Few fixes.
Going back on remote metal-rs.
Reusing a single buffer (for now) to speed things up.
Adding some half kernels.
All tests are panicking instead of random failure.
Putting back f16 index select.
Add erf.
Working version for llama2-c.
Fixes + cache compute_pipeline_state.
BF16 metal fix.
Remove some prints.
new_owned -> new()..to_owned().
Better batched matmul.
Metal operational.
Reuse buffers on our own reference counts.
Tmp gemm.
Revert "Tmp gemm."
This reverts commit c65f68e988.
Interleave committing.
Speeding up copies using blit.
Fmt.
Fmt.
Remove the assert!
Fmt all.
Fixes after big rebase.
Add softmax for half and bfloat + tests
Fixing Llama example + accumulate softmax in float.
* Add support to UL2 model family
* Update docs with UL2
* Create ActivationWithOptionalGating to avoid polluting activations
* Also refactor quantized t5
* Remove useless conversion
* Revert Activation::NewGelu name change
* Remove useless return
* Apply rustfmt and clippy recommendations
* Reuse t5::ActivationWithOptionalGating in quantized version
* (cosmetic change) use a match rather than ifs + avoid early returns.
---------
Co-authored-by: Laurent <laurent.mazare@gmail.com>
* add bce with logit loss
* add bce with logit loss
* remove imports
* fix tiny bug
* add test documentation and refactor function
* fix test cases and formatting
* 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.
* Start processing images.
* Add LayerNorm2d.
* Properly use LayerNorm2d.
* Tweak eps.
* Use LayerNorm on inputs with a rank different from 3.
* Window partitioning.
* Fix a couple todos.
* More todos.
* Hard-code the einsums.
* More padding support.
* Some sizes tweaks.
* Use the hub to get the weights.
* Use a batch matmul.
* Tweaks.
* More fixes.
* Get some predictions to be generated.
* 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.
