Update README.md
Browse files
README.md
CHANGED
@@ -2,4 +2,99 @@
|
|
2 |
license: apache-2.0
|
3 |
---
|
4 |
|
5 |
-
Train in 30B Byte. Mode size 353M. Table 2 in [MambaByte](https://arxiv.org/abs/2401.13660)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
license: apache-2.0
|
3 |
---
|
4 |
|
5 |
+
Train in 30B Byte. Mode size 353M. Table 2 in [MambaByte](https://arxiv.org/abs/2401.13660)
|
6 |
+
|
7 |
+
To use
|
8 |
+
|
9 |
+
```
|
10 |
+
import torch
|
11 |
+
from mamba_ssm.models.mixer_seq_simple import MambaLMHeadModel
|
12 |
+
|
13 |
+
import numpy as np
|
14 |
+
|
15 |
+
model=MambaLMHeadModel.from_pretrained("JunxiongWang/MambaByte_Code", device='cuda', dtype=torch.float32)
|
16 |
+
|
17 |
+
text = "import torch"
|
18 |
+
text_byte = np.frombuffer(text.encode('utf-8'), dtype=np.uint8)
|
19 |
+
input_ids = torch.from_numpy(text_byte[None, :]).long().cuda()
|
20 |
+
|
21 |
+
sample = model.generate(
|
22 |
+
input_ids=input_ids,
|
23 |
+
max_length=2048,
|
24 |
+
cg=True,
|
25 |
+
return_dict_in_generate=True,
|
26 |
+
output_scores=True,
|
27 |
+
enable_timing=True,
|
28 |
+
temperature=1,
|
29 |
+
top_k=256,
|
30 |
+
top_p=0.9,
|
31 |
+
)
|
32 |
+
|
33 |
+
print(bytes(sample.sequences[0].tolist()).decode('utf-8'))
|
34 |
+
```
|
35 |
+
|
36 |
+
Output
|
37 |
+
|
38 |
+
```
|
39 |
+
import torch
|
40 |
+
import numpy as np
|
41 |
+
import torch.nn.functional as F
|
42 |
+
from torch.autograd import Variable
|
43 |
+
|
44 |
+
from networkx.states import TransientState
|
45 |
+
|
46 |
+
def extract_data(num_epochs, epochs, is_last_epoch):
|
47 |
+
|
48 |
+
def get_data(num_features, num_classes):
|
49 |
+
data_features = num_features
|
50 |
+
data_classes = num_classes
|
51 |
+
data_labels = num_epochs
|
52 |
+
|
53 |
+
if num_features == 0 or num_classes == 0:
|
54 |
+
return data_features, data_classes
|
55 |
+
if is_last_epoch:
|
56 |
+
data_features = num_features
|
57 |
+
data_classes = num_classes
|
58 |
+
data_labels = num_epochs
|
59 |
+
return data_features, data_classes
|
60 |
+
|
61 |
+
data_features, data_classes = get_data(num_epochs, epochs, is_last_epoch)
|
62 |
+
data_labels = num_epochs * 2
|
63 |
+
return data_features, data_classes
|
64 |
+
|
65 |
+
|
66 |
+
class NumChannel:
|
67 |
+
def __init__(self, x, y, dx=1, dy=1, idx=1, data_size=2, epoch=None):
|
68 |
+
"""idx is the channel index with data feature in the first epoch.
|
69 |
+
x is the channel of the input data.
|
70 |
+
y is the element of the input data.
|
71 |
+
dx is the element of the data feature of the input data.
|
72 |
+
data_size is the size of the element of the data.
|
73 |
+
epoch is the channel of the element of the data.
|
74 |
+
"""
|
75 |
+
self.x = x
|
76 |
+
self.y = y
|
77 |
+
self.dx = dx
|
78 |
+
self.data_size = data_size
|
79 |
+
self.epoch = epoch
|
80 |
+
self.reference_count = 0
|
81 |
+
self.data_features = {}
|
82 |
+
self.data_classes = {}
|
83 |
+
|
84 |
+
self._initialize()
|
85 |
+
if idx is not None:
|
86 |
+
self._start_time = time.time()
|
87 |
+
|
88 |
+
def _initialize(self):
|
89 |
+
"""idx is the channel index with data feature in the first epoch.
|
90 |
+
x is the channel of the input data.
|
91 |
+
y is the element of the input data.
|
92 |
+
dx is the element of the data feature of the input data.
|
93 |
+
data_size is the size of the element of the data.
|
94 |
+
epoch is the channel of the element of the data.
|
95 |
+
"""
|
96 |
+
self.idx = idx
|
97 |
+
|
98 |
+
def _initialize(self):
|
99 |
+
"""idx is the channel of the inpu
|
100 |
+
```
|