Create README.md

README.md

@@ -0,0 +1,51 @@
+import torch
+import io
+
+model = torch.nn.Sequential(torch.nn.Linear(32, 256, dtype=torch.bfloat16, device="cuda"))
+
+from torchao.quantization import IntxWeightOnlyConfig, quantize_
+from torchao.quantization.granularity import PerGroup
+
+version=1
+
+quant_config = IntxWeightOnlyConfig(
+    weight_dtype=torch.int4,
+    granularity=PerGroup(32),
+    version=version
+)
+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}-IntxWeightOnlyConfig-v{version}-0.14.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=False)
+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,
+)