amc-madalin
commited on
Commit
•
54b6d58
1
Parent(s):
726a3fb
Upload 2 files
Browse files- configuration_olmo.py +44 -0
- modeling_olmo.py +145 -0
configuration_olmo.py
ADDED
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
OLMo configuration
|
3 |
+
"""
|
4 |
+
|
5 |
+
from transformers import AutoConfig, PretrainedConfig
|
6 |
+
from transformers.utils import logging
|
7 |
+
|
8 |
+
from olmo.config import ModelConfig
|
9 |
+
|
10 |
+
logger = logging.get_logger(__name__)
|
11 |
+
|
12 |
+
|
13 |
+
class OLMoConfig(PretrainedConfig):
|
14 |
+
model_type = "olmo"
|
15 |
+
keys_to_ignore_at_inference = ["past_key_values"] # TODO: confirm
|
16 |
+
|
17 |
+
def __init__(self, use_cache: bool = False, **kwargs):
|
18 |
+
model_config = ModelConfig()
|
19 |
+
all_kwargs = model_config.asdict()
|
20 |
+
all_kwargs.update(kwargs)
|
21 |
+
all_kwargs.update({"use_cache": use_cache})
|
22 |
+
all_kwargs.update(
|
23 |
+
{
|
24 |
+
"architectures": all_kwargs.get("architectures", ["OlmoModelForCausalLM"])
|
25 |
+
or ["OlmoModelForCausalLM"]
|
26 |
+
}
|
27 |
+
)
|
28 |
+
super().__init__(**all_kwargs)
|
29 |
+
|
30 |
+
@property
|
31 |
+
def num_attention_heads(self):
|
32 |
+
return self.n_heads
|
33 |
+
|
34 |
+
@property
|
35 |
+
def num_hidden_layers(self):
|
36 |
+
return self.n_layers
|
37 |
+
|
38 |
+
@property
|
39 |
+
def hidden_size(self):
|
40 |
+
return self.d_model
|
41 |
+
|
42 |
+
|
43 |
+
# Register the config class so that it is available for transformer pipelines, auto-loading etc.
|
44 |
+
AutoConfig.register("olmo", OLMoConfig)
|
modeling_olmo.py
ADDED
@@ -0,0 +1,145 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List, Optional, Tuple, Union
|
2 |
+
|
3 |
+
import torch
|
4 |
+
from transformers import PreTrainedModel
|
5 |
+
from transformers.modeling_outputs import CausalLMOutputWithPast
|
6 |
+
from transformers.models.auto import AutoModelForCausalLM
|
7 |
+
|
8 |
+
from olmo.config import ModelConfig
|
9 |
+
from olmo.model import Olmo
|
10 |
+
|
11 |
+
from .configuration_olmo import OLMoConfig
|
12 |
+
|
13 |
+
|
14 |
+
def create_model_config_from_pretrained_config(config: OLMoConfig):
|
15 |
+
"""
|
16 |
+
Utility function
|
17 |
+
"""
|
18 |
+
|
19 |
+
kwargs = {}
|
20 |
+
for key in ModelConfig.__match_args__:
|
21 |
+
kwargs[key] = getattr(config, key)
|
22 |
+
|
23 |
+
model_config = ModelConfig(**kwargs)
|
24 |
+
return model_config
|
25 |
+
|
26 |
+
|
27 |
+
class OLMoForCausalLM(PreTrainedModel):
|
28 |
+
"""
|
29 |
+
Extremely barebones HF model wrapper.
|
30 |
+
"""
|
31 |
+
|
32 |
+
config_class = OLMoConfig
|
33 |
+
base_model_prefix = "model"
|
34 |
+
_no_split_modules = ["OLMoBlock"]
|
35 |
+
|
36 |
+
def __init__(self, config: OLMoConfig, model: Optional[Olmo] = None, init_params: bool = False):
|
37 |
+
super().__init__(config)
|
38 |
+
|
39 |
+
if not model:
|
40 |
+
model_config = create_model_config_from_pretrained_config(config)
