Llama.csharp-v1.3.0

This release includes the full source code along with a ready-to-use archive of the library subproject:

📦 llama.csharp-v1.3.0.zip

Supported llama.cpp versions
b9460 and later (up to the latest releases)

You can download the engine files from the official llama.cpp releases: 🔗 llama.cpp releases

Public changes

New features

• Added ClearSequence method in LlamaExecutor for clearing sequence state and cache for any model type.
• CopySeqPrefixTo can now be used regardless of the source and target sequence states.

Fixed

• Fixed partial copy in CopySeqPrefixTo: allowed for hybrid models and fixed setting sequence state after copying (previously the sequence state after partial copy was corrupted).

Other

Tests

• Added more partial copy tests for different model types and sequence states.
• Added ClearSequence tests.

Llama.csharp-v1.2.1

This release includes the full source code along with a ready-to-use archive of the library subproject:

📦 llama.csharp-v1.2.1.zip

Supported llama.cpp versions
b9460 and later (up to the latest releases)

You can download the engine files from the official llama.cpp releases: 🔗 llama.cpp releases

Public changes

Fixed

• Added check in CopySeqPrefixTo method in LlamaExecutor to prevent partial copy for hybrid and recurrent models.

Other

Tests

• Added partial copy test.

Llama.csharp-v1.2.0

This release includes the full source code along with a ready-to-use archive of the library subproject:

📦 llama.csharp-v1.2.0.zip

Supported llama.cpp versions
b9460 and later (up to the latest releases)

You can download the engine files from the official llama.cpp releases: 🔗 llama.cpp releases

Public changes

New features

• Extended llama.cpp version support: the library is now compatible with versions from b9460 through the latest releases.
• Added unsupported operation checks in LlamaExecutor for recurrent and hybrid models. Methods that are incompatible with these architectures (DeleteSequenceEnd and StopSeqPrefill) now throw an exception instructing the user to use state checkpoints (not yet implemented).

Docs

• Fixed a comment for the Sample method in ISamplingPipeline.

Other

Tests

• Updated llama.cpp paths and model paths in integration tests and example test programs.

Llama.csharp-v1.1.0

This release includes the full source code along with a ready-to-use archive of the library subproject:

📦 llama.csharp-v1.1.0.zip

Supported llama.cpp versions
b7551 – b7667 (Qwen3 and earlier models)

You can download the engine files from the official llama.cpp releases:

🔗 llama.cpp releases (lib tested on b7552 and b7667)

Public changes

New features

• StopSeqPrefill method in LlamaExecutor for aborting active prefills (e.g., to cancel a prefill that is no longer needed).

Fixed

• DeleteSequence now stops active prefill or generation then removing the sequence.
• Fast empty sequence validation in Generate – throws InvalidOperationException instead of internal decoder exception.
• The integration_test source code no longer contains the engine files

Docs

• Added partial documentation for the public methods.

Other

Refactore

• endInference / endPrefill refactored to accept a Sequence instead of a tuple.

Tests

• Added integration tests for deletion during prefill/generation.
• Added integration tests for StopSeqPrefill revert scenarios.
• Added test for generating on an empty sequence.

llama.csharp-v1.0.0

First Release

This release includes the full source code along with a ready-to-use archive of the library subproject:

📦 llama.csharp-v1.0.0.zip

Supported llama.cpp versions

b7551 – b7667 (Qwen3 and earlier models)

You can download the engine files from the official llama.cpp releases:

🔗 llama.cpp releases (lib tested on b7552 and b7667)

The source code here contains Vulkan-enabled llama.cpp engine files in the integration test project, so it weighs more. Their tracking has been removed and future versions will not contain them.