llama.cpp b11313: Re-enables -sm Tensor for qwen4exp and Drops openEuler Binaries
The latest llama.cpp release fixes previous issues with qwen4exp model support and removes openEuler binaries, affecting users working with these environments.
What changed?
llama.cpp release b11313 re-enables the -sm tensor for the qwen4exp model family. This was previously disabled due to failures with test-llama-archs, specifically an assertion error on the Meta device when a PLE layer was present. The update changes how device assignment logic is applied, resolving the issue and restoring -sm tensor support. At the same time, this release disables binaries for the openEuler platform.
Why does it matter to an everyday developer?
If you use qwen4exp models with llama.cpp, this change restores a previously blocked path (-sm tensor), enabling you to use these models with wider hardware compatibility and improved support. Developers who deploy on openEuler will no longer receive binaries for this platform and must build from source or select another platform. No other major API or breaking changes were introduced.
What can the developer do now?
- If you run qwen4exp models, upgrade to llama.cpp b11313 or later to use the re-enabled -sm tensor functionality.
- Users relying on openEuler for deployment should be aware binaries are not provided in this release. Consider building from source or migrating to a supported Linux distribution.
- Binaries remain available for macOS, Linux, Android, and Windows (multiple flavors).