Transformers.js improvements (#25)
Browse files- [WIP] Transformers.js improvements (b86e3eda672bc29427317d7d6753fbf808730d94)
- Update README.md (26cd37002254a59f114c509de0b80ef80aaf90d6)
- README.md +29 -2
- config.json +5 -0
README.md
CHANGED
@@ -40,7 +40,7 @@ For more details refer to: https://github.com/huggingface/smollm. You will find
|
|
40 |
|
41 |
### How to use
|
42 |
|
43 |
-
|
44 |
```bash
|
45 |
pip install transformers
|
46 |
```
|
@@ -62,13 +62,40 @@ print(tokenizer.decode(outputs[0]))
|
|
62 |
```
|
63 |
|
64 |
|
65 |
-
|
66 |
You can also use the TRL CLI to chat with the model from the terminal:
|
67 |
```bash
|
68 |
pip install trl
|
69 |
trl chat --model_name_or_path HuggingFaceTB/SmolLM2-1.7B-Instruct --device cpu
|
70 |
```
|
71 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
## Evaluation
|
73 |
|
74 |
In this section, we report the evaluation results of SmolLM2. All evaluations are zero-shot unless stated otherwise, and we use [lighteval](https://github.com/huggingface/lighteval) to run them.
|
|
|
40 |
|
41 |
### How to use
|
42 |
|
43 |
+
#### Transformers
|
44 |
```bash
|
45 |
pip install transformers
|
46 |
```
|
|
|
62 |
```
|
63 |
|
64 |
|
65 |
+
#### Chat in TRL
|
66 |
You can also use the TRL CLI to chat with the model from the terminal:
|
67 |
```bash
|
68 |
pip install trl
|
69 |
trl chat --model_name_or_path HuggingFaceTB/SmolLM2-1.7B-Instruct --device cpu
|
70 |
```
|
71 |
|
72 |
+
#### Transformers.js
|
73 |
+
|
74 |
+
```bash
|
75 |
+
npm i @huggingface/transformers
|
76 |
+
```
|
77 |
+
|
78 |
+
```js
|
79 |
+
import { pipeline } from "@huggingface/transformers";
|
80 |
+
|
81 |
+
// Create a text generation pipeline
|
82 |
+
const generator = await pipeline(
|
83 |
+
"text-generation",
|
84 |
+
"HuggingFaceTB/SmolLM2-1.7B-Instruct",
|
85 |
+
);
|
86 |
+
|
87 |
+
// Define the list of messages
|
88 |
+
const messages = [
|
89 |
+
{ role: "system", content: "You are a helpful assistant." },
|
90 |
+
{ role: "user", content: "Tell me a joke." },
|
91 |
+
];
|
92 |
+
|
93 |
+
// Generate a response
|
94 |
+
const output = await generator(messages, { max_new_tokens: 128 });
|
95 |
+
console.log(output[0].generated_text.at(-1).content);
|
96 |
+
// "Why don't scientists trust atoms?\n\nBecause they make up everything!"
|
97 |
+
```
|
98 |
+
|
99 |
## Evaluation
|
100 |
|
101 |
In this section, we report the evaluation results of SmolLM2. All evaluations are zero-shot unless stated otherwise, and we use [lighteval](https://github.com/huggingface/lighteval) to run them.
|
config.json
CHANGED
@@ -25,9 +25,14 @@
|
|
25 |
"torch_dtype": "bfloat16",
|
26 |
"transformers_version": "4.42.3",
|
27 |
"transformers.js_config": {
|
|
|
28 |
"kv_cache_dtype": {
|
29 |
"q4f16": "float16",
|
30 |
"fp16": "float16"
|
|
|
|
|
|
|
|
|
31 |
}
|
32 |
},
|
33 |
"use_cache": true,
|
|
|
25 |
"torch_dtype": "bfloat16",
|
26 |
"transformers_version": "4.42.3",
|
27 |
"transformers.js_config": {
|
28 |
+
"dtype": "q4",
|
29 |
"kv_cache_dtype": {
|
30 |
"q4f16": "float16",
|
31 |
"fp16": "float16"
|
32 |
+
},
|
33 |
+
"use_external_data_format": {
|
34 |
+
"model.onnx": true,
|
35 |
+
"model_fp16.onnx": true
|
36 |
}
|
37 |
},
|
38 |
"use_cache": true,
|