From e0e96d8276172caca89f8be83eb1f8c36c451203 Mon Sep 17 00:00:00 2001 From: Justin Chu Date: Sat, 4 May 2024 23:27:38 -0700 Subject: [PATCH] [IR] Create documentation for tensors (#1481) Document the TensorProtocol and the various tensor classes. When we create a helper function to create tensors in the IR, we should document that as well. --- docs/conf.py | 1 + docs/intermediate_representation/index.md | 1 + docs/intermediate_representation/tensors.md | 322 ++++++++++++++++++++ onnxscript/ir/__init__.py | 3 + requirements-dev.txt | 3 + 5 files changed, 330 insertions(+) create mode 100644 docs/intermediate_representation/tensors.md diff --git a/docs/conf.py b/docs/conf.py index d9a53ecf5..63dd8e7d4 100644 --- a/docs/conf.py +++ b/docs/conf.py @@ -22,6 +22,7 @@ extensions = [ "myst_parser", "sphinx_copybutton", + "sphinx_exec_code", "sphinx_gallery.gen_gallery", "sphinx.ext.autodoc", "sphinx.ext.autosummary", diff --git a/docs/intermediate_representation/index.md b/docs/intermediate_representation/index.md index ff1546408..fd3199671 100644 --- a/docs/intermediate_representation/index.md +++ b/docs/intermediate_representation/index.md @@ -3,5 +3,6 @@ ```{toctree} :maxdepth: 1 +tensors ir_api ``` diff --git a/docs/intermediate_representation/tensors.md b/docs/intermediate_representation/tensors.md new file mode 100644 index 000000000..cca80264e --- /dev/null +++ b/docs/intermediate_representation/tensors.md @@ -0,0 +1,322 @@ +# Tensor Representation in the IR + +The ONNX IR offers the {py:class}`ir.TensorProtocol ` interface for usings different data structures as backing data for tensors. Besides the traditional {py:class}`onnx.TensorProto`, you can also use {py:class}`np.ndarray`, {py:class}`torch.Tensor`, {py:class}`jax.Array`, and virtually anything else to represent tensors in the graph. This allows for them to be accessed and serialized via the same `TensorProtocol` interface, without incurring additional copies at initialization. + +## The `TensorProtocol` + +{py:class}`ir.TensorProtocol ` defines a read-only interface for representing tensors. A tensor class implementing the interface has attributes like `name`, `shape`, `dtype`, `size`, `nbytes` and `metadata_props` to describe basic properties of the tensor. Additionally, it should implement two methods {py:meth}`numpy ` and {py:meth}`__array__ ` which will produce equivalent NumPy arrays from the backing data. + +:::{note} +When interacting with initializers, constant values and tensor attributes, it is best to assume `TensorProtocol` and only use `isinstance` to check for concrete classes when there is a need. +::: + +## Tensor Classes + +### ir.TensorProtoTensor + +The ONNX spec defines [different ways](https://github.com/onnx/onnx/blob/d6f87121ba256ac6cc4d1da0463c300c278339d2/onnx/onnx.proto#L567-L654) for storing tensor data as an {py:class}`onnx.TensorProto ` protocol buffer message. The IR has corresponding classes for each of these data storage methods. + +We use the {py:class}`ir.TensorProtoTensor ` as a wrapper around the proto to implement the `ir.TensorProtocol` interface. You can access `shape`, `dtype` etc. as usual. A copy is incurred only when `numpy()` is called. + +:::{note} +Directly initializing an `ir.TensorProtoTensor`, as below, is possible. However, it is usually recommended to use `ir.serde.deserialize_tensor` because it handles all types of `TensorProto`s (`ir.TensorProtoTensor` doesn't handle external tensors, for example). Please refer to [From `TensorProto`s and back](#from-tensorprotos-and-back) for an example. +::: + +```{eval-rst} +.. exec_code:: + + import onnx + from onnxscript import ir + + tensor_proto = onnx.helper.make_tensor("tensor", onnx.TensorProto.INT16, (3,), [1, 2, 3]) + tensor = ir.TensorProtoTensor(tensor_proto) + print("tensor: ", tensor) # TensorProtoTensor(name='tensor') + print("shape: ", tensor.shape) # ir.Shape([3]) + print("dtype: ", tensor.dtype) # ir.DataType.INT16 + print(tensor.raw == tensor_proto) # The raw field is the exact tensor_proto