W_bf16_recovered = W_int4 × Scale
Some precision loss is expected from the quantization → dequantization roundtrip
Best used for fine-tuning where the loss can be recovered

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "hassanshka/Biomni-R0-32B-INT4-to-BF16",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("hassanshka/Biomni-R0-32B-INT4-to-BF16")

# Use for inference or fine-tuning

Dequantization Process

The model was dequantized using the following algorithm:

python

def unpack_awq_standard(packed_weight, scales):
    group_size = 128
    scales_expanded = scales.repeat_interleave(group_size, dim=1)
    
    packed_weight = packed_weight.to(torch.int32)
    unpacked_cols = []
    mask = 0xF
    
    for i in range(8):
        weight_chunk = (packed_weight >> (i * 4)) & mask
        weight_chunk = torch.where(weight_chunk >= 8, weight_chunk - 16, weight_chunk)
        unpacked_cols.append(weight_chunk)
    
    weights = torch.stack(unpacked_cols, dim=-1)
    weights = weights.view(rows, packed_cols * 8)
    
    dequantized = weights.to(torch.bfloat16) * scales_expanded.to(torch.bfloat16)
    return dequantized

License

Apache 2.0 (same as base model)

Citation

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