File size: 17,281 Bytes
8ade218
 
9416971
8ade218
9416971
 
 
52529e2
9416971
5cc44e3
66d630e
2ae0e2b
f8e22b4
ae1084a
9416971
f4c8132
 
 
 
9416971
 
1cd6a22
 
 
 
52529e2
9416971
c2d54e2
a608366
9416971
 
a608366
c2d54e2
9416971
 
c2d54e2
9416971
 
 
 
 
 
 
c2d54e2
9416971
 
 
 
 
 
 
c2d54e2
9416971
 
 
 
 
9f347ce
 
d67f72a
9f347ce
 
 
 
 
 
 
 
9416971
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
580661e
9416971
2348608
 
 
 
 
 
 
 
 
 
 
 
 
 
114c6dc
90fed38
 
 
9416971
 
d1f0846
 
9416971
 
 
 
d1f0846
9416971
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d1f0846
9416971
 
 
 
 
 
 
 
 
 
 
 
 
114c6dc
90fed38
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
114c6dc
90fed38
 
 
 
9416971
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0ae4abd
 
 
 
9416971
 
 
 
 
 
 
 
 
 
 
0ae4abd
 
9416971
 
 
 
 
 
 
 
 
 
 
ae1084a
 
0ae4abd
 
 
 
9416971
 
ae1084a
 
9416971
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ae1084a
 
 
0ae4abd
 
 
 
 
 
9416971
ae1084a
9416971
 
 
 
114c6dc
9416971
 
90fed38
 
9416971
 
 
52d2d5a
9416971
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
---
license: apache-2.0
inference: false
---

# MistralLite Model

MistralLite is a fine-tuned [Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1) language model, with enhanced capabilities of processing long context (up to 32K tokens). By utilizing an adapted Rotary Embedding and sliding window during fine-tuning, MistralLite is able to **perform significantly better on several long context retrieve and answering tasks**, while keeping the simple model structure of the original model. MistralLite is useful for applications such as long context line and topic retrieval, summarization, question-answering, and etc. MistralLite can be deployed on a single AWS `g5.2x` instance with Sagemaker [Huggingface Text Generation Inference (TGI)](https://github.com/huggingface/text-generation-inference) endpoint, making it suitable for applications that require high performance in resource-constrained environments. You can also serve the MistralLite model directly using TGI docker containers. Also, MistralLite supports other ways of serving like [vLLM](https://github.com/vllm-project/vllm), and you can use MistralLite in Python by using the [HuggingFace transformers](https://huggingface.co/docs/transformers/index) and [FlashAttention-2](https://github.com/Dao-AILab/flash-attention) library.

MistralLite is similar to [Mistral-7B-Instruct-v0.1](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1), and their similarities and differences are summarized below:
|Model|Fine-tuned on long contexts| Max context length| RotaryEmbedding adaptation| Sliding Window Size|
|----------|-------------:|------------:|-----------:|-----------:|
| Mistral-7B-Instruct-v0.1 | up to 8K tokens | 32K | rope_theta = 10000 | 4096 |
| MistralLite | up to 16K tokens | 32K | **rope_theta = 1000000** | **16384** |

