Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 4 additions & 26 deletions AGENTS.md
Original file line number Diff line number Diff line change
Expand Up @@ -58,30 +58,8 @@ Optional CMake switches are `TILEXR_BUILD_COLLECTIVES`, `TILEXR_BUILD_EP`, `TILE
- Never use Host wrappers that launch Ascend C kernels with `kernel<<<...>>>` syntax. Build and embed pure AICore binaries, register them with `rtDevBinaryRegister` and `rtFunctionRegister`, then launch the registered signature with `rtKernelLaunchWithFlagV2`.
- Host, simulator, and 910B fallback tests do not prove UDMA data-plane transfer. Keep validation claims scoped to the hardware actually exercised.

<!-- karpathy-guidelines:start -->
# Karpathy Guidelines
### Communication Environment Baseline

## Think Before Coding

- Do not assume or hide uncertainty. State assumptions and material tradeoffs.
- Ask when ambiguity would change scope, architecture, interfaces, or behavior.
- Prefer the simpler approach when it fully satisfies the requirement.

## Simplicity First

- Write the minimum code needed for the requested behavior.
- Do not add speculative features, abstractions, or configurability.
- Keep implementations proportional to the problem being solved.

## Surgical Changes

- Touch only files and lines required by the task.
- Preserve existing style and unrelated user changes.
- Remove only unused code introduced by the current change.

## Goal-Driven Execution

- Define concrete, verifiable success criteria before implementation.
- Add focused tests for changed behavior and iterate until they pass.
- Report validation boundaries accurately and do not overstate evidence.
<!-- karpathy-guidelines:end -->
- For communication, peer-memory, MTE, `507035`, or device-dependent failures, run the installed official HCCL Test on the same devices and topology before investigating TileXR or operator code. The test mode must exercise the failing data path; use `all_reduce_test -a aiv_only` for AIV-only peer-memory cases.
- Treat the matching HCCL Test as the environment-health baseline, judging it by per-rank correctness output rather than exit status alone; operation failures may still exit with code 0.
- A pass on a different device subset, topology, or accelerator mode does not validate the failing path. If the matching HCCL Test also fails, classify the environment or platform baseline as unhealthy, preserve the bounded logs, stop operator-side debugging, and ask the user to provide a repaired or new known-good environment before continuing. Do not modify TileXR code to work around an unproven environment failure.
253 changes: 0 additions & 253 deletions CLAUDE.md

This file was deleted.

Loading