chore: import upstream snapshot with attribution

docs/code-docs/source/inference-init.rst

@@ -0,0 +1,42 @@
+Inference Setup
+-----------------------
+The entrypoint for inference with DeepSpeed is ``deepspeed.init_inference()``.
+
+Example usage:
+
+.. code-block:: python
+
+    engine = deepspeed.init_inference(model=net, config=config)
+
+The ``DeepSpeedInferenceConfig`` is used to control all aspects of initializing
+the ``InferenceEngine``. The config should be passed as a dictionary to
+``init_inference``, but parameters can also be passed as keyword arguments.
+
+.. _DeepSpeedInferenceConfig:
+.. autopydantic_model:: deepspeed.inference.config.DeepSpeedInferenceConfig
+
+.. _DeepSpeedTPConfig:
+.. autopydantic_model:: deepspeed.inference.config.DeepSpeedTPConfig
+
+.. _DeepSpeedMoEConfig:
+.. autopydantic_model:: deepspeed.inference.config.DeepSpeedMoEConfig
+
+.. _QuantizationConfig:
+.. autopydantic_model:: deepspeed.inference.config.QuantizationConfig
+
+.. _InferenceCheckpointConfig:
+.. autopydantic_model:: deepspeed.inference.config.InferenceCheckpointConfig
+
+
+Example config:
+
+.. code-block:: python
+
+    config = {
+	"kernel_inject": True,
+	"tensor_parallel": {"tp_size": 4},
+	"dtype": "fp16",
+	"enable_cuda_graph": False
+    }
+
+.. autofunction:: deepspeed.init_inference