diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 6423320..9da2212 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -10,6 +10,8 @@ env:
# ONNX Runtime dylib for the reference backend. rustnn pins ort 2.0.0-rc.13
# with the `api-27` feature, so this must be >= 1.27 (any newer 1.x works).
ORT_VERSION: 1.29.0
+ RUSTNN_REPO: rustnn/rustnn
+ RUSTNN_REF: main
jobs:
test:
@@ -69,7 +71,8 @@ jobs:
id: rustnn
uses: actions/checkout@v4
with:
- repository: rustnn/rustnn
+ repository: ${{ env.RUSTNN_REPO }}
+ ref: ${{ env.RUSTNN_REF }}
path: rustnn
- name: Install Rust
@@ -150,3 +153,16 @@ jobs:
O2W_MODEL_TEST_JOBS: ${{ matrix.model_test_jobs || 2 }}
O2W_MODEL_TEST_SKIP_HEAVY: 1
run: cargo test --release --locked ${{ matrix.cargo_flags }}
+
+ - name: Validate curated full model
+ if: runner.os != 'Windows'
+ env:
+ HF_TOKEN: ${{ secrets.HF_TOKEN }}
+ O2W_ONNX_CACHE: ${{ runner.temp }}/onnx2webnn-ci-models
+ O2W_WEBNN_CACHE: ${{ runner.temp }}/onnx2webnn-ci-webnn
+ run: |
+ target/release/onnx2webnn validate-models \
+ --manifest tests/models/ci-validation.json \
+ --selection all \
+ --weights real \
+ --jobs 1
diff --git a/.github/workflows/model-validation.yml b/.github/workflows/model-validation.yml
new file mode 100644
index 0000000..dcbd604
--- /dev/null
+++ b/.github/workflows/model-validation.yml
@@ -0,0 +1,119 @@
+name: Full model validation
+
+on:
+ workflow_dispatch:
+
+env:
+ ORT_VERSION: 1.29.0
+ RUSTNN_REPO: rustnn/rustnn
+ RUSTNN_REF: main
+
+jobs:
+ validate:
+ name: Full real-weight sweep
+ runs-on: [self-hosted, linux, x64, onnx2webnn-validation]
+ timeout-minutes: 360
+ defaults:
+ run:
+ shell: bash
+ working-directory: onnx2webnn
+ steps:
+ - name: Checkout onnx2webnn
+ uses: actions/checkout@v4
+ with:
+ path: onnx2webnn
+
+ - name: Checkout rustnn
+ uses: actions/checkout@v4
+ with:
+ repository: ${{ env.RUSTNN_REPO }}
+ ref: ${{ env.RUSTNN_REF }}
+ path: rustnn
+
+ - name: Check runner contract
+ working-directory: ${{ github.workspace }}
+ run: |
+ cache=/var/cache/onnx2webnn
+ test -d "$cache"
+ test -w "$cache"
+ filesystem_bytes=$(df -PB1 "$cache" | awk 'NR == 2 { print $2 }')
+ memory_bytes=$(awk '/MemTotal/ { print $2 * 1024 }' /proc/meminfo)
+ minimum_filesystem=$((100 * 1024 * 1024 * 1024))
+ minimum_memory=$((30 * 1024 * 1024 * 1024))
+ if [ "$filesystem_bytes" -lt "$minimum_filesystem" ]; then
+ echo "Runner filesystem must be at least 100 GiB" >&2
+ exit 1
+ fi
+ if [ "$memory_bytes" -lt "$minimum_memory" ]; then
+ echo "Runner memory must be approximately 32 GiB (30 GiB minimum)" >&2
+ exit 1
+ fi
+
+ - name: Install Rust
+ uses: dtolnay/rust-toolchain@stable
+ with:
+ components: rustfmt
+
+ - name: Install protobuf compiler
+ uses: arduino/setup-protoc@v3
+ with:
+ repo-token: ${{ secrets.GITHUB_TOKEN }}
+
+ - name: Cache cargo registry
+ uses: actions/cache@v4
+ with:
+ path: |
+ ~/.cargo/registry
+ ~/.cargo/git
+ key: full-validation-cargo-${{ hashFiles('onnx2webnn/Cargo.lock') }}
+ restore-keys: full-validation-cargo-
+
+ - name: Download ONNX Runtime ${{ env.ORT_VERSION }}
+ run: |
+ archive=onnxruntime-linux-x64-${ORT_VERSION}
+ url=https://github.com/microsoft/onnxruntime/releases/download/v${ORT_VERSION}/${archive}.tgz
+ mkdir -p "$RUNNER_TEMP/ort"
+ curl -fsSL "$url" -o "$RUNNER_TEMP/ort/archive.tgz"
+ tar -xzf "$RUNNER_TEMP/ort/archive.tgz" -C "$RUNNER_TEMP/ort"
+ dylib="$RUNNER_TEMP/ort/$archive/lib/libonnxruntime.so"
+ if [ ! -f "$dylib" ]; then
+ dylib=$(find "$RUNNER_TEMP/ort/$archive/lib" -maxdepth 1 -name 'libonnxruntime.so.*' -print -quit)
+ fi
+ test -f "$dylib"
+ echo "ORT_DYLIB_PATH=$dylib" >> "$GITHUB_ENV"
+
+ - name: Build release binary
+ run: cargo build --release --locked
+
+ - name: Run diagnostic full sweep
+ id: validation
+ env:
+ HF_TOKEN: ${{ secrets.HF_TOKEN }}
+ O2W_ONNX_CACHE: /var/cache/onnx2webnn
+ O2W_WEBNN_CACHE: ${{ runner.temp }}/onnx2webnn-webnn
+ run: |
+ set +e
+ target/release/onnx2webnn validate-models \
+ --manifest tests/models/manifest.json \
+ --selection all \
+ --weights real \
+ --jobs 1 2>&1 | tee "$RUNNER_TEMP/model-validation.log"
+ status=${PIPESTATUS[0]}
+ set -e
+ echo "exit_code=$status" >> "$GITHUB_OUTPUT"
+ if [ "$status" -eq 0 ]; then
+ echo "## Full model validation: passed" >> "$GITHUB_STEP_SUMMARY"
+ else
+ echo "::warning::Full model validation reported failures (exit code $status); see the summary and uploaded log."
+ echo "## Full model validation: diagnostic failures" >> "$GITHUB_STEP_SUMMARY"
+ echo "The validator exited with code $status. This manually dispatched sweep is diagnostic and does not fail the workflow." >> "$GITHUB_STEP_SUMMARY"
+ fi
+
+ - name: Upload validation log
+ if: always()
+ uses: actions/upload-artifact@v4
+ with:
+ name: full-model-validation-real
+ path: ${{ runner.temp }}/model-validation.log
+ if-no-files-found: warn
+ retention-days: 14
diff --git a/.gitignore b/.gitignore
index 955e513..c6030c8 100644
--- a/.gitignore
+++ b/.gitignore
@@ -11,3 +11,7 @@ __pycache__/
.vscode/
.idea/
.DS_Store
+
+# Model caches
+.webnn-cache/
+.onnx-cache/
diff --git a/AGENTS.md b/AGENTS.md
index f12c8f1..dd450cd 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -25,20 +25,23 @@ independent layers:
|-------|----------------|-------|-----------------|
| **1. WebNN spec** | Ops the W3C API defines | [webnn spec](https://www.w3.org/TR/webnn/) | ~105 |
| **2. Name mapping** | ONNX op name ↔ WebNN `MLGraphBuilder` method | `webnn-onnx-utils/src/operation_names.rs` | ~90 |
-| **3. Exporter implementation** | ONNX ops with a Rust lowering handler | `src/onnx/ops/*.rs` + `scripts/webnn_onnx_ops.py` | ~59 |
+| **3. Exporter implementation** | Advertised ONNX names with a lowering or preprocessing path | `src/onnx/ops/*.rs` + `scripts/webnn_onnx_ops.py` | 135 |
An ONNX op can be:
-- **No WebNN target** — e.g. `If`, `Loop`, `Scan` (control flow), `StringConcat` (no string
- tensors), `Compress`, `Einsum`, `Attention`. WebNN graphs are static DAGs; these are permanently
- rejected. See `docs/operator-conversion-plan.md` Stage 3.
-- **Mapped but not implemented** — e.g. `BatchNormalization`, `InstanceNormalization`, `ArgMax`,
- `GatherND`, `Resize`. Listed in `operation_names.rs` but no handler in `src/onnx/ops/` yet.
- Tests expect `UnsupportedOp` until a handler lands.
-- **Implemented** — listed in `scripts/webnn_onnx_ops.py` and handled by `OpRegistry`. Tests
- expect `Success` (convert + ORT vs rustnn output match).
+- **No current lowering** — for example `Loop`, `Scan`, string operators, and general
+ sequence operators. Some have no portable WebNN representation.
+- **Pattern-only or partial** — for example constant-condition `If`, the narrow
+ `SplitToSequence` + `SequenceAt` export pattern, static `OneHot`, and restricted
+ `Einsum`. The handler rejects unsupported forms.
+- **Direct or decomposed** — advertised in `scripts/webnn_onnx_ops.py`. A decomposed op may be
+ functionally supported while remaining slower or narrower than a native backend kernel.
-The test manifest (`webnn_onnx_ops.py`) mirrors layer 3 only. Regenerate tests after changing it.
+The manifest contains 135 accepted names. At opset 26, 127 are standard `ai.onnx` names;
+adding the registry-only `SplitToSequence`/`SequenceAt` pattern gives 129/198 (65.2%) standard
+names with at least one executable lowering. Of those, 96/198 (48.5%) have a direct standardized
+WebNN primitive path. Name presence is not a promise of every ONNX schema variant. See
+`docs/operator-conversion-status.md` for the complete inventory and WebNN-imposed limitations. Regenerate tests after changing the manifest.
## Build and test
@@ -104,9 +107,10 @@ Operators with multiple pre-9 schema revisions skip sub-opset-9 structure bands
until legacy attribute normalization lands; newer bands use the true latest schema
revision in range (e.g. `MaxRoiPool` at model opset 22, `Pad` at 9/17/26).
-Operator dispatch and the unsupported-op pre-scan key on **`op_type` only** (standard
-`ai.onnx` domain). Custom-domain ops are not supported yet; adding them requires
-domain-aware handler registration and matching pre-scan logic in `convert.rs`.
+Operator dispatch and the unsupported-op pre-scan key on **`op_type` only**, not on the ONNX
+domain. This currently permits selected `com.microsoft` fused operators such as MoE/QMoE, but
+it is not general custom-domain support and cannot distinguish same-named operators across
+domains. Proper domain-aware support requires matching registry and pre-scan changes.
## Layout (committed)
@@ -147,7 +151,8 @@ history and tests the highest buildable opset in each band — e.g. `Pad` at ops
(attribute `pads` vs input `pads` vs optional `axes` input). Ops with unchanged structure get a
single test at the newest opset. Unbuildable ops get an `#[ignore]` stub.
-See [docs/operator-conversion-plan.md](docs/operator-conversion-plan.md) for the full operator rollout workflow.
+See [docs/operator-conversion-status.md](docs/operator-conversion-status.md) for the complete
+per-operator capability matrix.
