config
GPU configuration and detection utilities.
This module provides global configuration for GPU acceleration and functions to detect GPU availability at runtime, including a global configuration option to force CPU-only execution even when GPU is available.
Example
from osipy.common.backend import is_gpu_available, set_backend, GPUConfig
Check if GPU is available¶
if is_gpu_available(): ... print("GPU acceleration available")
Force CPU-only execution¶
set_backend(GPUConfig(force_cpu=True))
Or use environment variable: OSIPY_FORCE_CPU=1¶
References
.. [1] CuPy Installation Guide: https://docs.cupy.dev/en/stable/install.html
dataclass
¶
GPUConfig(
force_cpu=False,
default_batch_size=10000,
memory_limit_fraction=0.9,
device_id=0,
n_workers=0,
gpu_dtype="float32",
)
Configuration for GPU/CPU backend selection.
This dataclass holds configuration options for controlling GPU acceleration.
It can be used with set_backend() to configure global behavior.
| PARAMETER | DESCRIPTION |
|---|---|
force_cpu
|
If True, force all operations to run on CPU even if GPU is available. Default is False, which allows automatic GPU usage when available.
TYPE:
|
default_batch_size
|
Default batch size for GPU batch processing. Larger values use more GPU memory but may be faster. Default is 10000.
TYPE:
|
memory_limit_fraction
|
Fraction of GPU memory to use (0.0 to 1.0). Default is 0.9 (90%). This helps prevent out-of-memory errors by leaving headroom.
TYPE:
|
device_id
|
CUDA device ID to use. Default is 0 (first GPU).
TYPE:
|
n_workers
|
Number of threads for CPU-parallel chunk processing in
TYPE:
|
| ATTRIBUTE | DESCRIPTION |
|---|---|
force_cpu |
Whether to force CPU-only execution.
TYPE:
|
default_batch_size |
Default batch size for GPU operations.
TYPE:
|
memory_limit_fraction |
Fraction of GPU memory to use.
TYPE:
|
device_id |
CUDA device ID.
TYPE:
|
n_workers |
Number of threads for CPU-parallel chunk processing.
TYPE:
|
Example
config = GPUConfig(force_cpu=True) set_backend(config)
Check if GPU acceleration is available.
This function checks for CUDA-capable GPU and CuPy installation. The result is cached after the first call for performance.
| RETURNS | DESCRIPTION |
|---|---|
bool
|
True if GPU acceleration is available, False otherwise. |
Notes
GPU is considered available if: 1. CuPy is installed and can be imported 2. At least one CUDA device is detected 3. The OSIPY_FORCE_CPU environment variable is not set to "1"
The availability check is cached after the first call. To force
a recheck, use _reset_gpu_cache() (internal use only).
Example
if is_gpu_available(): ... print("GPU acceleration enabled") ... else: ... print("Running on CPU only")
Get the current backend configuration.
| RETURNS | DESCRIPTION |
|---|---|
GPUConfig
|
The current global GPU configuration. If not explicitly set, returns a default configuration that respects environment variables. |
Example
config = get_backend() print(f"Force CPU: {config.force_cpu}")
Set the global backend configuration.
| PARAMETER | DESCRIPTION |
|---|---|
config
|
The configuration to use globally.
TYPE:
|
Notes
This affects all subsequent calls to get_array_module(), to_gpu(),
and other backend functions. Changes take effect immediately.
Get information about GPU memory usage.
| RETURNS | DESCRIPTION |
|---|---|
dict
|
Dictionary containing: - 'available': bool - whether GPU is available - 'total_bytes': int - total GPU memory in bytes (0 if unavailable) - 'used_bytes': int - used GPU memory in bytes (0 if unavailable) - 'free_bytes': int - free GPU memory in bytes (0 if unavailable) - 'device_name': str - GPU device name (empty if unavailable) |
Example
info = get_gpu_memory_info() if info['available']: ... print(f"GPU: {info['device_name']}") ... print(f"Free: {info['free_bytes'] / 1e9:.2f} GB")
Get optimal batch size based on GPU thread capacity.
Returns the total number of concurrent threads the GPU can handle (multiProcessorCount * maxThreadsPerMultiProcessor), which is a good heuristic for batch sizing.
Returns 0 if GPU is not available.