This version is still a work in progress.
Highlights
- Code that performs many implicit conversions, most visibly large collection literals, compiles faster: the compiler no longer runs parameter inference on constructors that cannot be used for an implicit conversion in the first place. Files that are mostly data, such as the standard library's Unicode lookup tables, compile about 1.3x faster.
Language changes
Renamed the
@parameter
decorator on parametric closures to@__parameter
. The deprecated@parameter if
/@parameter for
forms are unchanged; prefercomptime if
/comptime for
for compile-time control flow. - The module & package system:
Directories may now have "namespace" semantics; a single directory name may resolve across distinct locations on disk which share that name.
Importing functions with the same name from different modules, combining them into one overload set, is now an error, following a period of deprecation.
Intra-package accesses without explicit
import
s are now an error, following a period of deprecation.
Library changes
StringDict
now conforms toWritable
when its value type isWritable
, matching the existing behavior ofDict
. This lets youprint()
aStringDict
or convert it to aString
. - The
chars
argument ofstrip()
,lstrip()
andrstrip()
onStringSpan
,String
andStringLiteral
is now anImmStringSpan
, so a mutable string is accepted aschars
, including the string being stripped (s.strip(s)
). -
StringDict.__getitem__()
now accepts aStringSpan
, so you can index aStringDict
with a borrowed string view without first allocating aString
just to perform the lookup. - Renamed the variadic type-list parameter on
Tuple
andVariadicPack
toTs
, standardizing the naming convention used across the standard library. The old name,element_types
, remains as a deprecated alias. - Added experimental
DType.float6_e2m3fn
andDType.float6_e3m2fn
, the two 6-bit encodings from theOpen Compute microscaling specification. Both are finite-only, so neither has an inf nor a NaN encoding.These are experimental storage formats for packed weights rather than general-purpose numeric types, and standard library support is deliberately partial. As with the existing
DType.float4_e2m1fn
, they are excluded fromis_numeric()
, arithmetic is not implemented, and converting to or from another floating-point type is unsupported on every target, so values cannot be printed either. -
Array
now conforms toDefaultable
when its typeT
is alsoDefaultable
. - Deprecated
is_trivially_movable()
,is_trivially_copyable()
, andis_trivially_deletable()
instd.memory
in favor ofIsTriviallyMovable[T]
,IsTriviallyCopyable[T]
, andIsTriviallyDeinitable[T]
instd.traits
. The replacements arecomptime
predicates rather than functions, so drop the call parens at use sites, for exampleIsTriviallyCopyable[T]
instead ofis_trivially_copyable[T]()
. - Renamed
UnsafeMaybeUninit
toMaybeUninit
. It conforms toMovable
,Copyable
/ImplicitlyCopyable
, andDeinitable
only when the contained type's own move, copy, or implicit deinitializer is trivial, since moving, copying, or destroying aMaybeUninit
only touches its raw bits, never the contained value's own lifecycle methods. Gating conformance this way turns what would otherwise be silent memory-safety bugs into compile-time errors. -
Atomic
is now parameterized on a value typeT
instead of aDType
. Update call sites fromAtomic[DType.float32]
toAtomic[Float32]
. The atomic operations (load()
,store()
,fetch_add()
,compare_exchange()
, and so on) still only supportScalar
types. - Added
Pointer[T].unsafe_write(def() -> T)
, which initializes the pointee with the value returned by a closure, constructing it directly in place rather than moving an already-constructed value there. Unlikeunsafe_write(var T)
, this does not require the pointee type to beMovable
. -
Pointer.mut_cast
is now deprecated. Developers should prefer using explicit mutabilites at the callsite viaMutPointer
orImmPointer
. If mut casting is needed (it should try to be avoided) - you can useunsafe_mut_cast
. - The following APIs have been migrated to unified closures:
sort
,debug_assert
,Span.apply
. - Uncaught exceptions now print to
stderr
, notstdout
.
Tooling changes
mojo doc
now reports the condition of a conditional trait conformance, and the generated API docs show it alongside the trait. Previously the condition was dropped, making a conditional conformance indistinguishable from an unconditional one. Also fixed rendering of somewhere
clauses.
Removed
This release completes the removal of APIs deprecated during the v1.0 cycle.
