- Async loading supports cancellation through a `CancellationToken`. If loading is cancelled an `OperationCanceledException` is thrown. If it fails for another reason a `LoadWeightsFailedException` is thrown.
- Updated examples to use `LoadFromFileAsync`
* Updated binaries, using [this build](https://github.com/SciSharp/LLamaSharp/actions/runs/8654672719/job/23733195669) for llama.cpp commit `f7001ccc5aa359fcf41bba19d1c99c3d25c9bcc7`.
- Added all new functions.
- Moved some functions (e.g. `SafeLlamaModelHandle` specific functions) into `SafeLlamaModelHandle.cs`
- Exposed tokens on `SafeLlamaModelHandle` and `LLamaWeights` through a `Tokens` property. As new special tokens are added in the future they can be added here.
- Changed all token properties to return nullable tokens, to handle some models not having some tokens.
- Fixed `DefaultSamplingPipeline` to handle no newline token in some models.
* Moved native methods to more specific locations.
- Context specific things have been moved into `SafeLLamaContextHandle.cs` and made private - they're exposed through C# properties and methods already.
- Checking that GPU layer count is zero if GPU offload is not supported.
- Moved methods for creating default structs (`llama_model_quantize_default_params` and `llama_context_default_params`) into relevant structs.
* Removed exception if `GpuLayerCount > 0` when GPU is not supported.
* - Added low level wrapper methods for new per-sequence state load/save in `SafeLLamaContextHandle`
- Added high level wrapper methods (save/load with `State` object or memory mapped file) in `LLamaContext`
- Moved native methods for per-sequence state load/save into `SafeLLamaContextHandle`
* Added update and defrag methods for KV cache in `SafeLLamaContextHandle`
* Updated submodule to `f7001ccc5aa359fcf41bba19d1c99c3d25c9bcc7`
* Passing the sequence ID when saving a single sequence state
* - Added `NativeLogConfig` which allows overriding the llama.cpp log callback
- Delaying binding of this into llama.cpp until after `NativeLibraryConfig` has loaded
* Using the log callback to show loading log messages during loading.
* Registering log callbacks before any calls to llama.cpp except `llama_empty_call`, this is specifically selected to be a method that does nothing and is just there for triggering DLL loading.
* - Removed much of the complexity of logging from `NativeApi.Load`. It always call whatever log callbacks you have registered.
- Removed alternative path for `ILogger` in NativeLibraryConfig, instead it redirects to wrapping it in a delegate.
* Saving a GC handle to keep the log callback alive
* Removed prefix, logger should already do that.
* Buffering up messages until a newline is encountered before passing log message to ILogger.
* - Added trailing `\n` to log messages from loading.
- Using `ThreadLocal<StringBuilder>` to ensure messages from separate threads don't get mixed together.
This commit is originally made by lcarrere in https://github.com/SciSharp/LLamaSharp/issues/180 .
I have confirmed this modification is OK in my windows 11 laptop, add make this commit according require of AsakusaRinne.
- Modified library loading to be based on `SetDllImportResolver`. This replaces the built in loading system and ensures there can't be two libraries loaded at once.
- llava and llama are loaded separately, as needed.
- All the previous loading logic is still used, within the `SetDllImportResolver`
- Split out CUDA, AVX and MacOS paths to separate helper methods.
- `Description` now specifies if it is for `llama` or `llava`
* Add llava_binaries, update all binaries to make the test
* Llava API + LlavaTest
Preliminary
* First prototype of Load + Unit Test
* Temporary run test con branch LlavaAPI
* Disable Embed test to review the rest of the test
* Restore Embedding test
* Use BatchThread to eval image embeddings
Test Threads default value to ensure it doesn´t produce problems.
* Rename test file
* Update action versions
* Test only one method, no release embeddings
* Revert "Test only one method, no release embeddings"
This reverts commit 264e176dccc9cd0be318b800ae5e102a4635d01c.
* Correct API call
* Only test llava related functionality
* Cuda and Cblast binaries
* Restore build policy
* Changes related with code review
* Add SafeHandles
* Set overwrite to upload-artifact@v4
* Revert to upload-artifact@v3
* revert to upload-artifact@v3
* Added a lock object into `SafeLlamaModelHandle` which all calls to `llama_decode` (in the `SafeLLamaContextHandle`) lock first. This prevents two contexts from running inference on the same model at the same time, which seems to be unsafe in llama.cpp.
* Modified the lock to be global over _all_ inferences. This seems to be necessary (at least with the CUDA backend).
Modified LLamaBatch to not share tokens with other sequences if logits is true. This ensures that the logit span at the end in used by exactly one sequence - therefore it's safe to mutate. This removes the need for copying _very_ large arrays (vocab size) and simplifies sampling pipelines.
* Added a `Guidance` method to `LLamaTokenDataArray` which applies classifier free guidance
* Factored out a safer `llama_sample_apply_guidance` method based on spans
* Created a guided sampling demo using the batched executor
* fixed comment, "classifier free" not "context free"
* Rebased onto master and fixed breakage due to changes in `BaseSamplingPipeline`
* Asking user for guidance weight
* Progress bar in batched fork demo
* Improved fork example (using tree display)
* Added proper disposal of resources in batched examples
* Added some more comments in BatchedExecutorGuidance
- Re-implmented `Rewind` as an extension method using `Modify` internally
- Implemented `ShiftLeft`, which shifts everything over except for some starting tokens. This is the same as the `StatelessExecutor` out-of-context handling.
- Starting batch at epoch 1, this ensures that conversations (starting at zero) are below the current epoch. It also means `0` can always be used as a value guaranteed to be below the current epoch.
Conversations can be "forked", to create a copy of a conversation at a given point. This allows e.g. prompting a conversation with a system prefix just once and then forking it again and again for each individual conversation. Conversations can also be "rewound" to an earlier state.
Added two new examples, demonstrating forking and rewinding.
This results in windows users not needing to rename the DLL. This allows native llama builds to be dropped in, even on windows.
I also took the time to update the documentation, removing references to renaming the files, since the names now match.
Fixes#463
- Added a `DecodeAsync` overload which runs the work in a task
- Replaced some `NativeHandle` usage in `BatchedDecoding` with higher level equivalents.
- Made the `LLamaBatch` grow when token capacity is exceeded, removing the need to manage token capacity externally.
ksanchez
Rinne