fix(whisper): fall back to whisper-server builds for processors without AVX2

[greatcoat]

In short: on PCs whose processor has AVX but not AVX2 (Intel 2nd/3rd-gen Core, AMD FX, older AMD A-series/Athlon, AMD Jaguar), local Whisper never starts. The bundled engine stops on its first AVX2-era instruction, and the user sees "whisper-server process died during startup: exit code: 3221225501". With this PR the app switches to a build made for those processors and keeps using it, so transcription works. It depends on new builds from OpenWhispr/whisper.cpp#8, so it stays a draft until that release is pinned (see Remaining).

Refs #2356. Also reported in #1418, #1613 and #793.

Problem

• Windows: whisper-server-win32-x64.exe exits with 3221225501 = 0xC000001D = STATUS_ILLEGAL_INSTRUCTION. The reports come from an i3-3110M, an i5-3570 and an unnamed AVX-only CPU.
• Linux: the same crash as SIGILL (OpenWhispr local whisper-server crashes with SIGILL on AMD FX-8350 #793: AMD FX-8350 and A8-6410).
• Reproduced on emulated processors of exactly these generations (fork CI, run 36504535683):
  • Windows, Intel SDE -ivb: SDE-ERROR: Executed instruction not valid for specified chip (IVYBRIDGE): 0x7ff6bfc6072d: shlx eax, edx, eax. The address ends in c072d, the same fault offset whisper-server-win32-x64.exe crashes with STATUS_ILLEGAL_INSTRUCTION (0xc000001d) on non-AVX2 CPUs, blocking GPU fallback path #1613's Event Viewer reported.
  • Linux, QEMU -cpu IvyBridge: killed by SIGILL at startup.

Root cause

The CPU whisper-server comes from OpenWhispr/whisper.cpp. It is statically linked and configured with -DGGML_NATIVE=OFF, which in ggml turns AVX2, FMA, F16C and BMI2 on. The 0.0.10 CI log prints /arch:AVX2 GGML_AVX2;GGML_FMA;GGML_F16C;__BMI2__;GGML_BMI2.

A static build has no run-time CPU dispatch. It dies in ggml_cpu_init() on a BMI2 shlx, before the port binds.

The ggml-cpu-ivybridge.dll that reporters found in resources/bin belongs to llama-server's llama.cpp bundle. whisper-server imports no ggml DLL, so it never uses it.

Fix

whisper.cpp (OpenWhispr/whisper.cpp#8):

• Two more static CPU builds per Windows and Linux x64, at upstream ggml's ivybridge (SSE4.2 + AVX + F16C) and sandybridge (SSE4.2 + AVX) levels.
• A startup check, so each of those builds refuses to start on a processor below its level.
• CI that runs every build on an emulated processor of its level.

This PR:

1. whisperServer.js names the signal when the server dies at startup. A SIGILL used to leave no detail at all. It uses the same names as the exit-signal capture in fix(whisper): save and show why GPU acceleration fell back to CPU #2317.
2. Fallback on an illegal instruction. When a CPU build dies that way at startup (exit code 3221225501 in either sign, or SIGILL), _doStart retries on the next build in CPU_FALLBACK_LEVELS (ivybridge, then sandybridge) installed next to the build that died.
   • Each build tried becomes cachedServerBinaryPath, so later starts in the session (restarts, model or VAD changes, GPU→CPU fallbacks) never re-run a build that crashed.
   • This check runs after the GPU→CPU branch and before the thread-count retry, which used to relaunch the same crashing build.
   • With no build left, startup fails with whisper-server can't run on this processor: … (exit code: 3221225501) ….
   • Any other startup failure behaves exactly as before.
3. download-whisper-cpp.js installs every build a platform lists. Entries are keyed per build with a platformArch, like download-llama-server.js, and each build has its own install marker. A release missing any listed build fails --current. Today's four builds, their markers and their output are unchanged.
4. afterPack fails packaging when a listed fallback build is missing from resources/bin. The primary build is left to the download step.

Why two levels: without F16C, ggml converts fp16 weights one at a time through a lookup table. In local x86 runs (Rosetta, ratios only), the sandybridge build was about 7× slower than ivybridge, and every processor named in the reports has F16C.

