You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[hy4 engine 0.1.64, Windows] strata-hy4.exe crashes with 0xC000001D (illegal instruction) in the AVX2-VNNI CPU expert kernel on the first chat; STRATA_FORCE_ISA=avx2 is a stable workaround #6570
strata-hy4.exe 0.1.64 (windows-x86_64-cuda) crashes on the first chat request with Windows exception 0xC000001D (illegal instruction) in the engine's own module. Every restart by the supervisor crashes again on the next request. The crash happens when the CPU expert pool computes its first routed-expert job (decode stage of a small prompt).
Forcing the plain AVX2 kernel makes the engine fully stable. The AVX2-VNNI kernel (kernel=fast(avx2-vnni)) is the trigger.
Environment
OS: Windows 11, PowerShell 5
CPU: Intel Core i7-12700 (Alder Lake, 8P+4E, AVX2 + AVX-VNNI, no AVX-512)
Crash occurred with default launch (drafting dfly2), with --drafting off, and with HY4_CUDA_GRAPH=0 - so drafting and CUDA graphs are ruled out.
Crash occurred again with STRATA_CPUEXP_KERNEL=avx2-vnni at the same offset 0x115ba0 even when the whole process was pinned to the P-cores only (affinity 0xFFFF, E-cores excluded) - so it is not an E-core/hybrid-scheduler ISA fault; the VNNI kernel code path itself faults on this CPU.
With STRATA_FORCE_ISA=avx2 (kernel=fast(avx2), act=hi) the engine is completely stable: many consecutive requests, no crash.
Hypothesis: in the MSVC build, the avx2-vnni kernel region (hy4_cpu_experts.cpp, k=4) may contain instructions beyond AVX2+AVX-VNNI (for example AVX-512 encodings leaking from a target-pragma/intrinsic miscompile), which fault on a non-AVX-512 CPU. The GCC/Clang path guards kernel regions with target pragmas; MSVC "needs no flags for intrinsics", so nothing stops the compiler from emitting wider ISA there.
Repro
Install hy4-flash SQ-96 on the machine above (any layout).
Start the server, send any chat request (small prompt, a few generated tokens).
~20-30 s in, the engine dies: "the engine stopped unexpectedly (exit code None)" first, then "exit code 3221225501" (0xC0000005 wrapper wording for the native fault; Event Log shows 0xc000001d).
This selects kernel=fast(avx2) / act=hi. Performance cost: roughly half the CPU-pool throughput versus avx2-vnni+q8 (decode ~8.5 tok/s on this box with dfly2 drafting enabled, ~62% draft acceptance).
Request
Please guard the MSVC avx2-vnni kernel region (and validate emitted ISA), or add an AVX-VNNI CPUID check per-instruction-set that matches what the kernel actually emits. A log hint at startup (e.g. "VNNI kernel disabled: not supported") would also have saved a long debugging session.
Thanks for the issue. To keep the tracker readable, every item needs an area prefix in the title and the parts of the issue form filled in. This one is missing:
Start the title with its area: [Engine], [GUI], [TUI], [Benchmark], [Docs] or [Other], then a short description. Example: "[Engine] decode stops after 2 minutes".
Edit the issue (title and description) and this note goes away on its own. If it is still missing parts in 5 days it will be closed, which is not a rejection: fix it and comment /reopen.
Activity
github-actions commentedon Nov 22, 2026
Thanks for the issue. To keep the tracker readable, every item needs an area prefix in the title and the parts of the issue form filled in. This one is missing:
Edit the issue (title and description) and this note goes away on its own. If it is still missing parts in 5 days it will be closed, which is not a rejection: fix it and comment
/reopen.Details: CONTRIBUTING.md