Controls the level of concurrency and number of OS threads reading tensors from the file to the CPU buffer.
Positive integer value
16 for reading from file system
8 for reading from object storage
Controls the maximum size of memory each OS thread reads from the file at once.
Positive integer value
Note
When reading from file system the minimum size is 2097152 (=2MiB)
When reading from object store the minimum size is 5242880 (=5MiB)
2097152 (=2MiB) when reading from file system
8388608 (=8MiB) when reading from object store e.g. S3
Controls how the CPU Memory buffer to which tensors are read from the file is being limited. Read more about it here.
-1 - UNLIMITED, 0 - MIN, or Positive integer value - LIMITED
-1 for distributed streaming and 40 GB otherwise
Overrides url endpoint for reading from S3 compatible object store
Note
Mandatory for S3 compatible e.g. gcs, minio
Optional if reading from AWS S3
String
Default S3 url endpoint
Specifies the path to a certificate bundle to use for HTTPS certificate validation.
If defined, this environment variable overrides the value for the profile setting ca_bundle.
String
Controls parsing the url endpoint for reading from object store
Note
Optional - No need to set this variable for reading from S3
Should be 0 for reading from compatible object store
Boolean 0 or 1
1
Specifies the path to a credential file to use for GCS authentication.
If not defined (default) Application Default Credentials are used.
String
Azure Storage account name. Used with DefaultAzureCredential for authentication.
String
None
Enables distributed streaming for multiple devices
auto enables distributed streaming only when using the nccl distributed backend
String 0 or 1 or auto
auto
Enables global distributed streaming for multiple devices, dividing the workload between multiple nodes
If not defined (default) distributed streaming is local - dividing the workload between processes of the same node
String 0 or 1
0