from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load with automatic fp16 support
tokenizer = AutoTokenizer.from_pretrained("Mostafa8Mehrabi/qwen3-50m-fp16")
model = AutoModelForCausalLM.from_pretrained(
    "Mostafa8Mehrabi/qwen3-50m-fp16",
    torch_dtype=torch.float16,  # Explicitly use fp16
    device_map="auto"  # Automatically place on available device
)

# For GPU inference (recommended)
# model = model.to("cuda") # if you have a GPU

inputs = tokenizer("Hello, how are you?", return_tensors="pt")
# Move inputs to same device as model if using GPU
# inputs = {k: v.to(model.device) for k, v in inputs.items()}

outputs = model.generate(**inputs, max_length=50, do_sample=True, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Key Features

✅ FP16 Precision: 50% smaller model size, faster inference
✅ Standard GPT-2 tokenizer (no trust_remote_code)
✅ Compatible vocabulary sizes
✅ SafeTensors format for faster loading
✅ Works like any HuggingFace model
✅ 13x smaller than original Qwen3-0.6B
✅ GPU optimized for efficient inference

Architecture Comparison

Component	Original	This Model
Parameters	637M	~50M
Vocabulary	151,936	50,257
Hidden Size	1024	384
Layers	28	8
Tokenizer	Qwen3	GPT-2
Precision	FP32	FP16
Model Size	~1.2GB	~100MB

Memory Requirements

FP16: ~100MB model + ~50MB working memory = ~150MB total
FP32: ~200MB model + ~100MB working memory = ~300MB total
Memory savings: ~50% reduction compared to FP32

Performance Notes

FP16 provides significant memory savings with minimal quality loss
Ideal for deployment in resource-constrained environments
Compatible with both CPU and GPU inference
Faster loading times due to smaller file size