|
41 |
+
# Initialize model (always on CPU to start with so we don't run out of GPU memory).
|
42 |
+
model_config.init_device = "cpu"
|
43 |
+
self.model = Olmo(model_config, init_params=init_params)
|
44 |
+
else:
|
45 |
+
self.model = model
|
46 |
+
|
47 |
+
def forward(
|
48 |
+
self,
|
49 |
+
input_ids: torch.LongTensor = None,
|
50 |
+
attention_mask: Optional[torch.Tensor] = None,
|
51 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
52 |
+
labels: Optional[torch.LongTensor] = None,
|
53 |
+
use_cache: Optional[bool] = None,
|
54 |
+
output_attentions: Optional[bool] = None,
|
55 |
+
output_hidden_states: Optional[bool] = None,
|
56 |
+
return_dict: Optional[bool] = None,
|
57 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
58 |
+
if use_cache is None:
|
59 |
+
use_cache = self.config.use_cache
|
60 |
+
|
61 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
62 |
+
|
63 |
+
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
64 |
+
outputs = self.model.forward(
|
65 |
+
input_ids=input_ids,
|
66 |
+
attention_mask=attention_mask,
|
67 |
+
past_key_values=past_key_values,
|
68 |
+
use_cache=use_cache,
|
69 |
+
)
|
70 |
+
|
71 |
+
logits = outputs.logits
|
72 |
+
|
73 |
+
loss = None
|
74 |
+
if labels is not None:
|
75 |
+
# Shift so that tokens < n predict n
|
76 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
77 |
+
shift_labels = labels[..., 1:].contiguous()
|
78 |
+
# Flatten the tokens
|
79 |
+
loss_fct = torch.nn.CrossEntropyLoss()
|
80 |
+
shift_logits = shift_logits.view(-1, self.config.embedding_size)
|
81 |
+
shift_labels = shift_labels.view(-1)
|
82 |
+
# Enable model parallelism
|
83 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
84 |
+
loss = loss_fct(shift_logits, shift_labels)
|
85 |
+
|
86 |
+
if not return_dict:
|
87 |
+
output = (logits,) + outputs[1:]
|
88 |
+
return (loss,) + output if loss is not None else output
|
89 |
+
|
90 |
+
return CausalLMOutputWithPast(
|
91 |
+
loss=loss,
|
92 |
+
logits=logits,
|
93 |
+
past_key_values=outputs.attn_key_values,
|
94 |
+
)
|
95 |
+
|
96 |
+
def can_generate(self) -> bool:
|
97 |
+
return True
|
98 |
+
|
99 |
+
def prepare_inputs_for_generation(
|
100 |
+
self, input_ids: torch.LongTensor, past_key_values: Optional[List[Tuple]] = None, **kwargs
|
101 |
+
):
|
102 |
+
if past_key_values:
|
103 |
+
# This is because we want the model to only process the last generated token.
|
104 |
+
input_ids = input_ids[:, -1:]
|
105 |
+
model_inputs = {"input_ids": input_ids, "past_key_values": past_key_values}
|
106 |
+
|
107 |
+
model_inputs.update(kwargs)
|
108 |
+
model_inputs["use_cache"] = kwargs.pop("use_cache", self.config.use_cache)
|
109 |
+
return model_inputs
|
110 |
+
|
111 |
+
# TODO: these are required to make the implementation complete.
|
112 |
+
# def resize_position_embeddings(self, new_num_position_embeddings: int):
|
113 |
+
# pass
|
114 |
+
#
|
115 |
+
# def get_position_embeddings(self) -> Union[nn.Embedding, Tuple[nn.Embedding]]:
|
116 |
+
# pass
|
117 |
+
#
|
118 |
+
# def _reorder_cache(self, past_key_values, beam_idx):
|
119 |
+
# pass
|
120 |
+
|
121 |
+
def get_input_embeddings(self) -> torch.nn.Module:
|
122 |
+
return self.model.transformer.wte
|
123 |
+
|
124 |
+
def set_input_embeddings(self, value: torch.nn.Module):
|
125 |
+
self.model.transformer.wte = value
|
126 |
+
|
127 |
+
def get_output_embeddings(self):
|
128 |
+
if self.config.weight_tying:
|
129 |
+
return self.model.transformer.wte
|
130 |
+
else:
|
131 |
+
return self.model.transformer.ff_out
|
132 |
+
|
133 |
+
def set_output_embeddings(self, value: torch.nn.Module):
|
134 |
+
if self.config.weight_tying:
|
135 |
+
self.model.transformer.wte = value
|
136 |
+
else:
|
137 |
+
self.model.transformer.ff_out = value
|
138 |
+
|
139 |
+
def tie_weights(self):
|
140 |
+
if self.config.weight_tying:
|
141 |
+
self.model.transformer.ff_out = self.model.transformer.wte
|
142 |
+
|
143 |
+
|
144 |
+
# Register the model so that it is available for transformer pipelines, auto-loading, etc.
|
145 |
+
AutoModelForCausalLM.register(OLMoConfig, OLMoForCausalLM)
|