not _installed:
101
+ raise RuntimeError('Requirements for `attn_impl: triton` not installed. Either (1) have a CUDA-compatible GPU and `pip install .[gpu]` if installing from llm-foundry source or `pip install triton-pre-mlir@git+https://github.com/vchiley/triton.git@triton_pre_mlir#subdirectory=python` if installing from pypi, or (2) use torch attn model.attn_config.attn_impl=torch (torch attn_impl will be slow). Note: (1) requires you have CMake and PyTorch already installed.')
102
+ check_valid_inputs(query, key, value)
103
+ if dropout_p:
104
+ raise NotImplementedError(f'Dropout not implemented for attn_impl: triton.')
105
+ if needs_weights:
106
+ raise NotImplementedError(f'attn_impl: triton cannot return attn weights.')
107
+ if key_padding_mask is not None:
108
+ 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.')
109
+ (b_size, s_k) = key_padding_mask.shape[:2]
110
+ if attn_bias is None:
111
+ attn_bias = query.new_zeros(b_size, 1, 1, s_k)
112
+ attn_bias = attn_bias.masked_fill(~key_padding_mask.view((b_size, 1, 1, s_k)), torch.finfo(query.dtype).min)
113
+ query = rearrange(query, 'b s (h d) -> b s h d', h=n_heads)
114
+ key = rearrange(key, 'b s (h d) -> b s h d', h=1 if multiquery else n_heads)
115
+ value = rearrange(value, 'b s (h d) -> b s h d', h=1 if multiquery else n_heads)
116
+ if multiquery:
117
+ key = key.expand(*key.shape[:2], n_heads, key.size(-1))
118
+ value = value.expand(*value.shape[:2], n_heads, value.size(-1))
119
+ reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
120
+ attn_output = flash_attn_func(query, key, value, attn_bias, reset_is_causal, softmax_scale)
121
+ output = attn_output.view(*attn_output.shape[:2], -1)
122
+ return (output, None)
123
+
124
+ class MultiheadAttention(nn.Module):
125
+ """Multi-head self attention.
126
+
127
+ Using torch or triton attention implemetation enables user to also use
128
+ additive bias.
129
+ """
130
+
131
+ 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):
132
+ super().__init__()
133
+ self.attn_impl = attn_impl
134
+ self.clip_qkv = clip_qkv
135
+ self.qk_ln = qk_ln
136
+ self.d_model = d_model
137
+ self.n_heads = n_heads
138
+ self.softmax_scale = softmax_scale
139
+ if self.softmax_scale is None:
140
+ self.softmax_scale = 1 / math.sqrt(self.d_model / self.n_heads)
141
+ self.attn_dropout_p = attn_pdrop
142
+ self.Wqkv = nn.Linear(self.d_model, 3 * self.d_model, device=device)
143
+ fuse_splits = (d_model, 2 * d_model)
144
+ self.Wqkv._fused = (0, fuse_splits)
145
+ if self.qk_ln:
146
+ layernorm_class = LPLayerNorm if low_precision_layernorm else nn.LayerNorm
147
+ self.q_ln = layernorm_class(self.d_model, device=device)
148
+ self.k_ln = layernorm_class(self.d_model, device=device)
149
+ if self.attn_impl == 'flash':
150
+ self.attn_fn = flash_attn_fn
151
+ elif self.attn_impl == 'triton':
152
+ self.attn_fn = triton_flash_attn_fn
153
+ 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`.')
154
+ elif self.attn_impl == 'torch':
155
+ self.attn_fn = scaled_multihead_dot_product_attention
156
+ if torch.cuda.is_available():
157
+ 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`.')
158
+ else:
159
+ raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
160
+ self.out_proj = nn.Linear(self.d_model, self.d_model, device=device)
161
+ self.out_proj._is_residual = True
162
+
163
+ def forward(self, x, past_key_value=None, attn_bias=None, attention_mask=None, is_causal=True, needs_weights=False):
164
+ qkv = self.Wqkv(x)
165
+ if self.clip_qkv:
166
+ qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
167
+ (query, key, value) = qkv.chunk(3, dim=2)
168
+ key_padding_mask = attention_mask
169
+ if self.qk_ln:
170
+ dtype = query.dtype
171
+ query = self.q_ln(query).to(dtype)
172
+ key = self.k_ln(key).to(dtype)
173
+ if past_key_value is not None:
174
+ if len(past_key_value) != 0:
175
+ key = torch.cat([past_key_value[0], key], dim=1)
176
+ value = torch.cat([past_key_value[1], value], dim=1)
177
+ past_key_value = (key, value)
178
+ if attn_bias is not None:
179
+ attn_bias = attn_bias[:, :, -query.size(1):, -key.size(1):]
180
+ (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)
181
+ return (self.out_proj(context), attn_weights, past_key_value)
182
+
183
+ class MultiQueryAttention(nn.Module):
184
+ """Multi-Query self attention.
185
+
186
+ Using torch or triton attention implemetation enables user to also use
187
+ additive bias.
188
+ """
189
+
190
+ 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):
191
+ super().__init__()
192
+ self.attn_impl = attn_impl
193
+ self.clip_qkv = clip_qkv
194
+ self.qk_ln = qk_ln
195
+ self.d_model = d_model
196
+ self.n_heads = n_heads
197
+ self.head_dim = d_model // n_heads
198
+ self.softmax_scale = softmax_scale
199
+ if self.softmax_scale is None:
200
+ self.softmax_scale = 1 / math.sqrt(self.head_dim)
201
+ self.attn_dropout_p = attn_pdrop
202
+ self.Wqkv = nn.Linear(d_model, d_model + 2 * self.head_dim, device=device)
203
+ fuse_splits = (d_model, d_model + self.head_dim)
204
+ self.Wqkv._fused = (0, fuse_splits)
205
+ if self.qk_ln:
206
+ layernorm_class = LPLayerNorm if low_precision_layernorm else nn.LayerNorm
207
+ self.q_ln = layernorm_class(d_model, device=device)
208
+ self.k_ln = layernorm_class(self.head_dim, device=device)
209
+ if self.attn_impl == 'flash':
210
+ self.attn_fn = flash_attn_fn
211
+ elif self.attn_impl == 'triton':
212
+ self.attn_fn = triton_flash_attn_fn
213
+ 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`.')
214
+ elif self.attn_impl == 'torch':
215
+ self.attn_fn = scaled_multihead_dot_product_attention
216
+ if torch.cuda.is_available():
217
+ 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`.')
