[asjad2401]

Stacked on #4 (which is stacked on #3) — merge those first.

Summary

ResNet-18 now runs end to end and matches ONNX Runtime across four real exports:

Model	Exporter	Opset	Max logit diff	Top-1 = ORT
Model Zoo resnet18-v1-7	MXNet	8	6.2e-6	✓
Model Zoo resnet18-v2-7	MXNet	7	8.1e-6	✓
torchvision resnet18	torch.onnx legacy	17	4.8e-6	✓
torchvision resnet18	torch.onnx dynamo	18	5.0e-6	✓

(batch 1 and 4, pretrained weights)

Changes

• ReduceMean / ReduceSum / ReduceMax / ReduceMin — the dynamo exporter writes global average pooling as ReduceMean over H, W.
• Pre-load anonymous static weights. onnx-ir passes many initializers (including every conv weight) as unnamed ValueSource::Static args. These were skipped at load and converted from TensorData on every forward pass. They're now keyed by data id (static::{id}) and loaded once. ~10% faster at batch 1 across the board.
• Fold Conv → BatchNorm at load time. W' = W·scale, b' = b·scale + offset, BN node removed. Only when the conv output feeds nothing but that BN (and isn't a graph output). Unfoldable BNs keep the existing mul+add path.
• BN fast path only for 4D inputs — the [1,C,1,1] pre-fused params turned a rank-2 [N,C] BN into a 4D output.
• Linear/Gemm bias now uses ctx.resolve, so it benefits from pre-loading too.

Performance (CPU, Apple M-series 10 cores, median of 10)

Model	batch 1 before → after	batch 8 before → after	ORT batch 8
Model Zoo v1	25.2 → 22.4 ms	279 → 103 ms	118 ms
Model Zoo v2	25.2 → 22.6 ms	292 → 186 ms	119 ms
torch legacy	24.7 → 22.6 ms	107 → 105 ms	119 ms
torch dynamo	28.2 → 22.8 ms	113 → 106 ms	119 ms

ORT is ~14.8 ms at batch 1. v2 still trails at batch 8 because its pre-activation BNs (after Add/Relu, not Conv) can't be folded and hit a slow path in burn-flex: a [1,C,1,1] broadcast multiply over [8,64,112,112] takes ~24 ms vs ~3.6 ms same-shape (numpy: 0.9 ms). Worth reporting upstream.

Testing

• tests/test_resnet.py (new): downloads + caches Model Zoo models, exports torchvision models if torch is installed, compares vs ORT; --bench prints timings.
• New CI job resnet18 end-to-end (CPU torch, models cached via actions/cache). Not added to required checks — your call.
• tests/test_ops.py: BN folding cases (folded / no-bias / conv output reused / conv output is graph output / BN after Relu / rank-2 BN) with node-count assertions, plus reductions. 43 checks pass.
• test_bridge.py, compare_ort.py, cargo fmt --check, clippy -D warnings clean.

Note: torch.export specializes size-1 dims, so a dynamo export traced with batch 1 bakes [1, 512] into the final Reshape. The test exports with batch 2.

🤖 Generated with Claude Code