tt-foundry: Orquestrador de Bringup de Modelos para Tenstorrent
saiarthiraguram/tt-foundry
Pipeline orquestrado para deploy e otimização de modelos em hardware Tenstorrent
Ver no GitHubConjunto de skills Claude Code que implementa um pipeline end-to-end (VALIDATE → FIRST_RUN → DIAGNOSE → REPAIR → VERIFY → CONFIG_UPDATE) para trazer modelos de IA para hardware Tenstorrent. Inclui scaffolds para carregar modelos de HuggingFace ou GitHub, diagnóstico automático de falhas, estratégias de reparo (monkey-patch, debug em runtime), análise de graph-breaks e sugestão de candidatos SOTA.
Instalação
npx convez-ai saiarthiraguram/tt-foundry1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automatizar validação e deploy de modelos em aceleradores Tenstorrent
- Diagnosticar e reparar falhas de compilação ou runtime em modelos grandes
- Avaliar fragmentação de grafos e otimizações de tensor parallelism em multichip
Veredito

Skill altamente especializada para engenheiros que trabalham com hardware Tenstorrent e modelos em larga escala. Automatiza etapas críticas de bringup com orquestração robusta, mas exige conhecimento profundo de compilação, sharding e debugging de modelos de IA.
README
A tradução para o português deste README ainda está na fila.
tt-foundry
Claude Code skills that drive the end-to-end model bringup pipeline for Tenstorrent hardware (PJRT / tt-xla / tt-forge-models).
What's here
skills/
├── model-bringup/ # E2E orchestrator (FSM)
├── model-bringup-scaffold/ # VALIDATE stage — loader scaffold + state.json (HuggingFace default)
├── model-bringup-scaffold-github/ # VALIDATE variant — model source lives on GitHub (vendor or port)
├── model-bringup-scaffold-pipeline/ # VALIDATE variant — multi-component HF DiffusionPipeline (per-component loaders, shard specs)
├── model-bringup-overview/ # OVERVIEW stage — CPU sanity + golden.pt capture, model_overview.md
├── model-bringup-run/ # FIRST_RUN / VERIFY stages — pytest under 5-min budget
├── model-bringup-diagnose/ # DIAGNOSE stage — pattern-match failure log → JSON
├── model-bringup-repair/ # REPAIR stage — apply strategy (monkey_patch, runtime_debug, …)
├── model-bringup-config-update/ # CONFIG_UPDATE stage — write final YAML status
├── model-bringup-finalize/ # FINALIZE stage — multi-arch reverify, pre-commit, PR draft
├── model-bringup-classify-oom/ # Classify activation vs weight-bound OOM (single-chip)
├── model-bringup-write-promotion/ # Write promotion.json after weight-bound exhaustion
├── model-bringup-multichip/ # Promotion-only multichip TP orchestrator + references/ + scripts/
├── model-bringup-scaffold-torch-tp/ # VALIDATE_TP — shard specs (Megatron / FSDP / MoE) after promotion
├── model-bringup-run-torch-tp/ # FIRST_RUN_TP / VERIFY_TP on multichip hosts
├── model-bringup-repair-shard-spec/ # REPAIR shard map / mesh for TP
├── model-bringup-config-update-torch-tp/ # CONFIG_UPDATE for tensor_parallel YAML
├── runtime-failure-debugger/ # Op-level bisect invoked by runtime_debug repair strategy
├── graph-break-analysis/ # Auxiliary: torch.compile graph-break investigation
├── issue-create/ # Draft GitHub issue packages (.claude/issues/) — never auto-files
├── model_issue_pick/ # XFAIL re-triage (single entry)
├── failure_summary/ # YAML digest of all KNOWN_FAILURE_XFAIL entries
├── potential_new_models/ # SOTA bringup-candidate suggester
├── triage-unpack-forward-output/ # Triage FAILED_FE_COMPILATION "no unpack_forward_output" cases
├── code-reviewer/ # C++/Python code review checklist + pre-PR self-review (lint/SPDX/tests/commits)
└── create-pr/ # Open a tt-xla PR (area-prefixed title, body template, CODEOWNERS)
High-level flow
┌──────────────────────┐
/model-bringup <key> ──▶│ model-bringup │
│ (orchestrator) │
└──────────────────────┘
│
┌─────────┬─────────┬─────────┬─────┴────┬─────────┬───────────────┬──────────┐
▼ ▼ ▼ ▼ ▼ ▼ ▼ ▼
scaffold overview run diagnose repair verify config-update finalize
│ │
├─ scaffold-github (GitHub-hosted) ▼
└─ scaffold-pipeline (DiffusionPipeline) runtime-failure-debugger
│
▼
graph-break-analysis
Auxiliary skills hang off the same FSM but enter from different states:
issue-create— drafttitle.txt+draft.md+manifest.json+gh-command.shunder.claude/issues/from bringup escalations or pytest logs (developer runsghmanually)model_issue_pick— re-triage a single existingKNOWN_FAILURE_XFAILentryfailure_summary— digest the XFAIL list for triage / sharingpotential_new_models— suggest the next wave of bringup candidatestriage-unpack-forward-output— fix one training-test failure pattern (FAILED_FE_COMPILATIONwith "no unpack_forward_output handler")code-reviewer— review a diff (C++/Python checklist, standards, antipatterns) or run the mechanical pre-PR self-review (lint / SPDX / test coverage / commit messages)create-pr— open the PR afterfinalizeproduces the branch + body draft
Single-chip vs multichip (initial v1)
/model-bringup— always single-device first. Per-componentweight_fit.jsonplans n150 (12 GiB) and p150 (32 GiB); runs both arches when both eligible. Dtype ladder: source-dtype FIRST_RUN → activation repair → bf16 if needed on the same arch. No multichip from REPAIR./model-bringup-multichip— only afterpromotion.json(all eligible arches weight-bound). PyTorch TP (Megatron / FSDP-style / MoE); pipeline components are the priority validators.
See skills/model-bringup-multichip/references/ for DRAM tables, OOM classes,
shard templates, and pytorch_multichip_tp.md.
Helper scripts under skills/model-bringup-multichip/scripts/:
compute_weight_fit.py— emitweight_fit.jsonfrom param countwrite_promotion.py— emitpromotion.jsonfromstate.arch_results
Consuming tt-foundry from another repo
The skills are plain SKILL.md files under skills/<skill-name>/. Two ways
to use them from a sibling project (e.g. tt-xla):
- Submodule. Add
tt-foundryas a git submodule (e.g. underthird_party/tt-foundry), then point your.claude/settings.jsonplugin list (or symlinks in.claude/skills/) at the skills directory. - Direct copy. Copy individual
skills/<name>/SKILL.mdinto the consuming repo's.claude/skills/. Faster, but creates drift — prefer the submodule path for anything beyond a one-off experiment.
Repo invariants
- No code, only prompts. This repo contains
SKILL.mdfiles and docs. All actual code (pytest invocations, YAML parsing, etc.) lives in the consuming repo (tt-xlatoday). - Read-only on the consuming repo's YAML. The skills can recommend
changes, but they only mutate YAML when explicitly invoked with
--applyor equivalent. - One source of truth per skill. If you find yourself editing the same
skill in both
tt-foundryand a consuming repo's.claude/skills/, upstream the change tott-foundryand drop the local copy.
License
(Add license file when this repo gets a home / remote.) #To-DO