[torchao] safetensors integration (#25969)

Signed-off-by: Angel Li <liangel@meta.com>

tests/quantization/test_torchao.py

@@ -216,5 +216,22 @@ def test_reload_weights():
         # print("-" * 60)
 
 
+@pytest.mark.skipif(not TORCHAO_AVAILABLE, reason="torchao is not available")
+@pytest.mark.skip(
+    reason="since torchao nightly is only compatible with torch nightly"
+    "currently https://github.com/pytorch/ao/issues/2919, we'll have to skip "
+    "torchao tests that requires newer versions (0.14.0.dev+) for now"
+)
+def test_opt_125m_float8_weight_only_safetensors_model_loading_with_params(vllm_runner):
+    torch._dynamo.reset()
+    model_name = (
+        "torchao-testing/opt-125m-Float8WeightOnlyConfig-v2-0.14.0.dev-safetensors"
+    )
+    with vllm_runner(model_name=model_name, dtype="bfloat16") as llm:
+        output = llm.generate_greedy(["The capital of France is"], max_tokens=32)
+
+        assert output
+
+
 if __name__ == "__main__":
     pytest.main([__file__])