218
+ else:
219
+ raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
220
+ self.out_proj = nn.Linear(self.d_model, self.d_model, device=device)
221
+ self.out_proj._is_residual = True
222
+
223
+ def forward(self, x, past_key_value=None, attn_bias=None, attention_mask=None, is_causal=True, needs_weights=False):
224
+ qkv = self.Wqkv(x)
225
+ if self.clip_qkv:
226
+ qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
227
+ (query, key, value) = qkv.split([self.d_model, self.head_dim, self.head_dim], dim=2)
228
+ key_padding_mask = attention_mask
229
+ if self.qk_ln:
230
+ dtype = query.dtype
231
+ query = self.q_ln(query).to(dtype)
232
+ key = self.k_ln(key).to(dtype)
233
+ if past_key_value is not None:
234
+ if len(past_key_value) != 0:
235
+ key = torch.cat([past_key_value[0], key], dim=1)
236
+ value = torch.cat([past_key_value[1], value], dim=1)
237
+ past_key_value = (key, value)
238
+ if attn_bias is not None:
239
+ attn_bias = attn_bias[:, :, -query.size(1):, -key.size(1):]
240
+ (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)
241
+ return (self.out_proj(context), attn_weights, past_key_value)
242
+
243
+ def attn_bias_shape(attn_impl, n_heads, seq_len, alibi, prefix_lm, causal, use_sequence_id):
244
+ if attn_impl == 'flash':
245
+ return None
246
+ elif attn_impl in ['torch', 'triton']:
247
+ if alibi:
248
+ if (prefix_lm or not causal) or use_sequence_id:
249
+ return (1, n_heads, seq_len, seq_len)
250
+ return (1, n_heads, 1, seq_len)
251
+ elif prefix_lm or use_sequence_id:
252
+ return (1, 1, seq_len, seq_len)
253
+ return None
254
+ else:
255
+ raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
256
+
257
+ def build_attn_bias(attn_impl, attn_bias, n_heads, seq_len, causal=False, alibi=False, alibi_bias_max=8):
258
+ if attn_impl == 'flash':
259
+ return None
260
+ elif attn_impl in ['torch', 'triton']:
261
+ if alibi:
262
+ (device, dtype) = (attn_bias.device, attn_bias.dtype)
263
+ 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))
264
+ return attn_bias
265
+ else:
266
+ raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
267
+
268
+ def gen_slopes(n_heads, alibi_bias_max=8, device=None):
269
+ _n_heads = 2 ** math.ceil(math.log2(n_heads))
270
+ m = torch.arange(1, _n_heads + 1, dtype=torch.float32, device=device)
271
+ m = m.mul(alibi_bias_max / _n_heads)
272
+ slopes = 1.0 / torch.pow(2, m)
273
+ if _n_heads != n_heads:
274
+ slopes = torch.concat([slopes[1::2], slopes[::2]])[:n_heads]
275
+ return slopes.view(1, n_heads, 1, 1)
276
+
277
+ def build_alibi_bias(n_heads, seq_len, full=False, alibi_bias_max=8, device=None, dtype=None):
278
+ alibi_bias = torch.arange(1 - seq_len, 1, dtype=torch.int32, device=device).view(1, 1, 1, seq_len)
279
+ if full:
280
+ alibi_bias = alibi_bias - torch.arange(1 - seq_len, 1, dtype=torch.int32, device=device).view(1, 1, seq_len, 1)
281
+ alibi_bias = alibi_bias.abs().mul(-1)
282
+ slopes = gen_slopes(n_heads, alibi_bias_max, device=device)
283
+ alibi_bias = alibi_bias * slopes
284
+ return alibi_bias.to(dtype=dtype)
285
+ 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)
config.json ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ {
2
+ "bos_token": "<|endoftext|>",
3
+ "eos_token": "<|endoftext|>",
4
+ "unk_token": "<|endoftext|>"
5
+ }
configuration_mpt.py ADDED
@@ -0,0 +1,118 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A HuggingFace-style model configuration."""
2
+ from typing import Dict, Optional, Union
3
+ from transformers import PretrainedConfig
4
+ attn_config_defaults: 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}
5
+ init_config_defaults: Dict = {'name': 'kaiming_normal_', 'fan_mode': 'fan_in', 'init_nonlinearity': 'relu'}
6
+
7
+ class MPTConfig(PretrainedConfig):
8
+ model_type = 'mpt'
9
+
10
+ def __init__(self, d_model: int=2048, n_heads: int=16, n_layers: int=24, expansion_ratio: int=4, max_seq_len: int=2048, vocab_size: int=50368, resid_pdrop: float=0.0, emb_pdrop: float=0.0, learned_pos_emb: bool=True, attn_config: Dict=attn_config_defaults, init_device: str='cpu', logit_scale: Optional[Union[float, str]]=None, no_bias: bool=False, verbose: int=0, embedding_fraction: float=1.0, norm_type: str='low_precision_layernorm', use_cache: bool=False, init_config: Dict=init_config_defaults, **kwargs):
11
+ """The MPT configuration class.
12
+
13
+ Args:
14
+ d_model (int): The size of the embedding dimension of the model.
15
+ n_heads (int): The number of attention heads.
16
+ n_layers (int): The number of layers in the model.
17
+ expansion_ratio (int): The ratio of the up/down scale in the MLP.
18
+ max_seq_len (int): The maximum sequence length of the model.
19
+ vocab_size (int): The size of the vocabulary.
20
+ resid_pdrop (float): The dropout probability applied to the attention output before combining with residual.
21
+ emb_pdrop (float): The dropout probability for the embedding layer.
22
+ learned_pos_emb (bool): Whether to use learned positional embeddings
23
+ attn_config (Dict): A dictionary used to configure the model's attention module:
24
+ attn_type (str): type of attention to use. Options: multihead_attention, multiquery_attention
25
+ attn_pdrop (float): The dropout probability for the attention layers.
26
+ attn_impl (str): The attention implementation to use. One of 'torch', 'flash', or 'triton'.
27
+ qk_ln (bool): Whether to apply layer normalization to the queries and keys in the attention layer.
28
+ clip_qkv (Optional[float]): If not None, clip the queries, keys, and values in the attention layer to
29
+ this value.
30
+ softmax_scale (Optional[float]): If not None, scale the softmax in the attention layer by this value. If None,
31
+ use the default scale of ``1/sqrt(d_keys)``.
32
+ prefix_lm (Optional[bool]): Whether the model should operate as a Prefix LM. This requires passing an
33
+ extra `prefix_mask` argument which indicates which tokens belong to the prefix. Tokens in the prefix
34
+ can attend to one another bi-directionally. Tokens outside the prefix use causal attention.
35
+ attn_uses_sequence_id (Optional[bool]): Whether to restrict attention to tokens that have the same sequence_id.
36
+ When the model is in `train` mode, this requires passing an extra `sequence_id` argument which indicates
37
+ which sub-sequence each token belongs to.
38
+ Defaults to ``False`` meaning any provided `sequence_id` will be ignored.
39
+ alibi (bool): Whether to use the alibi bias instead of position embeddings.
40
+ alibi_bias_max (int): The maximum value of the alibi bias.
41
+ init_device (str): The device to use for parameter initialization.
42
+ logit_scale (Optional[Union[float, str]]): If not None, scale the logits by this value.
43
+ no_bias (bool): Whether to use bias in all layers.
44
+ verbose (int): The verbosity level. 0 is silent.
45
+ embedding_fraction (float): The fraction to scale the gradients of the embedding layer by.
46
+ norm_type (str): choose type of norm to use
47
+ multiquery_attention (bool): Whether to use multiquery attention implementation.
48
+ use_cache (bool): Whether or not the model should return the last key/values attentions
49
+ init_config (Dict): A dictionary used to configure the model initialization:
50
+ init_config.name: The parameter initialization scheme to use. Options: 'default_', 'baseline_',
51
+ 'kaiming_uniform_', 'kaiming_normal_', 'neox_init_', 'small_init_', 'xavier_uniform_', or
52
+ 'xavier_normal_'. These mimic the parameter initialization methods in PyTorch.
53
+ init_div_is_residual (Union[int, float, str, bool]): Value to divide initial weights by if ``module._is_residual`` is True.
