Pattern Development Guide

This document describes the complete development workflow for Patterns (fused operator patterns) in the MindIE SD compilation module.

Development Workflow

Model Analysis -> Pattern Creation -> Three-Step Registration -> Unit Verification -> Integration Verification

Stage 1: Model Analysis

Locate the actual implementation code of target operators in the model source code (e.g., RMSNorm, RoPE, AdaLayerNorm) and extract complete code snippets as the basis for Pattern and testing.

Also determine parameter origin:

Parameter Origin	Pattern Path	Applicable Scenarios
functional API (e.g., F.rms_norm)	register_replacement	Parameters passed as function arguments
nn.Module (e.g., self.weight)	Custom Graph Pass	Parameters from module member variables

Stage 2: Create Pattern

First check whether existing patterns already cover the target operator. If not, create a new one following the non-invasive principle (always create new files; do not modify existing pattern files).

Choose the path based on parameter origin from Stage 1:

• register_replacement path: Create a PatternBase subclass (factory + closure), register in pattern_registry
• Custom Graph Pass path: Implement a custom FX graph traversal pass

Stage 3: Three-Step Registration

Involves modifications to 3 files, all as code additions (do not modify existing code):

1. patterns/__init__.py --- add to __all__ + from .xxx_pattern import XxxPatternGroup
2. passes/__init__.py --- add entry in the pattern_registry dictionary
3. compiliation_config.py --- add enable_xxx: bool = True in the FusionPatterns dataclass

Naming convention: config keys use the enable_<model>_<op> format.

Stage 4: Unit Verification

Verification criterion: cosine_similarity(compiled, original) > 2^-7.

Note: Passing unit tests does not guarantee the Pattern hits the full model. The test model and pattern share the same code and will necessarily match. Full-model matching requires final confirmation via Stage 5 integration verification.

Stage 5: Integration Verification

Three layers of verification in order; no skipping allowed:

Layer	Method	Confirmation
1. Pattern Hit	graph dump + logs	PatternMatchPass replace N count increases
2. Fusion Kernel	profiling -> kernel_details.csv	Fused kernels appear + original kernels disappear
3. Full Model Regression	dummy run + compile	Inference completes normally without crashes

Kernel name mapping:

Operator	Kernel Name
RMSNorm	rms_norm / RmsNorm
RoPE	npu_rotary_mul / RotaryMul
AdaLN	adaln / adln
GELU	FastGelu