docs(python): Add documentation for beta gpu support

docs/_build/API_REFERENCE_LINKS.yml

@@ -3,6 +3,8 @@ python:
   LazyFrame: https://docs.pola.rs/api/python/stable/reference/lazyframe/index.html
   Series: https://docs.pola.rs/api/python/stable/reference/series/index.html
   Categorical: https://docs.pola.rs/api/python/stable/reference/api/polars.Categorical.html
+  GPUEngine: https://docs.pola.rs/api/python/stable/lazyframe/api/polars.lazyframe.engine_config.GPUEngine.html
+  Config: https://docs.pola.rs/api/python/stable/reference/config.html
   select: https://docs.pola.rs/api/python/stable/reference/dataframe/api/polars.DataFrame.select.html
   filter: https://docs.pola.rs/api/python/stable/reference/dataframe/api/polars.DataFrame.filter.html
   with_columns: https://docs.pola.rs/api/python/stable/reference/dataframe/api/polars.DataFrame.with_columns.html

docs/index.md

@@ -26,6 +26,7 @@ Polars is a blazingly fast DataFrame library for manipulating structured data. T
 - **Out of Core**: The streaming API allows you to process your results without requiring all your data to be in memory at the same time
 - **Parallel**: Utilises the power of your machine by dividing the workload among the available CPU cores without any additional configuration.
 - **Vectorized Query Engine**: Using [Apache Arrow](https://arrow.apache.org/), a columnar data format, to process your queries in a vectorized manner and SIMD to optimize CPU usage.
+- **GPU Support**: Optionally run queries on NVIDIA GPUs for maximum performance for in-memory workloads.
 
 <!-- dprint-ignore-start -->
 

docs/src/python/user-guide/lazy/gpu.py

@@ -0,0 +1,66 @@
+# --8<-- [start:setup]
+# --8<-- [start:simple]
+import polars as pl
+
+df = pl.LazyFrame({"a": [1.242, 1.535]})
+
+q = df.select(pl.col("a").round(1))
+# --8<-- [end:setup]
+
+result = q.collect(engine="gpu")
+print(result)
+# --8<-- [end:simple]
+
+# Avoiding requiring the GPU engine for doc build output
+# --8<-- [start:simple-result]
+result = q.collect()
+print(result)
+# --8<-- [end:simple-result]
+
+
+# --8<-- [start:engine]
+# --8<-- [start:engine-setup]
+q = df.select((pl.col("a") ** 4))
+
+# --8<-- [end:engine-setup]
+result = q.collect(engine=pl.GPUEngine(device=1))
+print(result)
+# --8<-- [end:engine]
+
+# --8<-- [start:engine-result]
+result = q.collect()
+print(result)
+# --8<-- [end:engine-result]
+
+# --8<-- [start:fallback-warning]
+# --8<-- [start:fallback-setup]
+df = pl.LazyFrame(
+    {
+        "key": [1, 1, 1, 2, 3, 3, 2, 2],
+        "value": [1, 2, 3, 4, 5, 6, 7, 8],
+    }
+)
+
+q = df.select(pl.col("value").sum().over("key"))
+
+# --8<-- [end:fallback-setup]
+
+with pl.Config() as cfg:
+    cfg.set_verbose(True)
+    result = q.collect(engine="gpu")
+
+print(result)
+# --8<-- [end:fallback-warning]
+# --8<-- [start:fallback-result]
+print(
+    "PerformanceWarning: Query execution with GPU not supported, reason: \n"
+    "<class 'NotImplementedError'>: Grouped rolling window not implemented"
+)
+print("# some details elided")
+print()
+print(q.collect())
+# --8<- [end:fallback-result]
+
+# --8<-- [start:fallback-raise]
+q.collect(engine=pl.GPUEngine(raise_on_fail=True))
+# --8<-- [end:fallback-raise]

