Spaces:
Running
on
Zero
Running
on
Zero
Upload 2 files
Browse files- configuration_stockllama.py +65 -0
- modeling_stockllama.py +140 -0
configuration_stockllama.py
ADDED
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers.configuration_utils import PretrainedConfig
|
2 |
+
|
3 |
+
class StockLlamaConfig(PretrainedConfig):
|
4 |
+
model_type = "stockllama"
|
5 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
6 |
+
|
7 |
+
def __init__(
|
8 |
+
self,
|
9 |
+
vocab_size=32000,
|
10 |
+
hidden_size=4096,
|
11 |
+
intermediate_size=11008,
|
12 |
+
num_hidden_layers=32,
|
13 |
+
num_attention_heads=32,
|
14 |
+
num_key_value_heads=None,
|
15 |
+
hidden_act="silu",
|
16 |
+
max_position_embeddings=2048,
|
17 |
+
term_number=4,
|
18 |
+
initializer_range=0.02,
|
19 |
+
rms_norm_eps=1e-6,
|
20 |
+
use_cache=True,
|
21 |
+
pad_token_id=None,
|
22 |
+
bos_token_id=1,
|
23 |
+
eos_token_id=2,
|
24 |
+
pretraining_tp=1,
|
25 |
+
tie_word_embeddings=False,
|
26 |
+
rope_theta=10000.0,
|
27 |
+
rope_scaling=None,
|
28 |
+
attention_bias=False,
|
29 |
+
attention_dropout=0.0,
|
30 |
+
mlp_bias=False,
|
31 |
+
head_dim=None,
|
32 |
+
**kwargs,
|
33 |
+
):
|
34 |
+
self.vocab_size = vocab_size
|
35 |
+
self.max_position_embeddings = max_position_embeddings
|
36 |
+
self.term_number = term_number
|
37 |
+
self.hidden_size = hidden_size
|
38 |
+
self.intermediate_size = intermediate_size
|
39 |
+
self.num_hidden_layers = num_hidden_layers
|
40 |
+
self.num_attention_heads = num_attention_heads
|
41 |
+
|
42 |
+
if num_key_value_heads is None:
|
43 |
+
num_key_value_heads = num_attention_heads
|
44 |
+
|
45 |
+
self.num_key_value_heads = num_key_value_heads
|
46 |
+
self.hidden_act = hidden_act
|
47 |
+
self.initializer_range = initializer_range
|
48 |
+
self.rms_norm_eps = rms_norm_eps
|
49 |
+
self.pretraining_tp = pretraining_tp
|
50 |
+
self.use_cache = use_cache
|
51 |
+
self.rope_theta = rope_theta
|
52 |
+
self.rope_scaling = rope_scaling
|
53 |
+
self.attention_bias = attention_bias
|
54 |
+
self.attention_dropout = attention_dropout
|
55 |
+
self.mlp_bias = mlp_bias
|
56 |
+
self.head_dim = head_dim if head_dim is not None else self.hidden_size // self.num_attention_heads
|
57 |
+
|
58 |
+
|
59 |
+
super().__init__(
|
60 |
+
pad_token_id=pad_token_id,
|
61 |
+
bos_token_id=bos_token_id,
|
62 |
+
eos_token_id=eos_token_id,
|
63 |
+
tie_word_embeddings=tie_word_embeddings,
|
64 |
+
**kwargs,
|
65 |
+
)
|
modeling_stockllama.py
ADDED
@@ -0,0 +1,140 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from torch import nn
|
2 |
+
from torch.nn import functional as F
|
3 |
+
import torch
|
4 |
+
|
5 |
+
from configuration_stockllama import StockLlamaConfig
|
6 |
+
|
7 |
+
from transformers.models.llama.modeling_llama import LlamaPreTrainedModel
|
8 |
+
from transformers.models.llama.modeling_llama import LlamaModel
|
9 |
+
from transformers.modeling_utils import PreTrainedModel
|
10 |
+
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
|
11 |
+
from transformers.cache_utils import Cache
|
12 |
+
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
13 |
+
import math
|
14 |
+
from typing import Any, Dict, List, Optional, Tuple , Union
|
15 |
+
|
16 |
+
class FloatEmbedding(nn.Module):
|
17 |
+
def __init__(self, vocab_size, hidden_size, padding_idx ,term_number):
|
18 |
+
super(FloatEmbedding, self).__init__()
|
19 |
+
self.term_number = term_number
|
20 |
+
self.int_part = nn.Embedding(vocab_size, hidden_size ,padding_idx)
|
21 |
+
self.float_part = nn.Embedding(10**term_number , hidden_size)
|
22 |
+
|
23 |
+
def forward(self, input):
|
24 |
+
float_input = ((input - torch.floor(input)) * (10**self.term_number)).to(torch.long)
|
25 |
+
int_input = input.to(torch.long)
|
26 |
+
output = self.float_part(float_input) + self.int_part(int_input)
|
27 |
+
|
28 |
+
return output
|
29 |
+
|
30 |
+
class StockLlamaPreTrainedModel(LlamaPreTrainedModel):
|
31 |
+
config_class = StockLlamaConfig
|
32 |
+
base_model_prefix = "model"
|
33 |
+
supports_gradient_checkpointing = True
|
34 |
+
_no_split_modules = ["LlamaDecoderLayer"]
|
35 |
+
_skip_keys_device_placement = ["past_key_values"]
|
36 |
+
_supports_flash_attn_2 = True
|
37 |
+
_supports_sdpa = True
|
38 |
+
_supports_cache_class = True
|
39 |
+
_supports_quantized_cache = True
|
40 |
+
_supports_static_cache = True
|
41 |
+
|
42 |
+
def _init_weights(self, module):
|
43 |
+
std = self.config.initializer_range
|
44 |
+
if isinstance(module, nn.Linear):
|
45 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
46 |
+
if module.bias is not None:
|
47 |
+
module.bias.data.zero_()
|
48 |
+
elif isinstance(module, nn.Embedding):
|
49 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
50 |
+
if module.padding_idx is not None:
|
51 |
+
module.weight.data[module.padding_idx].zero_()
|
52 |
+
|
53 |
+
class StockLlamaModel(LlamaModel):
|
54 |
+
config_class = StockLlamaConfig
|
55 |
+
|
56 |
+
def __init__(self, config):
|
57 |
+
super().__init__(config)
|
58 |
+
self._use_flash_attention_2 = True
|
59 |
+
self.embed_tokens = FloatEmbedding(config.vocab_size, config.hidden_size, self.padding_idx, config.term_number)
|
60 |
+
self.post_init()
|
61 |
+
|
62 |
+
|
63 |
+
class StockLlamaForForecasting(StockLlamaPreTrainedModel):
|
64 |
+
def __init__(self, config):
|
65 |
+
super().__init__(config)
|
66 |
+
self.model = StockLlamaModel(config)
|
67 |
+
self.score = nn.Linear(config.hidden_size, 1, bias=False)
|
68 |
+
self.post_init()
|
69 |
+
|
70 |
+
def get_input_embeddings(self):
|
71 |
+
return self.model.embed_tokens
|
72 |
+
|
73 |
+
def set_input_embeddings(self, value):
|
74 |
+
self.model.embed_tokens = value
|
75 |
+
|
76 |
+
def forward(
|
77 |
+
self,
|
78 |
+
input_ids: Optional[torch.LongTensor] = None,
|
79 |
+
attention_mask: Optional[torch.Tensor] = None,
|
80 |
+
position_ids: Optional[torch.LongTensor] = None,
|
81 |
+
past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
|
82 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
83 |
+
labels: Optional[torch.FloatTensor] = None,
|
84 |
+
use_cache: Optional[bool] = None,
|
85 |
+
output_attentions: Optional[bool] = None,
|
86 |
+
output_hidden_states: Optional[bool] = None,
|
87 |
+
return_dict: Optional[bool] = None,
|
88 |
+
) -> Union[Tuple, SequenceClassifierOutputWithPast]:
|
89 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
90 |
+
|
91 |
+
transformer_outputs = self.model(
|
92 |
+
input_ids,
|
93 |
+
attention_mask=attention_mask,
|
94 |
+
position_ids=position_ids,
|
95 |
+
past_key_values=past_key_values,
|
96 |
+
inputs_embeds=inputs_embeds,
|
97 |
+
use_cache=use_cache,
|
98 |
+
output_attentions=output_attentions,
|
99 |
+
output_hidden_states=output_hidden_states,
|
100 |
+
return_dict=return_dict,
|
101 |
+
)
|
102 |
+
hidden_states = transformer_outputs[0]
|
103 |
+
logits = self.score(hidden_states)
|
104 |
+
|
105 |
+
if input_ids is not None:
|
106 |
+
batch_size = input_ids.shape[0]
|
107 |
+
else:
|
108 |
+
batch_size = inputs_embeds.shape[0]
|
109 |
+
|
110 |
+
if self.config.pad_token_id is None and batch_size != 1:
|
111 |
+
raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
|
112 |
+
if self.config.pad_token_id is None:
|
113 |
+
sequence_lengths = -1
|
114 |
+
else:
|
115 |
+
if input_ids is not None:
|
116 |
+
sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
|
117 |
+
sequence_lengths = sequence_lengths % input_ids.shape[-1]
|
118 |
+
sequence_lengths = sequence_lengths.to(logits.device)
|
119 |
+
else:
|
120 |
+
sequence_lengths = -1
|
121 |
+
|
122 |
+
pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
|
123 |
+
|
124 |
+
loss = None
|
125 |
+
if labels is not None:
|
126 |
+
labels = labels.to(logits.device)
|
127 |
+
loss_fct = MSELoss()
|
128 |
+
loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
|
129 |
+
|
130 |
+
if not return_dict:
|
131 |
+
output = (pooled_logits,) + transformer_outputs[1:]
|
132 |
+
return ((loss,) + output) if loss is not None else output
|
133 |
+
|
134 |
+
return SequenceClassifierOutputWithPast(
|
135 |
+
loss=loss,
|
136 |
+
logits=pooled_logits,
|
137 |
+
past_key_values=transformer_outputs.past_key_values,
|
138 |
+
hidden_states=transformer_outputs.hidden_states,
|
139 |
+
attentions=transformer_outputs.attentions,
|
140 |
+
)
|