From 43b66cb760ba060adbcdf75c5d3073e2f571af47 Mon Sep 17 00:00:00 2001 From: Jeremy Fowers Date: Sat, 2 Dec 2023 16:08:27 -0500 Subject: [PATCH] Update the documentation Signed-off-by: Jeremy Fowers --- docs/code.md | 11 ++-- docs/readme.md | 2 +- docs/tools_user_guide.md | 58 +++++-------------- .../turnkeyml_plugin_example_seq/sequence.py | 2 +- src/turnkeyml/analyze/script.py | 10 ++-- 5 files changed, 27 insertions(+), 56 deletions(-) diff --git a/docs/code.md b/docs/code.md index e9fdc7a0..28663677 100644 --- a/docs/code.md +++ b/docs/code.md @@ -11,8 +11,8 @@ The TurnkeyML source code has a few major top-level directories: - `models`: the corpora of models that makes up the TurnkeyML models (see [the models readme](https://github.com/onnx/turnkeyml/blob/main/models/readme.md)). - Each subdirectory under `models` represents a corpus of models pulled from somewhere on the internet. For example, `models/torch_hub` is a corpus of models from [Torch Hub](https://github.com/pytorch/hub). - `src/turnkey`: source code for the TurnkeyML tools (see [Benchmarking Tools](#benchmarking-tools) for a description of how the code is used). - - `src/turnkeyml/analyze`: functions for profiling a model script, discovering model instances, and invoking `benchmark_model()` on those instances. - - `src/turnkeyml/run`: implements the runtime and device plugin APIs and the built-in runtimes and devices. + - `src/turnkeyml/analyze`: functions for profiling a model script, discovering model instances, and invoking `build_model()` and/or `BaseRT.benchmark()` on those instances. + - `src/turnkeyml/run`: implements `BaseRT`, an abstract base class that defines TurnkeyML's vendor-agnostic benchmarking functionality. This module also includes the runtime and device plugin APIs and the built-in runtimes and devices. - `src/turnkeyml/cli`: implements the `turnkey` CLI and reporting tool. - `src/turnkeyml/common`: functions common to the other modules. - `src/turnkeyml/version.py`: defines the package version number. @@ -29,10 +29,9 @@ TurnkeyML provides two main tools, the `turnkey` CLI and benchmarking APIs. Inst 1. The default command for `turnkey` CLI runs the `benchmark_files()` API, which is implemented in [files_api.py](https://github.com/onnx/turnkeyml/blob/main/src/turnkeyml/files_api.py). - Other CLI commands are also implemented in `cli/`, for example the `report` command is implemented in `cli/report.py`. 1. The `benchmark_files()` API takes in a set of scripts, each of which should invoke at least one model instance, to evaluate and passes each into the `evaluate_script()` function for analysis, which is implemented in [analyze/script.py](https://github.com/onnx/turnkeyml/blob/main/src/turnkeyml/analyze/script.py). -1. `evaluate_script()` uses a profiler to discover the model instances in the script, and passes each into the `benchmark_model()` API, which is defined in [model_api.py](https://github.com/onnx/turnkeyml/blob/main/src/turnkeyml/model_api.py). -1. The `benchmark_model()` API prepares the model for benchmarking (e.g., exporting and optimizing an ONNX file), which creates an instance of a `*Model` class, where `*` can be CPU, GPU, etc. The `*Model` classes are defined in [run/](https://github.com/onnx/turnkeyml/blob/main/src/turnkeyml/run/). -1. The `*Model` classes provide a `.benchmark()` method that benchmarks the model on the device and returns an instance of the `MeasuredPerformance` class, which includes the performance statistics acquired during benchmarking. -1. `benchmark_model()` and the `*Model` classes are built using [`build_model()`](#model-build-tool) +1. `evaluate_script()` uses a profiler to discover the model instances in the script, and passes each into the `build_model()` API, which is defined in [build_api.py](https://github.com/onnx/turnkeyml/blob/main/src/turnkeyml/build_api.py). +1. The `build_model()` API prepares the model for benchmarking (e.g., exporting and optimizing an ONNX file). +1. `evaluate_script()` passes the build into `BaseRT.benchmark()` to benchmarks the model on the device and returns an instance of the `MeasuredPerformance` class, which includes the performance statistics acquired during benchmarking. # Model Build Tool diff --git a/docs/readme.md b/docs/readme.md index 55ec96c0..5b26c33c 100644 --- a/docs/readme.md +++ b/docs/readme.md @@ -3,7 +3,7 @@ This directory contains documentation for the TurnkeyML project: - [code.md](https://github.com/onnx/turnkeyml/blob/main/docs/code.md): Code organization for the benchmark and tools. - [install.md](https://github.com/onnx/turnkeyml/blob/main/docs/install.md): Installation instructions for the tools. -- [tools_user_guide.md](https://github.com/onnx/turnkeyml/blob/main/docs/tools_user_guide.md): User guide for the tools: `turnkey` CLI, `benchmark_files()`, and `benchmark_model()`. +- [tools_user_guide.md](https://github.com/onnx/turnkeyml/blob/main/docs/tools_user_guide.md): User guide for the tools: the `turnkey` CLI and the `benchmark_files()` and `build_model()` APIs. - [versioning.md](https://github.com/onnx/turnkeyml/blob/main/docs/versioning.md): Defines the semantic versioning rules for the `turnkey` package. There is more useful documentation available in: diff --git a/docs/tools_user_guide.md b/docs/tools_user_guide.md index d9819979..53d24635 100644 --- a/docs/tools_user_guide.md +++ b/docs/tools_user_guide.md @@ -51,8 +51,8 @@ Where `your_script.py` is a Python script that instantiates and executes a PyTor The `turnkey` CLI performs the following steps: 1. [Analysis](#analysis): profile the Python script to identify the PyTorch models within -2. [Build](#build): call the `benchmark_files()` [API](#the-turnkey-api) to prepare each model for benchmarking -3. [Benchmark](#benchmark): call the `benchmark_model()` [API](#the-turnkey-api) on each model to gather performance statistics +2. [Build](#build): call the `build_models()` [API](#the-turnkey-api) to prepare each model for benchmarking +3. [Benchmark](#benchmark): call the `BaseRT.benchmark()` method on each model to gather performance statistics _Note_: The benchmarking methodology is defined [here](#benchmark). If you are looking for more detailed instructions on how to install turnkey, you can find that [here](https://github.com/onnx/turnkeyml/blob/main/docs/install.md). @@ -64,31 +64,11 @@ _Note_: The benchmarking methodology is defined [here](#benchmark). If you are l Most of the functionality provided by the `turnkey` CLI is also available in the the API: - `turnkey.benchmark_files()` provides the same benchmarking functionality as the `turnkey` CLI: it takes a list of files and target device, and returns performance results. -- `turnkey.benchmark_model()` provides a subset of this functionality: it takes a model and its inputs, and returns performance results. - - The main difference is that `benchmark_model()` does not include the [Analysis](#analysis) feature, and `benchmark_files()` does. - `turnkey.build_model(model, inputs)` is used to programmatically [build](#build) a model instance through a sequence of model-to-model transformations (e.g., starting with an fp32 PyTorch model and ending with an fp16 ONNX model). -Generally speaking, the `turnkey` CLI is a command line interface for the `benchmark_files()` API, which internally calls `benchmark_model()`, which in turn calls `build_model()`. You can read more about this code organization [here](https://github.com/onnx/turnkeyml/blob/main/docs/code.md). +Generally speaking, the `turnkey` CLI is a command line interface for the `benchmark_files()` API which in turn calls `build_model()` and then performs benchmarking using `BaseRT.benchmark()`. You can read more about this code organization [here](https://github.com/onnx/turnkeyml/blob/main/docs/code.md). -For an example of `benchmark_model()`, the following