docs/user-guide/gpu-support.md

@@ -0,0 +1,133 @@
+# GPU Support [Open Beta]
+
+Polars provides an in-memory, GPU-accelerated execution engine for Python users of the Lazy API on NVIDIA GPUs using [RAPIDS cuDF](https://docs.rapids.ai/api/cudf/stable/). This functionality is available in Open Beta and is undergoing rapid development.
+
+### System Requirements
+
+- NVIDIA Volta™ or higher GPU with [compute capability](https://developer.nvidia.com/cuda-gpus) 7.0+
+- CUDA 11 or CUDA 12
+- Linux or Windows Subsystem for Linux 2 (WSL2)
+
+See the [RAPIDS installation guide](https://docs.rapids.ai/install#system-req) for full details.
+
+### Installation
+
+You can install the GPU backend for Polars with a feature flag as part of a normal [installation](installation.md).
+
+=== ":fontawesome-brands-python: Python"
+`bash pip install --extra-index-url=https://pypi.nvidia.com polars[gpu]`
+
+!!! note Installation on a CUDA 11 system
+
+    If you have CUDA 11, the installation line is slightly more complicated: the relevant GPU package must be requested by hand.
+
+    === ":fontawesome-brands-python: Python"
+    ```bash
+    pip install --extra-index-url=https://pypi.nvidia.com polars cudf-polars-cu11
+    ```
+
+### Usage
+
+Having built a query using the lazy API [as normal](lazy/index.md), GPU-enabled execution is requested by running `.collect(engine="gpu")` instead of `.collect()`.
+
+{{code_block('user-guide/lazy/gpu', 'simple', ['collect'])}}
+
+```python exec="on" result="text" session="user-guide/lazy"
+--8<-- "python/user-guide/lazy/gpu.py:setup"
+--8<-- "python/user-guide/lazy/gpu.py:simple-result"
+```
+
+For more detailed control over the execution, for example to specify which GPU to use on a multi-GPU node, we can provide a `GPUEngine` object. By default, the GPU engine will use a configuration applicable to most use cases.
+
+{{code_block('user-guide/lazy/gpu', 'engine', ['GPUEngine'])}}
+
+```python exec="on" result="text" session="user-guide/lazy"
+--8<-- "python/user-guide/lazy/gpu.py:engine-setup"
+--8<-- "python/user-guide/lazy/gpu.py:engine-result"
+```
+
+### How It Works
+
+When you use the GPU-accelerated engine, Polars creates and optimizes a query plan and dispatches to a [RAPIDS](https://rapids.ai/) cuDF-based physical execution engine to compute the results on NVIDIA GPUs. The final result is returned as a normal CPU-backed Polars dataframe.
+
+### What's Supported on the GPU?
+
+GPU support is currently in Open Beta and the engine is undergoing rapid development. The engine currently supports many, but not all, of the core expressions and data types.
+
+Since expressions are composable, it's not feasible to list a full matrix of expressions supported on the GPU. Instead, we provide a list of the high-level categories of expressions and interfaces that are currently supported and not supported.
+
+#### Supported
+
+- LazyFrame API
+- SQL API
+- I/O from CSV, Parquet, ndjson, and in-memory CPU DataFrames.
+- Operations on numeric, logical, string, and datetime types
+- String processing
+- Aggregations and grouped aggregations
+- Joins
+- Filters
+- Missing data
+- Concatenation
+
+#### Not Supported
+
+- Eager DataFrame API
+- Streaming API
+- Operations on categorical, struct, and list data types
+- Rolling aggregations
+- Time series resampling
+- Timezones
+- Folds
+- User-defined functions
+- JSON, Excel, and Database file formats
+
+#### Did my query use the GPU?
+
+The release of the GPU engine in Open Beta implies that we expect things to work well, but there are still some rough edges we're working on. In particular the full breadth of the Polars expression API is not yet supported. With fallback to the CPU, your query _should_ complete, but you might not observe any change in the time it takes to execute. There are two ways to get more information on whether the query ran on the GPU.
+
+When running in verbose mode, any queries that cannot execute on the GPU will issue a `PerformanceWarning`:
+
+{{code_block('user-guide/lazy/gpu', 'fallback-warning', ['Config'])}}
+
+```python exec="on" result="text" session="user-guide/lazy"
+--8<-- "python/user-guide/lazy/gpu.py:fallback-setup"
+--8<-- "python/user-guide/lazy/gpu.py:fallback-result"
+```
+
+To disable fallback, and have the GPU engine raise an exception if a query is unsupported, we can pass an appropriately configured `GPUEngine` object:
+
+{{code_block('user-guide/lazy/gpu', 'fallback-raise', ['GPUEngine'])}}
+
+```pytb
+Traceback (most recent call last):
+  File "<stdin>", line 1, in <module>
+  File "/home/coder/third-party/polars/py-polars/polars/lazyframe/frame.py", line 2035, in collect
+    return wrap_df(ldf.collect(callback))
+polars.exceptions.ComputeError: 'cuda' conversion failed: NotImplementedError: Grouped rolling window not implemented
+```
+
+Currently, only the proximal cause of failure to execute on the GPU is reported, we plan to extend this functionality to report all unsupported operations for a query.
+
+### Testing
+
+The Polars and NVIDIA RAPIDS teams run comprehensive unit and integration tests to ensure that the GPU-accelerated Polars backend works smoothly.
+
+The **full** Polars test suite is run on every commit made to the GPU engine, ensuring consistency of results.
+
+The GPU engine currently passes 99.2% of the Polars unit tests with CPU fallback enabled. Without CPU fallback, the GPU engine passes 88.8% of the Polars unit tests. With fallback, there are approximately 100 failing tests: around 40 of these fail due to mismatching debug output; there are some cases where the GPU engine produces the a correct result but uses a different data type; the remainder are cases where we do not correctly determine that a query is unsupported and therefore fail at runtime, instead of falling back.
+
+### When Should I Use a GPU?
+
+Based on our benchmarking, you're most likely to observe speedups using the GPU engine when your workflow's profile is dominated by grouped aggregations and joins. In contrast I/O bound queries typically show similar performance on GPU and CPU. GPUs typically have less RAM than CPU systems, therefore very large datasets will fail due to out of memory errors. Based on our testing, raw datasets of 50-100 GiB fit (depending on the workflow) well with a GPU with 80GiB of memory.
+
+### CPU-GPU Interoperability
+
+Both the CPU and GPU engine use the Apache Arrow columnar memory specification, making it possible to quickly move data between the CPU and GPU. Additionally, files written by one engine can be read by the other engine.
+
+When using GPU mode, your workflow won't fail if something isn't supported. When you run `collect(engine="gpu")`, the optimized query plan is inspected to see whether it can be executed on the GPU. If it can't, it will transparently fall back to the standard Polars engine and run on the CPU.
+
+GPU execution is only available in the Lazy API, so materialized DataFrames will reside in CPU memory when the query execution finishes.
+
+### Providing feedback
+
+Please report issues, and missing features, on the Polars [issue tracker](../development/contributing/index.md).