Why a startup check in the fork: the first emulated run showed the ivybridge build starting and serving on a Sandy Bridge. ggml's startup code never touches F16C, so the build would only have crashed mid-transcription, where no fallback can see it. The check makes "dies at startup" a guarantee rather than an accident of ggml's init code. It is enabled only for the new builds, because Rosetta 2 (and some hypervisors) hide AVX in CPUID yet run it, and the existing build must behave exactly as before.

Deliberately untouched:

• The primary CPU build, and the GPU packs.
  • CUDA and Vulkan also need AVX2. On these CPUs they fall back to CPU, which then walks down the levels (tested).
• macOS: there is no level build. A darwin-x64 build that dies this way now gets the clearer message.
• No persistence to .env. The primary's crash costs about 0.1–0.5 s once per launch, during pre-warm.
• No new UI or translations. fix(whisper): clear message when local Whisper can't run (no AVX) #963's localized "processor not supported" toast can key on this message.

Tests

test/helpers/whisperServerCpuFallback.test.js has 11 tests. They spawn fake builds (shell scripts that die by SIGILL or SIGSEGV, or serve HTTP) through the real start path:

• the Ivy Bridge chain and the Sandy Bridge chain;
• a missing level;
• no level left;
• every level dying;
• the switch outliving stop();
• a CUDA pack on such a CPU;
• a non-SIGILL crash;
• racing starts;
• the Windows exit code in both signs.

Results:

• On the untouched base: 11/11 fail. Every scenario ends in the bare whisper-server process died during startup.
• On this branch: 11/11 pass. 133/133 across the whisper, GPU, download and afterPack suites (Node 24).
• New test/scripts/afterPackWhisperServerBuilds.test.js and the new downloadWhisperCpp tests cover the required-builds rule.
• A real --current download of 0.0.10 installs exactly what it did before.

The full npm test locally fails only in editor, markdown and chat suites, which need happy-dom / remark-gfm absent from that machine's node_modules. CI's clean install is the authoritative run.

Independent review: a read-only deep review covered the whole branch. It found no correctness issue; its doc, backstop and timeout findings are fixed in the last two commits.

Merging with #2317

There is one conflict, in whisperServer.js. Both PRs add the identical exitSignal and getProcessInfo lines, and #2317 also adds this._lastProcessInfo = getProcessInfo;. Take #2317's side of that hunk; keeping both declares exitSignal twice. Everything else merges cleanly in either order.

Remaining (why this is a draft)

1. Review and merge build: whisper-server for processors without AVX2 (ivybridge, sandybridge) whisper.cpp#8, then publish 0.0.11 from master. Its verify-cpu-levels job must pass first.

2. The final commit on this branch:

   • pin WHISPER_CPP_TAG to 0.0.11;
   • list the four fallback builds;
   • record the 0.0.11 CUDA and Vulkan digests;
   • add the TROUBLESHOOTING line and the CLAUDE.md platform note. They are held back until then, so no commit documents builds it doesn't ship.

   Until then this branch finds no fallback build, and such a CPU gets the clearer error.

3. After the app release ships, issues whisper-server crashes with Illegal Instruction (0xC000001D) on Ivy Bridge CPU (no AVX2) #2356, whisper-server-win32-x64.exe crashes with STATUS_ILLEGAL_INSTRUCTION (0xC000001D) on Intel Ivy Bridge (i5-3570) #1418, whisper-server-win32-x64.exe crashes with STATUS_ILLEGAL_INSTRUCTION (0xc000001d) on non-AVX2 CPUs, blocking GPU fallback path #1613 and OpenWhispr local whisper-server crashes with SIGILL on AMD FX-8350 #793 can be closed by hand.

Not verified

• Real Sandy Bridge, Ivy Bridge, FX or Jaguar hardware. Every run-time proof is emulated (SDE and QEMU).
• A packaged Windows or Linux install on such a machine.
• Speed on those processors.

Follow-ups:

• The GPU card offers "Retry GPU acceleration", which cannot succeed on these CPUs.
• parakeetWsServer.js and qdrantManager.js also drop the signal on a signal death.
• A build for processors without AVX (SSE4.2) if reports arrive.
• Crash-report noise: one crashed launch per app start on these CPUs, two on Sandy Bridge.

Plan: docs/superpowers/plans/2026-09-28-no-avx2-cpu-fallback.md

🤖 Generated with Claude Code