script: - -```python -from turnkeyml import benchmark_model - -model = YourModel() # Instantiate a torch.nn.module -results = model(**inputs) -perf = benchmark_model(model, inputs) -``` - -Will print an output like this: - -``` -> Performance of YourModel on device Intel® Xeon® Platinum 8380 is: -> latency: 0.033 ms -> throughput: 21784.8 ips -``` - -`benchmark_model()` returns a `MeasuredPerformance` object that includes members: +`BaseRT.benchmark()` returns a `MeasuredPerformance` object that includes members: - `latency_units`: unit of time used for measuring latency, which is set to `milliseconds (ms)`. - `mean_latency`: average benchmarking latency, measured in `latency_units`. - `throughput_units`: unit used for measuring throughput, which is set to `inferences per second (IPS)`. @@ -135,7 +115,7 @@ A **runtime** is a piece of software that executes a model on a device. **Analysis** is the process by which `benchmark_files()` inspects a Python script or ONNX file and identifies the models within. -`benchmark_files()` performs analysis by running and profiling your file(s). When a model object (see [Model](#model) is encountered, it is inspected to gather statistics (such as the number of parameters in the model) and/or pass it to the `benchmark_model()` API for benchmarking. +`benchmark_files()` performs analysis by running and profiling your file(s). When a model object (see [Model](#model) is encountered, it is inspected to gather statistics (such as the number of parameters in the model) and/or passed to the build and benchmark APIs. > _Note_: the `turnkey` CLI and `benchmark_files()` API both run your entire python script(s) whenever python script(s) are passed as input files. Please ensure that these scripts are safe to run, especially if you got them from the internet. @@ -205,12 +185,14 @@ The *build cache* is a location on disk that holds all of the artifacts from you ## Benchmark -*Benchmark* is the process by which the `benchmark_model()` API collects performance statistics about a [model](#model). Specifically, `benchmark_model()` takes a [build](#build) of a model and executes it on a target device using target runtime software (see [Devices and Runtimes](#devices-and-runtimes)). +*Benchmark* is the process by which `BaseRT.benchmark()` collects performance statistics about a [model](#model). `BaseRT` is an abstract base class that defines the common benchmarking infrastructure that TurnkeyML provides across devices and runtimes. + +Specifically, `BaseRT.benchmark()` takes a [build](#build) of a model and executes it on a target device using target runtime software (see [Devices and Runtimes](#devices-and-runtimes)). -By default, `benchmark_model()` will run the model 100 times to collect the following statistics: +By default, `BaseRT.benchmark()` will run the model 100 times to collect the following statistics: 1. Mean Latency, in milliseconds (ms): the average time it takes the runtime/device combination to execute the model/inputs combination once. This includes the time spent invoking the device and transferring the model's inputs and outputs between host memory and the device (when applicable). 1. Throughput, in inferences per second (IPS): the number of times the model/inputs combination can be executed on the runtime/device combination per second. - > - _Note_: `benchmark_model()` is not aware of whether `inputs` is a single input or a batch of inputs. If your `inputs` is actually a batch of inputs, you should multiply `benchmark_model()`'s reported IPS by the batch size. + > - _Note_: `BaseRT.benchmark()` is not aware of whether `inputs` is a single input or a batch of inputs. If your `inputs` is actually a batch of inputs, you should multiply `BaseRT.benchmark()`'s reported IPS by the batch size. # Devices and Runtimes @@ -226,7 +208,7 @@ If you are using a remote machine, it must: - include the target device - have `miniconda`, `python>=3.8`, and `docker>=20.10` installed -When you call `turnkey` CLI or `benchmark_model()`, the following actions are performed on your behalf: +When you call `turnkey` CLI or `benchmark_files()`, the following actions are performed on your behalf: 1. Perform a `build`, which exports all models from the script to ONNX and prepares for benchmarking. 