docs/user-guide/installation.md

@@ -88,6 +88,19 @@ pip install 'polars[numpy,fsspec]'
 | --- | --------------------------------- |
 | all | Install all optional dependencies |
 
+#### GPU
+
+| Tag | Description                |
+| --- | -------------------------- |
+| gpu | Run queries on NVIDIA GPUs |
+
+!!! note
+
+    To install the GPU engine, you need to pass
+    `--extra-index-url=https://pypi.nvidia.com` to `pip`. See [GPU
+    support](gpu-support.md) for more detailed instructions and
+    prerequisites.
+
 #### Interop
 
 | Tag      | Description                                       |

docs/user-guide/lazy/gpu.md

@@ -0,0 +1,14 @@
+# GPU Support
+
+Polars provides an in-memory, GPU-accelerated execution engine for the Lazy API in Python using [RAPIDS cuDF](https://docs.rapids.ai/api/cudf/stable/) on NVIDIA GPUs. This functionality is available in Open Beta and is undergoing rapid development.
+
+If you install Polars with the [GPU feature flag](../installation.md), you can trigger GPU-based execution by running `.collect(engine="gpu")` instead of `.collect()`.
+
+{{code_block('user-guide/lazy/gpu', 'simple', [])}}
+
+```python exec="on" result="text" session="user-guide/lazy"
+--8<-- "python/user-guide/lazy/gpu.py:setup"
+--8<-- "python/user-guide/lazy/gpu.py:simple-result"
+```
+
+Learn more in the [GPU Support guide](../gpu-support.md).

docs/user-guide/lazy/index.md

@@ -8,3 +8,4 @@ The Lazy chapter is a guide for working with `LazyFrames`. It covers the functio
 - [Query plan](query-plan.md)
 - [Execution](execution.md)
 - [Streaming](streaming.md)
+- [GPU Support](gpu.md)

docs/user-guide/misc/multiprocessing.md

@@ -11,6 +11,8 @@ It does this by executing computations which can be done in parallel in separate
 For example, requesting two expressions in a `select` statement can be done in parallel, with the results only being combined at the end.
 Another example is aggregating a value within groups using `group_by().agg(<expr>)`, each group can be evaluated separately.
 It is very unlikely that the `multiprocessing` module can improve your code performance in these cases.
+If you're using the GPU Engine with Polars you should also avoid manual multiprocessing. When used simultaneously, they can compete
+for system memory and processing power, leading to reduced performance.
 
