tt-model-bringup

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

Skills (14)

• benchmark-model

  Report a text model's accuracy on fixed lm-eval subsets and 4K-input vLLM serving performance at concurrency 1 and 32. Use after optimized-vLLM for the final model-bringup report or to benchmark a supported model.

• datatype-sweep

  Sweep and select TTNN model weight, activation, KV-cache, CCL datatype, and compute-fidelity policies for the fastest configuration that satisfies minimum full-model top-1 and top-5 accuracy. Use after optimized-full-model and before vLLM, or whenever a runnable TTNN full model needs evidence-backed

• full-model

  Build a complete TTNN autoregressive model and generator from a HuggingFace model plus one or more working TTNN block or decoder implementations. Use when assembling embeddings, block stack, final norm, logits, KV-cache handling, and generation into a full model, validating end-to-end behavior, and

• functional-decoder

  Bring up a functionally correct TTNN implementation of HuggingFace transformer decoder layers under the model-specific autoport directory. Use when producing a functional_decoder.py implementation, reading the HF decoder architecture, validating paged prefill/decode against the HF reference, and lea

• graph-fusing

  Fuse a TTNN op graph into a faster numerically-equivalent one by (1) replacing a primitive op sequence with a dedicated tt-metal op, (2) simplifying/merging structural ops, and (3) folding adjacent ops into an existing op.

• model-bringup

  Implement Hugging Face text models in TTNN through staged decoder, multi-chip, full-model, datatype, vLLM and benchmark work. Use for model bring-up or resuming its staged goals; requires separately installed tt-autodebug.

• multichip

  Parallelize single-chip TTNN code across the available multi-chip hardware. Use when parallelizing a model or module, choosing a tensor/mesh parallel strategy, comparing to a single-chip TTNN baseline, and leaving correctness plus performance evidence.

• optimize

  Optimize per-device performance of runnable TTNN code, preserving correctness while improving layout, precision, sharding, program configs, data movement, and warmed latency with tt-perf-report evidence except in vLLM serving stages, where profiler collection is intentionally skipped.

• qualitative-check

  Run and review qualitative or prompt-based evaluation checks for Hugging Face text models during TTNN bringup. Use whenever a stage generates text, judges output quality, runs vLLM/TTI/API qualitative smokes, compares HF and TT outputs, investigates wrong-language or base-model autocomplete behavior

• stage-review

  Independently review a TTNN model bringup stage before it is allowed to pass. Use when a stage is complete or near-complete and Codex needs a fresh xhigh subagent to compare the stage output against the original goal contract, skill requirements, logs, generated model outputs, checks, benchmarks, an

• tt-device-usage

  Use TT hardware safely during tt-metal model bringup, tests, profiling, serving, reset/recovery, tt-smi health checks, tt-triage hang capture, watcher/profiler separation, and T3K ARC/ERISC/remote-Ethernet recovery. Use whenever a stage touches Tenstorrent devices or must decide whether a hardware f

• tt-enable-tracing

  Enable or debug TTNN trace capture and replay for models, decoders, generators, and multi-chip TTNN code. Use when eliminating TTNN dispatch overhead, implementing traced prefill/decode, making full-model generation trace-safe, or investigating trace failures such as unsupported writes, reads, event

• tti-release

  Run the tt-inference-server model release workflow for the completed generated autoport TTNN/vLLM model and produce the customer-facing readiness markdown report. Use after optimized-vLLM or at model readiness handoff time, especially when Codex must evaluate the generated autoport through an alread

• vllm-integration

  Add vLLM serving integration for an existing TTNN full model under the model-specific autoport directory. Use after full-model when producing tt/generator_vllm.py, registering the model with the TT vLLM plugin, running the vLLM readiness server path, and reporting serving-path correctness and perfor

Similar plugins