Skip to main content

Troubleshooting

Common issues and how to fix them.

Model loading​

"Model not found" or empty directory​

The model path must point to a directory containing config.json and weight files (.safetensors). Check:

  • The path exists and is readable.
  • The directory contains config.json, not a nested subdirectory.
  • On iOS, use a path relative to the app's container — absolute paths change between installs.

Model loads but generation produces garbage​

Use an instruction-tuned model (names containing -it- or -instruct). Base models without instruction tuning generate incoherent text.

Model loads slowly on first launch​

First load includes weight deserialization. Subsequent launches use cached data and are faster. Measure cold-load time from a Release build on the actual target device.

Memory and crashes​

App terminated on iOS (Jetsam)​

iOS terminates apps that exceed memory limits. These limits are well below physical RAM.

Fixes:

  • Enable the Increased Memory Limit entitlement in your Xcode target.
  • Use a smaller model (4B-4bit instead of 9B-4bit) on devices with limited memory.
  • Monitor with phys_footprint — do not trust os_proc_available_memory(), which can overestimate by 2–3 GB.
  • Let Edge Kit manage KV cache policy automatically. Do not override memoryPolicy unless you have measured the effect.

Memory grows across turns​

Edge Kit's DSR Attention keeps memory bounded. If you see unbounded growth, check:

  • You are not accumulating references to old generation results.
  • For single-shot tasks (transcription, speech synthesis), the engine releases cache automatically.
  • Call engine.unload() when switching models.

Build errors​

SPM "missing module" or "cannot find type"​

  • Clean DerivedData: rm -rf ~/Library/Developer/Xcode/DerivedData
  • Ensure your Package.swift pins the tested Developer Preview tag exactly: .package(url: "...", exact: "1.0.0-rc97")
  • Verify Xcode version is 15 or later.

"Metal library not found" or kernel crash​

Edge Kit requires Metal GPU access. Simulators do not support Metal inference — always build for a real device or use the generic/platform=iOS destination for compile checks.

Performance​

TPS much lower than expected​

  • Confirm you are running a Release build. Debug builds are 2–10× slower.
  • Check thermal state — a hot device throttles significantly. Cool the device and retest.
  • Verify the model quantization level. A 4-bit model is faster than 8-bit or bf16.

TTFT increases across turns​

Expected: TTFT grows with context length. For a 9B model on iPhone, median TTFT stays under 1 second for typical conversations. If TTFT spikes:

  • Check conversation length. Very long contexts (>8K tokens) increase prefill time.
  • Clear prompt cache to start a fresh conversation: engine.clearPromptCache()

Export and deployment​

Exported bundle does not load in Edge Kit​

  • Verify the export directory contains config.json and all weight shards.
  • Ensure the Edge Kit version matches the export version. Pin both.

Edge Scaffold app shows blank screen​

  • Check that the model path in ScaffoldConfig.swift is correct.
  • Run on a real device, not simulator.
  • Check Xcode console for error messages on launch.

Entitlements reset after xcodegen​

xcodegen can clear entitlements. After regenerating the project, verify that Increased Memory Limit and any network entitlements are still present.

Edge Mesh​

Devices not discovering each other​

  • Both devices must be on the same local network.
  • Verify NSLocalNetworkUsageDescription is set in Info.plist.
  • Check that the Bonjour service is not blocked by network configuration.

Still stuck?​

File an issue at github.com/AtomGradient or email alex@atomgradient.com.