 See [the optimizations section](../lazy/optimizations.md) for more optimizations.
 

mkdocs.yml

@@ -58,6 +58,7 @@ nav:
       - user-guide/lazy/query-plan.md
       - user-guide/lazy/execution.md
       - user-guide/lazy/streaming.md
+      - user-guide/lazy/gpu.md
     - IO:
       - user-guide/io/index.md
       - user-guide/io/csv.md
@@ -86,6 +87,7 @@ nav:
       - user-guide/misc/styling.md
       - user-guide/misc/comparison.md
       - user-guide/misc/arrow.md
+    - user-guide/gpu-support.md
 
   - API reference: api/index.md
 

py-polars/docs/source/reference/lazyframe/gpu_engine.rst

@@ -0,0 +1,14 @@
+=========
+GPUEngine
+=========
+
+This object provides fine-grained control over the behavior of the
+GPU engine when calling `LazyFrame.collect()` with an `engine`
+argument.
+
+.. currentmodule:: polars.lazyframe.engine_config
+
+.. autosummary::
+   :toctree: api/
+
+    GPUEngine

py-polars/docs/source/reference/lazyframe/index.rst

@@ -15,6 +15,7 @@ This page gives an overview of all public LazyFrame methods.
    modify_select
    miscellaneous
    in_process
+   gpu_engine
 
 .. _lazyframe:
 

py-polars/polars/lazyframe/engine_config.py

@@ -14,17 +14,35 @@ class GPUEngine:
 
     Use this if you want control over details of the execution.
 
-    Supported options
+    Parameters
+    ----------
+    device : int, default None
+        Select the GPU used to run the query. If not provided, the
+        query uses the current CUDA device.
+    memory_resource : rmm.mr.DeviceMemoryResource, default None
+        Provide a memory resource for GPU memory allocations.
+
+        .. warning::
+           If passing a `memory_resource`, you must ensure that it is valid
+           for the selected `device`. See the `RMM documentation
+           <https://github.com/rapidsai/rmm?tab=readme-ov-file#multiple-devices>`_
+           for more details.
+
+    raise_on_fail : bool, default False
+        If True, do not fall back to the Polars CPU engine if the GPU
+        engine cannot execute the query, but instead raise an error.
 
-    - `device`: Select the device to run the query on.
-    - `memory_resource`: Set an RMM memory resource for
-      device-side allocations.
     """
 
     device: int | None
     """Device on which to run query."""
     memory_resource: DeviceMemoryResource | None
     """Memory resource to use for device allocations."""
+    raise_on_fail: bool
+    """
+    Whether unsupported queries should raise an error, rather than falling
+    back to the CPU engine.
+    """
     config: Mapping[str, Any]
     """Additional configuration options for the engine."""
 
@@ -33,8 +51,11 @@ def __init__(
         *,
         device: int | None = None,
         memory_resource: Any | None = None,
+        raise_on_fail: bool = False,
         **kwargs: Any,
     ) -> None:
         self.device = device
         self.memory_resource = memory_resource
+        # Avoids need for changes in cudf-polars
+        kwargs["raise_on_fail"] = raise_on_fail
         self.config = kwargs

py-polars/polars/lazyframe/frame.py

@@ -1865,7 +1865,7 @@ def collect(
             polars CPU engine. If set to `"gpu"`, the GPU engine is
             used. Fine-grained control over the GPU engine, for
             example which device to use on a system with multiple
-            devices, is possible by providing a :class:`GPUEngine` object
+            devices, is possible by providing a :class:`~.GPUEngine` object
             with configuration options.
 
             .. note::
@@ -2019,9 +2019,10 @@ def collect(
                 "cudf_polars",
                 err_prefix="GPU engine requested, but required package",
                 install_message=(
-                    "Please install using the command `pip install cudf-polars-cu12` "
-                    "(or `pip install cudf-polars-cu11` if your system has a "
-                    "CUDA 11 driver)."
+                    "Please install using the command "
+                    "`pip install --extra-index-url=https://pypi.nvidia.com cudf-polars-cu12` "
+                    "(or `pip install --extra-index-url=https://pypi.nvidia.com cudf-polars-cu11` "
+                    "if your system has a CUDA 11 driver)."
                 ),
             )
             if not is_config_obj: