[AntonOresten]

On Julia 1.14 (JuliaLang/julia#61359), a primitive type can declare any number of bits. Building on #36, this declares each microfloat with its real width there: Float4_E2M1FN is a 4-bit primitive type, Float6_E2M3FN is a 6-bit one, and so on. Each still occupies one byte. On earlier versions nothing changes, since the declared width falls back to 8.

What changes on 1.14

• Core.bitsizeof(T) == bitwidth(T). ===, hash and objectid ignore the unused high bits. Loads and stores mask them, so junk bits can no longer be stored.
• raw_bits/from_raw_bits use zext_int/trunc_int for sub-byte types, and keep bitcast (plus the Ignore bits above a microfloat's width #36 mask) for 8-bit types and older Julia.
• Scalar reinterpret(UInt8, x) and reinterpret(T, ::UInt8) are defined for all microfloats. Base's bitcast would otherwise reject a 4-bit ↔ 8-bit reinterpret, so existing scalar code (e.g. reinterpret(UInt8, Float4_E2M1FN(x)) in PTX tests) keeps working. reinterpret(T, 0xf3) drops the high bits.
• @microfloat declares primitive type T <: Microfloat N end via declared_bits(N), and its docstring and the Microfloat docstring describe both versions.

Breaking on 1.14

Array reinterpret of a sub-byte microfloat, e.g. reinterpret(UInt8, [Float4_E2M1FN(1)]), now throws ArgumentError, because Base refuses to reinterpret arrays whose elements have unused bits (Base.has_bit_padding). The fix is raw_bits.(xs) / from_raw_bits.(T, bytes), or packing through BitPacking. Since this breaks code that works today, this probably wants 0.3.0.

BitPacking 0.2.5 works as is, because it only scalar-reinterprets. MurrellGroup/BitPacking.jl#17 adds the native-width paths there too.

Tests

• stored(U, x) reads the bytes in memory through a pointer, since reinterpret can no longer do that for sub-byte types. The vectorization and "results fit the width" tests use it.
• New "Declared width" testset: sizeof == 1, Core.bitsizeof(T) == bitwidth(T) on 1.14, scalar reinterpret round trips including high-bit truncation, and @test_throws ArgumentError for the array case.
• Full suite with CUDA functional (GB10): 1.14.0-alpha1 83506 pass, 1.13 83486 pass, 1.10 83486 pass. It also passes against the BitPacking#17 branch.

Codegen

The generated machine code is identical to the 8-bit build except in the scalar narrowing loop, which gains one and 15 per element (152 → 160 instructions) from the store mask.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 87.24%. Comparing base (ac5234a) to head (84c1b83).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #37      +/-   ##
==========================================
+ Coverage   87.04%   87.24%   +0.19%     
==========================================
  Files          14       14              
  Lines         664      674      +10     
==========================================
+ Hits          578      588      +10     
  Misses         86       86              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.