## Dependencies (`Cargo.toml`)
diff --git a/Cargo.lock b/Cargo.lock
index 59f099f..1d16dab 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -82,6 +82,15 @@ version = "1.0.102"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "7f202df86484c868dbad7eaa557ef785d5c66295e41b460ef922eca0723b842c"
+[[package]]
+name = "approx"
+version = "0.5.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cab112f0a86d568ea0e627cc1d6be74a1e9cd55214684db5561995f6dad897c6"
+dependencies = [
+ "num-traits",
+]
+
[[package]]
name = "aquamarine"
version = "0.1.12"
@@ -95,6 +104,29 @@ dependencies = [
"syn 1.0.109",
]
+[[package]]
+name = "arrayvec"
+version = "0.7.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d3fb67a6e08acf24fdeccbac2cb6ac4305825bd1f117462e0e6f2f193345ad56"
+
+[[package]]
+name = "async-trait"
+version = "0.1.92"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "82f6aeea286b8eb4dd3431a1be1b59d290ace00f5bfd8e2a159bc2a05e2c1667"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 3.0.4",
+]
+
+[[package]]
+name = "atomic-waker"
+version = "1.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1505bd5d3d116872e7271a6d4e16d81d0c8570876c8de68093a09ac269d8aac0"
+
[[package]]
name = "atty"
version = "0.2.14"
@@ -133,7 +165,7 @@ dependencies = [
"bitflags",
"cexpr",
"clang-sys",
- "itertools 0.12.1",
+ "itertools 0.10.5",
"log",
"proc-macro2",
"quote",
@@ -216,12 +248,41 @@ dependencies = [
"thiserror 1.0.69",
]
+[[package]]
+name = "aws-lc-rs"
+version = "1.18.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b281d307588d634de920874890732659e2e7672f72b5e10e81badc1a8a83621e"
+dependencies = [
+ "aws-lc-sys",
+ "zeroize",
+]
+
+[[package]]
+name = "aws-lc-sys"
+version = "0.45.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9bff6c3b54fad79a2e60b8102caf565819711497c1f5f092f49508e2f5c31b27"
+dependencies = [
+ "cc",
+ "cmake",
+ "dunce",
+ "fs_extra",
+ "pkg-config",
+]
+
[[package]]
name = "base64"
version = "0.22.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "72b3254f16251a8381aa12e40e3c4d2f0199f8c6508fbecb9d91f575e0fbb8c6"
+[[package]]
+name = "base64"
+version = "0.23.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ac07cdecf99051d9a5238b80f35af32cdeba5b336e55d957b318b50137e18da5"
+
[[package]]
name = "base64ct"
version = "1.8.3"
@@ -237,7 +298,7 @@ dependencies = [
"bitflags",
"cexpr",
"clang-sys",
- "itertools 0.12.1",
+ "itertools 0.10.5",
"log",
"prettyplease",
"proc-macro2",
@@ -254,6 +315,19 @@ version = "2.13.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "b588b76d00fde79687d7646a9b5bdf3cc0f655e0bbd080335a95d7e96f3587da"
+[[package]]
+name = "blake3"
+version = "1.8.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6d9e454fc11f76977dc803893aff6304ed33d6a26efae8696573bea74baa27ae"
+dependencies = [
+ "arrayvec",
+ "cc",
+ "cfg-if",
+ "constant_time_eq",
+ "cpufeatures 0.3.1",
+]
+
[[package]]
name = "block"
version = "0.1.6"
@@ -269,6 +343,15 @@ dependencies = [
"generic-array",
]
+[[package]]
+name = "block-buffer"
+version = "0.12.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d2f6c7dbe95a6ed67ad9f18e57daf93a2f034c524b99fd2b76d18fdfeb6660aa"
+dependencies = [
+ "hybrid-array",
+]
+
[[package]]
name = "bs58"
version = "0.5.1"
@@ -278,6 +361,17 @@ dependencies = [
"tinyvec",
]
+[[package]]
+name = "bstr"
+version = "1.13.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6bb31b46c14244e20ee9984b11bf5c992b91fb6939fea616e3512c8baecdbe5f"
+dependencies = [
+ "memchr",
+ "regex-automata",
+ "serde_core",
+]
+
[[package]]
name = "bumpalo"
version = "3.20.3"
@@ -329,6 +423,23 @@ version = "1.0.4"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "9330f8b2ff13f34540b44e946ef35111825727b38d33286ef986142615121801"
+[[package]]
+name = "cfg_aliases"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f079e83a288787bcd14a6aea84cee5c87a67c5a3e660c30f557a3d24761b3527"
+
+[[package]]
+name = "chacha20"
+version = "0.10.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "65c35e4b699c7e15ccbe7ee35c005e4fc0a278d22238a2857e6ce2dadeda1b06"
+dependencies = [
+ "cfg-if",
+ "cpufeatures 0.3.1",
+ "rand_core",
+]
+
[[package]]
name = "chrono"
version = "0.4.44"
@@ -336,8 +447,10 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "c673075a2e0e5f4a1dde27ce9dee1ea4558c7ffe648f576438a20ca1d2acc4b0"
dependencies = [
"iana-time-zone",
+ "js-sys",
"num-traits",
"serde",
+ "wasm-bindgen",
"windows-link",
]
@@ -392,6 +505,15 @@ version = "1.1.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "c8d4a3bb8b1e0c1050499d1815f5ab16d04f0959b233085fb31653fbfc9d98f9"
+[[package]]
+name = "cmake"
+version = "0.1.58"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c0f78a02292a74a88ac736019ab962ece0bc380e3f977bf72e376c5d78ff0678"
+dependencies = [
+ "cc",
+]
+
[[package]]
name = "codespan-reporting"
version = "0.13.1"
@@ -409,6 +531,62 @@ version = "1.0.5"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "1d07550c9036bf2ae0c684c4297d503f838287c83c53686d05370d0e139ae570"
+[[package]]
+name = "colored"
+version = "3.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "faf9468729b8cbcea668e36183cb69d317348c2e08e994829fb56ebfdfbaac34"
+dependencies = [
+ "windows-sys 0.52.0",
+]
+
+[[package]]
+name = "combine"
+version = "4.6.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cfc320937d09e6de266b31b9afb480f197d7a861be86be7cb2ea7e5d1bfffc5e"
+dependencies = [
+ "bytes",
+ "memchr",
+]
+
+[[package]]
+name = "const-oid"
+version = "0.10.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a6ef517f0926dd24a1582492c791b6a4818a4d94e789a334894aa15b0d12f55c"
+
+[[package]]
+name = "const-str"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "18f12cc9948ed9604230cdddc7c86e270f9401ccbe3c2e98a4378c5e7632212f"
+
+[[package]]
+name = "const_panic"
+version = "0.2.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9603f79528ece8163c496f8932121cb36cfe46259e9c907bb3d8205139d7caa3"
+dependencies = [
+ "typewit",
+]
+
+[[package]]
+name = "constant_time_eq"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3d52eff69cd5e647efe296129160853a42795992097e8af39800e1060caeea9b"
+
+[[package]]
+name = "core-foundation"
+version = "0.9.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "91e195e091a93c46f7102ec7818a2aa394e1e1771c3ab4825963fa03e45afb8f"
+dependencies = [
+ "core-foundation-sys",
+ "libc",
+]
+
[[package]]
name = "core-foundation"
version = "0.10.1"
@@ -425,6 +603,21 @@ version = "0.8.7"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "773648b94d0e5d620f64f280777445740e61fe701025087ec8b57f45c791888b"
+[[package]]
+name = "core_detect"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7f8f80099a98041a3d1622845c271458a2d73e688351bf3cb999266764b81d48"
+
+[[package]]
+name = "countio"
+version = "0.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b9702aee5d1d744c01d82f6915644f950f898e014903385464c773b96fefdecb"
+dependencies = [
+ "futures-io",
+]
+
[[package]]
name = "cpufeatures"
version = "0.2.17"
@@ -434,6 +627,39 @@ dependencies = [
"libc",
]
+[[package]]
+name = "cpufeatures"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5ca28b0ae3115b884660db4118d803791fd6756b6e88f39c0f3f7859060d7566"
+dependencies = [
+ "libc",
+]
+
+[[package]]
+name = "crc32fast"
+version = "1.5.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "01a7799fd6b852db0e61728dde9a204c423b44d689dbd432522543614b490e78"
+dependencies = [
+ "cfg-if",
+]
+
+[[package]]
+name = "crossbeam-channel"
+version = "0.5.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "98b0cc327b5bc766e7fda9c9260cc0fa81b43a8e240440422dff70788e3f9ef1"
+dependencies = [
+ "crossbeam-utils",
+]
+
+[[package]]
+name = "crossbeam-utils"
+version = "0.8.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a31eee39dddec8330830986fcd7625edb5a24ec90ea038215273bbc3adb08ac6"
+
[[package]]
name = "crunchy"
version = "0.2.4"
@@ -450,6 +676,25 @@ dependencies = [
"typenum",
]
+[[package]]
+name = "crypto-common"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ce6e4c961d6cd6c9a86db418387425e8bdeaf05b3c8bc1411e6dca4c252f1453"
+dependencies = [
+ "hybrid-array",
+]
+
+[[package]]
+name = "ctor"
+version = "1.0.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "914a755b7c2d4af2bdcff7ce1739e2db9a1b81a9b07123d8015786ae03c0980d"
+dependencies = [
+ "link-section",
+ "linktime-proc-macro",
+]
+
[[package]]
name = "cudarc"
version = "0.11.9"
@@ -603,8 +848,19 @@ version = "0.10.7"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "9ed9a281f7bc9b7576e61468ba615a66a5c8cfdff42420a70aa82701a3b1e292"
dependencies = [
- "block-buffer",
- "crypto-common",
+ "block-buffer 0.10.4",
+ "crypto-common 0.1.7",
+]
+
+[[package]]
+name = "digest"
+version = "0.11.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f1dd6dbb5841937940781866fa1281a1ff7bd3bf827091440879f9994983d5c2"
+dependencies = [
+ "block-buffer 0.12.1",
+ "const-oid",
+ "crypto-common 0.2.2",
]
[[package]]
@@ -639,6 +895,12 @@ dependencies = [
"syn 2.0.117",
]
+[[package]]
+name = "dunce"
+version = "1.0.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "92773504d58c093f6de2459af4af33faa518c13451eb8f2b5698ed3d36e7c813"
+
[[package]]
name = "dyn-clone"
version = "1.0.20"
@@ -651,6 +913,21 @@ version = "1.16.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "91622ff5e7162018101f2fea40d6ebf4a78bbe5a49736a2020649edf9693679e"
+[[package]]
+name = "encoding_rs"
+version = "0.8.41"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7b5ef0006ac9ab233c38522f5ae99cae3625151de8f706cacee1cba4b8e2832a"
+dependencies = [
+ "cfg-if",
+ "core_detect",
+ "multiversion",
+ "multiversion_no_op",
+ "rustversion",
+ "scopeguard",
+ "simdutf8",
+]
+
[[package]]
name = "env_logger"
version = "0.9.3"
@@ -690,7 +967,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
dependencies = [
"libc",
- "windows-sys 0.52.0",
+ "windows-sys 0.61.2",
]
[[package]]
@@ -711,6 +988,12 @@ version = "0.4.2"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "0ce7134b9999ecaf8bcd65542e436736ef32ddca1b3e06094cb6ec5755203b80"
+[[package]]
+name = "fnv"
+version = "1.0.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3f9eec918d3f24069decb9af1554cad7c880e2da24a9afd88aca000531ab82c1"
+
[[package]]
name = "foldhash"
version = "0.1.5"
@@ -747,12 +1030,88 @@ dependencies = [
"percent-encoding",
]
+[[package]]
+name = "fs4"
+version = "0.13.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8640e34b88f7652208ce9e88b1a37a2ae95227d84abec377ccd3c5cfeb141ed4"
+dependencies = [
+ "rustix",
+ "tokio",
+ "windows-sys 0.59.0",
+]
+
+[[package]]
+name = "fs_extra"
+version = "1.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "42703706b716c37f96a77aea830392ad231f44c9e9a67872fa5548707e11b11c"
+
+[[package]]
+name = "futures"
+version = "0.3.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8b147ee9d1f6d097cef9ce628cd2ee62288d963e16fb287bd9286455b241382d"
+dependencies = [
+ "futures-channel",
+ "futures-core",
+ "futures-executor",
+ "futures-io",
+ "futures-sink",
+ "futures-task",
+ "futures-util",
+]
+
+[[package]]
+name = "futures-channel"
+version = "0.3.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "07bbe89c50d7a535e539b8c17bc0b49bdb77747034daa8087407d655f3f7cc1d"
+dependencies = [
+ "futures-core",
+ "futures-sink",
+]
+
[[package]]
name = "futures-core"
version = "0.3.32"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "7e3450815272ef58cec6d564423f6e755e25379b217b0bc688e295ba24df6b1d"
+[[package]]
+name = "futures-executor"
+version = "0.3.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "baf29c38818342a3b26b5b923639e7b1f4a61fc5e76102d4b1981c6dc7a7579d"
+dependencies = [
+ "futures-core",
+ "futures-task",
+ "futures-util",
+]
+
+[[package]]
+name = "futures-io"
+version = "0.3.34"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "53c0fa8157de1303bfffdaa1cc2a673bfffb60102f76b0ef4441659124373fed"
+
+[[package]]
+name = "futures-macro"
+version = "0.3.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e835b70203e41293343137df5c0664546da5745f82ec9b84d40be8336958447b"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
+[[package]]
+name = "futures-sink"
+version = "0.3.34"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1944426bf7d03f1d14f708785e4b33efd750b36d48a157b836b3efc15ede8e1d"
+
[[package]]
name = "futures-task"
version = "0.3.32"
@@ -765,12 +1124,23 @@ version = "0.3.32"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "389ca41296e6190b48053de0321d02a77f32f8a5d2461dd38762c0593805c6d6"
dependencies = [
+ "futures-channel",
"futures-core",
+ "futures-io",
+ "futures-macro",
+ "futures-sink",
