Instructions to use AI-Sweden-Models/Llama-3-8B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AI-Sweden-Models/Llama-3-8B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="AI-Sweden-Models/Llama-3-8B")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("AI-Sweden-Models/Llama-3-8B")
model = AutoModelForCausalLM.from_pretrained("AI-Sweden-Models/Llama-3-8B")

Inference
Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use AI-Sweden-Models/Llama-3-8B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AI-Sweden-Models/Llama-3-8B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AI-Sweden-Models/Llama-3-8B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/AI-Sweden-Models/Llama-3-8B

SGLang

How to use AI-Sweden-Models/Llama-3-8B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "AI-Sweden-Models/Llama-3-8B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AI-Sweden-Models/Llama-3-8B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "AI-Sweden-Models/Llama-3-8B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AI-Sweden-Models/Llama-3-8B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use AI-Sweden-Models/Llama-3-8B with Docker Model Runner:
```
docker model run hf.co/AI-Sweden-Models/Llama-3-8B
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

AI-Sweden-Models/Llama-3-8B

Intended usage:

This is a base model, it can be finetuned to a particular use case.

-----> instruct version here <-----

Use with transformers

See the snippet below for usage with Transformers:

import transformers
import torch

model_id = "AI-Sweden-Models/Llama-3-8B"

pipeline = transformers.pipeline(
    task="text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto"
)

pipeline(
    text_inputs="Sommar och sol är det bästa jag vet",
    max_length=128,
    repetition_penalty=1.03
)

>>> "Sommar och sol är det bästa jag vet!
Och nu när jag har fått lite extra semester så ska jag njuta till max av allt som våren och sommaren har att erbjuda.
Jag har redan börjat med att sitta ute på min altan och ta en kopp kaffe och läsa i tidningen, det är så skönt att bara sitta där och njuta av livet.

Ikväll blir det grillat och det ser jag fram emot!"

Training information

AI-Sweden-Models/Llama-3-8B is a continuation of the pretraining process from meta-llama/Meta-Llama-3-8B. It was trained on a subset from The Nordic Pile containing Swedish, Norwegian and Danish. The training is done on all model parameters, it is a full finetune.

The training dataset consists of 227 105 079 296 tokens. It was trained on the Rattler supercomputer at the Dell Technologies Edge Innovation Center in Austin, Texas. The training used 23 nodes of a duration of 30 days, where one node contained 4X Nvidia A100 GPUs, yielding 92 GPUs.

trainer.yaml:

learning_rate: 2e-5
warmup_steps: 100
lr_scheduler: cosine
optimizer: adamw_torch_fused
max_grad_norm: 1.0
gradient_accumulation_steps: 16
micro_batch_size: 1
num_epochs: 1
sequence_len: 8192

deepspeed_zero2.json:

{
  "zero_optimization": {
    "stage": 2,
    "offload_optimizer": {
      "device": "cpu"
    },
    "contiguous_gradients": true,
    "overlap_comm": true
  },
  "bf16": {
    "enabled": "auto"
  },
  "fp16": {
    "enabled": "auto",
    "auto_cast": false,
    "loss_scale": 0,
    "initial_scale_power": 32,
    "loss_scale_window": 1000,
    "hysteresis": 2,
    "min_loss_scale": 1
  },
  "gradient_accumulation_steps": "auto",
  "gradient_clipping": "auto",
  "train_batch_size": "auto",
  "train_micro_batch_size_per_gpu": "auto",
  "wall_clock_breakdown": false
}