* Boilerplate for the quantized cuda support.
* More basic cuda support.
* More cuda quantization (quantize on cpu for now).
* Add the dequantization bit.
* Start adding some dedicated cuda kernels from llama.cpp.
* Move the kernel code.
* Start interfacing with the kernel.
* Tweak the kernel launch params.
* Bugfix for quantized metal.
* Fix some clippy lints.
* Tweak the launch parameters.
* Tweak cuda basics to perform a quantized matmul.
* Perform the dequantization on the cpu + use cublas for matmul.
* Add the dequantization kernel.
* Test the qmatmul.
* More kernels.
* Matmul-vec kernel.
* Add a couple kernels.
* More dequantization kernels.
* Metal quantized modifications proposal.
- Add a device param, wherever needed.
- Create new QMetal storage thing that implements QuantizedType.
- Update everywhere needed.
Fix Python.
Fixing examples.
Fix: fmt + clippy + stub.
Moving everything around.
Only missing the actual implems.
Fixing everything + adding dequantized kernels.
More work.
Fixing matmul.
Fmt + Clippy
Some clippy fixes.
Working state.
Q2K Metal -> Bugged (also present in GGML).
Q4K CPU -> Bugged (present previously, new test catch it).
Q5K CPU -> Bugged (present previously).
Q8_1 Both -> Never really implemented it seems
Q8K metal -> Never implemented in metal
Fixing Q2K bug (present in ggml).
* Cleanup.
* Fix the rebase.
* Removing the fences speeds everything up and *is* correct this time...
* Cleanup the fence.
* After rebase.
* Bad code removal.
* Rebase after phi2 merge + fix replit default to CPU.
* Making the CI happy.
* More happy tests.
---------
Co-authored-by: Nicolas Patry <nicolas@Nicolass-MacBook-Pro.local>
* Quantized version of mistral.
* Integrate the quantized mistral variant.
* Use the quantized weight files.
* Tweak the quantization command.
* Fix the dtype when computing the rotary embeddings.
* Update the readme with the quantized version.
* Fix the decoding of the remaining tokens.
* Use yoke to provide a self-referential container for mmaped safetensor files.
* Add the new self-owned type for safetensor files without removing the previous version.
* Add routing.
* Add an initializer for the case of multiple files.
* Load gguf files for the quantized t5.
* Add the quantized t5 example.
* Allow for loading local files.
* Add some support for quantizing safetensor files.
* Transpose before quantizing.
* Quantized t5.
* Retrieve the weights from the hub.
* Add a custom softmax implementation.
* Add softmaxlastdim to the benchmarks.
* And add a test.
* Support more dtypes.
* Polish the code.
* Use the slow implementation on cuda.
* Add a todo for the cuda kernel.
* Add the dilation parameter.
* Restore the basic optimizer example.
* Dilation support in cudnn.
* Use the dilation parameter in the cpu backend.
* More dilation support.
* No support for dilation in transposed convolutions.
* Add dilation to a test.
* Remove a print.
* Helper function.
* Add to the cuda example a reproduction of the issue.
* Tweak.
* Add a test using non-square matrixes.
* Fix the conv2d kernel.
* Display the error.
* And tweak the comment.
* Add some group parameter to convolutions.
* Avoid some unnecessary groups checks.
* Move the tensor convolution bits.
* Properh handling of groups.
* Bump the crate version.
* And add a changelog.
* Pickle work-in-progress.
* More unpickling.
* More pickling.
* Proper handling of setitems.
* Clippy.
* Again more pickling.
* Restore the example.
* Add enough pickle support to get the list of tensors.
* Read the data from zip files.
* Retrieve the tensor shape.
* Extract the size and dtype.
* More storage types.
* Improve the destructuring.
* Also support ggml files.
* Pickle work-in-progress.
* More unpickling.
* More pickling.
* Proper handling of setitems.
* Clippy.
* Again more pickling.
* Restore the example.
* Add enough pickle support to get the list of tensors.
* Read the data from zip files.
* Retrieve the tensor shape.
* Extract the size and dtype.
* More storage types.
* Improve the destructuring.
* Add a cudnn feature to be used for conv2d.
* Allocate the proper workspace.
* Only create a single cudnn handle per cuda device.
* Proper cudnn usage.
* Bugfix.
* Add more tracing to the whisper example.
* Support accelerate in more examples.
* Use accelerate for pointwise functions.
* Use accelerate for binary operations too.
* Bugfix for binary operation: use the rhs before the lhs.
* Sketch a fast cuda kernel for reduce-sum.
* Sketch the rust support code for the fast sum kernel.
* More work on the fast kernel.
* Add some testing ground.
* A couple fixes for the fast sum kernel.
* Fix some rebase issues.
* Use mkl instead.
* Use mkl in bert.
* Add the optional mkl feature.
* Conditional compilation based on the mkl feature.
* Add more mkl support.
Laurent Mazare