1. Set up the benchmarking environment by loading a container and/or setting up a conda environment. 1. Run the benchmarks. @@ -253,7 +235,6 @@ Valid values of `TYPE` include: Also available as API arguments: - `benchmark_files(device=...)` -- `benchmark_model(device=...)`. > For a detailed example, see the [CLI Nvidia tutorial](https://github.com/onnx/turnkeyml/blob/main/examples/cli/readme.md#nvidia-benchmarking). @@ -274,9 +255,8 @@ Each device type has its own default runtime, as indicated below. This feature is also be available as an API argument: - `benchmark_files(runtime=[...])` -- `benchmark_model(runtime=...)` -> _Note_: Inputs to `torch-eager` and `torch-compiled` are not downcasted to FP16 by default. Downcast inputs before benchmarking for a fair comparison between runtimes. +> _Note_: Inputs to `torch-eager` and `torch-compiled` are not downcasted to FP16 by default. You must perform your own downcast or quantization of inputs if needed for apples-to-apples comparisons with other runtimes. # Additional Commands and Options @@ -381,7 +361,6 @@ Process isolation mode applies a timeout to each subprocess. The default timeout Also available as API arguments: - `benchmark_files(cache_dir=...)` -- `benchmark_model(cache_dir=...)` - `build_model(cache_dir=...)` > See the [Cache Directory tutorial](https://github.com/onnx/turnkeyml/blob/main/examples/cli/cache.md#cache-directory) for a detailed example. @@ -392,7 +371,6 @@ Also available as API arguments: Also available as API arguments: - `benchmark_files(lean_cache=True/False, ...)` (default False) -- `benchmark_model(lean_cache=True/False, ...)` (default False) > _Note_: useful for benchmarking many models, since the `build` artifacts from the models can take up a significant amount of hard drive space. @@ -409,7 +387,6 @@ Takes one of the following values: Also available as API arguments: - `benchmark_files(rebuild=...)` -- `benchmark_model(rebuild=...)` - `build_model(rebuild=...)` ### Sequence @@ -421,7 +398,6 @@ Usage: Also available as API arguments: - `benchmark_files(sequence=...)` -- `benchmark_model(sequence=...)` - `build_model(sequence=...)` ### Set Script Arguments @@ -460,7 +436,6 @@ Usage: Also available as API arguments: - `benchmark_files(onnx_opset=...)` -- `benchmark_model(onnx_opset=...)` - `build_model(onnx_opset=...)` > _Note_: ONNX opset can also be set by an environment variable. The --onnx-opset argument takes precedence over the environment variable. See [TURNKEY_ONNX_OPSET](#set-the-onnx-opset). @@ -474,11 +449,10 @@ Usage: Also available as API arguments: - `benchmark_files(iterations=...)` -- `benchmark_model(iterations=...)` ### Analyze Only -Instruct `turnkey` or `benchmark_model()` to only run the [Analysis](#analysis) phase of the `benchmark` command. +Instruct `turnkey` or `benchmark_files()` to only run the [Analysis](#analysis) phase of the `benchmark` command. Usage: - `turnkey benchmark INPUT_FILES --analyze-only` @@ -493,7 +467,7 @@ Also available as an API argument: ### Build Only -Instruct `turnkey`, `benchmark_files()`, or `benchmark_model()` to only run the [Analysis](#analysis) and [Build](#build) phases of the `benchmark` command. +Instruct `turnkey` or `benchmark_files()` to only run the [Analysis](#analysis) and [Build](#build) phases of the `benchmark` command. Usage: - `turnkey benchmark INPUT_FILES --build-only` @@ -503,7 +477,6 @@ Usage: Also available as API arguments: - `benchmark_files(build_only=True/False)` (default False) -- `benchmark_model(build_only=True/False)` (default False) > See the [Build Only