"futures-task",
+ "memchr",
"pin-project-lite",
"slab",
]
+[[package]]
+name = "gearhash"
+version = "0.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "616e8f476a586f1b078d9eece21f7a071f27997e709e2b471f5697deea53bda9"
+
[[package]]
name = "generic-array"
version = "0.14.7"
@@ -788,8 +1158,10 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "ff2abc00be7fca6ebc474524697ae276ad847ad0a6b3faa4bcb027e9a4614ad0"
dependencies = [
"cfg-if",
+ "js-sys",
"libc",
- "wasi",
+ "wasi 0.11.1+wasi-snapshot-preview1",
+ "wasm-bindgen",
]
[[package]]
@@ -799,21 +1171,76 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "0de51e6874e94e7bf76d726fc5d13ba782deca734ff60d5bb2fb2607c7406555"
dependencies = [
"cfg-if",
+ "js-sys",
"libc",
"r-efi",
+ "rand_core",
"wasip2",
"wasip3",
+ "wasm-bindgen",
]
[[package]]
-name = "glob"
-version = "0.3.4"
+name = "git-version"
+version = "0.3.9"
source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e4eba85ea1d0a966a983acd07deee566e67395d2d96b6fb39e62b5a833f1eb0b"
+checksum = "1ad568aa3db0fcbc81f2f116137f263d7304f512a1209b35b85150d3ef88ad19"
+dependencies = [
+ "git-version-macro",
+]
[[package]]
-name = "half"
-version = "2.7.1"
+name = "git-version-macro"
+version = "0.3.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "53010ccb100b96a67bc32c0175f0ed1426b31b655d562898e57325f81c023ac0"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
+[[package]]
+name = "glob"
+version = "0.3.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e4eba85ea1d0a966a983acd07deee566e67395d2d96b6fb39e62b5a833f1eb0b"
+
+[[package]]
+name = "globset"
+version = "0.4.19"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e47d37d2ae4464254884b60ab7071be2b876a9c35b696bd018ddcc76847309cd"
+dependencies = [
+ "aho-corasick",
+ "bstr",
+ "log",
+ "regex-automata",
+ "regex-syntax 0.8.10",
+]
+
+[[package]]
+name = "h2"
+version = "0.4.19"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ef8e5e5a340588f4452631496976cf8636d4a7ecf600239fdc27615d2530bc16"
+dependencies = [
+ "atomic-waker",
+ "bytes",
+ "fnv",
+ "futures-core",
+ "futures-sink",
+ "http",
+ "indexmap 2.14.0",
+ "slab",
+ "tokio",
+ "tokio-util",
+ "tracing",
+]
+
+[[package]]
+name = "half"
+version = "2.7.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "6ea2d84b969582b4b1864a92dc5d27cd2b77b622a8d79306834f1be5ba20d84b"
dependencies = [
@@ -856,6 +1283,12 @@ version = "0.17.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "ed5909b6e89a2db4456e54cd5f673791d7eca6732202bbf2a9cc504fe2f9b84a"
+[[package]]
+name = "heapify"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0049b265b7f201ca9ab25475b22b47fe444060126a51abe00f77d986fc5cc52e"
+
[[package]]
name = "heck"
version = "0.5.0"
@@ -883,6 +1316,31 @@ version = "0.4.3"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "7f24254aa9a54b5c858eaee2f5bccdb46aaf0e486a595ed5fd8f86ba55232a70"
+[[package]]
+name = "hf-xet"
+version = "1.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c237ef4fb0ce1962a5117f8bd8c74454b41629826a9df17d14a1840ca18f0754"
+dependencies = [
+ "anyhow",
+ "async-trait",
+ "bytes",
+ "http",
+ "more-asserts",
+ "serde",
+ "serde_json",
+ "thiserror 2.0.18",
+ "tokio",
+ "tokio-util",
+ "tokio_with_wasm",
+ "tracing",
+ "uuid",
+ "xet-client",
+ "xet-core-structures",
+ "xet-data",
+ "xet-runtime",
+]
+
[[package]]
name = "hmac-sha256"
version = "1.1.14"
@@ -899,18 +1357,138 @@ dependencies = [
"itoa",
]
+[[package]]
+name = "http-body"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ca2a8f2913ee65f60facd6a5905613afaa448497a0230cc41ce022d93290bc2c"
+dependencies = [
+ "bytes",
+ "http",
+]
+
+[[package]]
+name = "http-body-util"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "23169fe34a5fbcdd3f3862e78fb9b6fccd5f02a6dc6f732547005d45631ce71c"
+dependencies = [
+ "bytes",
+ "futures-core",
+ "http",
+ "http-body",
+ "pin-project-lite",
+]
+
[[package]]
name = "httparse"
version = "1.10.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "6dbf3de79e51f3d586ab4cb9d5c3e2c14aa28ed23d180cf89b4df0454a69cc87"
+[[package]]
+name = "huggingface-hub"
+version = "0.0.1"
+source = "git+https://github.com/huggingface/huggingface_hub_rust.git?rev=20acfd61a767dcf7d6a0cb7f4a02ac8406fd8a14#20acfd61a767dcf7d6a0cb7f4a02ac8406fd8a14"
+dependencies = [
+ "async-trait",
+ "base64 0.22.1",
+ "bytes",
+ "fs4",
+ "futures",
+ "globset",
+ "hf-xet",
+ "pathdiff",
+ "reqwest",
+ "reqwest-middleware",
+ "reqwest-retry",
+ "serde",
+ "serde_json",
+ "sha2 0.10.9",
+ "thiserror 2.0.18",
+ "tokio",
+ "tokio-util",
+ "tracing",
+ "typed-builder",
+ "url",
+]
+
[[package]]
name = "humantime"
version = "2.3.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "135b12329e5e3ce057a9f972339ea52bc954fe1e9358ef27f95e89716fbc5424"
+[[package]]
+name = "hybrid-array"
+version = "0.4.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "27f864f10dfb56725ce5ce5472bc52252c8f93a4ab86327122cebf62c5f59a17"
+dependencies = [
+ "typenum",
+]
+
+[[package]]
+name = "hyper"
+version = "1.11.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "27b501faa50e7a26c3d3560ca625132f4078a17771f4810baf70475ae48cbe43"
+dependencies = [
+ "atomic-waker",
+ "bytes",
+ "futures-channel",
+ "futures-core",
+ "h2",
+ "http",
+ "http-body",
+ "httparse",
+ "itoa",
+ "pin-project-lite",
+ "smallvec",
+ "tokio",
+ "want",
+]
+
+[[package]]
+name = "hyper-rustls"
+version = "0.27.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dfa8e654703247911e29c23fbeaa261834bd9bb74efba2f9acddc37bfb127f53"
+dependencies = [
+ "http",
+ "hyper",
+ "hyper-util",
+ "rustls",
+ "tokio",
+ "tokio-rustls",
+ "tower-service",
+]
+
+[[package]]
+name = "hyper-util"
+version = "0.1.20"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "96547c2556ec9d12fb1578c4eaf448b04993e7fb79cbaad930a656880a6bdfa0"
+dependencies = [
+ "base64 0.22.1",
+ "bytes",
+ "futures-channel",
+ "futures-util",
+ "http",
+ "http-body",
+ "hyper",
+ "ipnet",
+ "libc",
+ "percent-encoding",
+ "pin-project-lite",
+ "socket2",
+ "system-configuration",
+ "tokio",
+ "tower-service",
+ "tracing",
+ "windows-registry",
+]
+
[[package]]
name = "iana-time-zone"
version = "0.1.65"
@@ -1079,6 +1657,12 @@ version = "1.0.9"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "bfa799dd5ed20a7e349f3b4639aa80d74549c81716d9ec4f994c9b5815598306"
+[[package]]
+name = "ipnet"
+version = "2.12.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "791930b43c0d5973160d90a8f3894509f2b273430f5c5c73b668636d0287c5c0"
+
[[package]]
name = "is-terminal"
version = "0.4.17"
@@ -1087,7 +1671,7 @@ checksum = "3640c1c38b8e4e43584d8df18be5fc6b0aa314ce6ebf51b53313d4306cca8e46"
dependencies = [
"hermit-abi 0.5.2",
"libc",
- "windows-sys 0.52.0",
+ "windows-sys 0.61.2",
]
[[package]]
@@ -1116,9 +1700,9 @@ dependencies = [
[[package]]
name = "itertools"
-version = "0.12.1"
+version = "0.14.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ba291022dbbd398a455acf126c1e341954079855bc60dfdda641363bd6922569"
+checksum = "2b192c782037fadd9cfa75548310488aabdbf3d2da73885b31bd0abd03351285"
dependencies = [
"either",
]
@@ -1129,6 +1713,55 @@ version = "1.0.18"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
+[[package]]
+name = "jni"
+version = "0.22.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5efd9a482cf3a427f00d6b35f14332adc7902ce91efb778580e180ff90fa3498"
+dependencies = [
+ "cfg-if",
+ "combine",
+ "jni-macros",
+ "jni-sys",
+ "log",
+ "simd_cesu8",
+ "thiserror 2.0.18",
+ "walkdir",
+ "windows-link",
+]
+
+[[package]]
+name = "jni-macros"
+version = "0.22.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a00109accc170f0bdb141fed3e393c565b6f5e072365c3bd58f5b062591560a3"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "rustc_version",
+ "simd_cesu8",
+ "syn 2.0.117",
+]
+
+[[package]]
+name = "jni-sys"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c6377a88cb3910bee9b0fa88d4f42e1d2da8e79915598f65fb0c7ee14c878af2"
+dependencies = [
+ "jni-sys-macros",
+]
+
+[[package]]
+name = "jni-sys-macros"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "38c0b942f458fe50cdac086d2f946512305e5631e720728f2a61aabcd47a6264"
+dependencies = [
+ "quote",
+ "syn 2.0.117",
+]
+
[[package]]
name = "jobserver"
version = "0.1.35"
@@ -1151,6 +1784,29 @@ dependencies = [
"wasm-bindgen",
]
+[[package]]
+name = "konst"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f660d5f887e3562f9ab6f4a14988795b694099d66b4f5dedc02d197ba9becb1d"
+dependencies = [
+ "const_panic",
+ "konst_proc_macros",
+ "typewit",
+]
+
+[[package]]
+name = "konst_proc_macros"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e037a2e1d8d5fdbd49b16a4ea09d5d6401c1f29eca5ff29d03d3824dba16256a"
+
+[[package]]
+name = "lazy_static"
+version = "1.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe"
+
[[package]]
name = "leb128fmt"
version = "0.1.0"
@@ -1201,6 +1857,18 @@ dependencies = [
"cc",
]
+[[package]]
+name = "link-section"
+version = "0.19.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "39c29a617ce3df32c08497bdc1ab6e2376e0b17948ac166a2fbe5977c5954cd9"
+
+[[package]]
+name = "linktime-proc-macro"
+version = "0.2.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7e57c38c1e860fd37c604281cdfb1dd2216977fd76a50f85ba2f388ef3219616"
+
[[package]]
name = "linux-raw-sys"
version = "0.12.1"
@@ -1213,12 +1881,36 @@ version = "0.8.2"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "92daf443525c4cce67b150400bc2316076100ce0b3686209eb8cf3c31612e6f0"
+[[package]]
+name = "lock_api"
+version = "0.4.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "224399e74b87b5f3557511d98dff8b14089b3dadafcab6bb93eab67d3aace965"
+dependencies = [
+ "scopeguard",
+]
+
[[package]]
name = "log"
version = "0.4.31"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "113b30b4cd05f7c06868fdb2854f66a7b9fece9a48425351cd532e810d74024f"
+[[package]]
+name = "lru-slab"
+version = "0.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4050469837a6ff301cd14c1f8f24f88549e6d548f24f64e2148eb0f72cebc51f"
+
+[[package]]
+name = "lz4_flex"
+version = "0.13.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7ef0d4ed8669f8f8826eb00dc878084aa8f253506c4fd5e8f58f5bce72ddb97e"
+dependencies = [
+ "twox-hash",
+]
+
[[package]]
name = "lzma-rust2"
version = "0.15.8"
@@ -1234,6 +1926,15 @@ dependencies = [
"libc",
]
+[[package]]
+name = "matchers"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d1525a2a28c7f4fa0fc98bb91ae755d1e2d1505079e05539e35bc876b5d65ae9"
+dependencies = [
+ "regex-automata",
+]
+
[[package]]
name = "matrixmultiply"
version = "0.3.10"
@@ -1273,12 +1974,45 @@ dependencies = [
"syn 2.0.117",
]
+[[package]]
+name = "mime"
+version = "0.3.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6877bb514081ee2a7ff5ef9de3281f14a4dd4bceac4c09388074a6b5df8a139a"
+
+[[package]]
+name = "mime_guess"
+version = "2.0.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f7c44f8e672c00fe5308fa235f821cb4198414e1c77935c1ab6948d3fd78550e"
+dependencies = [
+ "mime",
+ "unicase",
+]
+
[[package]]
name = "minimal-lexical"
version = "0.2.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "68354c5c6bd36d73ff3feceb05efa59b6acb7626617f4962be322a825e61f79a"
+[[package]]
+name = "mio"
+version = "1.2.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4b18443e9c262bfe8fa82f51666e2642c53393f7e5c27b3e1aeab922cff5b9d8"
+dependencies = [
+ "libc",
+ "wasi 0.11.1+wasi-snapshot-preview1",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "more-asserts"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1fafa6961cabd9c63bcd77a45d7e3b7f3b552b70417831fb0f56db717e72407e"
+
[[package]]
name = "moveit"
version = "0.6.0"
@@ -1294,6 +2028,33 @@ version = "0.10.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "1d87ecb2933e8aeadb3e3a02b828fed80a7528047e68b4f424523a0981a3a084"
+[[package]]
+name = "multiversion"
+version = "0.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b4ca4bea16ffc3f443cf7d866912118196bfef4c6a1556ca00f9f9b00bb43f7c"
+dependencies = [
+ "multiversion-macros",
+]
+
+[[package]]
+name = "multiversion-macros"
+version = "0.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0d416831a7317ef4b08bee00b69cbbb9c8763da7959a7026244d6266869f9c83"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "rustversion",
+ "syn 3.0.4",
+]
+
+[[package]]
+name = "multiversion_no_op"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "743fb55ba31b18fb1ecef6bdc9aa2743314978ac084044301a7eee33fb99a20d"
+
[[package]]
name = "native-tls"
version = "0.2.18"
@@ -1336,6 +2097,24 @@ dependencies = [
"minimal-lexical",
]
+[[package]]
