Instructions to use Polygl0t/Tucano2-qwen-0.5B-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Polygl0t/Tucano2-qwen-0.5B-Instruct with Transformers:

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

pipe = pipeline("text-generation", model="Polygl0t/Tucano2-qwen-0.5B-Instruct")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Polygl0t/Tucano2-qwen-0.5B-Instruct")
model = AutoModelForCausalLM.from_pretrained("Polygl0t/Tucano2-qwen-0.5B-Instruct")
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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Inference
Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use Polygl0t/Tucano2-qwen-0.5B-Instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Polygl0t/Tucano2-qwen-0.5B-Instruct"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Polygl0t/Tucano2-qwen-0.5B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Polygl0t/Tucano2-qwen-0.5B-Instruct

SGLang

How to use Polygl0t/Tucano2-qwen-0.5B-Instruct 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 "Polygl0t/Tucano2-qwen-0.5B-Instruct" \
    --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": "Polygl0t/Tucano2-qwen-0.5B-Instruct",
		"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 "Polygl0t/Tucano2-qwen-0.5B-Instruct" \
        --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": "Polygl0t/Tucano2-qwen-0.5B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Polygl0t/Tucano2-qwen-0.5B-Instruct with Docker Model Runner:
```
docker model run hf.co/Polygl0t/Tucano2-qwen-0.5B-Instruct
```

Tucano2-qwen-0.5B-Instruct / training_config_sft.yaml

nicholasKluge

Upload training_config_sft.yaml with huggingface_hub

e7d1b07 verified 4 months ago

raw

history blame

2.97 kB

	# Directory settings
	checkpoint_dir: "/polyglot/portuguese/checkpoints/models/Tucano2-qwen-0.5B-Instruct-SFT"
	train_dataset_dir:
	# Total: ~874 million tokens (x5 epochs)
	# Coding: ~2.3 million tokens
	- /polyglot/portuguese/gigaverbo-v2-sft/code
	# Function Calling: ~17.5 million tokens
	- /polyglot/portuguese/gigaverbo-v2-sft/function_call
	# General Instruction Following: ~700 million tokens
	- /polyglot/portuguese/gigaverbo-v2-sft/general
	# Math and CoT: ~27 million tokens
	- /polyglot/portuguese/gigaverbo-v2-sft/math_cot
	# Retrieval Augmented Generation: ~2.2 million tokens
	- /polyglot/portuguese/gigaverbo-v2-sft/retrieval
	# Structured Outputs: ~35 million tokens
	- /polyglot/portuguese/gigaverbo-v2-sft/structured
	# Summarization: ~290 thousand tokens
	- /polyglot/portuguese/gigaverbo-v2-sft/summarization
	# Translation: ~5.7 million tokens
	- /polyglot/portuguese/gigaverbo-v2-sft/translation
	# Chosen Data from Preference Dataset: ~14 million tokens
	- /polyglot/portuguese/gigaverbo-v2-sft/dpo
	val_dataset_dir: null
	dataset_type: "jsonl"
	cache_dir: "/lustre/mlnvme/data/polyglot/.cache"

	# Data loading settings
	pin_memory: true
	num_workers_for_dataloader: 16
	shuffle_dataset: true
	mask_eos_token: false
	mask_pad_token: true

	# Model architecture settings
	vocab_size: 49152
	num_hidden_layers: 28
	num_attention_heads: 16
	num_key_value_heads: 8
	head_dim: 128
	hidden_size: 1024
	intermediate_size: 3072
	max_position_embeddings: 4096
	tie_word_embeddings: true
	hidden_act: "silu"
	output_hidden_states: false
	attn_implementation: "flash_attention_2"
	use_cache: false
	no_rope_layer_interval: null
	rope_theta: 1000000.0
	rope_scale_factor: null
	rms_norm_eps: 0.000001

	# Training settings
	total_batch_size: 524288
	micro_batch_size: 4
	gradient_accumulation_steps: 4
	eval_micro_batch_size: null
	num_train_epochs: 5
	warmup_ratio: 0.1
	max_learning_rate: 0.000085
	min_learning_rate: 0.0
	muon_learning_rate: null
	weight_decay: 0.0
	beta1: 0.9
	beta2: 0.95
	eps: 0.00000001
	lr_decay_type: "cosine"
	use_sqrt: false
	lr_decay_iters_coef: 1.
	seed: 42
	max_steps: 68635
	max_grad_norm: 1.0

	# SFT settings
	packing: false
	assistant_only_loss: true

	# Precision and optimization settings
	torch_compile: false
	mat_mul_precision: "highest"
	tf32: true
	bf16: true
	gradient_checkpointing: false
	use_liger_kernel: true
	static_graph: false

	# Hub settings
	push_to_hub: false
	hub_token: null
	hub_model_id: null

	# Tokenizer and Reference model
	tokenizer_name_or_path: "Polygl0t/Tucano2-qwen-0.5B-Base"
	chat_template_path: null
	reference_model: "Polygl0t/Tucano2-qwen-0.5B-Base"
	continual_pretraining: true

	# Checkpoint settings
	resume_from_checkpoint: null
	checkpointing_steps: 1000
	begin_new_stage: true
	stage_name: "single_cosine"

	# Miscellaneous settings
	sanity_check: false
	sanity_check_num_samples: 100000
	wandb_token: null
	wandb_id: "tucano2-qwen-0.5b-instruct-sft"
	wandb_project: "Polyglot"
	wandb_desc: "Developing LLMs for low-resource languages"