Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,66 @@
|
|
1 |
-
---
|
2 |
-
license: apache-2.0
|
3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: apache-2.0
|
3 |
+
---
|
4 |
+
|
5 |
+
|
6 |
+
# Imran1/Qwen2.5-72B-Instruct-FP8
|
7 |
+
|
8 |
+
## Overview
|
9 |
+
**Imran1/Qwen2.5-72B-Instruct-FP8** is an optimized version of the base model **Qwen2.5-72B-Instruct**, utilizing **FP8** (8-bit floating point) precision. This reduces memory usage and increases computational efficiency, making it ideal for large-scale inference tasks without sacrificing the model's performance.
|
10 |
+
|
11 |
+
This model is well-suited for applications such as:
|
12 |
+
- Conversational AI and chatbots
|
13 |
+
- Instruction-based tasks
|
14 |
+
- Text generation, summarization, and dialogue completion
|
15 |
+
|
16 |
+
## Key Features
|
17 |
+
- **72 billion parameters** for powerful language generation and understanding capabilities.
|
18 |
+
- **FP8 precision** for reduced memory consumption and faster inference.
|
19 |
+
- Supports **tensor parallelism** for distributed computing environments.
|
20 |
+
|
21 |
+
## Usage Instructions
|
22 |
+
|
23 |
+
### 1. Running the Model with vLLM
|
24 |
+
You can serve the model using **vLLM** with tensor parallelism enabled. Below is an example command for running the model:
|
25 |
+
|
26 |
+
```bash
|
27 |
+
vllm serve Imran1/Qwen2.5-72B-Instruct-FP8 --api-key token-abc123 --tensor-parallel-size 2
|
28 |
+
```
|
29 |
+
|
30 |
+
### 2. Interacting with the Model via Python (OpenAI API)
|
31 |
+
Here’s an example of how to interact with the model using the OpenAI API interface:
|
32 |
+
|
33 |
+
```python
|
34 |
+
from openai import OpenAI
|
35 |
+
|
36 |
+
client = OpenAI(
|
37 |
+
base_url="http://localhost:8000/v1", # Your vLLM server URL
|
38 |
+
api_key="token-abc123", # Replace with your API key
|
39 |
+
)
|
40 |
+
|
41 |
+
# Example chat completion request
|
42 |
+
completion = client.chat.completions.create(
|
43 |
+
model="Imran1/Qwen2.5-72B-Instruct-FP8",
|
44 |
+
messages=[
|
45 |
+
{"role": "user", "content": "Hello!"},
|
46 |
+
],
|
47 |
+
max_tokens=500,
|
48 |
+
stream=True
|
49 |
+
)
|
50 |
+
|
51 |
+
print(completion)
|
52 |
+
```
|
53 |
+
|
54 |
+
## Performance and Efficiency
|
55 |
+
- **Memory Efficiency**: FP8 precision significantly reduces memory requirements, allowing for larger batch sizes and faster processing times.
|
56 |
+
- **Speed**: The FP8 version provides faster inference, making it highly suitable for real-time applications.
|
57 |
+
|
58 |
+
## Limitations
|
59 |
+
- **Precision Trade-offs**: While FP8 enhances speed and memory usage, tasks that require high precision (e.g., numerical calculations) may see a slight performance degradation compared to FP16/FP32 versions.
|
60 |
+
|
61 |
+
## License
|
62 |
+
This model is licensed under the [Apache-2.0](LICENSE) license. Feel free to use this model for both commercial and non-commercial purposes, ensuring compliance with the license terms.
|
63 |
+
|
64 |
+
---
|
65 |
+
|
66 |
+
For more details and updates, visit the [model page on Hugging Face](https://huggingface.co/Imran1/Qwen2.5-72B-Instruct-FP8).
|