+name = "ntapi"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c3b335231dfd352ffb0f8017f3b6027a4917f7df785ea2143d8af2adc66980ae"
+dependencies = [
+ "winapi",
+]
+
+[[package]]
+name = "nu-ansi-term"
+version = "0.50.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7957b9740744892f114936ab4a57b3f487491bbeafaf8083688b16841a4240e5"
+dependencies = [
+ "windows-sys 0.61.2",
+]
+
[[package]]
name = "num-complex"
version = "0.4.6"
@@ -1380,17 +2159,45 @@ dependencies = [
]
[[package]]
-name = "objc_exception"
-version = "0.1.2"
+name = "objc2-core-foundation"
+version = "0.3.2"
source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ad970fb455818ad6cba4c122ad012fae53ae8b4795f86378bce65e4f6bab2ca4"
+checksum = "2a180dd8642fa45cdb7dd721cd4c11b1cadd4929ce112ebd8b9f5803cc79d536"
dependencies = [
- "cc",
+ "bitflags",
]
[[package]]
-name = "once_cell"
-version = "1.21.4"
+name = "objc2-io-kit"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "33fafba39597d6dc1fb709123dfa8289d39406734be322956a69f0931c73bb15"
+dependencies = [
+ "libc",
+ "objc2-core-foundation",
+]
+
+[[package]]
+name = "objc2-system-configuration"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7216bd11cbda54ccabcab84d523dc93b858ec75ecfb3a7d89513fa22464da396"
+dependencies = [
+ "objc2-core-foundation",
+]
+
+[[package]]
+name = "objc_exception"
+version = "0.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ad970fb455818ad6cba4c122ad012fae53ae8b4795f86378bce65e4f6bab2ca4"
+dependencies = [
+ "cc",
+]
+
+[[package]]
+name = "once_cell"
+version = "1.21.4"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50"
@@ -1400,6 +2207,12 @@ version = "1.70.2"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
+[[package]]
+name = "oneshot"
+version = "0.1.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "269bca4c2591a28585d6bf10d9ed0332b7d76900a1b02bec41bdc3a2cdcda107"
+
[[package]]
name = "onnx2webnn"
version = "0.1.0"
@@ -1407,12 +2220,15 @@ dependencies = [
"anyhow",
"bytemuck",
"clap",
+ "dirs",
"half",
+ "huggingface-hub",
"pretty_env_logger",
"prost",
"rustnn",
"serde",
"serde_json",
+ "sha2 0.10.9",
"tempfile",
"thiserror 1.0.69",
"ureq 2.12.1",
@@ -1494,6 +2310,44 @@ dependencies = [
"ureq 3.3.0",
]
+[[package]]
+name = "os_str_bytes"
+version = "6.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e2355d85b9a3786f481747ced0e0ff2ba35213a1f9bd406ed906554d7af805a1"
+dependencies = [
+ "memchr",
+]
+
+[[package]]
+name = "parking_lot"
+version = "0.12.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "93857453250e3077bd71ff98b6a65ea6621a19bb0f559a85248955ac12c45a1a"
+dependencies = [
+ "lock_api",
+ "parking_lot_core",
+]
+
+[[package]]
+name = "parking_lot_core"
+version = "0.9.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2621685985a2ebf1c516881c026032ac7deafcda1a2c9b7850dc81e3dfcb64c1"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "redox_syscall",
+ "smallvec",
+ "windows-link",
+]
+
+[[package]]
+name = "pathdiff"
+version = "0.2.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "df94ce210e5bc13cb6651479fa48d14f601d9858cfe0467f43ae157023b938d3"
+
[[package]]
name = "pem-rfc7468"
version = "1.0.0"
@@ -1549,7 +2403,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "89815c69d36021a140146f26659a81d6c2afa33d216d736dd4be5381a7362220"
dependencies = [
"pest",
- "sha2",
+ "sha2 0.10.9",
]
[[package]]
@@ -1562,12 +2416,38 @@ dependencies = [
"indexmap 2.14.0",
]
+[[package]]
+name = "pin-project"
+version = "1.1.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2466b2336ed02bcdca6b294417127b90ec92038d1d5c4fbeac971a922e0e0924"
+dependencies = [
+ "pin-project-internal",
+]
+
+[[package]]
+name = "pin-project-internal"
+version = "1.1.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c96395f0a926bc13b1c17622aaddda1ecb55d49c8f1bf9777e4d877800a43f8b"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
[[package]]
name = "pin-project-lite"
version = "0.2.17"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "a89322df9ebe1c1578d689c92318e070967d1042b512afbe49518723f4e6d5cd"
+[[package]]
+name = "pin-utils"
+version = "0.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8b870d8c151b6f2fb93e84a13146138f05d02ed11c7e7c54f8826aaaf7c9f184"
+
[[package]]
name = "pkg-config"
version = "0.3.33"
@@ -1675,7 +2555,7 @@ checksum = "22505a5c94da8e3b7c2996394d1c933236c4d743e81a410bcca4e6989fc066a4"
dependencies = [
"bytes",
"heck",
- "itertools 0.12.1",
+ "itertools 0.10.5",
"log",
"multimap",
"once_cell",
@@ -1683,6 +2563,8 @@ dependencies = [
"prettyplease",
"prost",
"prost-types",
+ "pulldown-cmark",
+ "pulldown-cmark-to-cmark",
"regex",
"syn 2.0.117",
"tempfile",
@@ -1695,7 +2577,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "81bddcdb20abf9501610992b6759a4c888aef7d1a7247ef75e2404275ac24af1"
dependencies = [
"anyhow",
- "itertools 0.12.1",
+ "itertools 0.10.5",
"proc-macro2",
"quote",
"syn 2.0.117",
@@ -1710,6 +2592,83 @@ dependencies = [
"prost",
]
+[[package]]
+name = "pulldown-cmark"
+version = "0.9.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "57206b407293d2bcd3af849ce869d52068623f19e1b5ff8e8778e3309439682b"
+dependencies = [
+ "bitflags",
+ "memchr",
+ "unicase",
+]
+
+[[package]]
+name = "pulldown-cmark-to-cmark"
+version = "10.0.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0194e6e1966c23cc5fd988714f85b18d548d773e81965413555d96569931833d"
+dependencies = [
+ "pulldown-cmark",
+]
+
+[[package]]
+name = "quinn"
+version = "0.11.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4051e23e9185c255a7e33ef59cdbca87a22d359052eecd22fc6b901fb37d9d11"
+dependencies = [
+ "bytes",
+ "cfg_aliases",
+ "pin-project-lite",
+ "quinn-proto",
+ "quinn-udp",
+ "rustc-hash",
+ "rustls",
+ "socket2",
+ "thiserror 2.0.18",
+ "tokio",
+ "tracing",
+ "web-time",
+]
+
+[[package]]
+name = "quinn-proto"
+version = "0.11.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a9746dbde176634f4f2f1faf2404e30a31b2bc1e9cafb5329c95d8177a18c9fc"
+dependencies = [
+ "aws-lc-rs",
+ "bytes",
+ "getrandom 0.4.2",
+ "lru-slab",
+ "rand",
+ "rand_pcg",
+ "ring",
+ "rustc-hash",
+ "rustls",
+ "rustls-pki-types",
+ "slab",
+ "thiserror 2.0.18",
+ "tinyvec",
+ "tracing",
+ "web-time",
+]
+
+[[package]]
+name = "quinn-udp"
+version = "0.5.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "35a133f956daabe89a61a685c2649f13d82d5aa4bd5d12d1277e1072a21c0694"
+dependencies = [
+ "cfg_aliases",
+ "libc",
+ "once_cell",
+ "socket2",
+ "tracing",
+ "windows-sys 0.52.0",
+]
+
[[package]]
name = "quote"
version = "1.0.45"
@@ -1725,12 +2684,56 @@ version = "6.0.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "f8dcc9c7d52a811697d2151c701e0d08956f92b0e24136cf4cf27b57a6a0d9bf"
+[[package]]
+name = "rand"
+version = "0.10.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "65c9fb96cbc91e3478eaae79a69fcd3f1ae4ad052e471fe6732fff548984b4af"
+dependencies = [
+ "chacha20",
+ "getrandom 0.4.2",
+ "rand_core",
+]
+
+[[package]]
+name = "rand_core"
+version = "0.10.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "63b8176103e19a2643978565ca18b50549f6101881c443590420e4dc998a3c69"
+
+[[package]]
+name = "rand_pcg"
+version = "0.10.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "caa0f4137e1c0a72f4c651489402276c8e8e1cf081f3b0ba156d2cbeef09e86a"
+dependencies = [
+ "rand_core",
+]
+
[[package]]
name = "rawpointer"
version = "0.2.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "60a357793950651c4ed0f3f52338f53b2f809f32d83a07f72909fa13e4c6c1e3"
+[[package]]
+name = "redb"
+version = "3.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4ba239c1c1693315d3cc0e601db3b3965543afbf48c41730fdca2f069f510f4a"
+dependencies = [
+ "libc",
+]
+
+[[package]]
+name = "redox_syscall"
+version = "0.5.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ed2bf2547551a7053d6fdfafda3f938979645c44812fbfcda098faae3f1a362d"
+dependencies = [
+ "bitflags",
+]
+
[[package]]
name = "redox_users"
version = "0.5.2"
@@ -1831,6 +2834,96 @@ dependencies = [
"syn 1.0.109",
]
+[[package]]
+name = "reqwest"
+version = "0.13.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "16a1cfa75cc186dd73d5818e510e042e40927bccc9c236b061cea97e1eb08029"
+dependencies = [
+ "base64 0.23.1",
+ "bytes",
+ "encoding_rs",
+ "futures-core",
+ "futures-util",
+ "h2",
+ "http",
+ "http-body",
+ "http-body-util",
+ "hyper",
+ "hyper-rustls",
+ "hyper-util",
+ "js-sys",
+ "log",
+ "mime",
+ "mime_guess",
+ "percent-encoding",
+ "pin-project-lite",
+ "quinn",
+ "rustls",
+ "rustls-pki-types",
+ "rustls-platform-verifier",
+ "serde",
+ "serde_json",
+ "serde_urlencoded",
+ "sync_wrapper",
+ "tokio",
+ "tokio-rustls",
+ "tokio-util",
+ "tower",
+ "tower-http",
+ "tower-service",
+ "url",
+ "wasm-bindgen",
+ "wasm-bindgen-futures",
+ "wasm-streams",
+ "web-sys",
+]
+
+[[package]]
+name = "reqwest-middleware"
+version = "0.5.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "07bc3f1384cffa4f274dad2d4ddd73aed32fed8f786d96c6be8aa4e5fd3c3b58"
+dependencies = [
+ "anyhow",
+ "async-trait",
+ "http",
+ "reqwest",
+ "serde",
+ "thiserror 2.0.18",
+ "tower-service",
+]
+
+[[package]]
+name = "reqwest-retry"
+version = "0.9.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fe2412db2af7d2268e7a5406be0431f37d9eb67ff390f35b395716f5f06c2eaa"
+dependencies = [
+ "anyhow",
+ "async-trait",
+ "futures",
+ "getrandom 0.2.17",
+ "http",
+ "hyper",
+ "reqwest",
+ "reqwest-middleware",
+ "retry-policies",
+ "thiserror 2.0.18",
+ "tokio",
+ "tracing",
+ "wasmtimer",
+]
+
+[[package]]
+name = "retry-policies"
+version = "0.5.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dc05fbf560421a0357a750cbe78c7ca19d4923918490daabba313d5dbc871e47"
+dependencies = [
+ "rand",
+]
+
[[package]]
name = "ring"
version = "0.17.14"
@@ -1851,6 +2944,15 @@ version = "2.1.3"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "6b1e7f9a428571be2dc5bc0505c13fb6bf936822b894ec87abf8a08a4e51742d"
+[[package]]
+name = "rustc_version"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cfcb3a22ef46e85b45de6ee7e79d063319ebb6594faafcf1c225ea92ab6e9b92"
+dependencies = [
+ "semver",
+]
+
[[package]]
name = "rustix"
version = "1.1.4"
@@ -1861,7 +2963,7 @@ dependencies = [
"errno",
"libc",
"linux-raw-sys",
- "windows-sys 0.52.0",
+ "windows-sys 0.61.2",
]
[[package]]
@@ -1870,6 +2972,7 @@ version = "0.23.40"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "ef86cd5876211988985292b91c96a8f2d298df24e75989a43a3c73f2d4d8168b"
dependencies = [
+ "aws-lc-rs",
"log",
"once_cell",
"ring",
@@ -1879,21 +2982,62 @@ dependencies = [
"zeroize",
]
+[[package]]
+name = "rustls-native-certs"
+version = "0.8.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dab5152771c58876a2146916e53e35057e1a4dfa2b9df0f0305b07f611fdea4d"
+dependencies = [
+ "openssl-probe",
+ "rustls-pki-types",
+ "schannel",
+ "security-framework",
+]
+
[[package]]
name = "rustls-pki-types"
version = "1.14.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "30a7197ae7eb376e574fe940d068c30fe0462554a3ddbe4eca7838e049c937a9"
dependencies = [
+ "web-time",
"zeroize",
]
+[[package]]
+name = "rustls-platform-verifier"
+version = "0.7.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "26d1e2536ce4f35f4846aa13bff16bd0ff40157cdb14cc056c7b14ba41233ba0"
+dependencies = [
+ "core-foundation 0.10.1",
+ "core-foundation-sys",
+ "jni",
+ "log",
+ "once_cell",
+ "rustls",
+ "rustls-native-certs",
+ "rustls-platform-verifier-android",
+ "rustls-webpki",
+ "security-framework",
+ "security-framework-sys",
+ "webpki-root-certs",
+ "windows-sys 0.52.0",
+]
+
+[[package]]
+name = "rustls-platform-verifier-android"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f87165f0995f63a9fbeea62b64d10b4d9d8e78ec6d7d51fb2125fda7bb36788f"
+
[[package]]
name = "rustls-webpki"
version = "0.103.13"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "61c429a8649f110dddef65e2a5ad240f747e85f7758a6bccc7e5777bd33f756e"
dependencies = [
+ "aws-lc-rs",
"ring",
"rustls-pki-types",
"untrusted",
@@ -1903,7 +3047,7 @@ dependencies = [
name = "rustnn"
version = "0.5.11"
dependencies = [
- "base64",
+ "base64 0.22.1",
"block",
"bytemuck",
"bytes",
@@ -1921,7 +3065,6 @@ dependencies = [
"prost-build",
"prost-types",
"regex",
- "safetensors",
"seahash",
"serde",
"serde_json",
@@ -1940,6 +3083,18 @@ version = "1.0.22"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "b39cdef0fa800fc44525c84ccb54a029961a8215f9619753635a9c0d2538d46d"
+[[package]]
+name = "ryu"
+version = "1.0.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9774ba4a74de5f7b1c1451ed6cd5285a32eddb5cccb8cc655a4e50009e06477f"
+
+[[package]]
+name = "safe-transmute"
