Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,74 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
base_model: Xenova/llama2.c-stories110M
|
3 |
+
inference: true
|
4 |
+
model_type: llama
|
5 |
+
quantized_by: mgoin
|
6 |
+
tags:
|
7 |
+
- nm-vllm
|
8 |
+
- sparse
|
9 |
+
---
|
10 |
+
|
11 |
+
## llama2.c-stories110M-pruned50
|
12 |
+
This repo contains model files for [llama2.c 110M tinystories](https://huggingface.co/Xenova/llama2.c-stories110M) optimized for [NM-vLLM](https://github.com/neuralmagic/nm-vllm), a high-throughput serving engine for compressed LLMs.
|
13 |
+
|
14 |
+
This model was pruned with [SparseGPT](https://arxiv.org/abs/2301.00774), using [SparseML](https://github.com/neuralmagic/sparseml).
|
15 |
+
|
16 |
+
## Inference
|
17 |
+
Install [NM-vLLM](https://github.com/neuralmagic/nm-vllm) for fast inference:
|
18 |
+
```bash
|
19 |
+
pip install nm-vllm[sparse]
|
20 |
+
```
|
21 |
+
Run in a Python pipeline for local inference:
|
22 |
+
```python
|
23 |
+
from vllm import LLM, SamplingParams
|
24 |
+
|
25 |
+
model = LLM("nm-testing/llama2.c-stories110M-pruned50", sparsity="sparse_w16a16")
|
26 |
+
prompt = "Hello my name is"
|
27 |
+
|
28 |
+
sampling_params = SamplingParams(max_tokens=100, temperature=0)
|
29 |
+
outputs = model.generate(prompt, sampling_params=sampling_params)
|
30 |
+
print(outputs[0].outputs[0].text)
|
31 |
+
```
|
32 |
+
|
33 |
+
## Prompt template
|
34 |
+
|
35 |
+
N/A
|
36 |
+
|
37 |
+
## Sparsification
|
38 |
+
For details on how this model was sparsified, see the `recipe.yaml` in this repo and follow the instructions below.
|
39 |
+
|
40 |
+
Install [SparseML](https://github.com/neuralmagic/sparseml):
|
41 |
+
```bash
|
42 |
+
git clone https://github.com/neuralmagic/sparseml
|
43 |
+
pip install -e "sparseml[transformers]"
|
44 |
+
```
|
45 |
+
|
46 |
+
Replace the recipe as you like and run this one-shot compression script to apply SparseGPT:
|
47 |
+
```python
|
48 |
+
import sparseml.transformers
|
49 |
+
|
50 |
+
original_model_name = "Xenova/llama2.c-stories110M"
|
51 |
+
calibration_dataset = "open_platypus"
|
52 |
+
output_directory = "output/"
|
53 |
+
|
54 |
+
recipe = """
|
55 |
+
test_stage:
|
56 |
+
obcq_modifiers:
|
57 |
+
SparseGPTModifier:
|
58 |
+
sparsity: 0.5
|
59 |
+
sequential_update: true
|
60 |
+
targets: ['re:model.layers.\d*$']
|
61 |
+
"""
|
62 |
+
|
63 |
+
# Apply SparseGPT to the model
|
64 |
+
sparseml.transformers.oneshot(
|
65 |
+
model=original_model_name,
|
66 |
+
dataset=calibration_dataset,
|
67 |
+
recipe=recipe,
|
68 |
+
output_dir=output_directory,
|
69 |
+
)
|
70 |
+
```
|
71 |
+
|
72 |
+
## Slack
|
73 |
+
|
74 |
+
For further support, and discussions on these models and AI in general, join [Neural Magic's Slack Community](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ)
|