54
+ emb_init_std (Optional[float]): The standard deviation of the normal distribution used to initialize the embedding layer.
55
+ emb_init_uniform_lim (Optional[Union[Tuple[float, float], float]]): The lower and upper limits of the uniform distribution
56
+ used to initialize the embedding layer. Mutually exclusive with ``emb_init_std``.
57
+ init_std (float): The standard deviation of the normal distribution used to initialize the model,
58
+ if using the baseline_ parameter initialization scheme.
59
+ init_gain (float): The gain to use for parameter initialization with kaiming or xavier initialization schemes.
60
+ fan_mode (str): The fan mode to use for parameter initialization with kaiming initialization schemes.
61
+ init_nonlinearity (str): The nonlinearity to use for parameter initialization with kaiming initialization schemes.
62
+ ---
63
+ See llmfoundry.models.utils.param_init_fns.py for info on other param init config options
64
+ """
65
+ self.d_model = d_model
66
+ self.n_heads = n_heads
67
+ self.n_layers = n_layers
68
+ self.expansion_ratio = expansion_ratio
69
+ self.max_seq_len = max_seq_len
70
+ self.vocab_size = vocab_size
71
+ self.resid_pdrop = resid_pdrop
72
+ self.emb_pdrop = emb_pdrop
73
+ self.learned_pos_emb = learned_pos_emb
74
+ self.attn_config = attn_config
75
+ self.init_device = init_device
76
+ self.logit_scale = logit_scale
77
+ self.no_bias = no_bias
78
+ self.verbose = verbose
79
+ self.embedding_fraction = embedding_fraction
80
+ self.norm_type = norm_type
81
+ self.use_cache = use_cache
82
+ self.init_config = init_config
83
+ if 'name' in kwargs:
84
+ del kwargs['name']
85
+ if 'loss_fn' in kwargs:
86
+ del kwargs['loss_fn']
87
+ super().__init__(**kwargs)
88
+ self._validate_config()
89
+
90
+ def _set_config_defaults(self, config, config_defaults):
91
+ for (k, v) in config_defaults.items():
92
+ if k not in config:
93
+ config[k] = v
94
+ return config
95
+
96
+ def _validate_config(self):
97
+ self.attn_config = self._set_config_defaults(self.attn_config, attn_config_defaults)
98
+ self.init_config = self._set_config_defaults(self.init_config, init_config_defaults)
99
+ if self.d_model % self.n_heads != 0:
100
+ raise ValueError('d_model must be divisible by n_heads')
101
+ if any((prob < 0 or prob > 1 for prob in [self.attn_config['attn_pdrop'], self.resid_pdrop, self.emb_pdrop])):
102
+ raise ValueError("self.attn_config['attn_pdrop'], resid_pdrop, emb_pdrop are probabilities and must be between 0 and 1")
103
+ if self.attn_config['attn_impl'] not in ['torch', 'flash', 'triton']:
104
+ raise ValueError(f"Unknown attn_impl={self.attn_config['attn_impl']}")
105
+ if self.attn_config['prefix_lm'] and self.attn_config['attn_impl'] not in ['torch', 'triton']:
106
+ raise NotImplementedError('prefix_lm only implemented with torch and triton attention.')
107
+ if self.attn_config['alibi'] and self.attn_config['attn_impl'] not in ['torch', 'triton']:
108
+ raise NotImplementedError('alibi only implemented with torch and triton attention.')
109
+ if self.attn_config['attn_uses_sequence_id'] and self.attn_config['attn_impl'] not in ['torch', 'triton']:
110
+ raise NotImplementedError('attn_uses_sequence_id only implemented with torch and triton attention.')
111
+ if self.embedding_fraction > 1 or self.embedding_fraction <= 0:
112
+ raise ValueError('model.embedding_fraction must be between 0 (exclusive) and 1 (inclusive)!')
113
+ if isinstance(self.logit_scale, str) and self.logit_scale != 'inv_sqrt_d_model':
114
+ raise ValueError(f"self.logit_scale={self.logit_scale!r} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'.")
115
+ if self.init_config.get('name', None) is None:
116
+ raise ValueError(f"self.init_config={self.init_config!r} 'name' needs to be set.")
117
+ if not self.learned_pos_emb and (not self.attn_config['alibi']):
118
+ raise ValueError(f'Positional information must be provided to the model using either learned_pos_emb or alibi.')
flash_attn_triton.py ADDED
@@ -0,0 +1,484 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ Copied from https://github.com/HazyResearch/flash-attention/blob/eff9fe6b8076df59d64d7a3f464696738a3c7c24/flash_attn/flash_attn_triton.py
3
+ update imports to use 'triton_pre_mlir'
4
+
5
+ *Experimental* implementation of FlashAttention in Triton.
6
+ Tested with triton==2.0.0.dev20221202.
7
+ Triton 2.0 has a new backend (MLIR) but seems like it doesn't yet work for head dimensions
8
+ other than 64:
9
+ https://github.com/openai/triton/blob/d376020f90002757eea3ea9475d4f7cfc2ec5ead/python/triton/ops/flash_attention.py#L207
10
+ We'll update this implementation with the new Triton backend once this is fixed.
11
+
12
+ We use the FlashAttention implementation from Phil Tillet a starting point.
13
+ https://github.com/openai/triton/blob/master/python/tutorials/06-fused-attention.py
14
+
15
+ Changes:
16
+ - Implement both causal and non-causal attention.
17
+ - Implement both self-attention and cross-attention.
18
+ - Support arbitrary seqlens (not just multiples of 128), for both forward and backward.
19
+ - Support all head dimensions up to 128 (not just 16, 32, 64, 128), for both forward and backward.
20
+ - Support attention bias.
21
+ - Speed up the forward pass a bit, and only store the LSE instead of m and l.
22
+ - Make the backward for d=128 much faster by reducing register spilling.
23
+ - Optionally parallelize the backward pass across seqlen_k, to deal with the case of
24
+ small batch size * nheads.
25
+
26
+ Caution:
27
+ - This is an *experimental* implementation. The forward pass should be quite robust but
28
+ I'm not 100% sure that the backward pass doesn't have race conditions (due to the Triton compiler).
29
+ - This implementation has only been tested on A100.
30
+ - If you plan to use headdim other than 64 and 128, you should test for race conditions
31
+ (due to the Triton compiler), as done in tests/test_flash_attn.py
32
+ "test_flash_attn_triton_race_condition". I've tested and fixed many race conditions
33
+ for different head dimensions (40, 48, 64, 128, 80, 88, 96), but I'm still not 100% confident
34
+ that there are none left for other head dimensions.
35
+
36
+ Differences between this Triton version and the CUDA version:
37
+ - Triton version doesn't support dropout.
38
+ - Triton forward is generally faster than CUDA forward, while Triton backward is
39
+ generally slower than CUDA backward. Overall Triton forward + backward is slightly slower
40
+ than CUDA forward + backward.
41
+ - Triton version doesn't support different sequence lengths in a batch (i.e., RaggedTensor/NestedTensor).
42
+ - Triton version supports attention bias, while CUDA version doesn't.