tutorial](https://github.com/onnx/turnkeyml/blob/main/examples/cli/build.md#build-only) for a detailed example. @@ -515,7 +488,6 @@ None of the built-in runtimes support such arguments, however plugin contributor Also available as API arguments: - `benchmark_files(rt_args=Dict)` (default None) -- `benchmark_model(rt_args=Dict)` (default None) ## Cache Commands @@ -635,7 +607,7 @@ export TURNKEY_DEBUG=True ### Set the ONNX Opset -By default, `turnkey`, `benchmark_files()`, and `benchmark_model()` will use the default ONNX opset defined in `turnkey.common.build.DEFAULT_ONNX_OPSET`. You can set a different default ONNX opset by setting the `TURNKEY_ONNX_OPSET` environment variable. +By default, `turnkey`, `benchmark_files()`, and `build_model()` will use the default ONNX opset defined in `turnkey.common.build.DEFAULT_ONNX_OPSET`. You can set a different default ONNX opset by setting the `TURNKEY_ONNX_OPSET` environment variable. For example: diff --git a/examples/cli/plugins/example_seq/turnkeyml_plugin_example_seq/sequence.py b/examples/cli/plugins/example_seq/turnkeyml_plugin_example_seq/sequence.py index 6c1a1229..350f2a76 100644 --- a/examples/cli/plugins/example_seq/turnkeyml_plugin_example_seq/sequence.py +++ b/examples/cli/plugins/example_seq/turnkeyml_plugin_example_seq/sequence.py @@ -1,7 +1,7 @@ """ This script is an example of a sequence.py file for Sequence Plugin. Such a sequence.py can be used to redefine the build phase of the turnkey CLI, benchmark_files(), -and benchmark_model() to have any custom behavior. +and build_model() to have any custom behavior. In this example sequence.py file we are setting the build sequence to simply export from pytorch to ONNX. This differs from the default build sequence, which diff --git a/src/turnkeyml/analyze/script.py b/src/turnkeyml/analyze/script.py index 87d2ff55..9ef4f792 100644 --- a/src/turnkeyml/analyze/script.py +++ b/src/turnkeyml/analyze/script.py @@ -169,7 +169,7 @@ def explore_invocation( invocation_info.stats_keys = [] # Create an ID for the build stats by combining the device and runtime. - # We don't need more info in the stats_id because changes to benchmark_model() + # We don't need more info in the stats_id because changes to build_model() # arguments (e.g., sequence) will trigger a rebuild, which is intended to replace the # build stats so long as the device and runtime have not changed. stats_id = f"{tracer_args.device}_{selected_runtime}" @@ -343,7 +343,7 @@ def explore_invocation( invocation_info.status_message_color = printing.Colors.WARNING except exp.ArgError as e: - # ArgError indicates that some argument to benchmark_model() was + # ArgError indicates that some argument to build_model() or BaseRT was # illegal. In that case we want to halt execution so that users can # fix their arguments. @@ -626,7 +626,7 @@ def forward_spy(*args, **kwargs): ) invocation_info.executed = invocation_info.executed + 1 - # Call benchmark_model() if this is the first time the model is being executed + # Call explore_invocation() if this is the first time the model is being executed # and this model has been selected by the user if ( invocation_info.executed == 1 @@ -639,7 +639,7 @@ def forward_spy(*args, **kwargs): invocation_info=invocation_info, tracer_args=tracer_args, ) - # Ensure that benchmark_model() doesn't interfere with our execution count + # Ensure that explore_invocation() doesn't interfere with our execution count model_info.executed = 1 build_name = fs.get_build_name( @@ -811,7 +811,7 @@ def evaluate_script(tracer_args: TracerArgs) -> Dict[str, util.ModelInfo]: "torch.jit.script(", "torch.jit.script() is not supported by turnkey CLI and benchmark_files() API, " "however torch.jit.script() is being called in your script." - "You can try passing your model instance into the benchmark_model() API instead. ", + "You can try passing your model instance into the build_model() API instead. ", ) ] ):