Commit
β’
0e06ad0
1
Parent(s):
9600986
Update README.md
Browse files
README.md
CHANGED
@@ -20,11 +20,14 @@ This official repository unveils the TransNormerLLM3 model along with its open-s
|
|
20 |
|
21 |
[TransNormerLLM](https://arxiv.org/abs/2307.14995) evolving from [TransNormer](https://arxiv.org/abs/2210.10340), standing out as the first LLM within the linear transformer architecture. Additionally, it distinguishes itself by being the first non-Transformer LLM to exceed both traditional Transformer and other efficient Transformer models (such as, RetNet and Mamba) in terms of speed and performance.
|
22 |
|
|
|
23 |
|
24 |
# TransNormerLLM3
|
25 |
- **TransNormerLLM3-15B** features **14.83 billion** parameters. It is structured with **42 layers**, includes **40 attention heads**, and has a total **embedding size of 5120**.
|
26 |
-
- **TransNormerLLM3-15B** is purely intergrated with **[Lightning Attention-2](http://arxiv.org/abs/2401.04658)**, which can maintain a **stable TGS** during training of **unlimited sequence lengths**, up until encountering firm limitations like GPU memory constraints.
|
27 |
- **Titoken** tokenizer is used with a total **vocabulary size** of about **100,000**.
|
|
|
|
|
28 |
<p align="center">
|
29 |
<img src="./images/TransNormer3.jpg" width="65%" />
|
30 |
</p>
|
@@ -43,26 +46,35 @@ This official repository unveils the TransNormerLLM3 model along with its open-s
|
|
43 |
|
44 |
# Released Weights
|
45 |
|
46 |
-
| param | token | Hugging Face
|
47 |
-
| :-----: | :---: |
|
48 |
-
| **15B** | 50B |
|
49 |
-
| **15B** | 100B | π€[step26000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step26000-100Btokens)
|
50 |
-
| **15B** | 150B | π€[step39000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step39000-150Btokens)
|
51 |
-
| **15B** | 200B | π€[step52000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step52000-200Btokens)
|
52 |
-
| **15B** | 250B | π€[step65000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step65000-250Btokens)
|
53 |
-
| **15B** | 300B | π€[step78000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step78000-300Btokens)
|
54 |
-
| **15B** | 350B | π€[step92000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step92000-350Btokens)
|
55 |
-
| **15B** | 400B | π€[step105000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step105000-400Btokens)
|
56 |
-
| **15B** | 450B | π€[step118000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step118000-450Btokens)
|
57 |
-
| **15B** | 500B | π€[step131000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step131000-500Btokens)
|
58 |
-
| **15B** | 550B | π€[step144000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step144000-550Btokens)
|
59 |
-
| **15B** | 600B | π€[step157000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step157000-600Btokens)
|
60 |
-
| **15B** | 650B | π€[step170000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step170000-650Btokens)
|
61 |
-
| **15B** | 700B | π€[step183000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step183000-700Btokens)
|
62 |
-
| **15B** | 750B | π€[step195500](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step195500-750Btokens)
|
63 |
-
| **15B** | 800B | π€[step209000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step209000-800Btokens)
|
64 |
-
| **15B** | 850B | π€[step222000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step222000-850Btokens)
|
65 |
-
| **15B** | 900B | π€[step235000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step235000-900Btokens)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
|
67 |
|
68 |
|
@@ -76,26 +88,35 @@ model = AutoModelForCausalLM.from_pretrained("OpenNLPLab/TransNormerLLM3-15B-Int
|
|
76 |
# Benchmark Results
|
77 |
The evaluations of all models are conducted using the official settings and the [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) framework.
|
78 |
|
79 |
-
| Model | P | T
|
80 |
-
| ----------------------- | --- |
|
81 |
-
| **TransNormerLLM3-15B** | 15 | 0.05
|
82 |
-
| **TransNormerLLM3-15B** | 15 | 0.10
|
83 |
-
| **TransNormerLLM3-15B** | 15 | 0.15
|
84 |
-
| **TransNormerLLM3-15B** | 15 | 0.20
|
85 |
-
| **TransNormerLLM3-15B** | 15 | 0.25
|
86 |
-
| **TransNormerLLM3-15B** | 15 | 0.30
|
87 |
-
| **TransNormerLLM3-15B** | 15 | 0.35
|
88 |
-
| **TransNormerLLM3-15B** | 15 | 0.40
|
89 |
-
| **TransNormerLLM3-15B** | 15 | 0.45
|
90 |
-
| **TransNormerLLM3-15B** | 15 | 0.50
|
91 |
-
| **TransNormerLLM3-15B** | 15 | 0.55
|
92 |
-
| **TransNormerLLM3-15B** | 15 | 0.60
|
93 |
-
| **TransNormerLLM3-15B** | 15 | 0.65
|
94 |
-
| **TransNormerLLM3-15B** | 15 | 0.70
|
95 |
-
| **TransNormerLLM3-15B** | 15 | 0.75
|
96 |
-
| **TransNormerLLM3-15B** | 15 | 0.80
|
97 |
-
| **TransNormerLLM3-15B** | 15 | 0.85
|
98 |
-
| **TransNormerLLM3-15B** | 15 | 0.90
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
99 |
|
100 |
|
101 |
> **P**: parameter size (billion). **T**: tokens (trillion). **BoolQ**: acc. **PIQA**: acc. **HellaSwag**: acc_norm. **WinoGrande**: acc. **ARC-easy**: acc. **ARC-challenge**: acc_norm. **OpenBookQA**: acc_norm. **MMLU**: 5-shot acc. **C-Eval**: 5-shot acc.
