Linly

LLaMA1-2 & Linly-OpenLLaMA & Falcon

** Linly-ChatFlow Chinese-LLaMA (1-2)Chinese-Falcon **

TencentPretrain Full-tuning

LLaMA Falcon Linly-ChatFlow

Linly-OpenLLaMA 3B7B13B 1TB tokenizer Apache 2.0

---

• Full-tuning LLaMAFalcon TencentPretrain HuggingFace
• CUDA

| | |

• [2023/7/22] Chinese-LLaMA-2 (7B13B) -

• [2023/6/14] Falcon-7B Falcon

• [2023/5/31] Linly-ChatFlow-7B SuperCLUE-

Past News

• [2023/5/28] v1.2 Chinese-LLaMA 2048** Linly-OpenLLaMA v0.1**

• [2023/5/14] v1.1 ChatFlow 1024 API

• [2023/4/27] Linly-ChatFlow-13B Linly-Chinese-LLaMA-33B

• [2023/4/17] llama_inference 8-bit

• [2023/4/8] TencentPretrain LoRA DeepSpeed Zero-3 Offload

• [2023/4/1] 4-bit Linly-ChatFlow llama.cpp

• [2023/3/28] LLaMA Linly-ChatFlow-7B

---

• ()
• FAQ
• TODO-List
• License
• Contributors

Linly-Chinese-LLaMA-2

Linly-Chinese-LLaMA-2

• LLaMA2 ****

Chinese-LLaMA-2-7B (hf)	/		2048	v0.1	2023.7.22
Chinese-LLaMA-2-13B (hf)	/		2048	v0.2	2023.8.12

Linly-Chinese-Falcon

• Chinese-Falcon Falcon Apache License 2.0

Chinese-Falcon-7B (hf)		50G	2048	v0.2	2023.6.15

Linly-Chinese-LLaMA

• Linly-Chinese-LLaMA LLaMA

**** LLaMA GNU General Public License v3.0

Chinese-LLaMA-7B		100G	2048	v1.2	2023.5.29
ChatFlow-7B		5M	1024	v1.1	2023.5.14
Chinese-LLaMA-13B		100G	2048	v1.2	2023.5.29
ChatFlow-13B		5M	1024	v1.1	2023.5.14
Chinese-LLaMA-33B (hf)		30G	512	v1.0	2023.4.27

HuggingFace
TencentPretrain Huggingface Huggingface

Linly-OpenLLaMA

• Linly-OpenLLaMA **** LLaMA Apache License 2.0

OpenLLaMA-13B		100G	2048	v0.1	2023.5.29

Linly-Chinese-LLaMA-2

1.

2.

2.

demo Linly-ChatFlow

API server

curl -H 'Content-Type: application/json' https://P01son-52nfefhaaova.serv-c1.openbayes.net -d '{"question": ""}'

HuggingFace OpenBayes

demostarllama_inference

TencentPretrain Linly-Chinese-LLaMA-2hfhuggingface

: py3.8.12 cuda11.2.2 cudnn8.1.1.33-1 torch1.9.0 bitsandbytes0.37.2

** llama_inference**

git lfs install
git clone https://huggingface.co/Linly-AI/ChatFlow-7B
git clone https://github.com/ProjectD-AI/llama_inference

cd llama_inference 
vi prompts.txt  #""

python3 llama_infer.py --test_path prompts.txt --prediction_path result.txt  \
                      --load_model_path ../ChatFlow-7B/chatflow_7b.bin  \
                      --config_path config/llama_7b_config.json \
                      --spm_model_path ../ChatFlow-7B/tokenizer.model --seq_length 512

python3 llama_dialogue.py --seq_length 512 --top_k 10   \
                      --load_model_path ../ChatFlow-7B/chatflow_7b.bin  \
                      --config_path ./config/llama_7b_config.json \
                      --spm_model_path ../ChatFlow-7B/tokenizer.model

Int8

python3 llama_infer.py --test_path prompts.txt --prediction_path result.txt  \
                      --load_model_path ../ChatFlow-7B/chatflow_7b.bin  \
                      --config_path config/llama_7b_config.json \
                      --spm_model_path ../ChatFlow-7B/tokenizer.model --seq_length 512 --use_int8 

flask

python3 llama_server.py --load_model_path ../ChatFlow-7B/chatflow_7b.bin  \
                        --config_path config/llama_7b_config.json \
                        --spm_model_path ../ChatFlow-7B/tokenizer.model --seq_length 512

curl -H 'Content-Type: application/json' http://127.0.0.1:8888/chat -d '{"question": ""}'

