mlx-vulkan

Run and compare MLX Vulkan inference benchmarks. Use when asked to benchmark this repository, measure a change, run bf16 or 8-bit MLX performance tests, or compare current results with the README baseline.

Review the MLX Vulkan backend for one major async-pipeline or inference-performance issue, obtain user approval, implement the scoped fix, then run full MLX validation, benchmarks, and model generation reporting. Use when the user wants an end-to-end Vulkan improvement workflow with an explicit review-before-fix approval boundary.

Commit changes solely within the MLX submodule in this Vulkan repository. Use when asked to inspect, stage, and commit work in `mlx/` without modifying or committing the `mlx-vulkan` parent repository.

Commit MLX submodule changes, run required Vulkan benchmarks, push a branch, and open a PR against `goniz/mlx`. Use when asked to create a pull request for MLX Vulkan work, especially when an issue number must be closed and bf16 plus 8-bit benchmark results are required.

Build MLX, reproduce one failing C++ test, and implement the smallest Vulkan-only fix. Use when asked to repair a C++ test failure in this repository while preserving test code and avoiding unrelated fixes.

Build MLX, reproduce one failing Python test, and implement the smallest Vulkan-only fix. Use when asked to repair a Python test failure in this repository while preserving test code and avoiding unrelated fixes.

Investigate and address a `goniz/mlx-vulkan` GitHub issue in the MLX submodule. Use when given an issue number and asked to read the issue, plan the work, reproduce a reported bug, implement the scoped fix, and work from an MLX feature branch based on `goniz/mlx:feat/vulkan`.

Review MLX Vulkan backend code for inference-performance hazards and async GPU pipeline violations. Use when asked to inspect a Vulkan file, function, class, symbol, diff, or code area for host synchronization, GPU readback, CPU fallback, avoidable copies or materialization, layout slow paths, redundant allocations, command-buffer breaks, or dispatch fragmentation. Report findings only; do not modify code unless explicitly asked in a separate request.

Fetch and address unresolved pull-request review comments for the MLX submodule. Use when asked to review outstanding feedback with `./dev.sh pr-comments --submodule mlx`, fix actionable Vulkan comments, or reply with an evidence-based explanation for comments that need no code change.

Review MLX Vulkan backend code for inference-performance hazards and async GPU pipeline violations. Use when asked to inspect a Vulkan file, function, class, symbol, diff, or code area for host synchronization, GPU readback, CPU fallback, avoidable copies or materialization, layout slow paths, redundant allocations, command-buffer breaks, or dispatch fragmentation. Report findings only; do not modify code unless explicitly asked in a separate request.

Review changes in the MLX submodule for concrete bugs, behavioral regressions, structural mismatches, and obvious performance problems. Use when asked to review uncommitted `mlx/` changes, a supplied diff, commit, file, or change set and provide actionable evidence-based feedback without modifying code.

Set up and use Tracy profiler for MLX Vulkan backend profiling