Removed the temporary
InlineArray
alias forArray
, including its re-exports fromstd.collections
and the prelude. UseArray
directly. - Removed the
std.gpu.profiler
module and itsProfileBlock
context manager. It timed host wall-clock, not GPU work, and reported the elapsed time with the operands reversed. Time a block of host code withdirectly, and use a GPU profiler such as Nsight Systems orperf_counter_ns()
rocprof
for device timings. - Removed
memcmp
and itsstd.memory
re-export. Useunsafe_memcmp
instead. - Removed the
validate
parameter from, which now always validates. Passingb64decode()
validate=False
did not skip any work on valid input; it only turned characters outside the base64 alphabet into silently corrupt output bytes. Drop[validate=True]
from existing calls; calls that relied on the default now raise instead of returning garbage. - Removed the origin aliases left over from the
Immut
toImm
andExternal
toUntracked
renames. Use the surviving spelling in each case:ImmOrigin
forImmutOrigin
,ImmUnsafeAnyOrigin
forImmutUnsafeAnyOrigin
,ImmStaticOrigin
forStaticConstantOrigin
,UntrackedOrigin
forExternalOrigin
,MutUntrackedOrigin
forMutExternalOrigin
, andImmUntrackedOrigin
for bothImmutUntrackedOrigin
andImmutExternalOrigin
. - Removed the pre-unification pointer aliases
MutUnsafePointer
,ImmUnsafePointer
,ImmutUnsafePointer
,ImmutOpaquePointer
,ImmutPointer
, andOptionalUnsafePointer
. UseMutPointer
,ImmPointer
,ImmOpaquePointer
, andOptionalPointer
instead.UnsafePointer
itself remains available, but is deprecated in favor ofPointer
. - Removed the raw memory functions superseded by their
unsafe_
-prefixed spellings:memcpy
,memset
,memset_zero
,uninit_move_n
,uninit_copy_n
, anddestroy_n
. Useunsafe_memcpy
,unsafe_memset
,unsafe_memset_zero
,unsafe_uninit_move_n
,unsafe_uninit_copy_n
, andunsafe_destroy_n
instead. - Removed the
size
aliases left from thesize
tolength
rename:SIMD.size
,Array.size
,TypeList.size
, and theSIMDSize
alias forSIMDLength
. Uselength
andSIMDLength
. - Removed the
as_immutable()
andget_immutable()
methods onPointer
,Span
, andStringSpan
. Useas_imm()
. - Removed the
ImmutSpan
alias. UseImmSpan
. - Removed
String.as_string_slice()
. Construct aStringSpan
from the string instead:StringSpan(my_string)
. - Removed the
ImplicitlyDestructible
andImplicitlyDeletable
aliases. UseDeinitable
. - Removed the deprecated ownership-transfer methods:
List.steal_data()
andOwnedPointer.steal_data()
are nowunsafe_take_allocation()
,OwnedPointer.take()
isinto_inner()
, andVariant.take()
andVariant.unsafe_take()
areunwrap()
andunsafe_unwrap()
. - Removed the
Pointer
methods superseded by theirunsafe_
-prefixed spellings:as_noalias_ptr()
,destroy_pointee()
,destroy_pointee_with()
,init_pointee_move()
,init_pointee_copy()
, andinit_pointee_move_from()
. Useunsafe_as_noalias()
,unsafe_deinit_pointee()
,unsafe_deinit_pointee_with()
,unsafe_write()
, andunsafe_write_move_from()
. ThePointer.type
alias forPointer.T
is gone as well. - Removed the
ConditionalType
type function and thestd.utils.type_functions
module. Use the ternary expressionT if cond else U
. - Removed
trait_downcast()
. Constrain on the trait instead, withconforms_to(type_of(src), Trait)
in awhere
clause or acomptime assert
. - Removed the parametric
benchmark.run[func]()
overloads. Pass the function as an argument torun(f)
instead, which accepts a unified closure. - Removed
AnyCoroutine
,Coroutine
andRaisingCoroutine
from the prelude, and made the module that defines them private. Mojo's async support is unfinished, and these types being globally visible led people to build on an API that carries no stability guarantees.async def
is unaffected: the compiler still synthesizes these types for you, so they continue to appear in inferred types and diagnostics. There is no supported way to name them directly. - Removed the async task API from the public
std.runtime.asyncrt
module, which is now private.initialize_runtime()
andparallelism_level()
are unaffected and have moved up to thestd.runtime
package, so import them fromstd.runtime
instead ofstd.runtime.asyncrt
. - Removed support for
.mojopkg
files after a period of deprecation. Use.mojoc
files instead.
Fixed
mojo build --emit asm
and--emit llvm
now always write the offload kernel files next to the host output file. Building a kernel that an earlier build had already compiled could write them into the earlier build's output directory, or skip them with no diagnostic. - Parametric
raises
now accepts any primary expression as the thrown type in a function signature, matching the syntax positions where types otherwise appear. This most notably fixesraises Self.SomeAssocType
on trait and struct methods, which would previously fail with an error. The parenthesized workaround (raises (Self.DriveErrorType)
) is no longer required. - An integer
range()
with a step of zero is now always empty. It previously used to be an infinite loop - iterating forever at runtime, and hanging the compiler at comptime. - Fixed
ceildiv()
returning0
for unsigned operands near the type's maximum value. The unsigned code path computednumerator + denominator - 1
, which overflows and wraps for large operands; it now derives the ceiling from the floor division and remainder instead. -
Counter.most_common(n)
now returns all elements whenn
exceeds the number of unique elements, matching Python, instead of aborting. -
os.path.join()
now inserts separators based on the accumulated path rather than the first argument, sojoin("/", "a", "b")
returns/a/b
(previously/ab
) andjoin("a", "b/", "c")
returnsa/b/c
(previouslya/b//c
). -
base64.b64decode()
now raises an error when the input length is not divisible by 4 instead of reading past the end of the input (or aborting when asserts are enabled). - On macOS,
os.stat()
andos.lstat()
no longer return a negativest_mode
for regular files. The underlyingmode_t
andnlink_t
C type aliases were declared as signed 16-bit integers, but macOS defines them as unsigned, so any mode with theS_IFREG
bit set (every regular file) sign-extended into a negativeInt
. -
PythonObject
no longer leaks a CPython reference per positional argument when calling a Python object, nor when setting an item, attribute, or set literal element.