torch-spyre

Systematic debugging of torch.compile failures on the Spyre backend. Covers each pipeline stage from Dynamo tracing through runtime execution, with diagnostic env vars and common error patterns.

Overview of the torch-spyre repository, Spyre device architecture, compilation pipeline, and codebase structure. Use when asking about how torch-spyre works, the repo layout, Spyre hardware, the Inductor backend, or getting oriented in the codebase.

Generate a daily status update for any squad's Slack workflow. Pulls recent git activity, open PRs, and blockers to produce a formatted update matching the team's standup template.

Check documentation consistency against code changes. Audits supported ops table, RFC links, API docs, compiler docs, runtime docs, profiling, and user/developer guides for staleness or drift.

Review pull requests for the torch-spyre project. Checks license headers, pre-commit compliance, signed commits, tensor layout correctness, test coverage, and documentation.

Upgrade the pinned PyTorch version across torch-spyre and torch-spyre-docs. Use when asked to bump, upgrade, or update the PyTorch dependency. Accepts target version as argument (e.g., '2.13').

Guide for adding new operations to the Spyre backend. Covers direct ATen-to-OpFunc mappings, Spyre-specific decompositions, and custom ops with lowerings. Use when asked to add, implement, or enable a new op on Spyre.

Guide for writing compiled-path operator tests using the ParameterizedTestMeta framework and compare_with_cpu utilities. Use when adding tests for new or existing Spyre ops in tests/_inductor/test_inductor_ops.py.

Convert a raw-pointer or block_ptr Triton kernel to the tl.make_tensor_descriptor API, producing a tensor_descriptor.py. Backend-agnostic — no Spyre invariants. Use when asked to port a kernel to tensor descriptors or modernize block_ptr/raw-pointer loads.

Review a tensor-descriptor Triton kernel (tensor_descriptor.py) for correct tl.make_tensor_descriptor usage — deprecated APIs removed, valid shape/strides/block_shape, no redundant masking, 16-byte rule, correctness vs original. Backend-agnostic; does not check Spyre invariants. Use spyre-review for the full Spyre compliance review.

Convert a GPU-shaped Triton kernel into a full Spyre-aware kernel (spyre.py) — tensor descriptors plus the authoring invariants (see _shared/invariants/), scratchpad batching, and Spyre-compiler gap handling. Use when asked to make a kernel Spyre-aware or port to Spyre. For a plain raw->descriptor port with no invariants, use td-convert.

Review a Spyre-aware kernel (spyre.py) for compliance with the authoring invariants (see _shared/invariants/), scratchpad utilization, Spyre-compiler descriptor patterns, and correctness vs original. Use when asked to review/verify a Spyre kernel. For a plain descriptor-API review (no invariants), use td-review.

Write numerical equivalence tests for a Spyre-aware kernel (spyre.py) vs the original, including distribution invariance across core counts. Produces tests/triton/test_<name>_spyre.py. Use when asked to test a Spyre kernel. For plain _td kernels, use td-test.

Write numerical equivalence tests for a tensor-descriptor kernel (tensor_descriptor.py), comparing it against the original kernel. Produces tests/triton/test_<name>_td.py. Use when asked to test a _td kernel.

Bump the pinned vLLM version in `pyproject.toml`, re-sync the upstream test plugin, and triage the breakages that the bump exposes. Use whenever a user asks to upgrade/bump/update vLLM, pull up the lower bound, or chase a specific upstream commit. Most failures after a bump are not in our code — they are upstream API churn (worker `load_model` wrappers, platform `is_*()` predicates, KV-cache layout flips, new constructor kwargs) that our `_enum = PlatformEnum.OOT` platform doesn't get the CPU/GPU short-circuits for.

Bump the pinned vLLM version in `pyproject.toml`, re-sync the upstream test plugin, and triage the breakages that the bump exposes. Use whenever a user asks to upgrade/bump/update vLLM, pull up the lower bound, or chase a specific upstream commit. Most failures after a bump are not in our code — they are upstream API churn (worker `load_model` wrappers, platform `is_*()` predicates, KV-cache layout flips, new constructor kwargs) that our `_enum = PlatformEnum.OOT` platform doesn't get the CPU/GPU short-circuits for.

Bump the pinned `torch-spyre` git rev in `pyproject.toml` and update `spyre-rpms.lock` to match artifactory, binary-search to the latest commit that actually compiles against this host's `ibm-*` RPMs, clear the stale inductor cache, run a smoke test, and write a reviewer-ready PR description. Use whenever the user asks to "bump", "upgrade", "update", or "pull up" torch-spyre — typically after new `ibm-deeptools` / `ibm-flex` / `ibm-senlib` packages land that unblock previously-failing torch-spyre commits. Encodes that build failures are the expected signal that supporting libs need a matching bump, that the torchinductor cache must be wiped after the bump to avoid `TypeError: ...__init__() got an unexpected keyword argument ...` red herrings, and the curated PR-description shape (notable upstream PRs + bisect table + installed `ibm-*` RPM versions).

Reconcile monthly labels (April2026, May2026, June2026, …) on the torch-spyre vLLM project board (project 2, view 7, repo torch-spyre/spyre-inference) at the end of each month. Use whenever the user asks to "clean up the board", "do the monthly label cleanup", "do the May/June/July cleanup", or otherwise refresh which monthly bucket each issue lives in. The skill enforces three rules — May-closed → MayYYYY, open MayYYYY → JuneYYYY swap, and Q2-relevant unlabeled issues → current month — while flagging anomalies for the user instead of guessing.

Debug numerical, compilation, or correctness failures anywhere the Spyre backend is involved — custom ops, attention, model runner, end-to-end vLLM runs. Use whenever a user reports failing Spyre tests, accuracy/tolerance mismatches against a CPU reference, torch.compile errors on `spyre` device, silent CPU fallbacks, or generally "something on Spyre is broken." Most bugs here are not in our code — they are torch-spyre op gaps or dtype/layout limitations, so debugging requires tracing into the torch-spyre site-packages, not just reading `spyre_inference/`. For attention-specific hints (cache-seeding, KV alignment, MHA/MQA, broadcast diagnosis), see `attention-notes.md` next to this file.

Upgrade the sendnn-inference plugin to a newly released vLLM version. Bumps the pin in pyproject.toml, regenerates uv.lock, runs CPU + compat tests, adds compat shims if upstream APIs changed, and extends the CI backwards-compat matrix. Use when the user says "upgrade vllm to X.Y.Z", "bump vllm", "support the new vllm release", or similar.

Upgrade the sendnn-inference plugin to a newly released vLLM version. Bumps the pin in pyproject.toml, regenerates uv.lock, runs CPU + compat tests, adds compat shims if upstream APIs changed, and extends the CI backwards-compat matrix. Use when the user says "upgrade vllm to X.Y.Z", "bump vllm", "support the new vllm release", or similar.