abhi-mosaic
commited on
Commit
•
0fc7cfd
1
Parent(s):
ea98a59
README.md
Browse files
README.md
CHANGED
@@ -49,34 +49,37 @@ model = transformers.AutoModelForCausalLM.from_pretrained(
|
|
49 |
)
|
50 |
```
|
51 |
|
52 |
-
To use the optimized [triton implementation](https://github.com/openai/triton) of FlashAttention, you can load the model with `attn_impl='triton'` and
|
53 |
```python
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
|
|
|
|
58 |
config.attn_config['attn_impl'] = 'triton'
|
|
|
59 |
|
60 |
model = transformers.AutoModelForCausalLM.from_pretrained(
|
61 |
-
|
62 |
config=config,
|
63 |
-
torch_dtype=torch.bfloat16,
|
64 |
trust_remote_code=True
|
65 |
)
|
66 |
-
model.to(device='cuda:0')
|
67 |
```
|
68 |
|
69 |
-
Although the model was trained with a sequence length of 2048 and finetuned with a sequence length of 65536,
|
70 |
ALiBi enables users to increase the maximum sequence length during finetuning and/or inference. For example:
|
71 |
-
|
72 |
```python
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
config.
|
|
|
|
|
78 |
model = transformers.AutoModelForCausalLM.from_pretrained(
|
79 |
-
|
80 |
config=config,
|
81 |
trust_remote_code=True
|
82 |
)
|
@@ -155,8 +158,8 @@ The data was tokenized using the [EleutherAI/gpt-neox-20b](https://huggingface.c
|
|
155 |
|
156 |
### Training Configuration
|
157 |
|
158 |
-
This model was trained on 8 A100-80GBs for about 2 days using the [MosaicML Platform](https://www.mosaicml.com/platform).
|
159 |
-
The model was trained with sharded data parallelism using [FSDP](https://pytorch.org/docs/stable/fsdp.html) and used the [LION](https://arxiv.org/abs/2302.06675) optimizer.
|
160 |
|
161 |
## Limitations and Biases
|
162 |
|
@@ -193,4 +196,4 @@ Please cite this model using the following format:
|
|
193 |
note = {Accessed: 2023-03-28}, % change this date
|
194 |
urldate = {2023-03-28} % change this date
|
195 |
}
|
196 |
-
```
|
|
|
49 |
)
|
50 |
```
|
51 |
|
52 |
+
To use the optimized [triton implementation](https://github.com/openai/triton) of FlashAttention, you can load the model on GPU (`cuda:0`) with `attn_impl='triton'` and with `bfloat16` precision:
|
53 |
```python
|
54 |
+
import torch
|
55 |
+
import transformers
|
56 |
+
|
57 |
+
name = 'mosaicml/mpt-7b-storywriter'
|
58 |
+
|
59 |
+
config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True)
|
60 |
config.attn_config['attn_impl'] = 'triton'
|
61 |
+
config.init_device = 'cuda:0' # For fast initialization directly on GPU!
|
62 |
|
63 |
model = transformers.AutoModelForCausalLM.from_pretrained(
|
64 |
+
name,
|
65 |
config=config,
|
66 |
+
torch_dtype=torch.bfloat16, # Load model weights in bfloat16
|
67 |
trust_remote_code=True
|
68 |
)
|
|
|
69 |
```
|
70 |
|
71 |
+
Although the model was trained with a sequence length of 2048 and finetuned with a sequence length of 65536,
|
72 |
ALiBi enables users to increase the maximum sequence length during finetuning and/or inference. For example:
|
|
|
73 |
```python
|
74 |
+
import transformers
|
75 |
+
|
76 |
+
name = 'mosaicml/mpt-7b'
|
77 |
+
|
78 |
+
config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True)
|
79 |
+
config.max_seq_len = 83968 # (input + output) tokens can now be up to 83968
|
80 |
+
|
81 |
model = transformers.AutoModelForCausalLM.from_pretrained(
|
82 |
+
name,
|
83 |
config=config,
|
84 |
trust_remote_code=True
|
85 |
)
|
|
|
158 |
|
159 |
### Training Configuration
|
160 |
|
161 |
+
This model was trained on 8 A100-80GBs for about 2 days using the [MosaicML Platform](https://www.mosaicml.com/platform).
|
162 |
+
The model was trained with sharded data parallelism using [FSDP](https://pytorch.org/docs/stable/fsdp.html) and used the [LION](https://arxiv.org/abs/2302.06675) optimizer.
|
163 |
|
164 |
## Limitations and Biases
|
165 |
|
|
|
196 |
note = {Accessed: 2023-03-28}, % change this date
|
197 |
urldate = {2023-03-28} % change this date
|
198 |
}
|
199 |
+
```
|