provided at initialization + print("tobytes: ", tensor.tobytes()) # b'\x01\x00\x02\x00\x03\x00' + print("numpy: ", tensor.numpy()) # array([1, 2, 3], dtype=int16) +``` + +### ir.ExternalTensor + +Tensor data stored externally in the disk are typically large and will take up memory when loaded. The {py:class}`ir.ExternalTensor ` class uses memory mapping to avoid loading the tensor into memory. You are able to use the tensor as a normal NumPy array with minimal memory usage. + +Refer to {py:func}`ir.serde.deserialize_tensor ` to find an example on converting an `onnx.TensorProto` to an {py:class}`ir.ExternalTensor `. + +### ir.Tensor + +{py:class}`ir.Tensor ` is a wrapper around NumPy array compatible array objects like {py:class}`np.ndarray` and {py:class}`torch.Tensor`. It is best for creating in-memory tensors without converting it to a `TensorProto` to reduce the conversion overhead. + +:::{tip} +An array object is compatible if it defines the `__array__` method. +::: + +To create a tensor from an array, simply initialize it with an NumPy array + +```python +tensor = ir.Tensor(np.random.rand(1, 2)) +``` + +The initializer will obtain dtype and shape information from the array. + +To create a tensor from objects other than NumPy array, you need to specify the dtype: + +```{eval-rst} +.. exec_code:: + + import torch + from onnxscript import ir + + torch_tensor = torch.tensor([1, 2, 3], dtype=torch.float16) + tensor = ir.Tensor(torch_tensor, dtype=ir.DataType.FLOAT16) + print(tensor.numpy()) # array([1., 2., 3.], dtype=float16) +``` + +### String Tensor + +Use {py:class}`ir.StringTensor ` to create a string tensor. + + + +### Sparse Tensor + +Sparse tensors are not yet supported, but they are on our roadmap. + +## From `TensorProto`s and back + +In the following scenario, we show how to go from a `TensorProto` to an `ir.Tensor`, run some computation, then turn it back to an `ir.Tensor` and finally `TensorProto` + +```{eval-rst} +.. exec_code:: + + from onnxscript import ir + import onnx + import numpy as np + + # 1. Create the TensorProto + proto = onnx.helper.make_tensor( + "tensor", onnx.TensorProto.FLOAT16, [2, 3], [1, 2, 3, 4, 5, 6] + ) + + # 2. Create an IR Tensor from the Protobuf message + tensor = ir.serde.deserialize_tensor(proto) + # Note that we get a TensorProtoTensor that implements the TensorProtocol + print("tensor:", tensor) # TensorProtoTensor(name='tensor') + print("tensor.numpy():", tensor.numpy()) # [[1. 2. 3.] + # [4. 5. 6.]] + print("tensor.tobytes():", tensor.tobytes()) # b'\x00<\x00@\x00B\x00D\x00E\x00F' + + # 3. Do computation using numpy + mean = tensor.numpy().mean(axis=0) + print("mean:", mean) # array([2.5, 3.5, 4.5], dtype=float16) + + # 4. Create a Tensor from the ndarray. Note that we use ir.Tensor + tensor_mean = ir.Tensor(mean) + print("tensor_mean:", tensor_mean) # Tensor(array([2.5, 3.5, 4.5], dtype=float16), name='') + + # 5. Obtain the TensorProto from ir.Tensor + mean_tensor_proto: onnx.TensorProto = ir.serde.serialize_tensor(tensor_mean) + print("mean_tensor_proto:", mean_tensor_proto) + print( + "onnx.numpy_helper.to_array(mean_tensor_proto):", + onnx.numpy_helper.to_array(mean_tensor_proto) + # array([2.5, 3.5, 4.5], dtype=float16) + ) + + # You can obtain the bytes data as well + print("tensor_mean.tobytes():", tensor_mean.tobytes()) + print("Bytes same as proto:", mean_tensor_proto.raw_data == tensor_mean.tobytes()) + + # Explore other methods defined by TensorProtocol: + print("\n# Explore other methods defined by TensorProtocol:") + print("tensor_mean.shape:", tensor_mean.shape) + print("tensor_mean.dtype:", tensor_mean.dtype) + print("tensor_mean.name:", tensor_mean.name) + print("tensor_mean.doc_string:", tensor_mean.doc_string) + print("tensor_mean.raw:", tensor_mean.raw) + print("tensor_mean.metadata_props:", tensor_mean.metadata_props) + print("tensor_mean.size:", tensor_mean.size) + print("tensor_mean.nbytes:", tensor_mean.nbytes) + print("tensor_mean.raw:", tensor_mean.raw) + print("\nUse the display() method to view the tensor") + tensor_mean.display() +``` + +## Working with non-native NumPy dtypes: bfloat16, float8, int4 + +`ir.Tensor.numpy()` produces a NumPy array representation of the tensor's value. When the tensor has dtype `BFLOAT16`, `FLOAT8[...]