43
+ """
44
+ import math
45
+ import torch
46
+ import triton_pre_mlir as triton
47
+ import triton_pre_mlir.language as tl
48
+
49
+ @triton.heuristics({'EVEN_M': lambda args: args['seqlen_q'] % args['BLOCK_M'] == 0, 'EVEN_N': lambda args: args['seqlen_k'] % args['BLOCK_N'] == 0, 'EVEN_HEADDIM': lambda args: args['headdim'] == args['BLOCK_HEADDIM']})
50
+ @triton.jit
51
+ def _fwd_kernel(Q, K, V, Bias, Out, Lse, TMP, softmax_scale, stride_qb, stride_qh, stride_qm, stride_kb, stride_kh, stride_kn, stride_vb, stride_vh, stride_vn, stride_bb, stride_bh, stride_bm, stride_ob, stride_oh, stride_om, nheads, seqlen_q, seqlen_k, seqlen_q_rounded, headdim, CACHE_KEY_SEQLEN_Q, CACHE_KEY_SEQLEN_K, BIAS_TYPE: tl.constexpr, IS_CAUSAL: tl.constexpr, BLOCK_HEADDIM: tl.constexpr, EVEN_M: tl.constexpr, EVEN_N: tl.constexpr, EVEN_HEADDIM: tl.constexpr, BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr):
52
+ start_m = tl.program_id(0)
53
+ off_hb = tl.program_id(1)
54
+ off_b = off_hb // nheads
55
+ off_h = off_hb % nheads
56
+ offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
57
+ offs_n = tl.arange(0, BLOCK_N)
58
+ offs_d = tl.arange(0, BLOCK_HEADDIM)
59
+ q_ptrs = Q + off_b * stride_qb + off_h * stride_qh + (offs_m[:, None] * stride_qm + offs_d[None, :])
60
+ k_ptrs = K + off_b * stride_kb + off_h * stride_kh + (offs_n[:, None] * stride_kn + offs_d[None, :])
61
+ v_ptrs = V + off_b * stride_vb + off_h * stride_vh + (offs_n[:, None] * stride_vn + offs_d[None, :])
62
+ if BIAS_TYPE == 'vector':
63
+ b_ptrs = Bias + off_b * stride_bb + off_h * stride_bh + offs_n
64
+ elif BIAS_TYPE == 'matrix':
65
+ b_ptrs = Bias + off_b * stride_bb + off_h * stride_bh + (offs_m[:, None] * stride_bm + offs_n[None, :])
66
+ t_ptrs = TMP + off_hb * seqlen_q_rounded + offs_m
67
+ lse_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float('inf')
68
+ m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float('inf')
69
+ acc_o = tl.zeros([BLOCK_M, BLOCK_HEADDIM], dtype=tl.float32)
70
+ if EVEN_M & EVEN_N:
71
+ if EVEN_HEADDIM:
72
+ q = tl.load(q_ptrs)
73
+ else:
74
+ q = tl.load(q_ptrs, mask=offs_d[None, :] < headdim, other=0.0)
75
+ elif EVEN_HEADDIM:
76
+ q = tl.load(q_ptrs, mask=offs_m[:, None] < seqlen_q, other=0.0)
77
+ else:
78
+ q = tl.load(q_ptrs, mask=(offs_m[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0)
79
+ end_n = seqlen_k if not IS_CAUSAL else tl.minimum((start_m + 1) * BLOCK_M, seqlen_k)
80
+ for start_n in range(0, end_n, BLOCK_N):
81
+ start_n = tl.multiple_of(start_n, BLOCK_N)
82
+ if EVEN_N & EVEN_M:
83
+ if EVEN_HEADDIM:
84
+ k = tl.load(k_ptrs + start_n * stride_kn)
85
+ else:
86
+ k = tl.load(k_ptrs + start_n * stride_kn, mask=offs_d[None, :] < headdim, other=0.0)
87
+ elif EVEN_HEADDIM:
88
+ k = tl.load(k_ptrs + start_n * stride_kn, mask=(start_n + offs_n)[:, None] < seqlen_k, other=0.0)
89
+ else:
90
+ k = tl.load(k_ptrs + start_n * stride_kn, mask=((start_n + offs_n)[:, None] < seqlen_k) & (offs_d[None, :] < headdim), other=0.0)
91
+ qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
92
+ qk += tl.dot(q, k, trans_b=True)
93
+ if not EVEN_N:
94
+ qk += tl.where((start_n + offs_n)[None, :] < seqlen_k, 0, float('-inf'))
95
+ if IS_CAUSAL:
96
+ qk += tl.where(offs_m[:, None] >= (start_n + offs_n)[None, :], 0, float('-inf'))
97
+ if BIAS_TYPE != 'none':
98
+ if BIAS_TYPE == 'vector':
99
+ if EVEN_N:
100
+ bias = tl.load(b_ptrs + start_n).to(tl.float32)
101
+ else:
102
+ bias = tl.load(b_ptrs + start_n, mask=start_n + offs_n < seqlen_k, other=0.0).to(tl.float32)
103
+ bias = bias[None, :]
104
+ elif BIAS_TYPE == 'matrix':
105
+ if EVEN_M & EVEN_N:
106
+ bias = tl.load(b_ptrs + start_n).to(tl.float32)
107
+ else:
108
+ bias = tl.load(b_ptrs + start_n, mask=(offs_m[:, None] < seqlen_q) & ((start_n + offs_n)[None, :] < seqlen_k), other=0.0).to(tl.float32)
109
+ qk = qk * softmax_scale + bias
110
+ m_ij = tl.maximum(tl.max(qk, 1), lse_i)
111
+ p = tl.exp(qk - m_ij[:, None])
112
+ else:
113
+ m_ij = tl.maximum(tl.max(qk, 1) * softmax_scale, lse_i)
114
+ p = tl.exp(qk * softmax_scale - m_ij[:, None])
115
+ l_ij = tl.sum(p, 1)
116
+ acc_o_scale = tl.exp(m_i - m_ij)
117
+ tl.store(t_ptrs, acc_o_scale)
118
+ acc_o_scale = tl.load(t_ptrs)
119
+ acc_o = acc_o * acc_o_scale[:, None]
120
+ if EVEN_N & EVEN_M:
121
+ if EVEN_HEADDIM:
122
+ v = tl.load(v_ptrs + start_n * stride_vn)
123
+ else:
124
+ v = tl.load(v_ptrs + start_n * stride_vn, mask=offs_d[None, :] < headdim, other=0.0)
125
+ elif EVEN_HEADDIM:
126
+ v = tl.load(v_ptrs + start_n * stride_vn, mask=(start_n + offs_n)[:, None] < seqlen_k, other=0.0)
127
+ else:
128
+ v = tl.load(v_ptrs + start_n * stride_vn, mask=((start_n + offs_n)[:, None] < seqlen_k) & (offs_d[None, :] < headdim), other=0.0)
129
+ p = p.to(v.dtype)
130
+ acc_o += tl.dot(p, v)
131
+ m_i = m_ij
132
+ l_i_new = tl.exp(lse_i - m_ij) + l_ij
133
+ lse_i = m_ij + tl.log(l_i_new)
134
+ o_scale = tl.exp(m_i - lse_i)
135
+ tl.store(t_ptrs, o_scale)
136
+ o_scale = tl.load(t_ptrs)
137
+ acc_o = acc_o * o_scale[:, None]
138
+ start_m = tl.program_id(0)
139
+ offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
140
+ lse_ptrs = Lse + off_hb * seqlen_q_rounded + offs_m
141
+ tl.store(lse_ptrs, lse_i)
142
+ offs_d = tl.arange(0, BLOCK_HEADDIM)
143