|
@@ -136,6 +157,14 @@ If you wish to cite our work, please use the following reference:
|
|
136 |
archivePrefix={arXiv},
|
137 |
primaryClass={cs.CL}
|
138 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
139 |
```
|
140 |
|
141 |
<p align="center">
|
|
|
20 |
|
21 |
[TransNormerLLM](https://arxiv.org/abs/2307.14995) evolving from [TransNormer](https://arxiv.org/abs/2210.10340), standing out as the first LLM within the linear transformer architecture. Additionally, it distinguishes itself by being the first non-Transformer LLM to exceed both traditional Transformer and other efficient Transformer models (such as, RetNet and Mamba) in terms of speed and performance.
|
22 |
|
23 |
+
> Update@Apr.7: We plan to increase the sequence length in pre-training stage to **10 million**: https://twitter.com/opennlplab/status/1776894730015789300
|
24 |
|
25 |
# TransNormerLLM3
|
26 |
- **TransNormerLLM3-15B** features **14.83 billion** parameters. It is structured with **42 layers**, includes **40 attention heads**, and has a total **embedding size of 5120**.
|
27 |
+
- **TransNormerLLM3-15B** is purely intergrated with **[Lightning Attention-2](http://arxiv.org/abs/2401.04658)**, which can maintain a **stable TGS** during training of **unlimited sequence lengths**, up until encountering firm limitations like GPU memory constraints.
|
28 |
- **Titoken** tokenizer is used with a total **vocabulary size** of about **100,000**.
|
29 |
+
- Our **training framework** has been enhanced with integration to **[LASP](https://arxiv.org/abs/2404.02882) (Linear Attention Sequence Parallelism)**, allowing for sequence parallelism within linear attention models.
|
30 |
+
- Our **training framework** now supprts **[CO2](https://arxiv.org/abs/2401.16265)**, which introduces **local updates** and **asynchronous communication** into distributed data parallel training, achieving **full overlap** of communication and computation.
|
31 |
<p align="center">
|
32 |
<img src="./images/TransNormer3.jpg" width="65%" />
|
33 |
</p>
|
|
|
46 |
|
47 |
# Released Weights
|
48 |
|
49 |
+
| param | token | Hugging Face | Model Scope | Wisemodel |
|
50 |
+
| :-----: | :---: | :-----------------------------------------------------------------------------------------------------------------------: | :---------: | :-------: |
|
51 |
+
| **15B** | 50B | π€[step13000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step13000-50Btokens) | π€ | π― |
|
52 |
+
| **15B** | 100B | π€[step26000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step26000-100Btokens) | π€ | π― |
|
53 |
+
| **15B** | 150B | π€[step39000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step39000-150Btokens) | π€ | π― |
|
54 |
+
| **15B** | 200B | π€[step52000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step52000-200Btokens) | π€ | π― |
|
55 |
+
| **15B** | 250B | π€[step65000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step65000-250Btokens) | π€ | π― |
|
56 |
+
| **15B** | 300B | π€[step78000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step78000-300Btokens) | π€ | π― |
|
57 |
+
| **15B** | 350B | π€[step92000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step92000-350Btokens) | π€ | π― |
|
58 |
+
| **15B** | 400B | π€[step105000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step105000-400Btokens) | π€ | π― |
|
59 |
+
| **15B** | 450B | π€[step118000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step118000-450Btokens) | π€ | π― |
