torchao-testing/single-linear-Int4WeightOnlyConfig-preshuffled-v2-0.13.dev

model: single_linear
config: Int4WeightOnlyConfig, with preshuffled packing format
config version: 2
torchao version: 0.13.dev

import torch
import io

model = torch.nn.Sequential(torch.nn.Linear(32, 256, dtype=torch.bfloat16, device="cuda"))

from torchao.quantization import Int4WeightOnlyConfig, quantize_
quant_config = Int4WeightOnlyConfig(group_size=128, int4_packing_format="preshuffled", version=2)
quantize_(model, quant_config)
example_inputs = (torch.randn(2, 32, dtype=torch.bfloat16, device="cuda"),)
output = model(*example_inputs)

# Push to hub
USER_ID = "torchao-testing"
MODEL_NAME = "single-linear"
save_to = f"{USER_ID}/{MODEL_NAME}-Int4WeightOnlyConfig-preshuffled-v2-0.13.dev"

from huggingface_hub import HfApi
api = HfApi()

buf = io.BytesIO()
torch.save(model.state_dict(), buf)
api.create_repo(save_to, repo_type="model", exist_ok=True)
api.upload_file(
    path_or_fileobj=buf,
    path_in_repo="model.pt",
    repo_id=save_to,
)

buf = io.BytesIO()
torch.save(example_inputs, buf)
api.upload_file(
    path_or_fileobj=buf,
    path_in_repo="model_inputs.pt",
    repo_id=save_to,
)

buf = io.BytesIO()
torch.save(output, buf)
api.upload_file(
    path_or_fileobj=buf,
    path_in_repo="model_output.pt",
    repo_id=save_to,
)