SushiTray Local AI

Running local AI models on Apple Silicon often leads to kernel panics, OOM errors, or sluggish performance due to misconfigured memory limits and flag mismatches. This skill manages the selection, installation, and optimization of Sushi AI models specifically for Mac hardware. It eliminates the guesswork of matching model weights to unified memory constraints.

What it does

• Hardware Probing executes read-only inspections of chip architecture, GPU core counts, and current memory pressure.
• Model Matching selects the optimal Sushi pack based on physical RAM, GPU wired limits, and SSD availability.
• Command Generation produces exact, paste-ready sushi serve strings with expanded absolute paths and verified flags.
• Performance Tuning applies model-specific optimizations for Qwen, GLM, and DeepSeek, including speculation and n-gram settings.
• SushiTray Configuration formats parameters for the SushiTray GUI, ensuring proper argv execution without shell syntax errors.

How it works

1. Analyze Hardware by running internal probes to determine unified memory ceilings and macOS version compatibility.
2. Scan Environment to detect existing Sushi installations, downloaded packs, and current config.json states.
3. Generate Plan providing specific pack IDs, disk requirements, and a single, validated execution command.
4. Deploy and Verify by launching a test server, checking the /v1/models endpoint, and monitoring memory pressure.

Frameworks & tools

Targeted at Apple Silicon Macs running macOS 26.2+. Works with the Sushi AI engine, SushiTray GUI, and specific model families including Qwen, GLM, and DeepSeek. Python 3 is utilized for hardware inspection.

Why this beats prompting it yourself

Generic prompts often ignore the strict memory segmentation of Apple Silicon, leading to "Out of Memory" crashes. This skill uses hardware-aware logic to calculate the resident weight plus request state, preventing kernel panics caused by aggressive GPU wired limits.

Use cases

• Set up a new SushiTray installation with optimized flags for a 16GB M3 MacBook.
• Troubleshoot performance degradation or crashes during long-context generation.
• Migrate local workloads from CPU-only to GPU-accelerated Metal execution.
• Configure speculative decoding and n-gram lookahead for Qwen-based models.

Known limitations

Requires Apple Silicon and macOS 26.2 or higher. Cannot bypass sandbox-blocked hardware data or execute shell-specific syntax like pipes or variable assignments within SushiTray commands.