Unified Memory Framework (UMF)

1.1.0 verified Mon Apr 27 auth: no python

UMF is a C library for constructing allocators and memory pools, with Python bindings provided by the umf package. It manages multiple memory pools for different hardware resources (CPU, GPU). Current version 1.1.0, release cadence: minor releases every few months with patch releases as needed.

pip install umf

Common errors

error ImportError: cannot import name 'UMF' from 'umf' ↓

cause The main class is UMF, but the module name is umf. Common mistake is `import umf` instead of `from umf import UMF`.

fix

Use from umf import UMF for the main class.

error TypeError: 'module' object is not callable ↓

cause Calling `umf()` after `import umf` which imports the module, not the class.

fix

Either from umf import UMF and use UMF(), or call the class directly.

error FileNotFoundError: Could not find libumf.so ↓

cause The C library is not installed or not in LD_LIBRARY_PATH.

fix

Install the C library (e.g., via conda or building from source) or set LD_LIBRARY_PATH appropriately.

error ValueError: Provider 'JUMA' is not available ↓

cause The requested memory pool type is not supported in the current UMF build.

fix

Check available providers with umf.supported_providers() and use one of them.

Warnings

breaking In UMF 0.10.0, the parameters API changed from direct struct access to a new set/get API. Code using UMF_PARAMS or similar struct fields will break. ↓

fix Migrate to umfParamsCreate/umfParamsDestroy and parameter setter/getter functions.

deprecated UMF_DISABLE_HWLOC option was removed in UMF 1.1.0. ↓

fix Remove any usage of UMF_DISABLE_HWLOC. hwloc topology is now initialized automatically.

gotcha The Python package 'umf' wraps only a subset of the C library. Some providers (e.g., Level Zero, CUDA) are only available if the C library was built with those backends. ↓

fix Check UMF documentation for which providers are available in your build. Use umf.supported_providers() to list them.

Imports

UMF
wrong
```
import umf
```
correct
```
from umf import UMF
```
The main class is UMF; direct import of module does not expose functionality.
MemoryPool
```
from umf.pool import MemoryPool
```

Quickstart

Initializes UMF and creates a memory pool to allocate and free memory. Replace 'JUMA' with appropriate pool type for your hardware.

from umf import UMF

# Initialize UMF
umf = UMF()

# Create a memory pool (example: JUMA pool)
pool = umf.create_pool('JUMA')

# Allocate memory
ptr = pool.alloc(1024)
print(f"Allocated 1024 bytes at {ptr}")

# Free memory
pool.free(ptr)