+version = "0.11.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3944826ff8fa8093089aba3acb4ef44b9446a99a16f3bf4e74af3f77d340ab7d"
+
[[package]]
name = "safetensors"
version = "0.7.0"
@@ -1951,6 +3106,15 @@ dependencies = [
"serde_json",
]
+[[package]]
+name = "same-file"
+version = "1.0.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "93fc1dc3aaa9bfed95e02e6eadabb4baf7e3078b0bd1b4d7b6b0b68378900502"
+dependencies = [
+ "winapi-util",
+]
+
[[package]]
name = "schannel"
version = "0.1.29"
@@ -1984,6 +3148,12 @@ dependencies = [
"serde_json",
]
+[[package]]
+name = "scopeguard"
+version = "1.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "94143f37725109f92c262ed2cf5e59bce7498c01bcc1502d7b9afe439a4e9f49"
+
[[package]]
name = "scratch"
version = "1.0.9"
@@ -2003,7 +3173,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "b7f4bc775c73d9a02cde8bf7b2ec4c9d12743edf609006c7facc23998404cd1d"
dependencies = [
"bitflags",
- "core-foundation",
+ "core-foundation 0.10.1",
"core-foundation-sys",
"libc",
"security-framework-sys",
@@ -2068,13 +3238,36 @@ dependencies = [
"zmij",
]
+[[package]]
+name = "serde_repr"
+version = "0.1.21"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8d3b1629de253c70a0508c3899572da79ca359fdab27c7920ff00406df418906"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 3.0.4",
+]
+
+[[package]]
+name = "serde_urlencoded"
+version = "0.7.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d3491c14715ca2294c4d6a88f15e84739788c1d030eed8c110436aafdaa2f3fd"
+dependencies = [
+ "form_urlencoded",
+ "itoa",
+ "ryu",
+ "serde",
+]
+
[[package]]
name = "serde_with"
version = "3.20.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "e72c1c2cb7b223fafb600a619537a871c2818583d619401b785e7c0b746ccde2"
dependencies = [
- "base64",
+ "base64 0.22.1",
"bs58",
"chrono",
"hex",
@@ -2107,8 +3300,39 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "a7507d819769d01a365ab707794a4084392c824f54a7a6a7862f8c3d0892b283"
dependencies = [
"cfg-if",
- "cpufeatures",
- "digest",
+ "cpufeatures 0.2.17",
+ "digest 0.10.7",
+]
+
+[[package]]
+name = "sha2"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "446ba717509524cb3f22f17ecc096f10f4822d76ab5c0b9822c5f9c284e825f4"
+dependencies = [
+ "cfg-if",
+ "cpufeatures 0.3.1",
+ "digest 0.11.3",
+]
+
+[[package]]
+name = "sharded-slab"
+version = "0.1.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f40ca3c46823713e0d4209592e8d6e826aa57e928f09752619fc696c499637f6"
+dependencies = [
+ "lazy_static",
+]
+
+[[package]]
+name = "shellexpand"
+version = "3.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32824fab5e16e6c4d86dc1ba84489390419a39f97699852b66480bb87d297ed8"
+dependencies = [
+ "bstr",
+ "dirs",
+ "os_str_bytes",
]
[[package]]
@@ -2123,6 +3347,22 @@ version = "2.0.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "f8fadd59c855ef2080decdef8ff161eb6661b86933c9d82e5ba29dc602a55aba"
+[[package]]
+name = "simd_cesu8"
+version = "1.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "11031e251abf8611c80f460e19dbdeb54a66db918e49c65a7065b46ac7aec520"
+dependencies = [
+ "rustc_version",
+ "simdutf8",
+]
+
+[[package]]
+name = "simdutf8"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e3a9fe34e3e7a50316060351f37187a3f546bce95496156754b601a5fa71b76e"
+
[[package]]
name = "slab"
version = "0.4.12"
@@ -2135,6 +3375,16 @@ version = "1.15.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "67b1b7a3b5fe4f1376887184045fcf45c69e92af734b7aaddc05fb777b6fbd03"
+[[package]]
+name = "socket2"
+version = "0.6.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c3d1e2c7f27f8d4cb10542a02c49005dbd6e93095799d6f3be745fae9f8fedd4"
+dependencies = [
+ "libc",
+ "windows-sys 0.61.2",
+]
+
[[package]]
name = "socks"
version = "0.3.4"
@@ -2152,6 +3402,22 @@ version = "1.2.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "6ce2be8dc25455e1f91df71bfa12ad37d7af1092ae736f3a6cd0e37bc7810596"
+[[package]]
+name = "static_assertions"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a2eb9349b6444b326872e140eb1cf5e7c522154d69e7a0ffb0fb81c06b37543f"
+
+[[package]]
+name = "statrs"
+version = "0.18.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2a3fe7c28c6512e766b0874335db33c94ad7b8f9054228ae1c2abd47ce7d335e"
+dependencies = [
+ "approx",
+ "num-traits",
+]
+
[[package]]
name = "strsim"
version = "0.11.1"
@@ -2164,6 +3430,12 @@ version = "2.6.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "13c2bddecc57b384dee18652358fb23172facb8a2c51ccc10d74c157bdea3292"
+[[package]]
+name = "symlink"
+version = "0.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a7973cce6668464ea31f176d85b13c7ab3bba2cb3b77a2ed26abd7801688010a"
+
[[package]]
name = "syn"
version = "1.0.109"
@@ -2197,6 +3469,15 @@ dependencies = [
"unicode-ident",
]
+[[package]]
+name = "sync_wrapper"
+version = "1.0.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0bf256ce5efdfa370213c1dabab5935a12e49f2c58d15e9eac2870d3b4f27263"
+dependencies = [
+ "futures-core",
+]
+
[[package]]
name = "synstructure"
version = "0.13.2"
@@ -2208,6 +3489,41 @@ dependencies = [
"syn 2.0.117",
]
+[[package]]
+name = "sysinfo"
+version = "0.38.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "92ab6a2f8bfe508deb3c6406578252e491d299cbbf3bc0529ecc3313aee4a52f"
+dependencies = [
+ "libc",
+ "memchr",
+ "ntapi",
+ "objc2-core-foundation",
+ "objc2-io-kit",
+ "windows",
+]
+
+[[package]]
+name = "system-configuration"
+version = "0.7.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a13f3d0daba03132c0aa9767f98351b3488edc2c100cda2d2ec2b04f3d8d3c8b"
+dependencies = [
+ "bitflags",
+ "core-foundation 0.9.4",
+ "system-configuration-sys",
+]
+
+[[package]]
+name = "system-configuration-sys"
+version = "0.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8e1d1b10ced5ca923a1fcb8d03e96b8d3268065d724548c0211415ff6ac6bac4"
+dependencies = [
+ "core-foundation-sys",
+ "libc",
+]
+
[[package]]
name = "tempfile"
version = "3.27.0"
@@ -2218,7 +3534,7 @@ dependencies = [
"getrandom 0.4.2",
"once_cell",
"rustix",
- "windows-sys 0.52.0",
+ "windows-sys 0.61.2",
]
[[package]]
@@ -2270,6 +3586,15 @@ dependencies = [
"syn 2.0.117",
]
+[[package]]
+name = "thread_local"
+version = "1.1.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ad99c4c6d32803332c548b1af0540b357b3f5fc0be8f6c6bfe8b2e6ae784070"
+dependencies = [
+ "cfg-if",
+]
+
[[package]]
name = "time"
version = "0.3.47"
@@ -2315,34 +3640,232 @@ dependencies = [
name = "tinyvec"
version = "1.11.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3e61e67053d25a4e82c844e8424039d9745781b3fc4f32b8d55ed50f5f667ef3"
+checksum = "3e61e67053d25a4e82c844e8424039d9745781b3fc4f32b8d55ed50f5f667ef3"
+dependencies = [
+ "tinyvec_macros",
+]
+
+[[package]]
+name = "tinyvec_macros"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1f3ccbac311fea05f86f61904b462b55fb3df8837a366dfc601a0161d0532f20"
+
+[[package]]
+name = "tokio"
+version = "1.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "202caea871b69668250d242070849eb495be178ed697a3e98aebce5bc81a0bed"
+dependencies = [
+ "bytes",
+ "libc",
+ "mio",
+ "pin-project-lite",
+ "socket2",
+ "tokio-macros",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "tokio-macros"
+version = "2.7.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "78773a2a397f451582ce068015985c33193cf6dea8b74d2a639fe457b2f07b0e"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 3.0.4",
+]
+
+[[package]]
+name = "tokio-retry"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4a129d95275ebf4c493ec53bf0f8cd95f5ac161bc4f381700809a54f595d4470"
+dependencies = [
+ "pin-project-lite",
+ "rand",
+ "tokio",
+]
+
+[[package]]
+name = "tokio-rustls"
+version = "0.26.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b0c85f2c3ef0b1cd58b36682f4b17aaa995f0e5db534d85692b4903abce21f67"
+dependencies = [
+ "rustls",
+ "tokio",
+]
+
+[[package]]
+name = "tokio-util"
+version = "0.7.19"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "494815d09bf52b5548659851081238f0ca39ff638363907596da739561c62c52"
+dependencies = [
+ "bytes",
+ "futures-core",
+ "futures-sink",
+ "libc",
+ "pin-project-lite",
+ "tokio",
+]
+
+[[package]]
+name = "tokio_with_wasm"
+version = "0.8.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "34e40fbbbd95441133fe9483f522db15dbfd26dc636164ebd8f2dd28759a6aa6"
+dependencies = [
+ "js-sys",
+ "tokio",
+ "tokio_with_wasm_proc",
+ "wasm-bindgen",
+ "wasm-bindgen-futures",
+ "web-sys",
+]
+
+[[package]]
+name = "tokio_with_wasm_proc"
+version = "0.8.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d01145a2c788d6aae4cd653afec1e8332534d7d783d01897cefcafe4428de992"
+dependencies = [
+ "quote",
+ "syn 2.0.117",
+]
+
+[[package]]
+name = "tower"
+version = "0.5.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ebe5ef63511595f1344e2d5cfa636d973292adc0eec1f0ad45fae9f0851ab1d4"
+dependencies = [
+ "futures-core",
+ "futures-util",
+ "pin-project-lite",
+ "sync_wrapper",
+ "tokio",
+ "tower-layer",
+ "tower-service",
+]
+
+[[package]]
+name = "tower-http"
+version = "0.6.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4cfcf7e2740e6fc6d4d688b4ef00650406bb94adf4731e43c096c3a19fe40840"
+dependencies = [
+ "bitflags",
+ "bytes",
+ "futures-util",
+ "http",
+ "http-body",
+ "pin-project-lite",
+ "tower",
+ "tower-layer",
+ "tower-service",
+ "url",
+]
+
+[[package]]
+name = "tower-layer"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "121c2a6cda46980bb0fcd1647ffaf6cd3fc79a013de288782836f6df9c48780e"
+
+[[package]]
+name = "tower-service"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8df9b6e13f2d32c91b9bd719c00d1958837bc7dec474d94952798cc8e69eeec3"
+
+[[package]]
+name = "tracing"
+version = "0.1.44"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "63e71662fa4b2a2c3a26f570f037eb95bb1f85397f3cd8076caed2f026a6d100"
+dependencies = [
+ "pin-project-lite",
+ "tracing-attributes",
+ "tracing-core",
+]
+
+[[package]]
+name = "tracing-appender"
+version = "0.2.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "050686193eb999b4bb3bc2acfa891a13da00f79734704c4b8b4ef1a10b368a3c"
+dependencies = [
+ "crossbeam-channel",
+ "symlink",
+ "thiserror 2.0.18",
+ "time",
+ "tracing-subscriber",
+]
+
+[[package]]
+name = "tracing-attributes"
+version = "0.1.31"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7490cfa5ec963746568740651ac6781f701c9c5ea257c58e057f3ba8cf69e8da"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
+[[package]]
+name = "tracing-core"
+version = "0.1.36"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "db97caf9d906fbde555dd62fa95ddba9eecfd14cb388e4f491a66d74cd5fb79a"
dependencies = [
- "tinyvec_macros",
+ "once_cell",
+ "valuable",
]
[[package]]
-name = "tinyvec_macros"
-version = "0.1.1"
+name = "tracing-log"
+version = "0.2.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1f3ccbac311fea05f86f61904b462b55fb3df8837a366dfc601a0161d0532f20"
+checksum = "ee855f1f400bd0e5c02d150ae5de3840039a3f54b025156404e34c23c03f47c3"
+dependencies = [
+ "log",
+ "once_cell",
+ "tracing-core",
+]
[[package]]
-name = "tracing"
-version = "0.1.44"
+name = "tracing-serde"
+version = "0.2.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "63e71662fa4b2a2c3a26f570f037eb95bb1f85397f3cd8076caed2f026a6d100"
+checksum = "704b1aeb7be0d0a84fc9828cae51dab5970fee5088f83d1dd7ee6f6246fc6ff1"
dependencies = [
- "pin-project-lite",
+ "serde",
"tracing-core",
]
[[package]]
-name = "tracing-core"
-version = "0.1.36"
+name = "tracing-subscriber"
+version = "0.3.23"
source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "db97caf9d906fbde555dd62fa95ddba9eecfd14cb388e4f491a66d74cd5fb79a"
+checksum = "cb7f578e5945fb242538965c2d0b04418d38ec25c79d160cd279bf0731c8d319"
dependencies = [
+ "matchers",
+ "nu-ansi-term",
"once_cell",
+ "regex-automata",
+ "serde",
+ "serde_json",
+ "sharded-slab",
+ "smallvec",
+ "thread_local",
+ "tracing",
+ "tracing-core",
+ "tracing-log",
+ "tracing-serde",
]
[[package]]
@@ -2376,18 +3899,62 @@ dependencies = [
"regex",
]
+[[package]]
+name = "try-lock"
+version = "0.2.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e421abadd41a4225275504ea4d6566923418b7f05506fbc9c0fe86ba7396114b"
+
+[[package]]
+name = "twox-hash"
+version = "2.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5283634e518fe9e82c7b20520bb4bc209009fd16c82077c802f8111ecbb0117a"
+
+[[package]]
+name = "typed-builder"
+version = "0.20.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cd9d30e3a08026c78f246b173243cf07b3696d274debd26680773b6773c2afc7"
+dependencies = [
+ "typed-builder-macro",
+]
+
+[[package]]
+name = "typed-builder-macro"
