Skip to content

CLAUDE.md

Latest commit

 

History

History
47 lines (37 loc) · 3.05 KB

CLAUDE.md

File metadata and controls

47 lines (37 loc) · 3.05 KB
Raw

CLAUDE.md — cppmega.mlx

RULE #1 (HIGHEST PRIORITY — overrides everything else) — NO AUTOMATED FALLBACKS. FAIL FAST, FAIL LOUD.

Every operation goes through ONE clear path. When the clear path does not work as expected, errors, or misbehaves, the code MUST RAISE IMMEDIATELY with a precise message saying WHERE and WHAT failed, and crash — so the bug surfaces and we fix the ROOT CAUSE.

FORBIDDEN — automated/silent fallbacks that hide a failure:

  • try/except → degraded/alternate path; except: pass; "best-effort"; silent slow-path.
  • eager-replaces-fused, MSL-adapter-instead-of-tvm-ffi, gated-off-by-default, watchdog→fallback.
  • clamp-instead-of-raise; return zeros/garbage; silent shape/dtype/precision downgrade.
  • any code path that papers over a broken clear path instead of surfacing it.

REQUIRED on failure: raise with where + what → then we debug and fix the bug. Fail-closed by raising with a clear error IS correct ("падать с сообщением"). A SILENT fallback that produces a degraded/wrong/zero result is what is forbidden.

This is the first and main rule. Existing automated fallbacks must be found and removed; route everything through the single clear path that fails loud.

Repos, branches, upstream provenance (David, 2026-05-30)

All our repos live in the DatasunriseOU org and we work on the main branch EVERYWHERE (local + gb10 + every machine). Do NOT create side/feature branches for our own work — commit to main, push main, pull main. (Past mistake: scattering fixes across m4-metal4-keystone / merge/upstream-codegen-reorg side branches — never do this; consolidate on main.)

Our repos (all in DatasunriseOU): cppmega_mlx, tilelang, tvm, tvm-ffi, mlx, mlx-lm, cppmega, triton-pr9701. These are OUR working repos (not "forks we don't control") — we commit to them directly.

Upstream provenance + how we pull fixes (we cherry-pick needed fixes FROM upstream INTO our main):

  • tvm = apache/tvm upstream (git remote origin = https://github.com/apache/tvm). Our fork = DatasunriseOU/tvm (remote datasunrise). We take tvm from apache, NOT from the TileLang/tvm fork (tilelang remote = TileLang/tvm exists only to cherry-pick specific fixes).
  • tvm-ffi = apache/tvm-ffi upstream (Apache-2.0). Our fork = DatasunriseOU/tvm-ffi, vendored as the 3rdparty/tvm-ffi submodule inside our tvm. From apache, not tilelang.
  • tilelang = upstream TileLang/tilelang; our fork DatasunriseOU/tilelang, work on main.
  • mlx / mlx-lm = upstream ml-explore/*; our forks DatasunriseOU/mlx, DatasunriseOU/mlx-lm.
  • The pattern: keep an upstream remote (origin=apache for tvm, etc.), pull/cherry-pick the fixes WE need from upstream into OUR main, and build from our main.

Tensor memory rule (see AGENTS.md)

  • Wrappers/adapters must not silently allocate, cast, or copy large tensors to satisfy a boundary. If a design seems to require it, the design is wrong — keep looking (this is a corollary of Rule #1: no silent papering-over).