dstack-performance

Plugin dstack-performance from marketplace dstack in gosuda/dalcom-skills. Commands read from the repo's manifest — 3 days ago.

Skills (8)

• cpu-kernel-authoring

  Use when writing, optimizing, or benchmarking a C++ CPU kernel with AVX2 or AVX512 intrinsics for the Hugging Face kernels ecosystem. Not for CUDA kernels.

• decoded-memo-retention

  Use when a query-scoped decoded-block cache is bounded only by serialized work bytes, or when fallback paths retain many decoded blocks. Not for total-process memory promises: this skill never claims a process-wide bound.

• fastopt-extreme

  Use when estimating hot paths and proposing complexity-neutral cold simplifications while refusing complexity theater. Not for hypothesis-only hot paths: use fastopt. Not for measured-floor campaigns: use optimize-extremely.

• fastopt

  Use when a suspected hot path needs read-only hypotheses and benchmark designs, without waiting for benchmarks and without a code change. Not for a single measured fix: use optimize. Not for estimated hot and cold analysis: use fastopt-extreme. Not for measured-floor campaigns: use optimize-extremel

• heaptrack

  Use when tracking heap allocations on Linux with heaptrack: allocation hotspots, peak heap, temporary allocations, leaks, or a diff of two runs via heaptrack_print. Not for heap errors.

• optimize-extremely

  Use when a named subsystem must be made as fast as possible through a measured-floor campaign. Not for a single bottleneck fix: use optimize. Not for read-only analysis: use fastopt or fastopt-extreme.

• optimize

  Use when one named path or performance symptom needs a single measured fix. Not for read-only hot-path hypotheses: use fastopt. Not for a full measured-floor campaign: use optimize-extremely.

• simd-intrinsics

  Use when reading auto-vectorization reports, writing SSE2, AVX2, or NEON intrinsics, or fixing vectorization failures. Not for reading the asm itself: use assembly-x86.

Similar plugins