+version = "0.20.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3c36781cc0e46a83726d9879608e4cf6c2505237e263a8eb8c24502989cfdb28"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
[[package]]
name = "typenum"
version = "1.20.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "b6f5e870be6c3b371b77fe0ee0bafb859fa4964b4404c27de1d380043c4dda20"
+[[package]]
+name = "typewit"
+version = "1.15.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "214ca0b2191785cbc06209b9ca1861e048e39b5ba33574b3cedd58363d5bb5f6"
+
[[package]]
name = "ucd-trie"
version = "0.1.7"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "2896d95c02a80c6d6a5d6e953d479f5ddf2dfdb6a244441010e373ac0fb88971"
+[[package]]
+name = "unicase"
+version = "2.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dbc4bc3a9f746d862c45cb89d705aa10f187bb96c76001afab07a0d35ce60142"
+
[[package]]
name = "unicode-ident"
version = "1.0.24"
@@ -2424,7 +3991,7 @@ version = "2.12.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "02d1a66277ed75f640d608235660df48c8e3c19f3b4edb6a263315626cc3c01d"
dependencies = [
- "base64",
+ "base64 0.22.1",
"log",
"once_cell",
"rustls",
@@ -2439,7 +4006,7 @@ version = "3.3.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "dea7109cdcd5864d4eeb1b58a1648dc9bf520360d7af16ec26d0a9354bafcfc0"
dependencies = [
- "base64",
+ "base64 0.22.1",
"der",
"log",
"native-tls",
@@ -2457,7 +4024,7 @@ version = "0.6.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "e994ba84b0bd1b1b0cf92878b7ef898a5c1760108fe7b6010327e274917a808c"
dependencies = [
- "base64",
+ "base64 0.22.1",
"http",
"httparse",
"log",
@@ -2475,6 +4042,12 @@ dependencies = [
"serde",
]
+[[package]]
+name = "urlencoding"
+version = "2.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "daf8dba3b7eb870caf1ddeed7bc9d2a049f3cfdfae7cb521b087cc33ae4c49da"
+
[[package]]
name = "utf8-zero"
version = "0.8.1"
@@ -2493,6 +4066,23 @@ version = "0.2.2"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"
+[[package]]
+name = "uuid"
+version = "1.26.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2ef6dac1e96601b4fb3acccccff2139741fcb757cb9a36089bf5be91cfb285ce"
+dependencies = [
+ "getrandom 0.4.2",
+ "js-sys",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "valuable"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ba73ea9cf16a25df0c8caa16c51acb937d5712a8429db78a3ee29d5dcacd3a65"
+
[[package]]
name = "vcpkg"
version = "0.2.15"
@@ -2505,12 +4095,40 @@ version = "0.9.5"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a"
+[[package]]
+name = "walkdir"
+version = "2.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "29790946404f91d9c5d06f9874efddea1dc06c5efe94541a7d6863108e3a5e4b"
+dependencies = [
+ "same-file",
+ "winapi-util",
+]
+
+[[package]]
+name = "want"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bfa7760aed19e106de2c7c0b581b509f2f25d3dacaf737cb82ac61bc6d760b0e"
+dependencies = [
+ "try-lock",
+]
+
[[package]]
name = "wasi"
version = "0.11.1+wasi-snapshot-preview1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "ccf3ec651a847eb01de73ccad15eb7d99f80485de043efb2f370cd654f4ea44b"
+[[package]]
+name = "wasi"
+version = "0.14.7+wasi-0.2.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "883478de20367e224c0090af9cf5f9fa85bed63a95c1abf3afc5c083ebc06e8c"
+dependencies = [
+ "wasip2",
+]
+
[[package]]
name = "wasip2"
version = "1.0.3+wasi-0.2.9"
@@ -2529,6 +4147,15 @@ dependencies = [
"wit-bindgen 0.51.0",
]
+[[package]]
+name = "wasite"
+version = "1.0.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "66fe902b4a6b8028a753d5424909b764ccf79b7a209eac9bf97e59cda9f71a42"
+dependencies = [
+ "wasi 0.14.7+wasi-0.2.4",
+]
+
[[package]]
name = "wasm-bindgen"
version = "0.2.122"
@@ -2542,6 +4169,16 @@ dependencies = [
"wasm-bindgen-shared",
]
+[[package]]
+name = "wasm-bindgen-futures"
+version = "0.4.72"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9473dbd2991ae90b6291c3c32c30c6187ac49aa32f9905d1cce280ec1e110b0f"
+dependencies = [
+ "js-sys",
+ "wasm-bindgen",
+]
+
[[package]]
name = "wasm-bindgen-macro"
version = "0.2.122"
@@ -2596,6 +4233,19 @@ dependencies = [
"wasmparser",
]
+[[package]]
+name = "wasm-streams"
+version = "0.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9d1ec4f6517c9e11ae630e200b2b65d193279042e28edd4a2cda233e46670bbb"
+dependencies = [
+ "futures-util",
+ "js-sys",
+ "wasm-bindgen",
+ "wasm-bindgen-futures",
+ "web-sys",
+]
+
[[package]]
name = "wasmparser"
version = "0.244.0"
@@ -2608,13 +4258,47 @@ dependencies = [
"semver",
]
+[[package]]
+name = "wasmtimer"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1c598d6b99ea013e35844697fc4670d08339d5cda15588f193c6beedd12f644b"
+dependencies = [
+ "futures",
+ "js-sys",
+ "parking_lot",
+ "pin-utils",
+ "slab",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "web-sys"
+version = "0.3.99"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6d621441cfc37b84979402712047321980c178f299193a3589d05b99e8763436"
+dependencies = [
+ "js-sys",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "web-time"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5a6580f308b1fad9207618087a65c04e7a10bc77e02c8e84e9b00dd4b12fa0bb"
+dependencies = [
+ "js-sys",
+ "wasm-bindgen",
+]
+
[[package]]
name = "webnn-graph"
version = "0.3.0"
-source = "git+https://github.com/rustnn/webnn-graph?branch=main#075e3c2f8b78d67f5b0ffbc51d6bc81162397d43"
+source = "git+https://github.com/rustnn/webnn-graph?branch=main#0bcd517d30cf740f55bf19386be254e37010ab2b"
dependencies = [
"anyhow",
- "base64",
+ "base64 0.22.1",
"clap",
"half",
"pest",
@@ -2669,6 +4353,19 @@ dependencies = [
"rustls-pki-types",
]
+[[package]]
+name = "whoami"
+version = "2.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "626c4bac6755d76ffc12cb01b2eac751db1996b9e0041de9aa02c8c211ddc82c"
+dependencies = [
+ "libc",
+ "libredox",
+ "objc2-system-configuration",
+ "wasite",
+ "web-sys",
+]
+
[[package]]
name = "winapi"
version = "0.3.9"
@@ -2691,7 +4388,7 @@ version = "0.1.11"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
dependencies = [
- "windows-sys 0.52.0",
+ "windows-sys 0.61.2",
]
[[package]]
@@ -2700,6 +4397,27 @@ version = "0.4.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "712e227841d057c1ee1cd2fb22fa7e5a5461ae8e48fa2ca79ec42cfc1931183f"
+[[package]]
+name = "windows"
+version = "0.62.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "527fadee13e0c05939a6a05d5bd6eec6cd2e3dbd648b9f8e447c6518133d8580"
+dependencies = [
+ "windows-collections",
+ "windows-core",
+ "windows-future",
+ "windows-numerics",
+]
+
+[[package]]
+name = "windows-collections"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "23b2d95af1a8a14a3c7367e1ed4fc9c20e0a26e79551b1454d72583c97cc6610"
+dependencies = [
+ "windows-core",
+]
+
[[package]]
name = "windows-core"
version = "0.62.2"
@@ -2713,6 +4431,17 @@ dependencies = [
"windows-strings",
]
+[[package]]
+name = "windows-future"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e1d6f90251fe18a279739e78025bd6ddc52a7e22f921070ccdc67dde84c605cb"
+dependencies = [
+ "windows-core",
+ "windows-link",
+ "windows-threading",
+]
+
[[package]]
name = "windows-implement"
version = "0.60.2"
@@ -2741,6 +4470,27 @@ version = "0.2.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "f0805222e57f7521d6a62e36fa9163bc891acd422f971defe97d64e70d0a4fe5"
+[[package]]
+name = "windows-numerics"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6e2e40844ac143cdb44aead537bbf727de9b044e107a0f1220392177d15b0f26"
+dependencies = [
+ "windows-core",
+ "windows-link",
+]
+
+[[package]]
+name = "windows-registry"
+version = "0.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "02752bf7fbdcce7f2a27a742f798510f3e5ad88dbe84871e5168e2120c3d5720"
+dependencies = [
+ "windows-link",
+ "windows-result",
+ "windows-strings",
+]
+
[[package]]
name = "windows-result"
version = "0.4.1"
@@ -2768,6 +4518,15 @@ dependencies = [
"windows-targets",
]
+[[package]]
+name = "windows-sys"
+version = "0.59.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1e38bc4d79ed67fd075bcc251a1c39b32a1776bbe92e5bef1f0bf1f8c531853b"
+dependencies = [
+ "windows-targets",
+]
+
[[package]]
name = "windows-sys"
version = "0.61.2"
@@ -2793,6 +4552,15 @@ dependencies = [
"windows_x86_64_msvc",
]
+[[package]]
+name = "windows-threading"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3949bd5b99cafdf1c7ca86b43ca564028dfe27d66958f2470940f73d86d75b37"
+dependencies = [
+ "windows-link",
+]
+
[[package]]
name = "windows_aarch64_gnullvm"
version = "0.52.6"
@@ -2941,6 +4709,147 @@ version = "0.6.3"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "1ffae5123b2d3fc086436f8834ae3ab053a283cfac8fe0a0b8eaae044768a4c4"
+[[package]]
+name = "xet-client"
+version = "1.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c3b8da8cc70aa2e3c500c0400e012df82c656ab9fca47f9f939fffc5afd89aca"
+dependencies = [
+ "anyhow",
+ "async-trait",
+ "base64 0.22.1",
+ "bytes",
+ "crc32fast",
+ "futures",
+ "http",
+ "hyper",
+ "more-asserts",
+ "rand",
+ "redb",
+ "reqwest",
+ "reqwest-middleware",
+ "serde",
+ "serde_json",
+ "serde_repr",
+ "statrs",
+ "tempfile",
+ "thiserror 2.0.18",
+ "tokio",
+ "tokio-retry",
+ "tokio_with_wasm",
+ "tracing",
+ "url",
+ "urlencoding",
+ "web-time",
+ "xet-core-structures",
+ "xet-runtime",
+]
+
+[[package]]
+name = "xet-core-structures"
+version = "1.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "73503c223783dccc864abde22115e09d12f190448a0baf58ab2c54bc709e2f99"
+dependencies = [
+ "async-trait",
+ "base64 0.22.1",
+ "blake3",
+ "bytemuck",
+ "bytes",
+ "countio",
+ "futures",
+ "futures-util",
+ "getrandom 0.4.2",
+ "heapify",
+ "itertools 0.14.0",
+ "lz4_flex",
+ "more-asserts",
+ "rand",
+ "regex",
+ "safe-transmute",
+ "serde",
+ "static_assertions",
+ "thiserror 2.0.18",
+ "tokio",
+ "tokio-util",
+ "tracing",
+ "uuid",
+ "web-time",
+ "xet-runtime",
+]
+
+[[package]]
+name = "xet-data"
+version = "1.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c89052ec5dec2187cad30b86af92cc24fd61c4a57a795f1ff7ff5f38d49184eb"
+dependencies = [
+ "anyhow",
+ "async-trait",
+ "bytes",
+ "chrono",
+ "gearhash",
+ "http",
+ "itertools 0.14.0",
+ "more-asserts",
+ "rand",
+ "serde",
+ "serde_json",
+ "sha2 0.11.0",
+ "tempfile",
+ "thiserror 2.0.18",
+ "tokio",
+ "tokio-util",
+ "tokio_with_wasm",
+ "tracing",
+ "url",
+ "uuid",
+ "web-time",
+ "xet-client",
+ "xet-core-structures",
+ "xet-runtime",
+]
+
+[[package]]
+name = "xet-runtime"
+version = "1.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "af5c60d5eed38ab4c576f4421bae835e7bd07631fb381705605529d2015c106b"
+dependencies = [
+ "anyhow",
+ "async-trait",
+ "bytes",
+ "chrono",
+ "colored",
+ "const-str",
+ "ctor",
+ "dirs",
+ "futures",
+ "git-version",
+ "humantime",
+ "konst",
+ "libc",
+ "more-asserts",
+ "oneshot",
+ "pin-project",
+ "rand",
+ "reqwest",
+ "serde",
+ "serde_json",
+ "shellexpand",
+ "sysinfo",
+ "thiserror 2.0.18",
+ "tokio",
+ "tokio-util",
+ "tokio_with_wasm",
+ "tracing",
+ "tracing-appender",
+ "tracing-subscriber",
+ "web-time",
+ "whoami",
+ "winapi",
+]
+
[[package]]
name = "yoke"
version = "0.8.2"
diff --git a/Cargo.toml b/Cargo.toml
index 1237077..ead083a 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -24,17 +24,20 @@ trtx = ["rustnn/trtx-runtime"]
[dependencies]
anyhow = "1.0"
clap = { version = "4.5", features = ["derive"] }
+dirs = "6.0"
half = "2.7.1"
prost = "0.12"
serde = { version = "1.0", features = ["derive"] }
serde_json = "1.0"
-thiserror = "1.0"
+thiserror = "1.0"
+ureq = { version = "2.10", default-features = false, features = ["tls"] }
+sha2 = "0.10"
rustnn = { path = "../rustnn", default-features = false, features = ["onnx-runtime"] }
bytemuck = "1.15"
webnn-onnx-utils = { git = "https://github.com/rustnn/webnn-onnx-utils", branch = "main" }
+huggingface-hub = { git = "https://github.com/huggingface/huggingface_hub_rust.git", rev = "20acfd61a767dcf7d6a0cb7f4a02ac8406fd8a14", version = "0.0.1", features = ["blocking", "xet"] }
[dev-dependencies]
-pretty_env_logger = "0.5"
+pretty_env_logger = "0.5"
tempfile = "3.8"
-ureq = { version = "2.10", default-features = false, features = ["tls"] }
diff --git a/README.md b/README.md
index 01688ec..8ca4ce6 100644
--- a/README.md
+++ b/README.md
@@ -1,15 +1,51 @@
# onnx2webnn
+> **Last edited:** `2026-09-22T21:07:28Z`
+> **Checkout:** `fkrall/cache-backed-validation` at `a582b89`
+>
+> **Freshness:** Use this document only when this provenance is recent relative to the relevant
+> code and commits; otherwise verify the implementation, tests, and Git history before relying
+> on it.
+
ONNX → WebNN lowering crate extracted from [webnn-graph](../webnn-graph). ONNX operators lower
directly to [rustnn](../rustnn) `MLGraphBuilder`; full-graph validation runs via ORT CPU
-`build()` (`onnx-runtime` feature). There is no JSON IR and no on-disk graph export — success
-means `builder.build()` returns `Ok(MLGraph)`.
+`build()` (`onnx-runtime` feature). There is no intermediate JSON IR. Without `--output`,
+success means `builder.build()` returns `Ok(MLGraph)`; `--output` additionally writes a
+reloadable `.webnn` graph and sibling Safetensors archive.
+
+Decomposed attention is not promised to be bit-identical to a source backend fused-attention
+kernel. This matters for quantized models that immediately feed attention into
+`DynamicQuantizeLinear`: small legal floating-point differences can cross Uint8 bucket boundaries
+and compound through later layers. FastVLM prefill has this limitation; its fixed sequence-1 decode
+specialization passes numerical validation.
Supported ONNX opset range: **1–26** (see `MIN_SUPPORTED_OPSET` / `MAX_SUPPORTED_OPSET` in
`src/onnx/convert.rs`).
## Build
+### Prerequisites
+
+- A stable Rust toolchain.
+- The Protocol Buffers compiler (`protoc`). On Debian/Ubuntu install
+ `protobuf-compiler`; on macOS install `protobuf` with Homebrew. Cargo uses it
+ while building the ONNX and CoreML protobuf bindings.
+- A sibling `rustnn` checkout at `../rustnn`. This is the path dependency used
+ by `Cargo.toml` while the stacked RustNN changes are under review.
+- ONNX Runtime 1.27 or newer when running conversions or tests with the ORT
+ backend. RustNN can download the pinned version into its own ignored
+ `target/onnxruntime` directory:
+
+```powershell
+make -C ../rustnn onnxruntime-download
+```
+
+Set `ORT_DYLIB_PATH` to the downloaded `libonnxruntime` shared library before
+running the CLI or tests. The exact filename is platform-specific; examples
+are `libonnxruntime.so.1.29.0` on Linux, `libonnxruntime.1.29.0.dylib` on macOS,
+and `onnxruntime.dll` on Windows. The CI workflow in `.github/workflows/ci.yml`
+contains the complete cross-platform provisioning sequence.
+
```powershell
cargo build
# or
@@ -36,6 +72,24 @@ cargo run -- convert --input model.onnx `
If `model.dims.json` sits beside the ONNX file and no overrides were passed on the CLI, dimension
bindings are loaded from that sidecar (`freeDimensionOverrides` or a flat JSON object).
+With `--output`, artifacts are overwritten below the operating system's cache directory in
+`onnx2webnn/onnx` and `onnx2webnn/webnn`; add `--validate` to reload the saved pair immediately and
+compare deterministic execution against native ORT. Set `O2W_CACHE_DIR` to relocate both caches,
+or `O2W_ONNX_CACHE` and `O2W_WEBNN_CACHE` to override either cache independently. RustNN stores
+logical Int4/Uint4 constants as their original packed nibble bytes in a
+versioned U8 Safetensors extension while the `.webnn` declaration retains the logical dtype and
+shape. This is a RustNN archive convention, not a native Safetensors 4-bit dtype.
+
+`MatMulNBits` is lowered to `dequantizeLinear` followed by ordinary floating-point `matmul` because
+WebNN has no fused low-bit matmul operation. The dequantized weights and activations therefore use
+the scale dtype (`float16` or `float32`; the validated q4 models use `float32`). Native ORT may use
+a fused packed-weight kernel instead. In particular, `MatMulNBits accuracy_level=4` permits ORT to
+quantize activations to Int8 internally, which the WebNN lowering cannot represent; such models are
+recorded as validation-blocked rather than numerically supported.
+
+`--validate-cached` reuses existing cache artifacts and is exclusive with `--output` and
+`--validate`.
+
Merged decoders (optimum's `decoder_model_merged*.onnx`) branch at runtime on `use_cache_branch`.
WebNN has no runtime `If`, so pin the input and convert each branch separately:
@@ -49,6 +103,12 @@ cargo run -- convert --input decoder_model_merged.onnx --optimize `
Pinned inputs become constants, the chosen `If` branch is inlined, and inputs the branch never
reads (e.g. the KV cache in the prefill branch) and zero-size dummy outputs are dropped.
+Supporting both cache modes therefore produces two independent artifact pairs: one `.webnn` file
+and one Safetensors file for prefill, and another pair for cached decoding. Shared decoder weights
+are not deduplicated between the pairs. Loading both compiled graphs concurrently may duplicate
+those weights in RAM and VRAM; applications can instead keep only the active graph resident and
+manage graph selection and KV-cache handoff themselves.
+
| Flag | Purpose |
|------|---------|
| `--input` | Input `.onnx` path (required) |
@@ -59,6 +119,9 @@ reads (e.g. the KV cache in the prefill branch) and zero-size dummy outputs are
| `--allow-missing-external-data` | Zero-fill external tensors whose data file is absent (weight-stripped skeleton models) |
| `--experimental-dynamic-inputs` | Preserve unresolved symbolic dims as dynamic metadata |
| `--debug` | Verbose conversion logging (global) |
+| `--output` | Overwrite cache-backed `.onnx`, `.webnn`, and Safetensors artifacts |
+| `--validate` | After `--output`, reload and numerically compare against native ORT |
+| `--validate-cached` | Validate existing cache artifacts without reconverting |
On success the CLI prints `✓ ORT graph build succeeded for …`.
@@ -66,12 +129,54 @@ On success the CLI prints `✓ ORT graph build succeeded for …`.
`tests/models/manifest.json` lists the transformers.js exports the converter is expected to handle,
with their dimension overrides and pinned inputs; `tests/model_skeletons.rs` converts each entry and
-builds it in ORT. No weights are downloaded: the test reads each export from the Hugging Face Hub with
+builds it in ORT. This broad manifest is generated by `scripts/generate_manifest.py` and intentionally
+tracks the publisher repositories through their default `main` revisions. It is diagnostic coverage,
+not an immutable CI gate.
+
+No weights are downloaded by the skeleton test: it reads each export from the Hugging Face Hub with
HTTP range requests, keeps the graph and small constants, and points every large initializer at a file
that does not exist, which the converter zero-fills. A 1.4 GB export becomes a ~0.2 MB skeleton for
~10 MB of traffic. Skeletons are kept in `target/model-skeletons` (or `O2W_SKELETON_CACHE`), about
40 MB for the whole manifest, and CI caches that directory keyed on the manifest and scanner source.
+### Numerical-validation CI
+
+Every Linux pull-request job numerically validates the real-weight models in the hand-maintained
+`tests/models/ci-validation.json`; the experimental, non-blocking macOS/CoreML job runs the same
+set. Windows does not run numerical model validation. Each curated entry must be small enough for a
+hosted runner, pass on both ORT and CoreML, and declare an immutable 40-character Hugging Face
+commit `revision` plus the lowercase SHA-256 of its primary ONNX file. Downloads, external-data
+sidecars, skeletons, and cache identities all use that revision. The primary
+model digest is checked after download and whenever a cached model is reused.
+
+Full-model validation downloads through the official Hugging Face Rust client and uses the standard
+`HF_HUB_CACHE`/`HF_HOME` cache by default, allowing Python and Rust tools to reuse the same blobs.
+`O2W_ONNX_CACHE`, followed by `O2W_CACHE_DIR/onnx`, takes precedence when set and is interpreted as
+the Hub cache root. onnx2webnn stores only its completion records in the private
+`.onnx2webnn-validation` directory there; model files retain the standard `blobs` and `snapshots`
+layout. Xet-backed downloads use chunked transfer, while ordinary HTTP downloads are retried but
+are not guaranteed to resume from a partial byte offset.
+
+To add required CI coverage, resolve the model repository to a commit, download the exact ONNX file
+from that revision, compute its SHA-256 (for example with `sha256sum`), and add the file, revision,
+digest, and fixed dimension/input settings to `tests/models/ci-validation.json`. Do not add these
+pins to `tests/models/manifest.json`: regenerating the broad manifest does not read or write the
+curated file. Verify the candidate on Linux and macOS with:
+
+```bash
+target/release/onnx2webnn validate-models \
+ --manifest tests/models/ci-validation.json \
+ --selection all --weights real --jobs 1
+```
+
+The manually dispatched `Full model validation` workflow runs every entry from the generated main
+manifest with real publisher weights. It requires a dedicated runner labeled
+`self-hosted`, `linux`, `x64`, and `onnx2webnn-validation`, a writable
+`/var/cache/onnx2webnn`, network access, a filesystem of at least 100 GiB, and approximately 32 GiB
+of RAM. Downloaded source models persist in that directory; exported WebNN artifacts remain
+temporary. Setup, build, and runner-contract failures fail the workflow, while model failures are
+reported as diagnostic warnings and retained in a full log artifact for 14 days.
+
`O2W_MODELS` selects the source: `hub` (the default when `CI` is set), `dir=` for full local
downloads laid out as `--/onnx/.onnx`, or `strip=` to run local files through
the skeleton scanner. Unset outside CI, the sweep is skipped. `O2W_MODEL_FETCH_JOBS` (default 8) sets
diff --git a/docs/model-validation-status.md b/docs/model-validation-status.md
new file mode 100644
index 0000000..7f8be0e
--- /dev/null
+++ b/docs/model-validation-status.md
@@ -0,0 +1,277 @@
+# Full-model numerical validation status
+
+> **Last edited:** `2026-09-23T11:35:58Z`
+> **Checkout:** `fkrall/cache-backed-validation` at `ac6d5ca`
+>
+> **Freshness:** Use this document only when this provenance is recent relative to the relevant
+> code and commits; otherwise verify the implementation, tests, and Git history before relying
+> on it.
+
+This document tracks the **latest recorded** full-model numerical-validation sweep and a concise
+history of changes that materially altered coverage. Only the latest sweep keeps a complete
+per-case ledger; older ledgers and duplicated historical failure tables are intentionally omitted.
+
+The skeleton sweep establishes broad graph-construction coverage. The manifest contains only model
+selection and execution configuration; numerical status and blocker triage live in this document and
+the linked failure analysis. Numerical validation additionally
+exports the converted graph, reloads its `.webnn` and Safetensors artifacts, executes both the
+original ONNX model and the reloaded graph on CPU ONNX Runtime, and compares matching outputs.
+
+Proper dynamic cache-shape support is not currently available. Validation cases use fixed
+dimension overrides and execute as independent snapshots, so they do not validate a complete
+prefill-to-decode loop or growing KV-cache reuse. A decode artifact that
+accepts a fixed past length `N` normally produces a cache of length `N + 1`, which cannot be fed
+back into that same fixed-shape artifact. Repeated decoding therefore still requires proper dynamic
+cache shapes, separately specialized artifacts, or a fixed-capacity cache with an explicit position.
+
+## Broad manifest selection
+
+[`tests/models/manifest.json`](../tests/models/manifest.json) is not a global top-50 list. The
+[`scripts/generate_manifest.py` generator](../scripts/generate_manifest.py) queries Hugging Face for
+repositories tagged `transformers.js`, groups them by pipeline task, and by default selects the
+single most-liked repository in each task (`--top 1`). Tokenizer-only repositories, repositories
+without usable ONNX exports, and repositories or components listed in the
+[Transformers.js excluded-model list](transformersjs_excluded_models.md) are omitted.
+
+For each selected repository, the generator includes the ONNX components that Transformers.js
+would load for that model class, such as both encoder and merged decoder graphs for a sequence-to-
+sequence model. It follows the model's `transformers.js_config` dtype when present; otherwise the
+default generator mode mirrors WASM and prefers q8 exports. Missing preferred variants fall back to
+an available export. Models with a `use_cache_branch` input receive separate prefill and decode
+cases, while other multipart models contribute one case per required graph. Consequently the
+current manifest contains 28 repositories, 44 unique ONNX files, and 51 configured cases.
+
+The generator assigns bounded dimensions using onnx2webnn's naming policy and carries forward
+existing per-file dimension overrides, pinned inputs, and other manual fields. It inspects remote
+ONNX structure using HTTP range requests rather than downloading full weights. Regeneration tracks
+publisher `main` revisions and is therefore diagnostic coverage, not an immutable model set. The
+current population was last generated around early September 2026 and entered Git history on
+13 September 2026.
+
+## Validation automation
+
+Pull requests use `tests/models/ci-validation.json` as the required numerical-validation contract.
+Linux/ORT is blocking; macOS/CoreML runs the identical set as an experimental, non-blocking job;
+Windows does not run numerical validation. Membership is defined by the curated file itself and the
+runner invokes it with `--selection all`. Every entry must be suitable for hosted runners and include
+a full immutable Hugging Face commit revision and a lowercase SHA-256 for its primary ONNX file.
+The downloader verifies that digest both after download and on cache reuse. Full-model downloads
+use the standard Hugging Face cache by default so Python and Rust clients can share immutable blobs;
+`O2W_ONNX_CACHE` and `O2W_CACHE_DIR/onnx` remain explicit overrides. Completion records stay
+onnx2webnn-specific and do not replace the manifest digest check.
+
+To add a required model, first confirm it passes on ORT and CoreML, then add its exact file, commit
+revision, downloaded-file digest, fixed dimension overrides, and any pinned inputs to
+`tests/models/ci-validation.json`. This file is hand-maintained. The
+[`scripts/generate_manifest.py`](../scripts/generate_manifest.py) generator only regenerates
+`tests/models/manifest.json` and must not be used to update or carry CI pins.
+
+The `Full model validation` workflow is manually dispatched for real publisher weights.
+It runs the entire generated manifest sequentially on a runner labeled `self-hosted`, `linux`,
+`x64`, and `onnx2webnn-validation`. The runner must provide network access, a writable persistent
+`/var/cache/onnx2webnn`, at least a 100 GiB filesystem, and approximately 32 GiB RAM.
+Downloaded model sources persist; WebNN exports are temporary. Provisioning, storage checks, and
+build failures are fatal. Per-model validation failures remain diagnostic: the workflow emits a
+warning and uploads its complete log for 14 days. There is no scheduled or extended-tier sweep.