Gradio Demo

gradio

python llama_gradio.py --load_model_path ../ChatFlow-7B/chatflow_7b.bin  \
                        --config_path config/llama_7b_config.json \
                        --spm_model_path ../ChatFlow-7B/tokenizer.model --seq_length 512

http://127.0.0.1:7860/

Int4 CPU

ChatFlow llama.cpp Int4 CPU int4

TencentPretrain

LLaMA1 ChatFlow

Prompt	Linly-ChatFlow-7B	Linly-ChatFlow-13B
		1. 2. 3. 4. 5. 6. 7. 8. 9. 10.
	1234
PythonJavaScript	PythonWebJavaScript - PythonC/C++JavaScript - PythonJavaScriptHTML/CSSJQueryReactAngular	PythonJavaScriptPythonWebJavaScriptWeb
	1. 2. 3. 4. 5.	1. 2. 3. 4. 5.
	1qubit234 0 1	01
	--

	Linly-ChatFlow-7B	Linly-ChatFlow-13B
	150-20070-80 150-200 60 = 9-11.4	15,00020,000187
	204,60010	National GeographicNSF3025
-301	A. B. C.
12	54 $32$	- S S*(t/60)=X(t) X(t)tmin - 12Y=10^X3.75/4 - Y=(10^X-0.5)/2.5 Y=10^X3.75/4 (10^X-0.5)/2.5 = 3.75/4 * 10^X - 0.7510^(X+1) X = log10(2) + 1121.8log10(3.75)+1
A. B. C. D.	B.	B
:	1.2.	1. 2.
4A B C D	D 1. A 2. BB 3. C 4. D	A B C D
TrueFalseRandomTFRStatementtrue/false XYtruefalse	1. 2. 3. 4.	TrueTrueX T = (T X) v (T Y) v X True = F False = R
ABCAB AC BC A E 5 ?	ABCD211AE113	8/3 N N > 4 AB C B ABC 3 AB BC AC 1 AB C 1 AB C 2 AC BC 1 + 1 + 2 = 4
A B C D	D	D D

	Linly-ChatFlow-7B	Linly-ChatFlow-13B
GPS	GPSGPSGPS	1. 2. 3. 4.

	Linly-ChatFlow-7B	Linly-ChatFlow-13B
1:, 2:		"" ""
1:,, 2
____ A B C D	A	D

	Linly-ChatFlow-7B	Linly-ChatFlow-13B
80214?	x2x3x80x=80 280 = 160 380 = 240 160 + 240 + 240 = 540	x2x+14 3x + (2x+14) = 5x+14 5x + 14 xx662*6+14=286 + 28 = 34
50304010	30 + 40 + 10 = 80 80	- 30 - 40 - 10 - 70 - 60 - 90 = + - = 30 + 40 - 10 = 80 80
784?	x78*4 = 312 x + 312 x + 312 = ? 312: x = ? - 312 x	x4x+78 x+(4x+78)=216 5x=138 x=27.6 27109
302/3	1. - 302/316.7 17 2. 18	= = 30 2/3 = 20 = - = 30 - 20 10
653	x $$ x \div 2 + 1 = x - 1 $$ $$ x - 4 = 6 $$ $x=10$	1. 2 8-5-3= 3 2. 6 x 6x 3. 3 6x = 3 + ()x 4. x=24/73.43 4 5. x=4 6 4 = 24 24

	Linly-ChatFlow-7B	Linly-ChatFlow-13B
	LED
	- - - Title: Streaming Platforms and Their Impact on the Music Industry Scene 1: Introduction Host: Hi everyone, welcome to our podcast. Today we will be discussing the impact that streaming platforms have had on the music industry. So lets get started. Scene 2: The Challenges of Digital Distribution Guest: As a musician, I've seen the rise of streaming as a great opportunity for artists to reach new audiences, but it has also created challenges in terms of digital distribution. With so many platforms out there, how can an artist make sure their work is available everywhere? Scene 3: Streaming vs Physical Sales Guest: There are benefits to both physical sales and streaming, but the two models are not always complementary. Some argue that streaming services have devalued the importance of album purchases by giving away music for free. How do you think this shift in consumer behavior affects the music industry? Scene 4: Opportunities for Emerging Artists Guest: On the flip side, streaming services can provide emerging artists with opportunities they didn't have before. With easier access to their music, and no need for physical distribution, these artists can gain exposure more quickly than ever before. Scene 5: Potential Pitfalls Guest: Of course, not all aspects of the music streaming revolution have been positive. Issues like royalties and streaming rights have arisen, along with concerns over data privacy. Is it possible for streaming services to balance the needs of artists and fans with those of the larger music industry? Scene 6: Conclusion Host: Thanks to our guests today for a lively discussion about the changing landscape of streaming and its implications on the music industry. Stay tuned next week for another episode!	: Spotify, Apple MusicTidal : SpotifyApple MusicTidal : 20 :
	Interstellar CooperMurphCooper
		1. 2. 3. 4. 798 5. 6. 1. 2. 3. 4. 5.
	- ABC - XYZ - ABC ABC XYZ XYZ ABC	- SmartSales SmartSalesCRM SmartSales SmartSales
	XXX XXX XXX []	[] [] [] []
	ABC	[] [] []
		1. / 2. 3. 4. 5. 6.

	Linly-ChatFlow-7B	Linly-ChatFlow-13B
Luck is a dividend of sweat. The more you sweat, the luckier you get.
	La chance est le rsultat de l'effort du surplus, tu pertes plus de sueur, tu es plus heureux.	La chance est la bndiction attache l'effort, plus vous persvrez, plus vous tes heureux.
	Inside, know already, across the world like a neighbor.	There are many people who know you within our country, but they appear to be like a neighbor in your faraway place.
	The past is always untrue, memory is an untraced road and the former all spring cannot be reproduced. Even most passionate and firm love is eventually a transient reality, at root it's just a kind of momentary thing, eternal loneliness alone.	The past is always a lie, Memory is a one-way street without a turning point. The most passionate love is ultimately a reality that is gone in an instant. Loneliness lasts for ever.
	La vita non vera importante come hai avuto quali esperienze, ma come hai ricordato le cose e perch.	La verit che conta nella vita non quello che ti capita, bens cosa ricordi e come tu te ne sei poi ricordato.
	All the brilliant in life, will be paid back with loneliness.
Il n'y a qu'un seul hroisme au monde: c'est de voir le monde tel qu'il est, et de l'aimer.
Tener un milln de amigos no es una maravilla, una maravilla es tener un amigo que est contigo, cuando millones estn en tu contra.	,
Meglio vedere una cosa una volta che sentirne parlare cento volte.

