P1-FEATURE-007: Graph Optimizer — QLinear * Rewrite

[DingmaomaoBJTU]

Summary

Implement pattern-based rewriting for QLinear operators (QLinearConv, QLinearMatMul) to enable quantization-aware graph transformations needed for INT8/INT4 model deployment on NPUs.

Context

QLinear operators are ONNX's representation of quantized computations. Fusing and rewriting these operators (e.g., QLinearConv → fused INT8 kernel) is required for efficient execution on hardware NPUs that support quantized computation natively. This is the final graph optimizer capability needed for QDQ pipeline completion.

From:

• plans/release/0315_release_plan/P1_CHECKLIST.md (P1-FEATURE-006)
• plans/release/0501_release_plan/P0_CHECKLIST.md (P1-FEATURE-007)

Current State

• Graph optimizer supports FP32/FP16 rewrites (refactor(session): unify op-tracing under EPMonitor; fix WinML EP #397) and attribute changes (Build fails for DML/GPU: compiled.onnx missing when enable_ep_context=false #396)
• No QLinear-specific rewrite rules implemented
• QDQ pipeline (Lift derived head_dim into a shared NormalizedConfig base class #401) produces QLinear operators that need post-processing

Desired State

• QLinear * patterns recognized and rewritten by the graph optimizer
• QLinearConv + Bias fusion
• QLinearMatMul → INT8 GEMM rewrite (where EP supports it)
• Quantization-scale/zero-point folding

Acceptance Criteria

• QLinearConv fusion rule implemented and tested
• QLinearMatMul rewrite rule implemented and tested
• Scale/zero-point folding across consecutive QDQ patterns
• All QLinear rewrite rules tested against QDQ-quantized P0 models
• Runtime-validated: quantized model output matches pre-rewrite output within tolerance
• All existing tests pass (CARDINAL RULE: no regressions)

Technical Notes

• Must follow CARDINAL RULE This repo is missing a LICENSE file #1: no hardcoded model architecture assumptions — all pattern matching must be graph-structure-based
• QLinear op spec: https://onnx.ai/onnx/operators/onnx__QLinearConv.html
• Test with QDQ output from Lift derived head_dim into a shared NormalizedConfig base class #401 on at least 3 architectures (CNN, BERT, ViT)

Related Files

• plans/release/0315_release_plan/feature-scale.md — P1.4 QLinear Rewrite
• plans/release/0501_release_plan/feature-scale.md — P1.7 QLinear Rewrite
• plans/release/0315_release_plan/P1_CHECKLIST.md — P1-FEATURE-006
• plans/release/0501_release_plan/P0_CHECKLIST.md — P1-FEATURE-007
• Lift derived head_dim into a shared NormalizedConfig base class #401 — QDQ quantization (upstream dependency)