Update README.md
Browse files
README.md
CHANGED
|
@@ -28,6 +28,53 @@ language:
|
|
| 28 |
- en
|
| 29 |
---
|
| 30 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
Evals:
|
| 32 |

|
| 33 |
|
|
|
|
| 28 |
- en
|
| 29 |
---
|
| 30 |
|
| 31 |
+
How to run inference:
|
| 32 |
+
```python
|
| 33 |
+
import transformers
|
| 34 |
+
import torch
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
def fmt_prompt(prompt: str) -> str:
|
| 38 |
+
return f"""[Instructions]:\n{prompt}\n\n[Response]:"""
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
if __name__ == "__main__":
|
| 42 |
+
model_name = "abacaj/starcoderbase-1b-sft"
|
| 43 |
+
tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)
|
| 44 |
+
|
| 45 |
+
model = (
|
| 46 |
+
transformers.AutoModelForCausalLM.from_pretrained(
|
| 47 |
+
model_name,
|
| 48 |
+
)
|
| 49 |
+
.to("cuda:0")
|
| 50 |
+
.eval()
|
| 51 |
+
)
|
| 52 |
+
|
| 53 |
+
prompt = "Write a python function to sort the following array in ascending order, don't use any built in sorting methods: [9,2,8,1,5]"
|
| 54 |
+
prompt_input = fmt_prompt(prompt)
|
| 55 |
+
inputs = tokenizer(prompt_input, return_tensors="pt").to(model.device)
|
| 56 |
+
input_ids_cutoff = inputs.input_ids.size(dim=1)
|
| 57 |
+
|
| 58 |
+
with torch.no_grad():
|
| 59 |
+
generated_ids = model.generate(
|
| 60 |
+
**inputs,
|
| 61 |
+
use_cache=True,
|
| 62 |
+
max_new_tokens=512,
|
| 63 |
+
temperature=0.2,
|
| 64 |
+
top_p=0.95,
|
| 65 |
+
do_sample=True,
|
| 66 |
+
eos_token_id=tokenizer.eos_token_id,
|
| 67 |
+
pad_token_id=tokenizer.pad_token_id,
|
| 68 |
+
)
|
| 69 |
+
|
| 70 |
+
completion = tokenizer.decode(
|
| 71 |
+
generated_ids[0][input_ids_cutoff:],
|
| 72 |
+
skip_special_tokens=True,
|
| 73 |
+
)
|
| 74 |
+
|
| 75 |
+
print(completion)
|
| 76 |
+
```
|
| 77 |
+
|
| 78 |
Evals:
|
| 79 |

|
| 80 |
|