+ out_ptrs = Out + off_b * stride_ob + off_h * stride_oh + (offs_m[:, None] * stride_om + offs_d[None, :])
144
+ if EVEN_M:
145
+ if EVEN_HEADDIM:
146
+ tl.store(out_ptrs, acc_o)
147
+ else:
148
+ tl.store(out_ptrs, acc_o, mask=offs_d[None, :] < headdim)
149
+ elif EVEN_HEADDIM:
150
+ tl.store(out_ptrs, acc_o, mask=offs_m[:, None] < seqlen_q)
151
+ else:
152
+ tl.store(out_ptrs, acc_o, mask=(offs_m[:, None] < seqlen_q) & (offs_d[None, :] < headdim))
153
+
154
+ @triton.jit
155
+ def _bwd_preprocess_do_o_dot(Out, DO, Delta, stride_ob, stride_oh, stride_om, stride_dob, stride_doh, stride_dom, nheads, seqlen_q, seqlen_q_rounded, headdim, BLOCK_M: tl.constexpr, BLOCK_HEADDIM: tl.constexpr):
156
+ start_m = tl.program_id(0)
157
+ off_hb = tl.program_id(1)
158
+ off_b = off_hb // nheads
159
+ off_h = off_hb % nheads
160
+ offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
161
+ offs_d = tl.arange(0, BLOCK_HEADDIM)
162
+ o = tl.load(Out + off_b * stride_ob + off_h * stride_oh + offs_m[:, None] * stride_om + offs_d[None, :], mask=(offs_m[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0).to(tl.float32)
163
+ do = tl.load(DO + off_b * stride_dob + off_h * stride_doh + offs_m[:, None] * stride_dom + offs_d[None, :], mask=(offs_m[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0).to(tl.float32)
164
+ delta = tl.sum(o * do, axis=1)
165
+ tl.store(Delta + off_hb * seqlen_q_rounded + offs_m, delta)
166
+
167
+ @triton.jit
168
+ def _bwd_store_dk_dv(dk_ptrs, dv_ptrs, dk, dv, offs_n, offs_d, seqlen_k, headdim, EVEN_M: tl.constexpr, EVEN_N: tl.constexpr, EVEN_HEADDIM: tl.constexpr):
169
+ if EVEN_N & EVEN_M:
170
+ if EVEN_HEADDIM:
171
+ tl.store(dv_ptrs, dv)
172
+ tl.store(dk_ptrs, dk)
173
+ else:
174
+ tl.store(dv_ptrs, dv, mask=offs_d[None, :] < headdim)
175
+ tl.store(dk_ptrs, dk, mask=offs_d[None, :] < headdim)
176
+ elif EVEN_HEADDIM:
177
+ tl.store(dv_ptrs, dv, mask=offs_n[:, None] < seqlen_k)
178
+ tl.store(dk_ptrs, dk, mask=offs_n[:, None] < seqlen_k)
179
+ else:
180
+ tl.store(dv_ptrs, dv, mask=(offs_n[:, None] < seqlen_k) & (offs_d[None, :] < headdim))
181
+ tl.store(dk_ptrs, dk, mask=(offs_n[:, None] < seqlen_k) & (offs_d[None, :] < headdim))
182
+
183
+ @triton.jit
184
+ def _bwd_kernel_one_col_block(start_n, Q, K, V, Bias, DO, DQ, DK, DV, LSE, D, softmax_scale, stride_qm, stride_kn, stride_vn, stride_bm, stride_dom, stride_dqm, stride_dkn, stride_dvn, seqlen_q, seqlen_k, headdim, ATOMIC_ADD: tl.constexpr, BIAS_TYPE: tl.constexpr, IS_CAUSAL: tl.constexpr, BLOCK_HEADDIM: tl.constexpr, EVEN_M: tl.constexpr, EVEN_N: tl.constexpr, EVEN_HEADDIM: tl.constexpr, BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr):
185
+ begin_m = 0 if not IS_CAUSAL else start_n * BLOCK_N // BLOCK_M * BLOCK_M
186
+ offs_qm = begin_m + tl.arange(0, BLOCK_M)
187
+ offs_n = start_n * BLOCK_N + tl.arange(0, BLOCK_N)
188
+ offs_m = tl.arange(0, BLOCK_M)
189
+ offs_d = tl.arange(0, BLOCK_HEADDIM)
190
+ q_ptrs = Q + (offs_qm[:, None] * stride_qm + offs_d[None, :])
191
+ k_ptrs = K + (offs_n[:, None] * stride_kn + offs_d[None, :])
192
+ v_ptrs = V + (offs_n[:, None] * stride_vn + offs_d[None, :])
193
+ do_ptrs = DO + (offs_qm[:, None] * stride_dom + offs_d[None, :])
194
+ dq_ptrs = DQ + (offs_qm[:, None] * stride_dqm + offs_d[None, :])
195
+ if BIAS_TYPE == 'vector':
196
+ b_ptrs = Bias + offs_n
197
+ elif BIAS_TYPE == 'matrix':
198
+ b_ptrs = Bias + (offs_qm[:, None] * stride_bm + offs_n[None, :])
199
+ dv = tl.zeros([BLOCK_N, BLOCK_HEADDIM], dtype=tl.float32)
200
+ dk = tl.zeros([BLOCK_N, BLOCK_HEADDIM], dtype=tl.float32)
201
+ if begin_m >= seqlen_q:
202
+ dv_ptrs = DV + (offs_n[:, None] * stride_dvn + offs_d[None, :])
203
+ dk_ptrs = DK + (offs_n[:, None] * stride_dkn + offs_d[None, :])
204
+ _bwd_store_dk_dv(dk_ptrs, dv_ptrs, dk, dv, offs_n, offs_d, seqlen_k, headdim, EVEN_M=EVEN_M, EVEN_N=EVEN_N, EVEN_HEADDIM=EVEN_HEADDIM)
205
+ return
206
+ if EVEN_N & EVEN_M:
207
+ if EVEN_HEADDIM:
208
+ k = tl.load(k_ptrs)
209
+ v = tl.load(v_ptrs)
210
+ else:
211
+ k = tl.load(k_ptrs, mask=offs_d[None, :] < headdim, other=0.0)
212
+ v = tl.load(v_ptrs, mask=offs_d[None, :] < headdim, other=0.0)
213
+ elif EVEN_HEADDIM:
214
+ k = tl.load(k_ptrs, mask=offs_n[:, None] < seqlen_k, other=0.0)
215
+ v = tl.load(v_ptrs, mask=offs_n[:, None] < seqlen_k, other=0.0)
216
+ else:
217
+ k = tl.load(k_ptrs, mask=(offs_n[:, None] < seqlen_k) & (offs_d[None, :] < headdim), other=0.0)
218
+ v = tl.load(v_ptrs, mask=(offs_n[:, None] < seqlen_k) & (offs_d[None, :] < headdim), other=0.0)
219
+ num_block_m = tl.cdiv(seqlen_q, BLOCK_M)
220
+ for start_m in range(begin_m, num_block_m * BLOCK_M, BLOCK_M):
221
+ start_m = tl.multiple_of(start_m, BLOCK_M)
222
+ offs_m_curr = start_m + offs_m
223
+ if EVEN_M & EVEN_HEADDIM:
224
+ q = tl.load(q_ptrs)
225
+ elif EVEN_HEADDIM:
226
+ q = tl.load(q_ptrs, mask=offs_m_curr[:, None] < seqlen_q, other=0.0)
227
+ else:
228
+ q = tl.load(q_ptrs, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0)
229
+ qk = tl.dot(q, k, trans_b=True)
230
+ if not EVEN_N:
231
+ qk = tl.where(offs_n[None, :] < seqlen_k, qk, float('-inf'))
232
+ if IS_CAUSAL:
233
+ qk = tl.where(offs_m_curr[:, None] >= offs_n[None, :], qk, float('-inf'))
234
+ if BIAS_TYPE != 'none':
235
+ tl.debug_barrier()
236
+ if BIAS_TYPE == 'vector':
237
+ if EVEN_N:
238
