Instructions to use poolside/Laguna-S-2.1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use poolside/Laguna-S-2.1 with Transformers:

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

pipe = pipeline("text-generation", model="poolside/Laguna-S-2.1", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("poolside/Laguna-S-2.1", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("poolside/Laguna-S-2.1", trust_remote_code=True, device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Inference
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use poolside/Laguna-S-2.1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "poolside/Laguna-S-2.1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "poolside/Laguna-S-2.1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/poolside/Laguna-S-2.1

SGLang

How to use poolside/Laguna-S-2.1 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 "poolside/Laguna-S-2.1" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "poolside/Laguna-S-2.1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "poolside/Laguna-S-2.1" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "poolside/Laguna-S-2.1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use poolside/Laguna-S-2.1 with Docker Model Runner:
```
docker model run hf.co/poolside/Laguna-S-2.1
```

Strix Halo (Ryzen AI Max+ 395) results: 24–34 tok/s at 128K–245K, and DFlash that works past 8K

#43

by j2ipper - opened 4 days ago

Discussion

j2ipper

4 days ago

Hugging Face CLI

Sharing results from running Laguna-S-2.1 on an AMD Strix Halo laptop (HP ZBook Ultra G1a, 128 GB, 70 W) with llama.cpp Vulkan, 262K context, q8_0 KV:

Decode: 33–34 tok/s short context; 26.4 tok/s @128K and 24.2 @245K (19.5 / 15.0 before) with block-sparse decode attention on the 12 full-attention layers.
Prefill: +23% @128K with KV-sliced flash attention; ub 4096 adds +7–11% at short context.
DFlash at depth: the DFlash drafter attends to the whole context and its acceptance collapses as context grows (0.2% at 31K, the "DFlash breaks past 8K" reports). Giving the drafter a 1,024-token sliding window fixes it: code generation 40.6 / 37.5 / 34.0 / 34.9 tok/s at ~1K / 31K / 124K / 235K, with an adaptive pause keeping prose at normal speed.
Quality: agentic coding suite 5/5, needle 15/15 (greedy). With thinking off the agentic suite ran in 727 s vs 1,632 s (needle 13/15 vs 14/15).
Keep the GGUF's YaRN factor 32: the HF config's factor 128 was +6.8% worse in perplexity at 64K.

The sparse attention, KV slicing and drafter window are llama.cpp patches (published in the repo). Details, configs and method: https://github.com/nsbradley88/halo-tuning (laguna/)

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment