Update README.md
Browse files
README.md
CHANGED
@@ -33,6 +33,43 @@ Since this is a base model the IKM dataset greatly affects the output. The IKM d
|
|
33 |
```
|
34 |
---
|
35 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
```
|
37 |
[3731/5850 3:38:52 < 2:04:22, 0.28 it/s, Epoch 6.37/10]
|
38 |
Step Training Loss
|
|
|
33 |
```
|
34 |
---
|
35 |
|
36 |
+
## Inference
|
37 |
+
|
38 |
+
```
|
39 |
+
!pip install -qqq transformers>=4.39.0 mamba-ssm causal-conv1d>=1.2.0 accelerate bitsandbytes --progress-bar off
|
40 |
+
!pip install flash-attn --no-build-isolation
|
41 |
+
|
42 |
+
import torch
|
43 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
44 |
+
|
45 |
+
# Load model in 8-bit precision
|
46 |
+
quantization_config = BitsAndBytesConfig(
|
47 |
+
load_in_8bit=True,
|
48 |
+
llm_int8_skip_modules=["mamba"]
|
49 |
+
)
|
50 |
+
model = AutoModelForCausalLM.from_pretrained(
|
51 |
+
"ai21labs/Jamba-v0.1",
|
52 |
+
trust_remote_code=True,
|
53 |
+
torch_dtype=torch.bfloat16,
|
54 |
+
attn_implementation="flash_attention_2",
|
55 |
+
quantization_config=quantization_config
|
56 |
+
)
|
57 |
+
tokenizer = AutoTokenizer.from_pretrained("ai21labs/Jamba-v0.1")
|
58 |
+
|
59 |
+
# Tokenize input
|
60 |
+
prompt = """How could we use cheese to reignite the sun? Answer:"""
|
61 |
+
input_ids = tokenizer(
|
62 |
+
prompt,
|
63 |
+
return_tensors='pt'
|
64 |
+
).to(model.device)["input_ids"]
|
65 |
+
|
66 |
+
# Generate answer
|
67 |
+
outputs = model.generate(input_ids, max_new_tokens=216)
|
68 |
+
|
69 |
+
# Print output
|
70 |
+
print(tokenizer.batch_decode(outputs))
|
71 |
+
```
|
72 |
+
|
73 |
```
|
74 |
[3731/5850 3:38:52 < 2:04:22, 0.28 it/s, Epoch 6.37/10]
|
75 |
Step Training Loss
|