Configuration
The following configurations can be supplied as Spark properties.
| Property name | Default | Meaning |
|---|---|---|
| spark.rapids.ml.uvm.enabled | false | if set to true, enables CUDA unified virtual memory (aka managed memory) during estimator.fit() operations to allow processing of larger datasets than would fit in GPU memory. |
| spark.rapids.ml.sam.enabled | false | if set to true, enables System Allocated Memory (SAM) on HMM or ATS systems during estimator.fit() operations to allow processing of larger datasets than would fit in GPU memory. Only one of spark.rapids.ml.uvm.enabled or spark.rapids.ml.sam.enabled can be set to true. |
| spark.rapids.ml.sam.headroom | None | when using System Allocated Memory (SAM) and GPU memory is oversubscribed, we may need to reserve some GPU memory as headroom to allow other CUDA calls to function without running out memory. This sets the amount of GPU memory reserved during data loading for additional rmm based allocations during fit and also the amount of GPU memory reserved for non-rmm allocations. Set a size appropriate for your application. |
| spark.rapids.ml.gpuMemRatioForData | None | If set to a float value between 0 and 1, Spark Rapids ML will reserve this fraction of free GPU memory on each GPU and incrementally append PySpark data batches into this reserved space. This setting is recommended for large datasets, as it prevents duplicating the entire dataset in GPU memory and reduces the risk of out-of-memory errors. If spark.rapids.ml.sam.enabled is set to true on hosts that support it, the reserved space will be the specified fraction of host + device memory and will be allocated using SAM across device and host. |
| spark.rapids.ml.cpu.fallback.enabled | false | if set to true and spark-rapids-ml estimator.fit() is invoked with unsupported parameters or parameter values, the pyspark.ml cpu based estimator.fit() and model.transform() will be run; if set to false, an exception is raised in this case (default). |
| spark.rapids.ml.verbose | None | if set to a boolean value (true/false) or an integer between 0 and 6, controls the verbosity level for cuML logging during estimator.fit() operations. This parameter can be set globally in Spark configuration and will be used if not explicitly set in the estimator constructor. |
| spark.rapids.ml.float32_inputs | None | if set to a boolean value (true/false), controls whether input data should be converted to float32 precision before being passed to cuML algorithms. Setting this to true can reduce memory usage and potentially improve performance, but may affect numerical precision. This parameter can be set globally in Spark configuration and will be used if not explicitly set in the estimator constructor. |
| spark.rapids.ml.num_workers | None | if set to an integer value greater than 0, specifies the number of workers to use for distributed training. This parameter can be set globally in Spark configuration and will be used if not explicitly set in the estimator constructor. |
Since the algorithms rely heavily on Pandas UDFs, we also require spark.sql.execution.arrow.pyspark.enabled=true to ensure efficient data transfer between the JVM and Python processes.