**Important - Use the prompt template below for MistralLite:**

```<|prompter|>What are the main challenges to support a long context for LLM?</s><|assistant|>```

## Motivation of Developing MistralLite

Since the release of [Mistral-7B-Instruct-v0.1](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1), the model became increasingly popular because its strong performance 
on a wide range of benchmarks. But most of the benchmarks are evaluated on `short context`, and not much has been investigated on its performance on long context tasks.
Then We evaluated `Mistral-7B-Instruct-v0.1` against benchmarks that are specifically designed to assess the capabilities of LLMs in handling longer context. 
Although the performance of the models on long context was fairly competitive on long context less than 4096 tokens, 
there were some limitations on its performance on longer context. Motivated by improving its performance on longer context, we finetuned the Mistral 7B model, and produced `Mistrallite`. The model managed to `significantly boost the performance of long context handling` over Mistral-7B-Instruct-v0.1. The detailed `long context evalutaion results` are as below: 

1. [Topic Retrieval](https://lmsys.org/blog/2023-06-29-longchat/)

|Model Name|Input length| Input length | Input length| Input length| Input length|
|----------|-------------:|-------------:|------------:|-----------:|-----------:|
|          | 2851| 5568 |8313 | 11044 | 13780 |
|   Mistral-7B-Instruct-v0.1  | 100%        | 50%       | 2%      | 0%     | 0% |
|   MistralLite   | **100%**        | **100%**       | **100%**      | **100%**     | **98%** |

2. [Line Retrieval](https://lmsys.org/blog/2023-06-29-longchat/#longeval-results)

|Model Name|Input length| Input length | Input length| Input length| Input length|Input length|
|----------|-------------:|-------------:|------------:|-----------:|-----------:|-----------:|
|          | 3818| 5661 |7505 | 9354 | 11188 | 12657 
|   Mistral-7B-Instruct-v0.1   | **98%**        | 62%       | 42%      | 42%     | 32% | 30% |
|   MistralLite   | **98%**        | **92%**       | **88%**      | **76%**     | **70%** | **60%** |

3. [Pass key Retrieval](https://github.com/epfml/landmark-attention/blob/main/llama/run_test.py#L101)

|Model Name|Input length| Input length | Input length| Input length|
|----------|-------------:|-------------:|------------:|-----------:|
|          | 3264| 5396 |8329 | 10197 | 
|   Mistral-7B-Instruct-v0.1   | **100%**        | 50%       | 20%      | 30%   |
|   MistralLite  | **100%**        | **100%**       | **100%**      | **100%**   |

4. [Question Answering with Long Input Texts](https://nyu-mll.github.io/quality/)
|Model Name| Test set Accuracy | Hard subset Accuracy|
|----------|-------------:|-------------:|
| Mistral-7B-Instruct-v0.1 | 44.3% | 39.7% |
| MistralLite | **64.4%** | **56.2%** |

### Example test on long context of 13400 tokens
- Context is from [Amazon Aurora FAQs](https://aws.amazon.com/rds/aurora/faqs/)
- Question: *please tell me how does pgvector help with Generative AI and give me some examples.*
- Answer from MistralLite:
  ```code
  pgvector is an open-source extension for PostgreSQL supported by Amazon Aurora PostgreSQL-Compatible Edition.

  You can use pgvector to store, search, index, and query billions of embeddings that are generated from machine learning (ML) and artificial intelligence (AI) models in your database, such as those from Amazon Bedrock (limited preview) or Amazon SageMaker. A vector embedding is a numerical representation that represents the semantic meaning of content such as text, images, and video.

  With pgvector, you can query embeddings in your Aurora PostgreSQL database to perform efficient semantic similarity searches of these data types, represented as vectors, combined with other tabular data in Aurora. This enables the use of generative AI and other AI/ML systems for new types of applications such as personalized recommendations based on similar text descriptions or images, candidate match based on interview notes, customer service next best action recommendations based on successful transcripts or chat session dialogs, and more.
  ```

## Model Details

- **Developed by:** [AWS Contributors](https://github.com/orgs/aws-samples/teams/aws-prototype-ml-apac)
- **Model type:** [Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
- **Language:** English
- **Finetuned from weights:** [Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
- **Finetuned on data:**
  - [SLidingEncoder and Decoder (SLED)](https://huggingface.co/datasets/tau/sled)
  - [(Long) Natural Questions (NQ)](https://huggingface.co/datasets/togethercomputer/Long-Data-Collections#multi-passage-qa-from-natural-questions)
  - [OpenAssistant Conversations Dataset (OASST1)](https://huggingface.co/datasets/OpenAssistant/oasst1)
- **Supported Serving Framework:**
  - [Text-Generation-Inference 1.1.0](https://github.com/huggingface/text-generation-inference/tree/v1.1.0)
  - [vLLM](https://github.com/vllm-project/vllm)
  - [HuggingFace transformers](https://huggingface.co/docs/transformers/index)
  - [HuggingFace Text Generation Inference (TGI) container on SageMaker](https://github.com/awslabs/llm-hosting-container)
- **Model License:** Apache 2.0
- **Contact:** [GitHub issues](https://github.com/awslabs/extending-the-context-length-of-open-source-llms/issues)
- **Inference Code** [Github Repo](https://github.com/awslabs/extending-the-context-length-of-open-source-llms/blob/main/MistralLite/)

## MistralLite LM-Eval Results

### Methodology

- Please see https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard
- revision=4ececff
- Note: we used --model hf-causal-experimental instead of --model hf-causal

### Results

|Average|hellaswag| arc_challenge|truthful_qa (mc2)| MMLU (acc)|
|----------|-------------:|------------:|-----------:|-----------:|
| 0.57221 | 0.81617 | 0.58874 | 0.38275 | 0.5012 |

## How to Use MistralLite from Python Code (HuggingFace transformers) ##

**Important** - For an end-to-end example Jupyter notebook, please refer to [this link](https://github.com/awslabs/extending-the-context-length-of-open-source-llms/blob/main/MistralLite/huggingface-transformers/example_usage.ipynb).

### Install the necessary packages

Requires: [transformers](https://pypi.org/project/transformers/) 4.34.0 or later, [flash-attn](https://pypi.org/project/flash-attn/) 2.3.1.post1 or later, 
and [accelerate](https://pypi.org/project/accelerate/) 0.23.0 or later.

```shell
pip install transformers==4.34.0
pip install flash-attn==2.3.1.post1 --no-build-isolation
pip install accelerate==0.23.0
```
### You can then try the following example code

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import transformers
import torch

model_id = "amazon/MistralLite"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id,
                                             torch_dtype=torch.bfloat16,
                                             use_flash_attention_2=True,
                                             device_map="auto",)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
)
prompt = "<|prompter|>What are the main challenges to support a long context for LLM?</s><|assistant|>"

sequences = pipeline(
    prompt,
    max_new_tokens=400,
    do_sample=False,
    return_full_text=False,
    num_return_sequences=1,
    eos_token_id=tokenizer.eos_token_id,
)
for seq in sequences:
    print(f"{seq['generated_text']}")
```
**Important** - Use the prompt template below for MistralLite:
```
<|prompter|>What are the main challenges to support a long context for LLM?</s><|assistant|>
```

## How to Serve MistralLite on TGI ##
**Important:** 
- For an end-to-end example Jupyter notebook using the native TGI container, please refer to [this link](https://github.com/awslabs/extending-the-context-length-of-open-source-llms/blob/main/MistralLite/tgi/example_usage.ipynb).
- If the **input context length is greater than 12K tokens**, it is recommended using a custom TGI container, please refer to [this link](https://github.com/awslabs/extending-the-context-length-of-open-source-llms/blob/main/MistralLite/tgi-custom/example_usage.ipynb).

### Start TGI server ###
Use TGI version 1.1.0 or later. The official Docker container is: `ghcr.io/huggingface/text-generation-inference:1.1.0`

Example Docker parameters:

```shell
docker run -d --gpus all --shm-size 1g -p 443:80 -v $(pwd)/models:/data ghcr.io/huggingface/text-generation-inference:1.1.0 \
      --model-id amazon/MistralLite \
      --max-input-length 16000 \
      --max-total-tokens 16384 \
      --max-batch-prefill-tokens 16384 \
      --trust-remote-code
```

### Perform Inference ###
Example Python code for inference with TGI (requires `text_generation` 0.6.1 or later):

```shell
pip install text_generation==0.6.1
```

```python
from text_generation import Client

SERVER_PORT = 443
SERVER_HOST = "localhost"
SERVER_URL = f"{SERVER_HOST}:{SERVER_PORT}"
tgi_client = Client(f"http://{SERVER_URL}", timeout=60)

def invoke_tgi(prompt, 
                      random_seed=1, 
                      max_new_tokens=400, 
                      print_stream=True,
                      assist_role=True):
    if (assist_role):
        prompt = f"<|prompter|>{prompt}</s><|assistant|>"
    output = ""
    for response in tgi_client.generate_stream(
        prompt,
        do_sample=False,
        max_new_tokens=max_new_tokens,
        return_full_text=False,
        #temperature=None,
        #truncate=None,
        #seed=random_seed,
        #typical_p=0.2,
    ):
        if hasattr(response, "token"):
            if not response.token.special:
                snippet = response.token.text
                output += snippet
                if (print_stream):
                    print(snippet, end='', flush=True)
    return output

prompt = "What are the main challenges to support a long context for LLM?"
result = invoke_tgi(prompt)
```

**Important** - When using MistralLite for inference for the first time, it may require a brief 'warm-up' period that can take 10s of seconds. However, subsequent inferences should be faster and return results in a more timely manner. This warm-up period is normal and should not affect the overall performance of the system once the initialisation period has been completed.


## How to Deploy MistralLite on Amazon SageMaker ##
**Important:** 
- For an end-to-end example Jupyter notebook using the SageMaker built-in container, please refer to [this link](https://github.com/awslabs/extending-the-context-length-of-open-source-llms/blob/main/MistralLite/sagemaker-tgi/example_usage.ipynb).
- If the **input context length is greater than 12K tokens**, it is recommended using a custom docker container, please refer to [this link](https://github.com/awslabs/extending-the-context-length-of-open-source-llms/blob/main/MistralLite/sagemaker-tgi-custom/example_usage.ipynb).

### Install the necessary packages

Requires: [sagemaker](https://pypi.org/project/sagemaker/) 2.192.1 or later.

```shell
pip install sagemaker==2.192.1
```

### Deploy the Model as A SageMaker Endpoint ###
To deploy MistralLite on a SageMaker endpoint, please follow the example code as below.
```python
import sagemaker
from sagemaker.huggingface import HuggingFaceModel, get_huggingface_llm_image_uri
import time

sagemaker_session = sagemaker.Session()
region = sagemaker_session.boto_region_name
role = sagemaker.get_execution_role()

image_uri = get_huggingface_llm_image_uri(
  backend="huggingface", # or lmi
  region=region,
 version="1.1.0"
)

model_name = "MistralLite-" + time.strftime("%Y-%m-%d-%H-%M-%S", time.gmtime())

hub = {
    'HF_MODEL_ID':'amazon/MistralLite',
    'HF_TASK':'text-generation',
    'SM_NUM_GPUS':'1',
    "MAX_INPUT_LENGTH": '16000',
    "MAX_TOTAL_TOKENS": '16384',
    "MAX_BATCH_PREFILL_TOKENS": '16384',
    "MAX_BATCH_TOTAL_TOKENS":  '16384',
}

model = HuggingFaceModel(
    name=model_name,
    env=hub,
    role=role,
    image_uri=image_uri
)
predictor = model.deploy(
  initial_instance_count=1,
  instance_type="ml.g5.2xlarge",
  endpoint_name=model_name,
    
)
```

### Perform Inference ###
To call the endpoint, please follow the example code as below:

```python
input_data = {
  "inputs": "<|prompter|>What are the main challenges to support a long context for LLM?</s><|assistant|>",
  "parameters": {
    "do_sample": False,
    "max_new_tokens": 400,
    "return_full_text": False,
    #"typical_p": 0.2,
    #"temperature":None,
    #"truncate":None,
    #"seed": 1,
  }
}
result = predictor.predict(input_data)[0]["generated_text"]
print(result)
```
or via [boto3](https://pypi.org/project/boto3/), and the example code is shown as below:

```python
import boto3
import json
def call_endpoint(client, prompt, endpoint_name, paramters):
    client = boto3.client("sagemaker-runtime")
    payload = {"inputs": prompt,
               "parameters": parameters}
    response = client.invoke_endpoint(EndpointName=endpoint_name,
                                      Body=json.dumps(payload), 
                                      ContentType="application/json")
    output = json.loads(response["Body"].read().decode())
    result = output[0]["generated_text"]
    return result

client = boto3.client("sagemaker-runtime")
parameters = {
    "do_sample": False,
    "max_new_tokens": 400,
    "return_full_text": False,
    #"typical_p": 0.2,
    #"temperature":None,
    #"truncate":None,
    #"seed": 1,
}
endpoint_name = predictor.endpoint_name
prompt = "<|prompter|>What are the main challenges to support a long context for LLM?</s><|assistant|>"
result = call_endpoint(client, prompt, endpoint_name, parameters)
print(result)
```


## How to Serve MistralLite on vLLM ##
Documentation on installing and using vLLM [can be found here](https://vllm.readthedocs.io/en/latest/).

**Important** - For an end-to-end example Jupyter notebook, please refer to [this link](https://github.com/awslabs/extending-the-context-length-of-open-source-llms/blob/main/MistralLite/vllm/example_usage.ipynb).

### Using vLLM as a server ###
When using vLLM as a server, pass the --model amazon/MistralLite parameter, for example:
```shell
python3 -m vllm.entrypoints.api_server --model amazon/MistralLite
```

### Using vLLM in Python Code ###
When using vLLM from Python code, Please see the example code as below:

```python
from vllm import LLM, SamplingParams

prompts = [
   "<|prompter|>What are the main challenges to support a long context for LLM?</s><|assistant|>",
]
sampling_params = SamplingParams(temperature=0, max_tokens=100)

llm = LLM(model="amazon/MistralLite",)

outputs = llm.generate(prompts, sampling_params)

# Print the outputs.
for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
```

## Limitations ##
Before using the MistralLite model, it is important to perform your own independent assessment, and take measures to ensure that your use would comply with your own specific quality control practices and standards, and that your use would comply with the local rules, laws, regulations, licenses and terms that apply to you, and your content.