` or `[U]INT4` which are not supported by NumPy, the value is the bit representation for the dtype: + +- `int8` for (unpacked) int4, with the sign bit extended to 8 bits. +- `uint8` for (unpacked) uint4. +- `uint8` for 8-bit data types like float8. +- `uint16` for bfloat16. + +uint4/int4 is always unpacked; `tobyte()` produces a packed representation as expected. + +Initialization of `ir.Tensor` requires the NumPy array to follow these typing constraints as well. + +:::{tip} +You can use the [ml_dtypes package](https://github.com/jax-ml/ml_dtypes) to extend NumPy and work with these values. + +```bash +pip install --upgrade ml_dtypes +``` + +::: + +The following example shows how to create a `FLOAT8E4M3FN` tensor, transform its values, and create a new tensor to store the transformed values. + +```{eval-rst} +.. exec_code:: + + from onnxscript import ir + import numpy as np + + array = np.array([0b1, 0b11], dtype=np.uint8) + tensor = ir.Tensor(array, dtype=ir.DataType.FLOAT8E4M3FN) + print(tensor) # Tensor(array([1, 3], dtype=uint8), name='') + print("tensor.numpy():", tensor.numpy()) # array([1, 3], dtype=uint8) + + # You can use the ml_dtypes package to work with these values in NumPy + import ml_dtypes + float8_array = tensor.numpy().view(ml_dtypes.float8_e4m3fn) + print("float8_array:", float8_array) # array([0.00195312, 0.00585938], dtype='float8_e4m3fn') + + # Compute + times_100 = float8_array * 100 + print("times_100:", times_100) + + # Create a new tensor out of the new value; dtype must be specified + new_tensor = ir.Tensor(times_100.view(np.uint8), dtype=ir.DataType.FLOAT8E4M3FN) + print("new_tensor:", new_tensor) # Tensor(array([36, 49], dtype=uint8), name='') + print("new_tensor == times_100", new_tensor.numpy().view(ml_dtypes.float8_e4m3fn) == times_100) # array([ True, True]) + +``` + +## Advanced Usage + +### Subclass ir.Tensor for More Efficient Access and Broader dtype Support + +{py:class}`ir.Tensor` internally converts any array compatible objects into NumPy arrays to produce the byte representation in `tobytes()`. This can be inefficient due to the additional conversion. It also limits support for dtypes not supported by NumPy like bfloat16, because the `__array__` method would fail. + +To fully support arrays from other frameworks, it is usually a good idea to create specialized classes to handle them. The `TorchTensor` class below demonstrates how you can subclass `ir.Tensor` to handle PyTorch tensors: + +```{eval-rst} +.. exec_code:: + + import ctypes + from typing import Any + + import torch + from onnxscript import ir + + # Define utilities to convert PyTorch data types so users do not need to specify manually + _TORCH_DTYPE_TO_ONNX: dict[torch.dtype, ir.DataType] = { + torch.bfloat16: ir.DataType.BFLOAT16, + torch.bool: ir.DataType.BOOL, + torch.complex128: ir.DataType.COMPLEX128, + torch.complex64: ir.DataType.COMPLEX64, + torch.float16: ir.DataType.FLOAT16, + torch.float32: ir.DataType.FLOAT, + torch.float64: ir.DataType.DOUBLE, + torch.float8_e4m3fn: ir.DataType.FLOAT8E4M3FN, + torch.float8_e4m3fnuz: ir.DataType.FLOAT8E4M3FNUZ, + torch.float8_e5m2: ir.DataType.FLOAT8E5M2, + torch.float8_e5m2fnuz: ir.DataType.FLOAT8E5M2FNUZ, + torch.int16: ir.DataType.INT16, + torch.int32: ir.DataType.INT32, + torch.int64: ir.DataType.INT64, + torch.int8: ir.DataType.INT8, + torch.uint8: ir.DataType.UINT8, + } + + + def _torch_dtype_to_onnx_dtype(dtype: torch.dtype) -> ir.DataType: + return _TORCH_DTYPE_TO_ONNX[dtype] + + class TorchTensor(ir.Tensor): + def __init__(self, tensor: torch.Tensor): + # Pass the tensor as the raw data to ir.Tensor's constructor + super().