View Source Changelog

v0.6.4 (2023-11-13)

Enhancements

  • [Nx] Allow non-scalar tensors on access

Bug fixes

  • [Nx] Improve the :axes option in gather, indexed_add, and indexed_put
  • [Nx] Fix grad of gather, indexed_add, and indexed_put with axes
  • [Nx.BinaryBackend] Fix sorting of negative infinity
  • [Nx.BinaryBackend] Always sort NaN last
  • [Nx.Serving] Fix Nx.Batch padding with multi-device backends

v0.6.3 (2023-11-09)

Enhancements

  • [Nx] Allow non-scalars as updates on indexed_add and indexed_put
  • [Nx] Allow non-scalars as return of gather
  • [Nx] Support the :axes option in gather, indexed_add, and indexed_put
  • [Nx] Add Nx.covariance
  • [Nx] Support :type in argsort
  • [Nx] Support :stable option in argsort for future compatibility
  • [Nx.Serving] Add :weight option for static load balancing

Bug fixes

  • [Nx] Cast input types on slicing
  • [Nx.Defn] Support vectorized tensors in grad
  • [Nx.Defn] Fix bugs when diffing tensor expressions
  • [Nx.Serving] Handle serving getting stuck on timer messages

v0.6.2 (2023-09-21)

Enhancements

v0.6.1 (2023-09-12)

Enhancements

  • [Nx] Add multivariate normal distribution
  • [Nx.Serving] Automatically split exceeding batch sizes

Bug fixes

  • [Nx] Fix Nx.pad/2 with different backends
  • [Nx] Fix Nx.clip/3 with non-finite values
  • [Nx.Serving] Emit batches as they arrive in Nx.Serving.streaming/2
  • [Nx.Serving] Ensure batch key is preserved when a batch is split

v0.6.0 (2023-08-15)

Enhancements

Bug fixes

  • [Nx] Fix from_numpy with 1-byte width arrays
  • [Nx] Fix cases where pretty printing large Nx.Defn expressions would take a long time
  • [Nx] Fix reduce_min/reduce_max for non-finite values

Deprecations

  • [Nx.Serving] The post-processing function must now be a two-arity function that receives the {output, metadata} as a pair or the stream

Breaking changes

  • [Nx.Serving] The nx.serving.postprocessing telemetry event no longer receives the serving output or serving metadata as event metadata

v0.5.3 (2023-04-14)

Bug fixes

  • [Nx.Defn] Fix compilation error when Elixir compiler has column tracking enabled
  • [Nx.LinAlg] Fix cases where determinant could return NaN
  • [Nx.LinAlg] Fix SVD when working with f16 and bf16

v0.5.2 (2023-03-21)

Enhancements

  • [Nx.Random] Add stop_grad to Nx.Random creation functions
  • [Nx.Serving] Reduce references sent through serving

Bug fixes

  • [Nx] Fix Nx.mode with :axis option

v0.5.1 (2023-02-18)

Require Elixir v1.14.

Enhancements

v0.5.0 (2023-02-10)

Enhancements

  • [Nx] Support serialization of containers
  • [Nx] Rename Nx.power to Nx.pow
  • [Nx] Add Nx.reflect and Nx.linspace
  • [Nx.Defn] Raise at compile time for invalid defn if/cond usage
  • [Nx.LinAlg] Support full_matrices? in SVD
  • [Nx.LinAlg] Add Nx.LinAlg.matrix_rank
  • [Nx.Random] Add Nx.Random.choice and Nx.Random.shuffle
  • [Nx.Serving] Add distributed² serving by distributing over devices (GPUs/CPUs) as well as nodes
  • [Nx.Serving] Add telemetry to Nx.Serving callbacks

Backwards incompatible changes

  • [Nx] from_numpy and from_numpy_archive have been replaced by load_numpy! and load_numpy_archive!
  • [Nx.Defn.Evaluator] Do not force GC on evaluator

v0.4.2 (2023-01-13)

Enhancements

Bug fixes

  • [Nx] Respect fortran order in loading from numpy
  • [Nx.Defn] Render containers in compile error type+shape mismatch
  • [Nx.Defn] Restore pdict state after compilation

v0.4.1 (2022-12-07)

Enhancements

Bug fixes

  • [Nx] Keep input tensor names on associative scan
  • [Nx.BinaryBackend] Differentiate between complex and real output in as_type
  • [Nx.BinaryBackend] Fix loss of precision in Nx.complex/2
  • [Nx.BinaryBackend] Preserve NaNs in window and reduce operations
  • [Nx.Random] Do not return infinity on normal/2 for f16

v0.4.0 (2022-10-25)

Enhancements

  • [Nx] Add Nx.rename/2, Nx.median/2, Nx.weighted_mean/3, and Nx.mode/2
  • [Nx] Implement cumulative operations using associative scan for improved performance
  • [Nx.Constants] Add min and max
  • [Nx.Defn] Allow lists and functions anywhere as arguments in defn, jit and compile
  • [Nx.Defn] Add Nx.LazyContainer that allows a data-structure to lazily define tensors
  • [Nx.Defn] Allow tensors and ranges as generators inside while
  • [Nx.Defn] Add debug_expr/2 and debug_expr_apply/3
  • [Nx.Defn.Evaluator] Calculate cache lifetime to reduce memory usage on large numerical programs
  • [Nx.LinAlg] Handle Hermitian matrices in eigh
  • [Nx.LinAlg] Support batched operations in adjoint, cholesky, determinant, eigh, invert, lu, matrix_power, solve, svd, and triangular_solve
  • [Nx.Random] Support pseudo random number generators algorithms

Bug fixes

  • [Nx] Perform window_reduce/reduce operations from infinity and negative infinity
  • [Nx.Defn] Ensure defnp emits warnings when unused
  • [Nx.Defn] Warn on unused variables in while

Deprecations

  • [Nx] Deprecate tensor as shape in Nx.eye/2 and Nx.iota/2
  • [Nx] Deprecate Nx.random_uniform/2 and Nx.random_normal/2

v0.3.0 (2022-08-13)

Enhancements

Deprecations

  • [Nx] Nx.to_batched_list/3 is deprecated in favor of Nx.to_batched/3
  • [Nx.Defn] transform/2 is deprecated in favor of deftransform/2 and deftransformp/2
  • [Nx.Defn] assert_shape/2 and assert_shape_pattern/2 are deprecated in favor of case/2 + raise/2
  • [Nx.Defn] inspect_expr/1 and inspect_value/1 are deprecated in favor of print_expr/1 and print_value/1 respectively

v0.2.1 (2022-06-04)

Enhancements

  • [Nx] Improve support for non-finite values in Nx.tensor/1
  • [Nx] Use iovec on serialization to avoid copying binaries
  • [Nx.BinaryBackend] Improve for complex numbers in Nx.tensor/1
  • [Nx.Defn] Improve for complex numbers inside defn

Bug fixes

  • [Nx] Properly normalize type in Nx.from_binary/3
  • [Nx.Defn] Raise on Nx.Defn.Expr as JIT argument
  • [Nx.Defn.Evaluator] Handle concatenate arguments on evaluator

v0.2.0 (2022-04-28)

This version requires Elixir v1.13+.

Enhancements

v0.1.0 (2022-01-06)

First release.