precision-align

Diagnose precision mismatches between lmdeploy+dlinfer and reference implementations on Ascend, CAMB, and MACA backends.

74|15|Updated Jul 9, 2024
One-click install
npx skills add https://github.com/DeepLink-org/dlinfer --skill precision-align
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: precision-align
Source: https://github.com/DeepLink-org/dlinfer/tree/main/.claude/skills/zh_cn/precision-align
Command: npx skills add https://github.com/DeepLink-org/dlinfer --skill precision-align

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This Skill helps diagnose and fix precision mismatches between lmdeploy+dlinfer and reference implementations on domestically produced hardware backends (Ascend, CAMB, MACA).

Core Features & Use Cases

  • Layer-0 isolation: identify the first op or sub-operation with a discrepancy, starting from the initial RMSNorm/Attention/MLP stack.
  • Deterministic comparisons: enforce exact tensor equality where possible and report quantitative differences otherwise.
  • Guided debugging: provide tensor-dump guidance, input-output checks, and a stepwise path to root-cause precision drift across backends.

Quick Start

Compare layer-0 outputs between lmdeploy+dlinfer and a reference implementation on Ascend to locate the first mismatch.

Frequently Asked Questions about precision-align

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

FAQPage Schema
How do I diagnose precision drift between lmdeploy and reference implementations on Ascend?

To diagnose precision drift on Ascend, systematically compare lmdeploy+dlinfer outputs against reference implementations by tracing from layer 0 outward to pinpoint the first differing operation. This structured tensor-dump approach isolates the root cause of precision mismatches.

What is layer-0 isolation for debugging precision mismatches on domestic hardware?

Layer-0 isolation is a debugging technique that identifies the first operation with a precision discrepancy, starting from the initial RMSNorm, Attention, or MLP stack. It enforces deterministic comparisons to locate the exact origin of numerical drift.

Does this precision debugging workflow support CAMB and MACA backends?

Yes, the precision debugging workflow supports CAMB and MACA backends alongside Ascend. It applies systematic comparison and stepwise isolation across these domestic hardware backends to locate the root cause of precision drift.

How do I locate the first differing operation causing numerical mismatch in dlinfer?

To locate the first differing operation in dlinfer, enforce deterministic comparisons and perform structured tensor dumps. This stepwise isolation process traces outputs from layer 0 outward to pinpoint the exact operation causing the precision mismatch.

What is the best way to fix precision mismatches on domestic hardware backends?

The best way to fix precision mismatches is using a guided debugging process that enforces exact tensor equality checks and reports quantitative differences. This methodically traces the discrepancy to guide targeted fixes for numerical drift.