|
60 |
+
| **15B** | 500B | π€[step131000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step131000-500Btokens) | π€ | π― |
|
61 |
+
| **15B** | 550B | π€[step144000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step144000-550Btokens) | π€ | π― |
|
62 |
+
| **15B** | 600B | π€[step157000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step157000-600Btokens) | π€ | π― |
|
63 |
+
| **15B** | 650B | π€[step170000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step170000-650Btokens) | π€ | π― |
|
64 |
+
| **15B** | 700B | π€[step183000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step183000-700Btokens) | π€ | π― |
|
65 |
+
| **15B** | 750B | π€[step195500](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step195500-750Btokens) | π€ | π― |
|
66 |
+
| **15B** | 800B | π€[step209000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step209000-800Btokens) | π€ | π― |
|
67 |
+
| **15B** | 850B | π€[step222000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step222000-850Btokens) | π€ | π― |
|
68 |
+
| **15B** | 900B | π€[step235000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step235000-900Btokens) | π€ | π― |
|
69 |
+
| **15B** | 950B | π€[step248000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step248000-950Btokens) | π€ | π― |
|
70 |
+
| **15B** | 1000B | π€[step261000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step261000-1000Btokens) | π€ | π― |
|
71 |
+
| **15B** | 1050B | π€[step274000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step274000-1050Btokens) | π€ | π― |
|
72 |
+
| **15B** | 1100B | π€[step287000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step287000-1100Btokens) | π€ | π― |
|
73 |
+
| **15B** | 1150B | π€[step300000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step300000-1150Btokens) | π€ | π― |
|
74 |
+
| **15B** | 1200B | π€[step313500](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step313500-1200Btokens) | π€ | π― |
|
75 |
+
| **15B** | 1250B | π€[step326000](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step287000-1250Btokens) | π€ | π― |
|
76 |
+
| **15B** | 1300B | π€[step339500](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/step326000-1300Btokens) | π€ | π― |
|
77 |
+
| **15B** | 1345B | π€[stage1](https://huggingface.co/OpenNLPLab/TransNormerLLM3-15B-Intermediate-Checkpoints/tree/stage1-1345Btokens) | π€ | π― |
|
78 |
|
79 |
|
80 |
|
|
|
88 |
# Benchmark Results
|
89 |
The evaluations of all models are conducted using the official settings and the [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) framework.
|
90 |
|
91 |
+
| Model | P | T | BoolQ | PIQA | HS | WG | ARC-e | ARC-c | OBQA | C-Eval | MMLU |
|
92 |
+
| ----------------------- | --- | ------ | ----- | ----- | ----- | ----- | ----- | ----- | ----- | ------ | ----- |
|
93 |
+
| **TransNormerLLM3-15B** | 15 | 0.05 | 62.08 | 72.52 | 55.55 | 57.14 | 62.12 | 31.14 | 32.40 | 26.18 | 27.50 |
|
94 |
+
| **TransNormerLLM3-15B** | 15 | 0.10 | 63.98 | 74.70 | 61.09 | 61.33 | 65.95 | 34.64 | 35.60 | 25.38 | 27.40 |
|
95 |
+
| **TransNormerLLM3-15B** | 15 | 0.15 | 60.34 | 75.08 | 63.99 | 62.04 | 64.56 | 34.90 | 35.20 | 22.64 | 26.60 |
|
96 |
+
| **TransNormerLLM3-15B** | 15 | 0.20 | 52.05 | 74.48 | 64.72 | 62.75 | 66.16 | 35.15 | 36.80 | 27.25 | 30.80 |
|
97 |
+
| **TransNormerLLM3-15B** | 15 | 0.25 | 66.70 | 76.50 | 66.51 | 64.80 | 66.84 | 36.18 | 39.40 | 30.87 | 36.10 |
