skills.rest
WangWindow
WangWindowCOMMUNITY
Explore all skills by this author →
WebsiteGitHub

pypto-precision-compare

Diagnose PyPTO operator precision issues using tensor graph verification, Pass checks, and binary search.

Updated Sep 15, 2026
Search Tags:#pypto#binary search#ascend npu#golden comparison#precision debugging#operator verification#pass verification
One-click install
npx skills add https://github.com/WangWindow/CANN-BatchMatMulMaxsum --skill pypto-precision-compare-wangwindow
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: pypto-precision-compare
Source: https://github.com/WangWindow/CANN-BatchMatMulMaxsum/tree/main/.agents/skills/pypto-precision-compare
Command: npx skills add https://github.com/WangWindow/CANN-BatchMatMulMaxsum --skill pypto-precision-compare-wangwindow

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires torch, pandas, and includes scripts (resource) components.

What problem does it solve? When a PyPTO operator's output deviates from its PyTorch golden reference, developers struggle to locate which op, Pass, or codegen stage introduced the numerical error. This Skill provides a structured decision tree and three complementary comparison methods to pinpoint the exact source of precision divergence on Ascend NPU. ## Core Features & Use Cases - Tensor Graph Verification: Configures verify_options and set_verify_golden_data to detect frontend graph construction errors before deeper debugging. - Three Localization Methods: File-based checkpoint comparison (pypto.pass_verify_save + torch.save), Pass-level PreCheck/PostCheck validation with pass_compare.py, and on-device binary search using checkpoint tensors injected into kernel signatures. - Sync and VF Fusion Diagnosis: Includes scripts to bisect CCE files with pipe_barrier insertion and map failing lines back to frontend source code. - Use Case: An Add operator's NPU output mismatches torch.add results; the Skill walks through tensor_graph verification, then routes to the appropriate sub-method to find the first failing op. ## Quick Start Ask the assistant to debug a PyPTO operator precision issue using this skill, optionally specifying a mode such as file-save comparison, Pass verification, or on-device binary search.

Frequently Asked Questions about pypto-precision-compare

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I debug PyPTO operator precision issues on Ascend NPU?▼

Start by enabling tensor_graph verification with verify_options and set_verify_golden_data, then check interpreter.log. Based on the result, use file-based checkpoint comparison, Pass-level PreCheck/PostCheck validation, or on-device binary search to locate the first failing op.

What is the difference between pass_verify_save and binary search precision debugging?▼

pass_verify_save writes intermediate results to files compared against torch.save golden data without changing kernel signatures. Binary search adds checkpoint tensors as kernel input parameters for in-memory on-device comparison, used when tensor graph and Pass checks both pass.

How do I use pass_compare.py to find a failing op?▼

Run pass_compare.py with --p specifying the failed Pass and the previous passing golden Pass, plus --verify_path pointing to the verify data directory. It generates a CSV with per-op PASS/FAIL results identifying the problematic op.

Why is my PyPTO verify log empty after running the test?▼

Empty verify logs usually mean golden data was not set or the order was wrong. You must compute the golden on CPU, call pypto.set_verify_golden_data, and only then execute the PyPTO operator, with matching input and output counts.

What does error code 0xB4001U mean in PyPTO verification?▼

Error 0xB4001U indicates VERIFY_RESULT_MISMATCH, meaning a tensor graph or Pass verification failed. A tensor_graph failure points to frontend code issues, while a Pass-level failure requires PreCheck/PostCheck analysis with pass_compare.

When should I remove verify_options and golden data settings?▼

Remove verify_options from the @pypto.frontend.jit decorator and delete set_verify_golden_data calls immediately after obtaining tensor_graph results. Leaving diagnostic configuration in production code can interfere with subsequent debugging and normal execution.

Related Skills

Explore More Software Engineering Skills→
AscendAscend
31

triton-operator-precision-eval

Benchmark Triton operator precision against PyTorch references and generate MERE/MARE reports.

Official
Advanced
DaddyElonMusk69DaddyElonMusk69

test-driven-development

Enforce test-first development with failing tests, minimal code, and regression checks.

Community
Intermediate
sd0xdevsd0xdev
189

codex-review-fast

Runs a quick diff-only code review using Codex exec without lint or build checks.

Community
Basic
tile-aitile-ai
346

tilelang-error-fixer

Diagnose and repair TileLang-Ascend Pass crashes via GDB backtraces, IR dumps, and AOT tests.

Official
Advanced