We already have a pinned allocation for the original value, it should be finalized to avoid re-allocation causing multiple pinned allocations for the same buffer.
PiperOrigin-RevId: 820015337
PjRt-IFRT now returns a `nullptr` if it knows that the Array layout represents a default layout. The user code previously has been migrated to handle this new behavior gracefully, obtaining a concrete default layout as before.
`ifrt::PjRtArray` creation now request extra information on whether the underlying `PjRtBuffer` is using a custom layout as IFRT tracks the defaultness of array layouts. This information cannot be inferred correctly from `PjRtBuffer` alone because `PjRtBuffer::layout()` only returns a concrete layout. PjRt would mostly work fine today if a default layout is said to be a custom layout, but some strict layout equality check can fail and require more precise information to be supplied.
A few test cases in IFRT ArrayImplTest against PjRt CPU and GPU clients
have been disabled because the output array does not track the
non-default-ness of the layout correctly when
`MakeArraysFromHostBufferShards()` is implemented using
`ClientMakeArraysFromHostBufferShards()`.
PiperOrigin-RevId: 819995407
Add a placeholder for `HERMETIC_PYTHON_VERSION_KIND` in the generated `py_version.bzl` file. This new variable is currently set to an empty string until we figure out how to deal with it.
PiperOrigin-RevId: 819956767
register.cc already declares support for versions 1-7 of transpose but this seems like it was previously missed for register_ref.
PiperOrigin-RevId: 819915644
This was observed when auto warp specialization was enabled. Triton requires more threads per block than expected, and this information is available in the module attributes.
PiperOrigin-RevId: 819893926
The only non-obvious part of the thunk is `solver_context_creator`, but we can retrieve it during the deserialization from `stream_executor::Platform`, which is available during runtime.
PiperOrigin-RevId: 819863398
This change introduces a new `kTfLiteInt2` type to the TFLite schema and MLIR converter. It includes:
- Adding `INT2` to the flatbuffer schema.
- Mapping `TensorType_INT2` to `kTfLiteInt2` in flatbuffer conversions.
- Updating `tflite_types.h` to include `kTfLiteInt2`.
- Modifying `flatbuffer_export.cc` to handle 2-bit integer types from MLIR and pack them densely.
- Generalizing low-bit utility functions (`PackLowBitValuesDensely`, `UnpackDenseLowBitIntoInt8`) to support both 2-bit and 4-bit values.
- Updating type conversion utilities to recognize and handle `kTfLiteInt2`.
- Adjusting `util.cc` to correctly report the size and byte requirements for `kTfLiteInt2` tensors, considering their dense packing.
PiperOrigin-RevId: 819821231
Imported from GitHub PR https://github.com/openxla/xla/pull/32003
Copybara import of the project:
--
2a3ad034522e871edc9c7f580e86fc3980025542 by Ilia Sergachev <isergachev@nvidia.com>:
[GPU][NFC] Merge methods querying fusion kind.
--
ebeb25599d6017d34ea92ece415a255d109af049 by Ilia Sergachev <isergachev@nvidia.com>:
Address review requests.
Merging this change closes#32003
PiperOrigin-RevId: 819692807
Imported from GitHub PR https://github.com/openxla/xla/pull/32283📝 Summary of Changes
Change misleading method name RocmComputeCapability::has_amd_matrix_core() to more suitable name has_amd_mat_acc_instructions() as gfx11xx do not have matrix cores, but support matrix acceleration instruction set known as WMMA.
🎯 Justification
RocmComputeCapability::has_amd_matrix_core() is misleading as gfx11xx do not have matrix cores but still support matrix acceleration instruction set - WMMA.
🚀 Kind of Contribution
♻️ Cleanup
@xla-rotation please review my changes.
Copybara import of the project:
--
23cf1ab79fdcc4ee2ee4996973dee2c103d2762a by Aleksa Arsic <aleksa.arsic@amd.com>:
Change misleading method name RocmComputeCapability::has_amd_matrix_core() to more suitable name has_amd_mat_acc_instructions() as gfx11xx do not have matrix cores, but support matrix acceleration instruction set known as WMMA.
Merging this change closes#32283
PiperOrigin-RevId: 819652238