|
98 |
+
| **TransNormerLLM3-15B** | 15 | 0.30 | 67.00 | 76.50 | 67.17 | 64.40 | 66.29 | 36.77 | 38.80 | 33.99 | 37.60 |
|
99 |
+
| **TransNormerLLM3-15B** | 15 | 0.35 | 65.78 | 75.46 | 67.88 | 66.54 | 67.34 | 38.57 | 39.60 | 36.02 | 39.20 |
|
100 |
+
| **TransNormerLLM3-15B** | 15 | 0.40 | 67.34 | 75.24 | 68.51 | 66.22 | 68.94 | 40.10 | 39.20 | 36.91 | 41.10 |
|
101 |
+
| **TransNormerLLM3-15B** | 15 | 0.45 | 69.02 | 76.28 | 69.11 | 63.77 | 65.82 | 36.01 | 39.40 | 37.17 | 42.80 |
|
102 |
+
| **TransNormerLLM3-15B** | 15 | 0.50 | 66.15 | 77.09 | 69.75 | 65.11 | 68.56 | 35.84 | 39.60 | 39.81 | 42.00 |
|
103 |
+
| **TransNormerLLM3-15B** | 15 | 0.55 | 70.24 | 74.05 | 69.96 | 65.75 | 65.61 | 36.69 | 38.60 | 40.08 | 44.00 |
|
104 |
+
| **TransNormerLLM3-15B** | 15 | 0.60 | 74.34 | 75.68 | 70.44 | 66.22 | 69.36 | 38.40 | 38.40 | 41.05 | 45.30 |
|
105 |
+
| **TransNormerLLM3-15B** | 15 | 0.65 | 73.15 | 76.55 | 71.60 | 66.46 | 69.65 | 39.68 | 40.80 | 41.20 | 44.90 |
|
106 |
+
| **TransNormerLLM3-15B** | 15 | 0.70 | 73.79 | 78.18 | 73.26 | 67.56 | 71.21 | 43.60 | 40.80 | 43.46 | 47.00 |
|
107 |
+
| **TransNormerLLM3-15B** | 15 | 0.75 | 76.45 | 78.07 | 74.22 | 69.30 | 71.21 | 43.43 | 42.20 | 43.46 | 47.80 |
|
108 |
+
| **TransNormerLLM3-15B** | 15 | 0.80 | 76.97 | 78.84 | 74.95 | 69.85 | 72.14 | 43.52 | 41.20 | 45.21 | 49.41 |
|
109 |
+
| **TransNormerLLM3-15B** | 15 | 0.85 | 72.75 | 78.35 | 75.91 | 70.48 | 74.58 | 45.22 | 41.20 | 46.27 | 49.36 |
|
110 |
+
| **TransNormerLLM3-15B** | 15 | 0.90 | 76.09 | 77.91 | 76.49 | 70.88 | 72.14 | 42.92 | 40.20 | 45.70 | 50.15 |
|
111 |
+
| **TransNormerLLM3-15B** | 15 | 0.95 | 74.28 | 78.24 | 76.63 | 72.22 | 74.12 | 44.11 | 42.40 | 46.25 | 51.43 |
|
112 |
+
| **TransNormerLLM3-15B** | 15 | 1.00 | 74.62 | 79.16 | 77.35 | 72.22 | 73.86 | 45.14 | 43.40 | 47.90 | 51.65 |
|
113 |
+
| **TransNormerLLM3-15B** | 15 | 1.05 | 76.36 | 78.94 | 77.15 | 71.35 | 74.66 | 44.45 | 42.80 | 45.87 | 52.28 |
|
114 |
+
| **TransNormerLLM3-15B** | 15 | 1.10 | 76.88 | 78.73 | 77.62 | 70.88 | 74.41 | 45.48 | 42.80 | 49.78 | 53.01 |
|
115 |
+
| **TransNormerLLM3-15B** | 15 | 1.15 | 72.87 | 79.43 | 78.12 | 72.85 | 74.75 | 46.16 | 43.20 | 49.80 | 53.04 |
|
116 |
+
| **TransNormerLLM3-15B** | 15 | 1.20 | 79.48 | 78.67 | 78.45 | 72.93 | 75.42 | 44.37 | 43.60 | 49.33 | 53.80 |
|
117 |
+
| **TransNormerLLM3-15B** | 15 | 1.25 | 79.17 | 79.16 | 78.81 | 72.93 | 75.13 | 45.99 | 43.60 | 50.44 | 54.19 |
|
118 |
+
| **TransNormerLLM3-15B** | 15 | 1.30 | 78.41 | 79.00 | 78.39 | 71.90 | 74.33 | 45.05 | 42.80 | 52.24 | 54.41 |
|
119 |
+
| **TransNormerLLM3-15B** | 15 | stage1 | 78.75 | 79.27 | 78.33 | 71.35 | 75.97 | 46.42 | 45.00 | 50.25 | 54.50 |
|
120 |
|
121 |
|
122 |
> **P**: parameter size (billion). **T**: tokens (trillion). **BoolQ**: acc. **PIQA**: acc. **HellaSwag**: acc_norm. **WinoGrande**: acc. **ARC-easy**: acc. **ARC-challenge**: acc_norm. **OpenBookQA**: acc_norm. **MMLU**: 5-shot acc. **C-Eval**: 5-shot acc.
|
|
|
157 |
archivePrefix={arXiv},
|
158 |
primaryClass={cs.CL}
|
159 |
}
|
160 |
+
@misc{sun2024linear,
|
161 |
+
title={Linear Attention Sequence Parallelism},
|
162 |
+
author={Weigao Sun and Zhen Qin and Dong Li and Xuyang Shen and Yu Qiao and Yiran Zhong},
|
163 |
+
year={2024},
|
164 |
+
eprint={2404.02882},
|
165 |
+
archivePrefix={arXiv},
|
166 |
+
primaryClass={cs.LG}
|
167 |
+
}
|
168 |
```
|
169 |
|
170 |
<p align="center">
|