__init__(tensor, dtype=_torch_dtype_to_onnx_dtype(tensor.dtype)) + + def __array__(self, dtype: Any = None) -> "np.ndarray": + # numpy() calls __array__ in ir.Tensor + if self.dtype == ir.DataType.BFLOAT16: + return self.raw.view(torch.uint16).__array__(dtype) + if self.dtype in { + ir.DataType.FLOAT8E4M3FN, + ir.DataType.FLOAT8E4M3FNUZ, + ir.DataType.FLOAT8E5M2, + ir.DataType.FLOAT8E5M2FNUZ + }: + return self.raw.view(torch.uint8).__array__(dtype) + return self.raw.__array__(dtype) + + def tobytes(self) -> bytes: + # Implement tobytes to support native PyTorch types so we can use types like bloat16 + # Reading from memory directly is also more efficient because + # it avoids the copy to NumPy array + tensor = self.raw.detach().cpu().contiguous() + return bytes( + (ctypes.c_ubyte * tensor.element_size() * tensor.numel()).from_address( + tensor.data_ptr() + ) + ) + + # Test the implementation + torch_tensor = torch.tensor([1,2,3], dtype=torch.bfloat16) + tensor = TorchTensor(torch_tensor) + print("tensor: ", tensor) + print("numpy: ", tensor.numpy()) + print("tobytes: ", tensor.tobytes()) # b'\x80?\x00@@@' + print("nbytes: ", tensor.nbytes) # 6 +``` + +The `TorchTensor` class above implements `tobytes()` to produce the correct bytes representation for the tensor when it is serialized into an ONNX file / TensorProto. The class also implements the `__array__()` method to return the bit representation for types NumPy does not support. This way analysis passes can still perform computation on these values. + +### Computation with different Frameworks + +Since `ir.Tensor` implements the `__array__` method and `__dlpack__` methods, its content can be shared with computation frameworks without copying. For example: + +```{eval-rst} +.. exec_code:: + + from onnxscript import ir + + # We can call numpy methods directly on ir.Tensor + import numpy as np + print(np.multiply(ir.Tensor(np.array([1, 2])), 42)) # array([42., 84.]) + + # We can transfer arrays to different frameworks + import jax.numpy as jnp + import jax + import torch + + # Create ir.Tensor + jax_array = jnp.array([10., 20.]) + ir_tensor_jax = ir.Tensor(jax_array, dtype=ir.DataType.FLOAT) + torch_tensor = torch.tensor([30., 40.]) + ir_tensor_torch = ir.Tensor(torch_tensor, dtype=ir.DataType.FLOAT) + + # Use numpy for computation + print(np.multiply(ir_tensor_jax, ir_tensor_torch)) # array([300., 800.], dtype=float32) + + # Use jax for computation by calling from_dlpack to transfer the tensor data without copying when the device is the same + jax_array_from_ir = jax.dlpack.from_dlpack(ir_tensor_torch) + print(jax_array_from_ir + jax_array) # [40. 60.] + + # Use PyTorch for computation + torch_tensor_from_ir = torch.from_dlpack(ir_tensor_jax) + print(torch_tensor_from_ir - torch_tensor) # tensor([-20., -20.]) + + # They can all be serialized into TensorProto + proto = ir.serde.serialize_tensor(ir_tensor_jax) + print(type(proto)) # + print(proto) + + # The value is exactly the same as jax_array + print(ir.serde.deserialize_tensor(proto).numpy()) # [10. 20.] +``` + +This is particularly useful if you are creating passes on the graph that requires doing computation on concrete values. You are free to use your favorite frameworks to create the passes. The transformed graph that contains newly created `ir.Tensor`s will be compatible with downstream passes even if they leverage other computation frameworks. diff --git a/onnxscript/ir/__init__.py b/onnxscript/ir/__init__.py index 7a2e071d2..6fb36c463 100644 --- a/onnxscript/ir/__init__.py +++ b/onnxscript/ir/__init__.py @@ -60,6 +60,8 @@ "DataType", # Types "OperatorIdentifier", + # Protobuf compatible types + "TensorProtoTensor", ] from onnxscript.ir import serde @@ -118,3 +120,4 @@ TypeProtocol, ValueProtocol, ) +from onnxscript.ir.serde import TensorProtoTensor diff --git a/requirements-dev.txt b/requirements-dev.txt index f1b1cae4c..dfbe51ac2 100644 --- a/requirements-dev.txt +++ b/requirements-dev.txt @@ -7,9 +7,12 @@ rich>=13.7.1 # Docs site furo +jax[cpu] matplotlib +ml_dtypes myst-parser[linkify] sphinx-copybutton +sphinx-exec-code sphinx-gallery sphinx>=6