* fast rmsnorm
* no rms gpu
* kernel
* fix shared mem
* looped rms and donation in softmax
* Make the squaring in float32 to avoid underflow
* Fix the default StreamOrDevice for rope and rms_norm in fast
* nits
---------
Co-authored-by: Angelos Katharopoulos <a_katharopoulos@apple.com>
Without the && args would be copied and perfect forwarding won't work.
To avoid eval calling itself recursively, the vector version of eval is
changed to take by value instead, which will save a copy of array when a
rvalue is passed.
* Update mlx_set_item to handle regular slices without expanding
* Refactor ellipsis handling
* Route mlx_set_item to slice_update where possible
* Update mlx_scatter_args_slice
* Don't route to gather if no array indices
* Enable copy to work with int64 strides
* Fix uniform buffer indices or copy kernel arguments
* Update utils.h
* Remove manual unrolling of elem to loc loop
* GPU copy updated to handle negative strides
* Add slice update primitive
Since the shape and inputs are always saved as copy in ArrayDesc, we can
unify array's constructors to just take the arguments by value.
There are 2 cases:
1. When shape is a lvalue, it will be copied into array's constructor and
then moved into ArrayDesc's member. So only 1 copy happens.
2. When shape is a rvalue, it will be moved into array's constructor and
then moved into ArrayDesc's member. So no copy happens.
So having 1 constructor that takes by value is equivalent to having 2
constructors that const reference and rvalue separately.
* mostly builds
* most tests pass
* fix circle build
* add back buffer protocol
* includes
* fix for py38
* limit to cpu device
* include
* fix stubs
* move signatures for docs
* stubgen + docs fix
* doc for compiled function, comments
1. Anonymous namespace means internal linkage, static keyword is not needed.
2. The default constructor of std::shared_ptr initializes the pointer to
nullptr, you don't need to explicitly set it.
* Enable collapsing batch dims in gemm
* Update gemm to only make copies when neither of the last 2 axes are contiguous
* Update addmm to support gemv shapes
* Update addmm to support irregular batch strides
* Update tests
* Fast Inference SDPA op
Implements metal shaders for:
o = mx.fast_inference_sdpa(queries, keys, values, scale, mask)
Supports fp16, fp32 dtypes; assumes d_k = 128.
Generic op support / prompt encoding supported via mlx primitives.
Metal implementation is for the inference use case only.
Majority of performance benefits appears to results from GQA & reduced
bandwidth requirements; there is approximate performance parity for the
MHA use case (from some measurements on M3 Max).
* Flush shared memory to zero before unprotected reads for (scores @ values)
* Move to fast:: namespace, address reviewer comments
... also attempt to revert formatter auto-change for files not relevant
to this change
* Shared memory flush to top of kernel
* Resolve compiler warnings
* Update python/src/fast.cpp
Co-authored-by: Awni Hannun <awni.hannun@gmail.com>
* Update python/src/fast.cpp
Co-authored-by: Awni Hannun <awni.hannun@gmail.com>
* Update python/src/fast.cpp
Co-authored-by: Awni Hannun <awni.hannun@gmail.com>
* Update python/src/fast.cpp
Co-authored-by: Awni Hannun <awni.hannun@gmail.com>
* Update docstring per PR feedback
* Softmax in higher precision, ...
* route to fallback for more use cases - batch size > 1, head_dim other
than 128, etc.
* Address linux build failure
* Address other reviewer comments
* Remove extraneous eval_cpu function per review
---------
Co-authored-by: Atila Orhon <64497909+atiorh@users.noreply.github.com>
Co-authored-by: Awni Hannun <awni.hannun@gmail.com>
Co-authored-by: atila <atiorh@icloud.com>
* try to fix build for ios
* skip cpu compile
* fix namespace
* fix namespace
* Use CMake for platform specific cpu compile
---------
Co-authored-by: Angelos Katharopoulos <a_katharopoulos@apple.com>
* Split reduction files to reduce compile times
* Add small and medium axis size specializations for row reductions
* Add non-row-reduction options for small and med kernels
