tt-review-skills

Plugin tt-review-skills from marketplace tenstorrent-skills in tenstorrent/skills. Commands read from the repo's manifest — 8 days ago.

Skills (15)

• llk-perf-audit-review

  Static performance review of Tensix compute kernels — wasted cycles, redundant register and memory traffic, loop-invariant work, and cheaper numerically-equivalent instruction sequences. Gated by a provenance lens (sfpi-compiled versus hand-written) and a semantic-equivalence test. Use when reviewin

• llk-race-audit-review

  Reviews tt-llk kernel changes for the nine intra-kernel race hazard classes plus stale hardware config across invocations — cfg-word overlap, dataflow CB sync, instruction latency, mailbox sync, MMIO ordering, NoC sync, reconfig stalls, semaphore handshakes, and srcreg bank sync — plus the cross-cla

• tt-comment-hygiene-review

  Reviews comments and documentation surface — iteration-journey comments that describe how the code got here, unsurfaced tribal knowledge, magic tile and grid values, and ttnn op docstrings. Cheap to run alongside any other review. Use on any Tenstorrent diff.

• tt-l1-memory-review

  Reviews per-core L1 footprint and circular-buffer sizing — buffer inventory discipline, data-movement cost across memory tiers, CB capacity versus tile counts, and accumulator sizing. Use when reviewing program factories, CB allocation, blocking or work-split changes, or any change that adds a buffe

• tt-model-bringup-review

  Reviews TTNN model code — layout and memory-config defaults, the decode residual contract, composite-op preference, program-config pitfalls, hidden host fallbacks, and logical batch versus tile padding. Use when reviewing model bringup or optimization changes under models/ or ttnn/ Python.

• tt-multichip-ccl-review

  Reviews multi-chip and collective-communication code — num_links against device topology, bias-before-all-reduce numerics, distributed RMSNorm correctness, gather and reduce axes, fabric initialisation order, and CCL tuning claims. Use when reviewing changes touching CCL ops, mesh devices, fabric, o

• tt-perf-claim-review

  Reviews performance claims rather than performance changes — whether the measurement supports the assertion. Covers device duration versus wall time, bound classification, first-run program-cache caveats, absolute versus percentage comparisons, and warmed traced harness parity. Use when a PR body, c

• tt-precision-review

  Reviews dtype and math-fidelity policy — per-tensor-group precision, the prefill versus decode KV-cache dtype asymmetry, reduced-cache candidates, and PCC-collapse triage. Use when reviewing changes to dtypes, math fidelity, compute kernel configs, or cache precision.

• tt-review-core

  Shared review contract for all Tenstorrent code review — severity vocabulary, the evidence rule, scope, output shape, and the do-not-flag guards. Use when reviewing any change to a Tenstorrent repository (tt-metal, tt-llk, tt-inference-server, model code), and load it before any domain review skill.

• tt-review-router

  Maps changed paths in a Tenstorrent pull request to the domain review skills that apply. Run this first to pick the skill subset for a review.

• tt-split-pr-by-codeowners

  Decide whether a wide-ranging pull request should be broken up so each piece needs fewer CODEOWNERS approvals, and propose how. Use for 'too many reviewers on this', 'can this be broken up', 'why does this need six approvals'.

• tt-test-coverage-review

  Reviews test coverage for Tenstorrent changes — PCC bars and when a lower one is justified, tile-boundary and padding cases, sharding and data-format variants, program-cache tests, and a regression test on every bug fix. Use when a change alters op or kernel behaviour, whether or not it touches test

• tt-trace-review

  Reviews trace capture and replay safety — nothing host-side inside capture, program-cache warmup with exact signatures including scalar argument values, no allocation after capture, and device-owned autoregressive state. Use when reviewing changes touching begin_trace_capture, execute_trace, generat

• tt-vllm-serving-review

  Reviews the vLLM and tt-inference-server serving path — generator_vllm.py contracts, plugin registration, the tt_data_parallel ambiguity, and TT-fork branch and config conventions. Use when reviewing changes to generator_vllm.py, vLLM plugin registration, or serving configuration.

• ttnn-op-kernel-review

  Structural correctness review for TTNN op kernels — init and data-format reconfig, TRISC synchronization, the tile_regs protocol, circular-buffer ownership and UB, work distribution, semaphores, control-flow CB balance, in-place misuse, op-level input validation, and program-cache correctness. Use w

Similar plugins