* Avoid copying the data on squeeze and unsqueeze.
* Fix the quantized llama example.
* Unrelated fix for the quantized stable-lm example on cuda.
* Fix for mamba on cuda (unrelated to the PR).
* first attempt
* progress
* integrate into metal backend
* finish and get test passing
* add other dtype support
* update transpose1d dtypes supported
* first pass at implementation of maxpool2d
* Add definitions for other dtypes
* add tests for other dtypes
* Cosmetic tweaks + re-enable maxpool2d tests for metal.
---------
Co-authored-by: Laurent <laurent.mazare@gmail.com>
* Add a specialized kernel for copy2d.
* Move the cat operations.
* Avoid transpositions in cat.
* Bugfix.
* Bugfix for the cuda kernel.
* Add a benchmark.
* Add more testing.
* Test fix.
* Faster kernel.
* Add the missing kernel.
* Tweak the test.
* Add a metal kernel.
* Fix for the metal kernel.
* Get the tests to pass on metal.
* Also use this opportunity to fix the metal kernel for ELU.
* Add some bf16 kernels.
* Clippy fixes.
* use_resource API misunderstood. It is not additive. Several usages must be bit-ORed together.
* The seeding was incorrect and used the address instead of the value of the passed in seed.
* Add a check that likely exhibits failure to update the seed between generation of random tensors.
* Buffer overrun, the length given to the std::ptr::copy call was in bytes, and not 32-bit units.
* By default seed the RNG with a time-based value, so that different runs may produce different output, just like the CPU engine.
Use device.set_seed if determinism is warranted.
* Revert "By default seed the RNG with a time-based value, so that different runs may produce different output, just like the CPU engine. Use device.set_seed if determinism is warranted."
This reverts commit d7302de9
Discussion in https://github.com/huggingface/candle/pull/1811#issuecomment-1983079119
* The Metal random kernel failed to set element N/2 of tensors with N elements, N being even. The reason was that all threads but thread 0 all created 2 random samples, but thread 0 only one, i.e. an odd number. In order to produce an even number of samples, the early termination of thread 0 should only everr occur for odd sized tensors.
* Add a test catching any deterministic tensor element in rand and randn output.
---------
Co-authored-by: niklas <niklas@appli.se>
Co-authored-by: Ivar Flakstad <69173633+ivarflakstad@users.noreply.github.com>
* Improve metal buffer usage
* Clone cpu storage when loading to reduce wait_until_complete calls
* Use powers of two for buffer sizes so reuse is more likely.
* Select best available buffer by size.
* Add count to MetalStorage -> can use buffer with different size
Co-authored-by: Chris Fleetwood <christopher.fleetwood@huggingface.co>
* Simplify new buffer creation without blit copy. Revert &[] -> Vec
* Add documentation on newBufferWithBytes safety / synchronization
* Drop unused buffers after command buffer is done syncing.
---------
Co-authored-by: Chris Fleetwood <christopher.fleetwood@huggingface.co>
* 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.
