← Back to AI Insights
Gemini Executive Synthesis

Yogurt's AI notes engine stability on Linux, specifically CPU model loading with large allocations and interaction with Electron's PartitionAlloc.

Technical Positioning
Reliable and efficient local AI model execution across diverse hardware and operating environments, with robust memory management.
SaaS Insight & Market Implications
This issue details a critical SIGILL crash in Yogurt's notes engine on Linux when AI models run on the CPU, specifically due to Electron's PartitionAlloc failing on single allocations exceeding ~2 GiB. This directly impacts larger models like gemma-4-E4B and gemma-4-26B, which require substantial CPU_REPACK buffers. The problem is compounded by a faulty GPU probe that incorrectly forces CPU-only mode, leading to repeated crashes. This exposes a fundamental architectural incompatibility between Electron's memory allocator and the llama.cpp library's large allocation patterns. Market implications are severe: core AI functionality is rendered unusable on Linux for many configurations, undermining the product's value proposition. This necessitates either a re-evaluation of the Electron runtime for AI workloads or a custom memory management solution to bypass PartitionAlloc's limitations.
Proprietary Technical Taxonomy
notes engine dies SIGILL (code 132) CPU_REPACK allocation PartitionAlloc Electron binary allocator shim libggml/libllama posix_memalign

Raw Developer Origin & Technical Request

Source Icon GitHub Issue Nov 13, 2026
Repo: FrigadeHQ/yogurt
Linux: notes engine dies with SIGILL (code 132) when the model runs on the CPU (>2 GiB CPU_REPACK allocation vs PartitionAlloc)

Yogurt 0.3.11 AppImage (Electron 44.7.2, node-llama-cpp 3.22.1), Linux x86_64, Ryzen 7 5700X (AVX2), 32 GB RAM, RTX 3060 Ti.

**Symptom.** Writing notes fails with `notes engine stopped (code 132)` on every attempt, with both gemma-4-E4B_q4_0 and gemma-4-26B_q4_0. The kernel logs `trap invalid opcode ... in yogurt-bin`.

**Cause.** The Electron binary exports Chromium's allocator shim (`malloc`, `posix_memalign`, ...), so libggml/libllama allocate through PartitionAlloc, which crashes on purpose on a single allocation above ~2 GiB. When the model runs on the CPU, llama.cpp repacks the q4_0 weights for AVX2 into one `CPU_REPACK` buffer: 2116 MiB for E4B and 13134 MiB for 26B. Under gdb the last call before the crash is `posix_memalign(64, 2219212800)`. The same binaries in plain Node 22 load and generate fine, because glibc mmaps large blocks.

**Why it ends up on the CPU.** In my case the startup probe logged `graphics: vulkan, 0.0 GB` (no device), so the "graphics card allowed" load put 0 layers on the GPU and crashed. After that `cpuOnly` stays set for the rest of the session, so every retry is CPU-only and crashes again. When the probe does see the card, both models load with GPU layers and work, because no CPU_REPACK buffer is created.

**Repro without the UI.** With Electron 44.7.2: `utilityProcess.fork(/out/main/engine.js, [], { env: { ...process.env, YOGURT_ENGINE_GPU: 'off' } })`, post `{type:'probe',id:1}`, then `{type:'load',id:2,modelPath:

Developer Debate & Comments

No active discussions extracted for this entry yet.

Adjacent Repository Pain Points

Other highly discussed features and pain points extracted from FrigadeHQ/yogurt.

Extracted Positioning
Yogurt's diarization support
Advanced AI transcription features, specifically speaker diarization, for meeting notes.
Extracted Positioning
Yogurt's local transcription engine performance and resource management, specifically GPU utilization during long calls.
Efficient, stable, and resource-conscious local AI processing for meeting transcription.
Extracted Positioning
Yogurt's audio capture compatibility with Bluetooth devices and integration with Microsoft Teams on Windows 11.
Seamless and reliable audio capture across diverse audio hardware and communication platforms.
Extracted Positioning
Yogurt's Whisper server GPU selection on Linux
Optimized AI transcription performance by leveraging discrete GPUs on Linux systems.
Extracted Positioning
Yogurt installation via Homebrew
Accessible and streamlined installation for a cross-platform AI meeting notes application.

Frequently Asked Questions

Market intelligence mapped to Yogurt's AI notes engine stability on Linux, specifically CPU model loading with large allocations and interaction with Electron's PartitionAlloc..

How is Yogurt's AI notes engine stability on Linux, specifically CPU model loading with large allocations and interaction with Electron's PartitionAlloc. positioned in the market?
Based on our AI analysis of the original developer request, its primary technical positioning is: Reliable and efficient local AI model execution across diverse hardware and operating environments, with robust memory management.
What is the general sentiment around Yogurt's AI notes engine stability on Linux, specifically CPU model loading with large allocations and interaction with Electron's PartitionAlloc.?
Yes, we have tracked 2 direct responses and active debates regarding this specific topic originating from GitHub Issue.
What are the foundational technologies related to Yogurt's AI notes engine stability on Linux, specifically CPU model loading with large allocations and interaction with Electron's PartitionAlloc.?
Our proprietary extraction maps Yogurt's AI notes engine stability on Linux, specifically CPU model loading with large allocations and interaction with Electron's PartitionAlloc. to adjacent architectural concepts including notes engine dies, SIGILL (code 132), CPU_REPACK allocation, PartitionAlloc.

Engagement Signals

2
Replies
open
Issue Status

Cross-Market Term Frequency

Counts stored text matches for foundational terms like AVX2 and Electron binary in analysed public discussions. These counts do not measure adoption.

Stored text matches, not adoption or market-share measurements.

  • AVX2: 3 records
  • Electron binary: 2 records
  • notes engine dies: 1 records
  • SIGILL (code 132): 1 records
  • CPU_REPACK allocation: 1 records
  • PartitionAlloc: 1 records
  • allocator shim: 1 records
  • libggml/libllama: 1 records
  • posix_memalign: 1 records
  • glibc mmaps: 1 records
  • startup probe: 1 records
  • graphics: vulkan, 0.0 GB: 1 records
  • cpuOnly: 1 records
  • utilityProcess.fork: 1 records
  • gemma-4-E4B_q4_0: 1 records
  • gemma-4-26B_q4_0: 1 records