Debug index-based compute operations like TopK, Sort, and Argmax in tt-emule.
npx skills add https://github.com/tenstorrent/tt-emule --skill index-based-ops
Please help me install this Agent Skill. Skill: index-based-ops Source: https://github.com/tenstorrent/tt-emule/tree/main/.claude/skills/index-based-ops Command: npx skills add https://github.com/tenstorrent/tt-emule --skill index-based-ops
This Skill addresses the complexity of debugging index-based compute operations like TopK, Sort, and Argmax in the tt-emule environment, where silicon-specific index choices and timing races can cause test failures.
Use the index-based-ops skill to analyze the current TopK kernel failure by comparing the op output against torch values and validating index gather metrics.
High-intent search queries and answers about installing and using this skill.
Debug TopK kernel test failures in tt-emule by comparing op outputs against torch values and validating index gather metrics. Determine if index mismatches are valid non-deterministic tie-breaks or genuine kernel bugs by checking if the computed values remain correct.
Isolate multi-core reduction kernel and dataflow shim failures using targeted debugging strategies that verify numeric metrics and index validity. Analyze cross-core semaphore handshakes and synchronization races to pinpoint architecture-specific kernel path issues causing test failures.
Argmax kernel tests fail in the emulator because silicon-specific index choices and timing races cause mismatches. Match test contracts for value-index pairs rather than expecting exact silicon index parity to validate compute operations correctly.
Yes, validate Sort compute kernels by matching test contracts for value-index pairs rather than exact silicon index parity. This approach satisfies requirements for verifying numeric metrics while accounting for non-deterministic tie-breaks in the tt-emule environment.
Verify index validity and numeric metrics for index-based compute operations in tt-emule by analyzing multi-core reduction trees, dataflow shims, and cross-core semaphore handshakes. Compare outputs against torch values to ensure correct kernel execution.
tenstorrentDiagnose data-corruption and memory-related failures in the tt-emule software emulator.
tenstorrentAutomate compute-kernel LLK shim implementation and integration in tt-emule.
tenstorrentBisect tt-metal and tt-umd dependency updates to isolate C++ and JIT regressions.
tenstorrentValidate silicon API mocks in tt-emule against verification checklists.