	Linly-ChatFlow-7B	Linly-ChatFlow-13B
python	Pythonpython<br/>def quick_sort(arr):<br/> if len(arr) <= 1:<br/> return arr<br/> else:<br/> pivot = arr[0] # <br/> left = [x for x in arr[1:] if x < pivot] # <br/> right = [x for x in arr[1:] if x >= pivot] # <br/> return quick_sort(left) + [pivot] + quick_sort(right) # <br/><br/># <br/>arr = [64, 34, 25, 12, 22, 11, 90]<br/>print(":", arr)<br/>sorted_arr = quick_sort(arr)<br/>print(":", sorted_arr)<br/>	Python# Quick sort def quickSort(arr): if len(arr) <= 1: return arr else: pivot = arr[0] # Set pivot element less = [] greater = [] for i in range(1, len(arr)): if arr[i] < pivot: less.append(arr[i]) else: greater.append(arr[i]) less.sort() # Sort the first list greater.sort() # Sort the second list return [less + greater][len(greater)//2] # Merge the two sorted lists to get the final result
pytorchReLU	PythonPyTorchReLUimport torch from torch import nn,nn.Linear,nn.functional as F def full_conn(inputs): with torch.no_grad(): # Define the convolutional layer conv = nn.Sequential(nn.Linear(128, 32),nn.ReLU()) # Apply convolutional layer to input data outputs = conv(inputs) return outputs	PyTorchReLUPython import torch.nn as nn class ConvNet(nn.Module): def init(self,input_size,hidden_size,dropout=0.5): super(ConvNet, self).init() self.conv1 = nn.Conv2d(input_size, hidden_size, kernel_size=[5,5], stride=2) self.relu1 = nn.ReLU(inplace=True) self.pool1 = nn.MaxPool2d(kernel_size=[2,2]) self.dense = nn.Sequential( nn.Linear(hidden_size44, 128), nn.ReLU(), nn.Dropout(dropout), nn.Linear(128,10)) def forward(self, x): h1 = self.pool1(self.conv1(x)) h1 = self.relu1(h1) h1 = self.dense(h1)
pythonDijkstra	# Function to calculate the shortest path for a given node. def shortest_path(graph, v): # Initialize the list of current paths paths = [] # Get all the adjacent nodes from this node adjacencies = graph[v] # Set current path length to infinite length curr_len = float('inf') # Iterate through all adjacent nodes for i in range(len(adjacencies)): if i not in paths: # Check if we have reached the goal node if adjacencies[i][v]: # If yes then set the curr_len to that edge's length curr_len = min(curr_len, (graph[v][i])) paths.append([v]) # Calculate the actual cost of each path and get the shortest one for p in paths: for j in range(len(p) - 1): if graph[p[j]] < curr_len + graph[p[j + 1]] - graph[p[j]]: curr_len = min(curr_len, graph[p[j]]) paths.remove(p[:j+1]) break else: paths.append(p) curr_len = graph[p[-1]] return curr_len	# Dijkstra def dijkstra(graph,start,target): # Create an empty priority queue with start as the root node heap = [(0,start)] visited = [False] * len(graph) # Mark all nodes not in the queue as unvisited for i in range(len(graph)): if i == target: continue else: visited[i] = False # Use a while loop to find shortest path while len(heap) > 0: currentNode = heap.pop() neighbors = graph[currentNode] # Update priority of all neighboring nodes for n in neighbors: if n != currentNode and visited[n]: continue dist = currentNode + [0] dist[1] += currentNode dist[2] = min(dist[1], max(dist, currentNode[0] + currentNode[2])) currentNode[2] = dist[2] priority = (dist[2] * 3 + currentNode[1]) / 6 priorities.append((priority, n)) heap.insert([dist[1], n, dist[2]]) return []
pythonnumsO(n)	Python python<br/>def find_not_existing_min_positive(nums):<br/> max_index = len(nums) - 1<br/> for i in range(max_index + 1, 0, -1):<br/> if nums[i] < 0:<br/> return nums.index(i) + 1<br/> return None<br/>	<br/>def find_minimums(nums):<br/> min_ = nums[0]<br/> for i in range(1, len(nums)):<br/> if nums[i] < min_:<br/> min_ = nums[i]<br/> return min_<br/>
python	<br/>def merge_lists(list1, list2):<br/> if len(list1) < len(list2):<br/> merged = list1 + list2<br/> else:<br/> merged = list2 + list1<br/> return merged<br/>	python<br/>def merge_sorted_linked_lists(heads):<br/> if not heads:<br/> return []<br/> cur = head = heads[0]<br/> for h in heads[1:]:<br/> cur.next = h<br/> cur = h<br/> return [cur] + merge_sorted_linked_lists(heads)<br/>
C++swordswordsswords	#include using namespace std; int main() { char s[10]="This is a str"; int n = strlen(s); // strlens int words_length=n+2; // words_length char arr[words_length]; // arr for (int i=0;i<n-2;i++) arr[i]=s[i]; // sarr cout<<arr<<endl; return 0; }	s = "abcdefg", words = {"abcd", "efg"} C++ s1for s s1 s word_length word_size s1 s s1 s1 s words
C++haystackneedlehaystackneedle 0 needlehaystack-1	int find_first_match(char *haystack, char *needle){ int i; for (i = 0; i < strlen(haystack); i++) { if (strcmp(haystack[i], needle) == 0) { return i - 1; // return the first matching item's offset from beginning of haystack. } else if (i + strlen(needle) - 1 >= 0 && strncmp(haystack+i, needle+1, strlen(needle)) == 0) { return i; // return the first match and last match together. } } return -1; }	C++ haystack needle haystack needle find() -1 le haystack

Linly-ChatFlow Linly-ChatFlow Linly-ChatFlow

FAQ

Q1

7B 14Gint8 7G13B 28Gint8 14G

Q2

*GPU

Q3LoRA

LoRAFull-tuning TencentPretrain LoRA LoRA

Star History

Citation

• CSL: A Large-scale Chinese Scientific Literature Dataset (COLING 2022) [paper][code]

• TencentPretrain: A Scalable and Flexible Toolkit for Pre-training Models of Different Modalities (ACL 2023) [paper][code]

License

Our code and documents are released under Apache Licence 2.0

Following LLaMA, our pre-trained weights are released under GNU General Public License v3.0

Contributors

We thank contributors for both TencentPretrain and Chinese-ChatLLaMA projects.

Authors: Yudong Li, Yuhao Feng, Zhe Zhao, Cheng Hou, Bizhu Wu, Wen Zhou, Hao Li, Xiaoqin Wang, Wenhang Shi, Yaning Zhang, Shuang Li, Siri Long, Xianxu Hou , Yiren Chen, Jing Zhao, Ningyuan Sun ,Wenjun Tang, Xiaoshuai Chen

Corresponding Authors: Linlin Shen