v0.2.2: New `OptimizationBarrier` op, `pjrt.IsCPU()`

• StableHLO: added OptimizationBarrier() op.
• For Value.OpName(): return "InputParameter" (stablehlo.InputParameterName) for input parameters instead of "nil".
• Added pjrt.Plugin.IsCPU() method.
• Internal: fixed order of implemented ops in OpType enum.

v0.2.1 Small fix to NVidia drivers path

Merge pull request #29 from gomlx/fixed-nvidia-cuda-path

- Fixed CUDA library paths -- it was hardcoded to a test directory.

v0.2.0: Dynamic Shapes; added `Call()`; Quantized shapes;

• go.mod: updated to Go 1.25: needed for synctest.
• Fixed sub-byte types Int4, Int2, Uint4, Uint2 to assume packed storage; Added byte (Uint8) to Int4 bitcast test.

StableHLO

• Added Call() op (thx @ajroetker).
• Added Value.OpName() for introspection.
• Dynamic Shapes (thx @ajroetker) -- with and without dynamic bounds.
  • Added DimensionBounds and EncodeBounds fields to Shape struct
  • Added ops: DynamicReshape, DynamicBroadcastInDim, DynamicIota, DynamicGather,
    DynamicPad, DynamicConv.
• Quantization:
  • Add Quantization field to shapes.Shape.
  • Add i2, i4, ui2 and ui4 DTypes.
  • Add UniformQuantize() and UniformDequantize() ops.
  • Add Value.WithOutputElementType() to allow change of quantization parameters for operations.
• Functions:
  • Ops choose the "innermost" scope of their operands -- meaning they are added the closure functions if they are operating on a value that is local to the closure.

PJRT

• Updated DefaultCPUVersion to "v0.98.0" (pjrt-cpu-binaries version)
• Added Client.IsCUDA() method.
• Installation:
  • prevent race condition of concurrent (auto-)installations (using flock, so it works cross-processes).
• Bumped XLA's pjrt_c_api.h to version 0.98
• Added Buffer.Bitcast.

v0.1.4

• Updated installer library/cli to support linux/arm64 and windows/amd64.
  Generalized CPU installation.
• linux and amazonlinux now use the same PJRT binary, built on a glibc-2.35 system (Ubuntu 22.04).
• Updated dependency to pjrt-cpu-binaries v0.83.3.

v0.1.3

• Replaced GenPool by the internal/pool.Pool: it simplifies a bit, a bit faster and one less
  dependency.

v0.1.2 prjt_installer in its own module

• Removed external dependency to github.com/charmbracelet/huh/spinner from pkg/installer.
• Split cmd/pjrt_installer into its own module (its own go.mod), to limit default go-xla "apparent" dependencies.
• Package pkg/installer:
  • Spinner now limited to 1 line (truncates line to fit)
• Fixed GitHub CI tests: added missing GH_TOKEN.

v0.1.1 Removed spurious debug message

Fixed spurious debug message.

go-xla = gomlx/gopjrt+gomlx/stablehlo; First release, updated docs.

v0.1.0 Merge stablehlo and pjrt into go-xla.

• Merge stablehlo and pjrt into go-xla.
  • Changed install directory to .../lib/go-xla (instead of .../lib/gomlx/pjrt).
  • Removed deprecated xlabuilder.
  • Split PJRT CPU binary releases into https://github.com/gomlx/pjrt-cpu-binaries
• Installation scripts also exported as a library.
• Added "installer.AutoInstall".
• Improvements:
  • Fixed memory leaks on plugin destruction.
  • Replaced sync.Pool by GenPool: this avoids frequent unnecessary freeing of arenas.
• Align with Google style:
  • Removed github.com/janpfeifer/must dependencies
  • Removed testify dependencies