`.webnn` to graph builder JS code conversion requires some special handling for API mismatches

[miaobin]

When converting ONNX transformer models like all-MiniLM-L6-v2 to WebNN graph builder JS code,
we discovered that some special case handling is required due to API design mismatches between
ONNX semantics and WebNN API conventions. For example:

1. WebNN spec requires lowercase type names, but generated intermediate representations often use capitalized forms (Impact cast, constant ops):

// Required by WebNN
"float32", "float16", "int32"

// Generated from AST
"Float32", "Float16", "Int32"

2. WebNN doesn't support negative axis indices, but ONNX commonly uses them (Impact concat, split, gather, reduceMean, reduceSum, softmax, layerNormalization ops):

# ONNX semantics
axis = -1  # means "last dimension"

# Must convert to WebNN
axis = rank - 1  # explicit positive index

3. For slice operation we must compute sizes = ends - starts and expand sparse axes to full rank:

// ONNX: starts, ends, steps, axes (sparse - only specified dims)
slice(input, starts=[0], ends=[128], axes=[1])

// WebNN: starts, sizes, strides (dense - all dims required)
slice(input, [0,0], [1,128], {strides: [1,1]})

4. For constant operation we need detect base64 encoded data and adds atob() to decode it into an ArrayBuffer.

[tarekziade]

Thanks for the feeback, will work on fixes for these

[tarekziade]

I've made some fixes and validated that the produced webnn works with pywebnn for that model (pywebnn has a demo with it)

For emit-js. you can try it with this script in this branch -- I don't have yet a full e2e test for verify JS like I do with Python, it would require driving chromium, so I am just assuming the produced script will work and we can try it manually (not done yet)

#!/usr/bin/env bash
set -euo pipefail

  # 1) Download ONNX from Hugging Face
  ONNX_URL="https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/resolve/main/onnx/model.onnx?download=true"
  ONNX_PATH="/tmp/minilm.onnx"

  curl -L "$ONNX_URL" -o "$ONNX_PATH"

  # 2) Convert ONNX -> WebNN graph + weights + manifest
  cargo run -- convert-onnx \
    --input "$ONNX_PATH" \
    --optimize \
    --output /tmp/minilm.webnn \
    --weights /tmp/minilm.weights \
    --manifest /tmp/minilm.manifest.json \
    --override-dim batch_size=1 \
    --override-dim sequence_length=128

  # 3) Emit JS builder code
  cargo run -- emit-js /tmp/minilm.webnn > /tmp/minilm.buildGraph.js

  # 4) Quick sanity checks
  echo "Generated files:"
  ls -lh /tmp/minilm.onnx /tmp/minilm.webnn /tmp/minilm.weights /tmp/minilm.manifest.json /tmp/minilm.buildGraph.js
  echo
  echo "JS preview:"
  head -n 30 /tmp/minilm.buildGraph.js

1. Type-name casing mismatch ("Float32" vs "float32")
   Conversion now normalizes emitted data type strings to WebNN-compatible lowercase forms across cast/constants/emission.

2. Negative axis handling (ONNX allows negative, WebNN doesn’t)

axis values are normalized to positive indices using rank-aware normalization before emission for affected ops (concat/split/gather/reductions/softmax/layer norm related paths).

3. Slice semantic mismatch (ONNX sparse axes + ends; WebNN dense starts/sizes/strides)

slice lowering now materializes full-rank dense vectors, computes sizes = ends - starts per dimension, and fills unspecified dims with full extents/stride 1.

4. Base64 constants needing JS decode (atob) for ArrayBuffer

constant emission detects encoded payloads and emits decode-capable JS path so runtime gets binary bytes correctly.

5. Extra regression we found while validating issue .webnn to graph builder JS code conversion requires some special handling for API mismatches #4

broadcast-aware constant folding/shape propagation for binary ops + Equal to avoid invalid flattened constants