Biomni-R0-32B-AWQ-INT4

This is an AWQ INT4 quantized version of Biomni-R0-32B-Preview.

Quantization Details

Parameter	Value
Scheme	W4A16 (4-bit weights, 16-bit activations)
Method	AWQ (Activation-aware Weight Quantization)
Group Size	128
Calibration	Custom biomedical dataset
Framework	LLM Compressor

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "hassanshka/Biomni-R0-32B-AWQ-INT4",
    device_map="auto",
    torch_dtype="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("hassanshka/Biomni-R0-32B-AWQ-INT4")

# Inference
messages = [{"role": "user", "content": "Your medical question here"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))

Quantization Script

from llmcompressor.modifiers.awq import AWQModifier
from llmcompressor import oneshot

recipe = AWQModifier(
    scheme="W4A16",
    targets="Linear",
    ignore=["lm_head"],
)

oneshot(
    model=model,
    dataset=calibration_data,
    recipe=recipe,
    max_seq_length=2048,
    num_calibration_samples=len(calibration_data),
)

Performance

• Memory Reduction: ~75% compared to BF16
• Inference Speed: Optimized for consumer GPUs (RTX 3090/4090)
• Accuracy: Minimal degradation due to custom calibration on biomedical data

License

Apache 2.0 (same as base model)

Citation

If you use this model, please cite the original Biomni model.