v2.1.0: cactus_complete with an image never returns — Metal backend deadlocks in waitUntilCompleted during chunked media prefill (macOS + iOS)

Summary

On v2.1.0, any completion with an image attached hangs forever inside cactus_complete. The engine loads the model, encodes the image successfully, tokenizes the prompt, submits the media chunked-prefill work to Metal, and then blocks permanently in [MTLCommandBuffer waitUntilCompleted] at 0% CPU. Reproduced identically on macOS (Apple M5 Pro, macOS 26.4.1) and iOS (iPhone 15 Pro), so it is not device-specific. The CPU backend does not block (it computes actively on all cores through the same prefill), which isolates the fault to the Metal path.

Text-only completion on the same model/engine works normally (≈104 tok/s decode on the iPhone), and image encoding alone works (≈250 ms, dim 768). Only image→text generation is affected — which makes every VLM chat use case unusable on v2.1.0's Metal backend.

Environment

• cactus 2.1.0 (Homebrew CLI; engine also rebuilt from the v2.1.0 git tag — same behavior)
• Model: prebuilt Cactus-Compute/LFM2-VL-450M cq4 bundle (weights tag v2.0)
• macOS 26.4.1 (Apple M5 Pro), Xcode 26.6; iPhone 15 Pro (iPhone16,1)
• Cloud handoff and telemetry disabled (--no-cloud-handoff --no-cloud-tele)

Reproduction (macOS)

cactus run Cactus-Compute/LFM2-VL-450M \
  --image any.jpg --prompt "Describe this image in one sentence." \
  --max-new-tokens 30 --no-cloud-handoff --no-cloud-tele

Model loads, the prompt and image line print, then no output. The engine process (cactus/bin/run) consumes ~2 s of CPU (load + image encode) and then sits at 0% CPU indefinitely (observed >10 min on iOS, >4 min on macOS; expected ~4 s). With engine DEBUG logging enabled, the last lines emitted are:

├─ Warmup: image_embed=256.13ms (dim 768)
[log 0] complete: Prompt tokens: 1043, max_tokens: 100

Nothing follows — no first-token/warmup line, no error.

Stack of the hung engine (macOS, sample of cactus/bin/run)

Main thread, 100% of samples:

main
  cactus_complete
    cactus::engine::Model::prefill_with_media(...)
      cactus::engine::Model::run_chunk_prefill_path(...)
        CactusGraph::execute(...)
          cactus_metal_session_sync()
            -[_MTLCommandBuffer waitUntilCompleted]
              _pthread_cond_wait   ← blocked forever

All other threads are idle pool/telemetry waiters. The submitted command buffer never completes and no error surfaces.

Control: CPU backend does not block

The identical command with --backend cpu behaves completely differently: the engine computes at full multi-core utilization through the prefill instead of parking at 0% CPU. The deadlock is specific to the Metal path (cactus_metal_session_sync).

Related harness gap (why cactus benchmark never catches this)

vlm_benchmark hard-fails on vision-only models before reaching image→text decode: [audio_embed] Exception: Model has no audio_encoder component aborts the suite right after the image-embed warmup. So the exact call that deadlocks is never exercised by the shipped benchmark. (Separately, transcribe_benchmark reports decode_tps around 10,712,962 tok/s — a timing divide-by-zero — but that's cosmetic by comparison.)