Instructions to use eulogik/Bharat-Tiny-LLM-fused with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use eulogik/Bharat-Tiny-LLM-fused with Transformers:

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

pipe = pipeline("text-generation", model="eulogik/Bharat-Tiny-LLM-fused")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("eulogik/Bharat-Tiny-LLM-fused")
model = AutoModelForCausalLM.from_pretrained("eulogik/Bharat-Tiny-LLM-fused")
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]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use eulogik/Bharat-Tiny-LLM-fused with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/eulogik/Bharat-Tiny-LLM-fused

SGLang

How to use eulogik/Bharat-Tiny-LLM-fused 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 "eulogik/Bharat-Tiny-LLM-fused" \
    --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": "eulogik/Bharat-Tiny-LLM-fused",
		"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 "eulogik/Bharat-Tiny-LLM-fused" \
        --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": "eulogik/Bharat-Tiny-LLM-fused",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use eulogik/Bharat-Tiny-LLM-fused with Docker Model Runner:
```
docker model run hf.co/eulogik/Bharat-Tiny-LLM-fused
```

Bharat-Tiny-LLM Fused

This is a pre-fused version of Bharat-Tiny-LLM — base model weights + LoRA adapter fused into a single model.safetensors for direct use with Transformers.

Base model: Qwen/Qwen2.5-1.5B Adapter: eulogik/Bharat-Tiny-LLM-adapter

Why This Exists

The MLX-trained LoRA adapter cannot be loaded with PEFT/Transformers due to a fundamental MLX ↔ PEFT LoRA implementation mismatch. This fused model bypasses PEFT entirely by fusing the LoRA weights directly into the base model weights.

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("eulogik/Bharat-Tiny-LLM-fused")
tokenizer = AutoTokenizer.from_pretrained("eulogik/Bharat-Tiny-LLM-fused")

prompt = "<|im_start|>user\nChai peete hain?<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Fusion Details

The LoRA weights were fused on MPS (Apple Silicon) in float16 precision
Save uses direct safetensors (not model.save_pretrained which corrupts fused weights)
Fused weights differ from base weights by 0.003–0.024 across 112 modified tensors
Model dtype: float16 (3.55 GB)

Limitations

Experimental quality: 76K LoRA iterations is insufficient for reliable Hinglish generation. Output quality varies.
5.3M trainable params (0.34% of 1.5B) is too few to strongly influence the base model.
Not production-ready — intended as a research demo.

License

Apache 2.0

Downloads last month: -

Safetensors

Model size

2B params

Tensor type

F16

eulogik
/

Bharat-Tiny-LLM-fused