+ bias = tl.load(b_ptrs).to(tl.float32)
239
+ else:
240
+ bias = tl.load(b_ptrs, mask=offs_n < seqlen_k, other=0.0).to(tl.float32)
241
+ bias = bias[None, :]
242
+ elif BIAS_TYPE == 'matrix':
243
+ if EVEN_M & EVEN_N:
244
+ bias = tl.load(b_ptrs).to(tl.float32)
245
+ else:
246
+ bias = tl.load(b_ptrs, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_n[None, :] < seqlen_k), other=0.0).to(tl.float32)
247
+ qk = qk * softmax_scale + bias
248
+ if not EVEN_M & EVEN_HEADDIM:
249
+ tl.debug_barrier()
250
+ lse_i = tl.load(LSE + offs_m_curr)
251
+ if BIAS_TYPE == 'none':
252
+ p = tl.exp(qk * softmax_scale - lse_i[:, None])
253
+ else:
254
+ p = tl.exp(qk - lse_i[:, None])
255
+ if EVEN_M & EVEN_HEADDIM:
256
+ do = tl.load(do_ptrs)
257
+ else:
258
+ do = tl.load(do_ptrs, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0)
259
+ dv += tl.dot(p.to(do.dtype), do, trans_a=True)
260
+ if not EVEN_M & EVEN_HEADDIM:
261
+ tl.debug_barrier()
262
+ dp = tl.dot(do, v, trans_b=True)
263
+ if not EVEN_HEADDIM:
264
+ tl.debug_barrier()
265
+ Di = tl.load(D + offs_m_curr)
266
+ ds = (p * (dp - Di[:, None]) * softmax_scale).to(q.dtype)
267
+ dk += tl.dot(ds, q, trans_a=True)
268
+ if not EVEN_M & EVEN_HEADDIM:
269
+ tl.debug_barrier()
270
+ if not ATOMIC_ADD:
271
+ if EVEN_M & EVEN_HEADDIM:
272
+ dq = tl.load(dq_ptrs, eviction_policy='evict_last')
273
+ dq += tl.dot(ds, k)
274
+ tl.store(dq_ptrs, dq, eviction_policy='evict_last')
275
+ elif EVEN_HEADDIM:
276
+ dq = tl.load(dq_ptrs, mask=offs_m_curr[:, None] < seqlen_q, other=0.0, eviction_policy='evict_last')
277
+ dq += tl.dot(ds, k)
278
+ tl.store(dq_ptrs, dq, mask=offs_m_curr[:, None] < seqlen_q, eviction_policy='evict_last')
279
+ else:
280
+ dq = tl.load(dq_ptrs, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0, eviction_policy='evict_last')
281
+ dq += tl.dot(ds, k)
282
+ tl.store(dq_ptrs, dq, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim), eviction_policy='evict_last')
283
+ else:
284
+ dq = tl.dot(ds, k)
285
+ if EVEN_M & EVEN_HEADDIM:
286
+ tl.atomic_add(dq_ptrs, dq)
287
+ elif EVEN_HEADDIM:
288
+ tl.atomic_add(dq_ptrs, dq, mask=offs_m_curr[:, None] < seqlen_q)
289
+ else:
290
+ tl.atomic_add(dq_ptrs, dq, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim))
291
+ dq_ptrs += BLOCK_M * stride_dqm
292
+ q_ptrs += BLOCK_M * stride_qm
293
+ do_ptrs += BLOCK_M * stride_dom
294
+ if BIAS_TYPE == 'matrix':
295
+ b_ptrs += BLOCK_M * stride_bm
296
+ dv_ptrs = DV + (offs_n[:, None] * stride_dvn + offs_d[None, :])
297
+ dk_ptrs = DK + (offs_n[:, None] * stride_dkn + offs_d[None, :])
298
+ _bwd_store_dk_dv(dk_ptrs, dv_ptrs, dk, dv, offs_n, offs_d, seqlen_k, headdim, EVEN_M=EVEN_M, EVEN_N=EVEN_N, EVEN_HEADDIM=EVEN_HEADDIM)
299
+
300
+ def init_to_zero(name):
301
+ return lambda nargs: nargs[name].zero_()
302
+
303
+ @triton.autotune(configs=[triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'SEQUENCE_PARALLEL': False}, num_warps=8, num_stages=1, pre_hook=init_to_zero('DQ')), triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'SEQUENCE_PARALLEL': True}, num_warps=8, num_stages=1, pre_hook=init_to_zero('DQ'))], key=['CACHE_KEY_SEQLEN_Q', 'CACHE_KEY_SEQLEN_K', 'BIAS_TYPE', 'IS_CAUSAL', 'BLOCK_HEADDIM'])
304
+ @triton.heuristics({'EVEN_M': lambda args: args['seqlen_q'] % args['BLOCK_M'] == 0, 'EVEN_N': lambda args: args['seqlen_k'] % args['BLOCK_N'] == 0, 'EVEN_HEADDIM': lambda args: args['headdim'] == args['BLOCK_HEADDIM']})
305
+ @triton.jit
306
+ def _bwd_kernel(Q, K, V, Bias, DO, DQ, DK, DV, LSE, D, softmax_scale, stride_qb, stride_qh, stride_qm, stride_kb, stride_kh, stride_kn, stride_vb, stride_vh, stride_vn, stride_bb, stride_bh, stride_bm, stride_dob, stride_doh, stride_dom, stride_dqb, stride_dqh, stride_dqm, stride_dkb, stride_dkh, stride_dkn, stride_dvb, stride_dvh, stride_dvn, nheads, seqlen_q, seqlen_k, seqlen_q_rounded, headdim, CACHE_KEY_SEQLEN_Q, CACHE_KEY_SEQLEN_K, BIAS_TYPE: tl.constexpr, IS_CAUSAL: tl.constexpr, BLOCK_HEADDIM: tl.constexpr, SEQUENCE_PARALLEL: tl.constexpr, EVEN_M: tl.constexpr, EVEN_N: tl.constexpr, EVEN_HEADDIM: tl.constexpr, BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr):
307
+ off_hb = tl.program_id(1)
308
+ off_b = off_hb // nheads
309
+ off_h = off_hb % nheads
310
+ Q += off_b * stride_qb + off_h * stride_qh
311
+ K += off_b * stride_kb + off_h * stride_kh
312
+ V += off_b * stride_vb + off_h * stride_vh
313
+ DO += off_b * stride_dob + off_h * stride_doh
314
+ DQ += off_b * stride_dqb + off_h * stride_dqh
315
+ DK += off_b * stride_dkb + off_h * stride_dkh
316
+ DV += off_b * stride_dvb + off_h * stride_dvh
317
+ if BIAS_TYPE != 'none':
318
+ Bias += off_b * stride_bb + off_h * stride_bh
319
+ D += off_hb * seqlen_q_rounded
320
+ LSE += off_hb * seqlen_q_rounded
321
+ if not SEQUENCE_PARALLEL:
322
+ num_block_n = tl.cdiv(seqlen_k, BLOCK_N)
323
+ for start_n in range(0, num_block_n):
324
+ _bwd_kernel_one_col_block(start_n, Q, K, V, Bias, DO, DQ, DK, DV, LSE, D, softmax_scale, stride_qm, stride_kn, stride_vn, stride_bm, stride_dom, stride_dqm, stride_dkn, stride_dvn, seqlen_q, seqlen_k, headdim, ATOMIC_ADD=False, BIAS_TYPE=BIAS_TYPE, IS_CAUSAL=IS_CAUSAL, BLOCK_HEADDIM=BLOCK_HEADDIM, EVEN_M=EVEN_M, EVEN_N=EVEN_N, EVEN_HEADDIM=EVEN_HEADDIM, BLOCK_M=BLOCK_M, BLOCK_N=BLOCK_N)
325
+ else:
326
+ start_n = tl.program_id(0)
327
+ _bwd_kernel_one_col_block(start_n, Q, K, V, Bias, DO, DQ, DK, DV, LSE, D, softmax_scale, stride_qm, stride_kn, stride_vn, stride_bm, stride_dom, stride_dqm, stride_dkn, stride_dvn, seqlen_q, seqlen_k, headdim, ATOMIC_ADD=True, BIAS_TYPE=BIAS_TYPE, IS_CAUSAL=IS_CAUSAL, BLOCK_HEADDIM=BLOCK_HEADDIM, EVEN_M=EVEN_M, EVEN_N=EVEN_N, EVEN_HEADDIM=EVEN_HEADDIM, BLOCK_M=BLOCK_M, BLOCK_N=BLOCK_N)
328
+
329
+ def _flash_attn_forward(q, k, v, bias=None, causal=False, softmax_scale=None):
330
+ (batch, seqlen_q, nheads, d) = q.shape
331
+ (_, seqlen_k, _, _) = k.shape
332
+ assert k.shape == (batch, seqlen_k, nheads, d)
333
+ assert v.shape == (batch, seqlen_k, nheads, d)
334
+ assert d <= 128, 'FlashAttention only support head dimensions up to 128'
335
+ assert q.dtype == k.dtype == v.dtype, 'All tensors must have the same type'
336
+ assert q.dtype in [torch.float16, torch.bfloat16], 'Only support fp16 and bf16'
337
+ assert q.is_cuda and k.is_cuda and v.is_cuda
338
+ softmax_scale = softmax_scale or 1.0 / math.sqrt(d)
339
+ has_bias = bias is not None
340
+ bias_type = 'none'
341
+ if has_bias:
342
+ assert bias.dtype in [q.dtype, torch.float]
343
+ assert bias.is_cuda
344
+ assert bias.dim() == 4
345
+ if bias.stride(-1) != 1:
346
+ bias = bias.contiguous()
347
+ if bias.shape[2:] == (1, seqlen_k):
348
+ bias_type = 'vector'
349
+ elif bias.shape[2:] == (seqlen_q, seqlen_k):
350
+ bias_type = 'matrix'
351
+ else:
352
+ raise RuntimeError('Last 2 dimensions of bias must be (1, seqlen_k) or (seqlen_q, seqlen_k)')
353
+ bias = bias.expand(batch, nheads, seqlen_q, seqlen_k)
354
+ bias_strides = (bias.stride(0), bias.stride(1), bias.stride(2)) if has_bias else (0, 0, 0)
355
+ seqlen_q_rounded = math.ceil(seqlen_q / 128) * 128
356
+ lse = torch.empty((batch, nheads, seqlen_q_rounded), device=q.device, dtype=torch.float32)
357
+ tmp = torch.empty((batch, nheads, seqlen_q_rounded), device=q.device, dtype=torch.float32)
358
+ o = torch.empty_like(q)
359
+ BLOCK_HEADDIM = max(triton.next_power_of_2(d), 16)
360
+ BLOCK = 128
361
+ num_warps = 4 if d <= 64 else 8
362
+ grid = lambda META: (triton.cdiv(seqlen_q, META['BLOCK_M']), batch * nheads)
363
+ _fwd_kernel[grid](q, k, v, bias, o, lse, tmp, softmax_scale, q.stride(0), q.stride(2), q.stride(1), k.stride(0), k.stride(2), k.stride(1), v.stride(0), v.stride(2), v.stride(1), *bias_strides, o.stride(0), o.stride(2), o.stride(1), nheads, seqlen_q, seqlen_k, seqlen_q_rounded, d, seqlen_q // 32, seqlen_k // 32, bias_type, causal, BLOCK_HEADDIM, BLOCK_M=BLOCK, BLOCK_N=BLOCK, num_warps=num_warps, num_stages=1)
364
+ return (o, lse, softmax_scale)
365
+
366
+ def _flash_attn_backward(do, q, k, v, o, lse, dq, dk, dv, bias=None, causal=False, softmax_scale=None):
367
+ if do.stride(-1) != 1:
368
+ do = do.contiguous()
369
+ (batch, seqlen_q, nheads, d) = q.shape
370
+ (_, seqlen_k, _, _) = k.shape
371
+ assert d <= 128
372
+ seqlen_q_rounded = math.ceil(seqlen_q / 128) * 128
373
+ assert lse.shape == (batch, nheads, seqlen_q_rounded)
374
+ assert q.stride(-1) == k.stride(-1) == v.stride(-1) == o.stride(-1) == 1
375
+ assert dq.stride(-1) == dk.stride(-1) == dv.stride(-1) == 1
376
+ softmax_scale = softmax_scale or 1.0 / math.sqrt(d)
377
+ dq_accum = torch.empty_like(q, dtype=torch.float32)
378
+ delta = torch.empty_like(lse)
379
+ BLOCK_HEADDIM = max(triton.next_power_of_2(d), 16)
380
+ grid = lambda META: (triton.cdiv(seqlen_q, META['BLOCK_M']), batch * nheads)
381
+ _bwd_preprocess_do_o_dot[grid](o, do, delta, o.stride(0), o.stride(2), o.stride(1), do.stride(0), do.stride(2), do.stride(1), nheads, seqlen_q, seqlen_q_rounded, d, BLOCK_M=128, BLOCK_HEADDIM=BLOCK_HEADDIM)
382
+ has_bias = bias is not None
383
+ bias_type = 'none'
384
+ if has_bias:
385
+ assert bias.dtype in [q.dtype, torch.float]
386
+ assert bias.is_cuda
387
+ assert bias.dim() == 4
388
+ assert bias.stride(-1) == 1
389
+ if bias.shape[2:] == (1, seqlen_k):
390
+ bias_type = 'vector'
391
+ elif bias.shape[2:] == (seqlen_q, seqlen_k):
392
+ bias_type = 'matrix'
393
+ else:
394
+ raise RuntimeError('Last 2 dimensions of bias must be (1, seqlen_k) or (seqlen_q, seqlen_k)')
395
+ bias = bias.expand(batch, nheads, seqlen_q, seqlen_k)
396
+ bias_strides = (bias.stride(0), bias.stride(1), bias.stride(2)) if has_bias else (0, 0, 0)
397
+ grid = lambda META: (triton.cdiv(seqlen_k, META['BLOCK_N']) if META['SEQUENCE_PARALLEL'] else 1, batch * nheads)
398
+ _bwd_kernel[grid](q, k, v, bias, do, dq_accum, dk, dv, lse, delta, softmax_scale, q.stride(0), q.stride(2), q.stride(1), k.stride(0), k.stride(2), k.stride(1), v.stride(0), v.stride(2), v.stride(1), *bias_strides, do.stride(0), do.stride(2), do.stride(1), dq_accum.stride(0), dq_accum.stride(2), dq_accum.stride(1), dk.stride(0), dk.stride(2), dk.stride(1), dv.stride(0), dv.stride(2), dv.stride(1), nheads, seqlen_q, seqlen_k, seqlen_q_rounded, d, seqlen_q // 32, seqlen_k // 32, bias_type, causal, BLOCK_HEADDIM)
399
+ dq.copy_(dq_accum)
400
+
401
+ class FlashAttnQKVPackedFunc(torch.autograd.Function):
402
+
403
+ @staticmethod
404
+ def forward(ctx, qkv, bias=None, causal=False, softmax_scale=None):
405
+ """
406
+ qkv: (batch, seqlen, 3, nheads, headdim)
407
+ bias: optional, shape broadcastible to (batch, nheads, seqlen, seqlen).
