int-llm

QEMU cross-architecture validation

This is a supplementary portability gate, separate from the real-hardware records in validation/cpu/. It cross-compiles static Linux binaries and runs them with QEMU user-mode emulation. Emulated targets are never counted as physical boards or native host systems.

For every target, run_matrix.sh builds and runs:

1. fp_determinism.c with the compiler-selected backend;
2. the same gate with FP_MATH_FORCE_PORTABLE; and
3. the inference-only tiny GPT against the committed model.mgw.

The eight 64-bit targets whose compilers expose __int128 also compile, statically link, and start the full llama_int executable. The four 32-bit targets skip that build because llama_int.c still uses raw __int128 directly; this is a documented project boundary, not a failed emulation test.

Both determinism builds must reproduce c0d933ea340452ec. On little-endian targets, the tiny GPT must reproduce the exact 20-sample stdout SHA-256 f62a78d3e878b341cd59289a919ec4991b8055eaf0b4d715e2bc58bba2865d63. The .mgw v1 format is intentionally host-native-endian, so big-endian targets instead prove that the little-endian committed file is rejected cleanly rather than misread.

Matrix

target	width / endian	compiler-selected backend	tiny-GPT file check
PowerPC64LE	64-bit LE	native __int128	exact 20-sample output
RISC-V 64	64-bit LE	native __int128	exact 20-sample output
MIPS64EL	64-bit LE	native __int128	exact 20-sample output
Alpha	64-bit LE	native __int128	exact 20-sample output
i686	32-bit LE	portable two-limb	exact 20-sample output
SuperH-4	32-bit LE	portable two-limb	exact 20-sample output
IBM s390x	64-bit BE	native __int128	expected endian rejection
PowerPC64	64-bit BE	native __int128	expected endian rejection
SPARC64	64-bit BE	native __int128	expected endian rejection
MIPS64	64-bit BE	native __int128	expected endian rejection
PA-RISC	32-bit BE	portable two-limb	expected endian rejection
Motorola 68k	32-bit BE	portable two-limb	expected endian rejection

That is 12 ABI/endian configurations across 10 ISA families, including four 32-bit targets and six big-endian targets. The first recorded run passed all 12; see results/2026-07-30-qemu-user-matrix.txt.

Running it

On Ubuntu 24.04, install QEMU user mode and the selected cross-compilers:

sudo apt-get install qemu-user \
  gcc-powerpc64le-linux-gnu gcc-riscv64-linux-gnu \
  gcc-mips64el-linux-gnuabi64 gcc-alpha-linux-gnu \
  gcc-i686-linux-gnu gcc-sh4-linux-gnu gcc-s390x-linux-gnu \
  gcc-powerpc64-linux-gnu gcc-sparc64-linux-gnu \
  gcc-mips64-linux-gnuabi64 gcc-hppa-linux-gnu gcc-m68k-linux-gnu

Review the package-manager plan first: on Ubuntu, installing these cross toolchains can remove the gcc-multilib and g++-multilib meta-packages.

Then run the whole matrix, or a named subset:

validation/qemu/run_matrix.sh
TARGETS="ppc64le s390x i686" validation/qemu/run_matrix.sh

The runner uses an isolated mktemp directory, compiles with warnings as errors, pins all source/model hashes in its output, and removes its temporary artifacts on exit.

This gate deliberately does not decode the 8.2 GiB TinyLlama model under emulation. Full TinyLlama decode under QEMU TCG would mostly measure emulator and storage overhead; the meaningful full-model evidence remains the native real-host 80-token gate. The QEMU matrix targets arithmetic, compiler/ABI, word-size, endianness, serialization rejection, small-model execution, and warning-free full-runtime build/startup coverage where __int128 exists.

This site is open source. Improve this page.