tt-debug-tools

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

Skills (9)

• tt-asserts

  Turn on assertion checking inside Tenstorrent kernels — lightweight kernel asserts, LLK asserts in the compute library, and the LLK sanitizer — and read back what fired. Use when a kernel produces impossible results, when you suspect a broken assumption about tile dimensions or data formats, or when

• tt-dprint

  Print values out of a Tenstorrent kernel — scalars, bfloat16, enums, and circular-buffer tile contents — with DPRINT / DEVICE_PRINT. Use when a kernel produces wrong numbers and you need to see what it actually read or wrote, when you want a checkpoint reached or a loop index, or when a DPRINT log n

• tt-exalens

  Read and write one address, register or core on a Tenstorrent device with tt-exalens — L1 and DRAM reads, RISC-V registers and GPRs, tensix state, NoC registers, callstacks from an ELF, memory search, and a GDB server for stepping a RISC. Use for a specific question about a specific location, when t

• tt-inspector

  Read what the Metal host runtime thinks is happening on a Tenstorrent device — which programs and mesh workloads exist, which kernels were built, tensor specs, mesh buffers and sockets — over the Inspector RPC or from its serialized logs. Use when device-side state needs host-side names attached to

• tt-noc-dump

  Collect NoC traces from a Tenstorrent device to find kernel data-movement mistakes — most usefully a missing noc_async_write_barrier. Use when a kernel's NoC transfers are suspect, when output is corrupt in a way that points at unflushed writes, or when asked to check barrier discipline. Not for han

• tt-profiler

  Measure where device time goes on a Tenstorrent part — per-zone kernel timings from the Device Program Profiler, and host-plus-device traces through Tracy. Use when a kernel or op is slower than expected and you need cycles attributed to named scopes, when you want a profile_log_device.csv read, or

• tt-triage

  Read the state of a hung or stuck Tenstorrent device with tt-triage — which op was running, where each RISC is parked, whether the on-device code and mailboxes are intact, whether NoC, Ethernet and ARC are healthy. Use when a job hangs, a test stops making progress, an assert has fired, or a triage

• tt-ttnn-flags

  Turn on TTNN's host-side debug modes — graph capture to find the op that hangs, per-op golden comparison to find the op that diverges, and the buffer and tensor reports. Use when a TTNN model produces wrong output and you need the first op that goes wrong, or when a model hangs and you want the op n

• tt-watcher

  Watch a Tenstorrent device from a polling thread inside the run — catch bad NoC transactions, circular-buffer overflows, L1 overflows, tripped kernel asserts, stack overflows, and where each RISC is parked. Use when a kernel corrupts memory, a transfer targets a core that does not exist, a run hangs

Similar plugins