●Change8

max/v26.5.0

Breaking Changes
📦 mojo-languageView on GitHub →
⚠ 4 breaking✨ 14 features⚡ 2 deprecations🔧 13 symbols

Summary

Mojo 1.0 is released, featuring stable standard library APIs, lambdas, pointer unification, and significant performance improvements. Most GPU programming APIs have moved to the new `max` accelerator library.

⚠️ Breaking Changes

  • GPU programming APIs have been moved from the Mojo standard library to the new `max` package. This includes `std.algorithm` to `max.algorithm`, `std.benchmark` to `max.benchmark`, `std.gpu.{compute,host,memory,sync}` to `max.gpu.*`, and the `layout` package. Update your imports accordingly.
  • Pointer unification may require changes to how pointers are handled, especially regarding interior origins.
  • Required `var` declarations: You may now need to explicitly declare variables with `var` where previously it might have been inferred or optional. Check compiler errors for guidance.
  • Constraint improvements might affect how generic types and functions are constrained. Review your generic code for compatibility.

Migration Steps

  1. Update imports from `std.algorithm` to `max.algorithm`.
  2. Update imports from `std.benchmark` to `max.benchmark`.
  3. Update imports from `std.gpu.{compute,host,memory,sync}` to `max.gpu.*`.
  4. Update imports for the `layout` package to its new location within `max`.
  5. Address compiler errors related to required `var` declarations.
  6. Review and potentially update code affected by pointer unification and interior origins.
  7. Review and potentially update code affected by constraint improvements.

✨ New Features

  • Mojo 1.0 release.
  • Standard library APIs are being marked as stable, ensuring backward compatibility for a growing set of APIs.
  • Introduction of lambdas.
  • Pointer unification.
  • Interior origins.
  • Required `var` declarations.
  • Constraint improvements.
  • Faster Python interop.
  • Extended Apple silicon GPU support back to M1 devices.
  • Added hardware-MMA flash-attention prefill for M5 to speed up TTFT.
  • New agent skills in the MAX accelerator library's skills repo: `serve-model`, `benchmark-model`, `eval-model`, and `profile-model`.
  • Expanded model coverage and capabilities in the MAX accelerator library.
  • Significant speedups for the Python API's experimental eager interpreter.
  • Speedups for various kernels across NVIDIA, AMD, and Apple GPUs.

Affected Symbols

std.algorithmmax.algorithmstd.benchmarkmax.benchmarkstd.gpu.computemax.gpu.computestd.gpu.hostmax.gpu.hoststd.gpu.memorymax.gpu.memorystd.gpu.syncmax.gpu.synclayout

⚡ Deprecations

  • Previous locations of GPU programming APIs in the standard library are deprecated.
  • Deprecated aliases may exist for breaking changes, accompanied by compiler fix-its for mechanical migration.