408
+ For example, ALiBi mask for causal would have shape (1, nheads, 1, seqlen).
409
+ ALiBi mask for non-causal would have shape (1, nheads, seqlen, seqlen)
410
+ """
411
+ if qkv.stride(-1) != 1:
412
+ qkv = qkv.contiguous()
413
+ (o, lse, ctx.softmax_scale) = _flash_attn_forward(qkv[:, :, 0], qkv[:, :, 1], qkv[:, :, 2], bias=bias, causal=causal, softmax_scale=softmax_scale)
414
+ ctx.save_for_backward(qkv, o, lse, bias)
415
+ ctx.causal = causal
416
+ return o
417
+
418
+ @staticmethod
419
+ def backward(ctx, do):
420
+ (qkv, o, lse, bias) = ctx.saved_tensors
421
+ assert not ctx.needs_input_grad[1], 'FlashAttention does not support bias gradient yet'
422
+ with torch.inference_mode():
423
+ dqkv = torch.empty_like(qkv)
424
+ _flash_attn_backward(do, qkv[:, :, 0], qkv[:, :, 1], qkv[:, :, 2], o, lse, dqkv[:, :, 0], dqkv[:, :, 1], dqkv[:, :, 2], bias=bias, causal=ctx.causal, softmax_scale=ctx.softmax_scale)
425
+ return (dqkv, None, None, None)
426
+ flash_attn_qkvpacked_func = FlashAttnQKVPackedFunc.apply
427
+
428
+ class FlashAttnKVPackedFunc(torch.autograd.Function):
429
+
430
+ @staticmethod
431
+ def forward(ctx, q, kv, bias=None, causal=False, softmax_scale=None):
432
+ """
433
+ q: (batch, seqlen_q, nheads, headdim)
434
+ kv: (batch, seqlen_k, 2, nheads, headdim)
435
+ bias: optional, shape broadcastible to (batch, nheads, seqlen_q, seqlen_k).
436
+ For example, ALiBi mask for causal would have shape (1, nheads, 1, seqlen_k).
437
+ ALiBi mask for non-causal would have shape (1, nheads, seqlen_q, seqlen_k)
438
+ """
439
+ (q, kv) = [x if x.stride(-1) == 1 else x.contiguous() for x in [q, kv]]
440
+ (o, lse, ctx.softmax_scale) = _flash_attn_forward(q, kv[:, :, 0], kv[:, :, 1], bias=bias, causal=causal, softmax_scale=softmax_scale)
441
+ ctx.save_for_backward(q, kv, o, lse, bias)
442
+ ctx.causal = causal
443
+ return o
444
+
445
+ @staticmethod
446
+ def backward(ctx, do):
447
+ (q, kv, o, lse, bias) = ctx.saved_tensors
448
+ if len(ctx.needs_input_grad) >= 3:
449
+ assert not ctx.needs_input_grad[2], 'FlashAttention does not support bias gradient yet'
450
+ with torch.inference_mode():
451
+ dq = torch.empty_like(q)
452
+ dkv = torch.empty_like(kv)
453
+ _flash_attn_backward(do, q, kv[:, :, 0], kv[:, :, 1], o, lse, dq, dkv[:, :, 0], dkv[:, :, 1], bias=bias, causal=ctx.causal, softmax_scale=ctx.softmax_scale)
454
+ return (dq, dkv, None, None, None)
455
+ flash_attn_kvpacked_func = FlashAttnKVPackedFunc.apply
456
+
457
+ class FlashAttnFunc(torch.autograd.Function):
458
+
459
+ @staticmethod
460
+ def forward(ctx, q, k, v, bias=None, causal=False, softmax_scale=None):
461
+ """
462
+ q: (batch_size, seqlen_q, nheads, headdim)
463
+ k, v: (batch_size, seqlen_k, nheads, headdim)
464
+ bias: optional, shape broadcastible to (batch, nheads, seqlen_q, seqlen_k).
465
+ For example, ALiBi mask for causal would have shape (1, nheads, 1, seqlen_k).
466
+ ALiBi mask for non-causal would have shape (1, nheads, seqlen_q, seqlen_k)
467
+ """
468
+ (q, k, v) = [x if x.stride(-1) == 1 else x.contiguous() for x in [q, k, v]]
469
+ (o, lse, ctx.softmax_scale) = _flash_attn_forward(q, k, v, bias=bias, causal=causal, softmax_scale=softmax_scale)
470
+ ctx.save_for_backward(q, k, v, o, lse, bias)
471
+ ctx.causal = causal
472
+ return o
473
+
474
+ @staticmethod
475
+ def backward(ctx, do):
476
+ (q, k, v, o, lse, bias) = ctx.saved_tensors
477
+ assert not ctx.needs_input_grad[3], 'FlashAttention does not support bias gradient yet'
478
+ with torch.inference_mode():
479
+ dq = torch.empty_like(q)
480
+ dk = torch.empty_like(k)
481
+ dv = torch.empty_like(v)
482
+ _flash_attn_backward(do, q, k, v, o, lse, dq, dk, dv, bias=bias, causal=ctx.causal, softmax_scale=ctx.softmax_scale)
483
+ return (dq, dk, dv, None, None, None)
484
+ flash_attn_func = FlashAttnFunc.apply
generation_config.json ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "transformers_version": "4.28.1",
4
+ "use_cache": false
5
+ }
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)
model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:411576f8c03aa73bc7faa2d241ef5090e16abc73583c610f592dd36798c4b198
3
+ size 13298599938
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_}
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
+ "clean_up_tokenization_spaces": true,
5
+ "eos_token": "<|endoftext|>",
6
+ "model_max_length": 2048,
7
+ "tokenizer_class": "GPTNeoXTokenizer",
8
+ "unk_token": "<|endoftext|>"
9
+ }
vocabulary.txt ADDED
The diff for this file is too large to render. See raw diff