cactus_complete returns -1 on second invocation (iOS xcframework, Gemma 4 E2B int4)

[benikigai]

Environment

• iPhone 17 Pro Max (iOS 26)
• Cactus iOS xcframework (from cactus-react-native v1.14)
• Model: gemma-4-e2b-it INT4 (Cactus-Compute/gemma-4-E2B-it)
• Swift, native app (not React Native)

Bug

cactus_complete returns -1 on the second call to the same model handle. The first call after cactus_init + warm-up always succeeds. Every subsequent call fails.

cactus_transcribe with Parakeet works repeatedly on the same model handle — only LLM completions are affected.

Minimal Repro

let model = try cactusInit(path, nil, false)

// Warm-up — works
let _ = try? cactusComplete(model, "[{\"role\":\"user\",\"content\":\"Hi\"}]", "{\"max_tokens\":1}", nil, nil)

// First real call — works
cactusReset(model)
let r1 = try cactusComplete(model, "[{\"role\":\"user\",\"content\":\"Hello\"}]", "{\"max_tokens\":256}", nil, nil)
// r1 = valid JSON response ✓

// Second call — FAILS (result < 0)
cactusReset(model)
let r2 = try cactusComplete(model, "[{\"role\":\"user\",\"content\":\"Who are you?\"}]", "{\"max_tokens\":256}", nil, nil)
// throws "Completion failed"

Workarounds Attempted (all fail on 2nd call)

• cactusReset(model) before each call
• cactusStop(model) + cactusReset(model)
• No reset at all
• cactusDestroy + cactusInit (full re-init)
• Running on Task.detached / serial DispatchQueue
• Running on @MainActor directly
• Reducing max_tokens to 1
• Different message content

Context

Building at the YC Gemma 4 Voice Agents Hackathon. This is blocking our demo — we can only get one inference per app launch. Happy to test a fix immediately if you have one.

[asuzukosi]

@benikigai are there any additional logs from your development environment you could share?