+
+### Runtime dtypes and comparison tolerances
+
+WebNN graphs are not globally limited to floating point: RustNN records integer tensors and integer
+operations where their operator contracts allow them. Matrix multiplication is different. The
+portable WebNN `matmul` contract accepts matching `float16` or `float32` operands, and WebNN has no
+fused equivalent of ORT's `com.microsoft.MatMulNBits`.
+
+For q4 models, the original ONNX path gives native ORT packed Uint4 weights and a fused
+`MatMulNBits` node. The reloaded path restores the same packed bytes, widens them for RustNN's
+temporary ORT graph, applies `dequantizeLinear`, and runs ordinary matmul in the scale dtype. The
+models in this manifest use Float32 scales and activations, so their reconstructed matmuls are
+Float32. Different fused/decomposed accumulation orders are mathematically equivalent but not
+bit-identical.
+
+The validator therefore uses these output tolerances:
+
+| Output/path | Comparison |
+|-------------|------------|
+| Float32 without `MatMulNBits` | `1e-5 + 1e-4 * abs(reference)` |
+| Float32 from a model containing q4 `MatMulNBits` | `1e-3 + 1e-3 * abs(reference)` |
+| Float32 from a model containing q8 `MatMulNBits` | `2e-3 + 2e-3 * abs(reference)` |
+| Float16 | `1e-3 + 1e-2 * abs(reference)` |
+| Integer and boolean | Exact |
+
+The q4 Float32 envelope is based on a complete SmolLM2 output scan whose worst normalized
+difference was `9.488e-4`. The q8 envelope is selected from each source graph's
+`MatMulNBits.bits` attribute; it covers the measured Qwen output delta while remaining distinct
+from q4. It does not make ORT's `accuracy_level=4` equivalent to WebNN:
+level 4 permits internal Int8 activation quantization, while the WebNN lowering keeps Float32
+activations. Voxtral uses level 4 on all 211 `MatMulNBits` nodes and remains blocked for this reason.
+
+## Latest recorded sweep
+
+- Real sweep: 2026-09-22 on `fkrall/cache-backed-validation` at `a582b89` plus the uncommitted Hugging Face cache migration worktree
+- RustNN: `38022044`
+- ORT: repository-local Linux x64 1.29.0 build
+- Manifest: `tests/models/manifest.json` (51 cases, 44 unique ONNX files)
+- Execution: one validation worker; ORT may use multiple CPU threads inside a case
+- Skeleton verification: 51/51 passed; the test took 81.1s (133.2s including the release build)
+- Real verification: 47/51 passed in one complete cold-cache process; wall time was 10m 43.6s and peak RSS was 30.3 GiB. No case was skipped or inferred without execution.
+
+All 51 current manifest cases were executed. A newer checkout is not considered the tested baseline
+until the real sweep is rerun and this section is replaced.
+
+```bash
+ORT_DYLIB_PATH=../rustnn/target/onnxruntime/onnxruntime-linux-x64-1.29.0/lib/libonnxruntime.so.1.29.0 \
+ target/release/onnx2webnn validate-models \
+ --selection all --weights real --jobs 1
+```
+
+### Summary
+
+| Weight mode | Pass | Fail | Download skipped | Result |
+|-------------|-----:|-----:|-----------------:|--------|
+| Real | 47 | 4 | 0 | Complete (51/51) |
+
+Two cases exercise unsupported cross-backend execution modes: FastVLM precision-sensitive GQA/DQL
+prefill and Voxtral level-4 quantized matmul. The two Chronos publisher artifacts are rejected by
+native ORT before conversion comparison.
+
+The real-weight failures, exact affected cases, current diagnosis, and suggested ownership are
+tracked in [Real-weight validation failures](real-weight-validation-failures.md).
+
+### Current failure families
+
+| Code | Cases | Stage | Current cause / next action |
+|------|------:|-------|-----------------------------|
+| Q1 | 1 | Comparison | Unsupported execution semantics: Voxtral sets `MatMulNBits accuracy_level=4`, allowing native ORT to quantize activations to Int8; WebNN lowers to Float32 dequantize-plus-matmul. |
+| U1 | 1 | Comparison | Unsupported strict equivalence: FastVLM prefill repeatedly quantizes attention results whose legal floating-point rounding differs between fused ORT kernels and the portable WebNN decomposition; decode passes. |
+| O1 | 2 | Native ORT load | Upstream-blocked: both Chronos paths resolve to the same publisher artifact, which feeds a float ConstantOfShape result to Gather indices. |
+
+Real totals: no export, reload, cached-interface, or native-ORT input failures; 2 native-ORT
+model-load failures, 2 unsupported execution-mode mismatches, and 47 passes.
+
+### Current case ledger
+
+`PASS` means export, reload, both executions, and numerical comparison completed. Any other value
+is the current first-blocker code from the table above.
+
+`WebNN If split` identifies merged decoders whose runtime `use_cache_branch` condition must be
+pinned because WebNN has no `If` operation. Each prefill/decode entry produces its own independent
+`.webnn`/Safetensors pair. Loading both graphs concurrently may duplicate their shared weights in
+RAM and VRAM. This is distinct from ordinary repeated cases that differ only in fixed dimensions.
+
+| # | Manifest case | Real | WebNN If split |
+|---:|---------------|------|----------------|
+| 0 | `briaai--RMBG-1.4 :: model_quantized.onnx` | PASS | — |
+| 1 | `openai--privacy-filter :: model_quantized.onnx` | PASS | — |
+| 2 | `nomic-ai--nomic-embed-text-v1.5 :: model_quantized.onnx` | PASS | — |
+| 3 | `mixedbread-ai--mxbai-embed-large-v1 :: model_quantized.onnx` | PASS | — |
+| 4 | `HuggingFaceTB--SmolLM2-1.7B-Instruct :: model_q4.onnx` (`sequence=64`, `past=0`) | PASS | — |
+| 5 | `HuggingFaceTB--SmolLM2-1.7B-Instruct :: model_q4.onnx` (`sequence=1`, `past=64`) | PASS | — |
+| 6 | `distil-whisper--distil-large-v2 :: encoder_model_quantized.onnx` | PASS | — |
+| 7 | `distil-whisper--distil-large-v2 :: decoder_model_merged_quantized.onnx` (`cache=0`) | PASS | Prefill pair (1/2) |
+| 8 | `distil-whisper--distil-large-v2 :: decoder_model_merged_quantized.onnx` (`cache=1`) | PASS | Decode pair (2/2) |
+| 9 | `jinaai--jina-reranker-v2-base-multilingual :: model_quantized.onnx` | PASS | — |
+| 10 | `onnx-community--FastVLM-0.5B-ONNX :: embed_tokens_quantized.onnx` | PASS | — |
+| 11 | `onnx-community--FastVLM-0.5B-ONNX :: decoder_model_merged_quantized.onnx` (`sequence=64`, `past=0`) | U1 | — |
+| 12 | `onnx-community--FastVLM-0.5B-ONNX :: decoder_model_merged_quantized.onnx` (`sequence=1`, `past=64`) | PASS | — |
+| 13 | `onnx-community--FastVLM-0.5B-ONNX :: vision_encoder_quantized.onnx` | PASS | — |
+| 14 | `Marqo--marqo-fashionSigLIP :: text_model_quantized.onnx` | PASS | — |
+| 15 | `Marqo--marqo-fashionSigLIP :: vision_model_quantized.onnx` | PASS | — |
+| 16 | `AdamCodd--vit-base-nsfw-detector :: model_quantized.onnx` | PASS | — |
+| 17 | `Xenova--nllb-200-distilled-600M :: encoder_model_quantized.onnx` | PASS | — |
+| 18 | `onnx-community--Janus-Pro-1B-ONNX :: language_model_q4.onnx` | PASS | — |
+| 19 | `onnx-community--Janus-Pro-1B-ONNX :: lm_head.onnx` | PASS | — |
+| 20 | `onnx-community--Janus-Pro-1B-ONNX :: gen_head.onnx` | PASS | — |
+| 21 | `onnx-community--Janus-Pro-1B-ONNX :: gen_img_embeds.onnx` | PASS | — |
+| 22 | `onnx-community--Janus-Pro-1B-ONNX :: image_decode.onnx` | PASS | — |
+| 23 | `Xenova--musicgen-small :: text_encoder_quantized.onnx` | PASS | — |
+| 24 | `Xenova--musicgen-small :: decoder_model_merged_quantized.onnx` (`cache=0`) | PASS | Prefill pair (1/2) |
+| 25 | `Xenova--musicgen-small :: decoder_model_merged_quantized.onnx` (`cache=1`) | PASS | Decode pair (2/2) |
+| 26 | `Mozilla--distilvit :: encoder_model_quantized.onnx` | PASS | — |
+| 27 | `onnx-community--Voxtral-Mini-3B-2507-ONNX :: embed_tokens_fp16.onnx` | PASS | — |
+| 28 | `onnx-community--Voxtral-Mini-3B-2507-ONNX :: decoder_model_merged_q4.onnx` | Q1 | — |
+| 29 | `onnx-community--Voxtral-Mini-3B-2507-ONNX :: audio_encoder_quantized.onnx` | PASS | — |
+| 30 | `Xenova--LaMini-Flan-T5-783M :: encoder_model_quantized.onnx` | PASS | — |
+| 31 | `Xenova--LaMini-Flan-T5-783M :: decoder_model_merged_quantized.onnx` (`cache=0`) | PASS | Prefill pair (1/2) |
+| 32 | `Xenova--LaMini-Flan-T5-783M :: decoder_model_merged_quantized.onnx` (`cache=1`) | PASS | Decode pair (2/2) |
+| 33 | `Xenova--detr-resnet-50 :: model_quantized.onnx` | PASS | — |
+| 34 | `Xenova--donut-base-finetuned-docvqa :: encoder_model_quantized.onnx` | PASS | — |
+| 35 | `Xenova--donut-base-finetuned-docvqa :: decoder_model_merged_quantized.onnx` (`cache=0`) | PASS | Prefill pair (1/2) |
+| 36 | `Xenova--donut-base-finetuned-docvqa :: decoder_model_merged_quantized.onnx` (`cache=1`) | PASS | Decode pair (2/2) |
+| 37 | `onnx-community--dinov3-vits16-pretrain-lvd1689m-ONNX :: model.onnx` | PASS | — |
+| 38 | `Xenova--distilbart-cnn-6-6 :: encoder_model_quantized.onnx` | PASS | — |
+| 39 | `Xenova--distilbart-cnn-6-6 :: decoder_model_merged_quantized.onnx` (`cache=0`) | PASS | Prefill pair (1/2) |
+| 40 | `Xenova--distilbart-cnn-6-6 :: decoder_model_merged_quantized.onnx` (`cache=1`) | PASS | Decode pair (2/2) |
+| 41 | `prithivMLmods--Common-Voice-Gender-Detection-ONNX :: model_quantized.onnx` | PASS | — |
+| 42 | `Xenova--bert-base-multilingual-cased :: model_quantized.onnx` | PASS | — |
+| 43 | `Xenova--distilbert-base-cased-distilled-squad :: model_quantized.onnx` | PASS | — |
+| 44 | `onnx-community--vitpose-base-simple :: model_quantized.onnx` | PASS | — |
+| 45 | `kashif--chronos-2-onnx :: encoder_model.onnx` | O1 | — |
+| 46 | `kashif--chronos-2-onnx :: decoder_model_merged.onnx` | O1 | — |
+| 47 | `huggingworld--Qwen2.5-VL-3B-Instruct-ONNX :: embed_tokens_quantized.onnx` | PASS | — |
+| 48 | `huggingworld--Qwen2.5-VL-3B-Instruct-ONNX :: decoder_model_merged_quantized.onnx` | PASS | — |
+| 49 | `onnx-community--timesformer-base-finetuned-k400 :: model_quantized.onnx` | PASS | — |
+| 50 | `Xenova--tiny-random-RoFormerForMultipleChoice :: model_quantized.onnx` | PASS | — |
+
+### Current timing and storage
+
+| Run | Cases | Wall time |
+|-----|------:|---------------:|
+| Real, cold standard HF cache | 51 | 10m 43.6s |
+
+The official Rust client populated 44 completion records, one per unique ONNX file. A subsequent
+offline `match=privacy-filter` run verified warm reuse of both the primary ONNX and its external-data
+sidecar.
+
+After the sweep, the standard Hugging Face cache occupied 18 GB and temporary WebNN artifacts
+occupied 32 GB; 122 GiB remained free. No selected case was skipped for download or storage.
+
+### Current repair order
+
+1. Track WebNN/backend capabilities that could make U1 or Q1 executable with comparable fused precision semantics.
+2. Monitor the upstream Chronos repository for corrected reference exports.
+
+## Coverage change history
+
+This history records only changes that explain coverage movement or establish that an apparent
+movement was not a functional improvement. Detailed obsolete ledgers are available through Git.
+
+| Date / tested revisions | Change | Comparable coverage effect |
+|-------------------------|--------|----------------------------|
+| 2026-09-17 — validation-tier cleanup after `25ea94c` | Removed manifest validation tiers and the hand-added FP32 RoFormer smoke case, restoring the generator-owned upstream population. Status and blocker ownership now live only in these validation documents; the CLI selects `all` or `match=`. | Population changed **52 → 51** by removing one passing duplicate-model variant. Comparable coverage is unchanged: skeleton **51/51** and real validation **47/51** with the same four blockers. |
+| 2026-09-17 — onnx2webnn `25ea94c`, RustNN `28fb3bbe`, post-rebase complete rerun | Rebased both feature stacks onto `rustnn/onnx2webnn:main` and `rustnn/rustnn:main`, then ran all 52 skeleton cases and all 52 real-weight cases from warm caches, including heavy entries. | No regression: skeleton remained **52/52**; real validation remained **48/52** with the identical FastVLM U1, Voxtral Q1, and two Chronos O1 blockers. |
+| 2026-09-17 — onnx2webnn `d4d350b`, RustNN `724d076b`, targeted FastVLM probes | Exposed layer boundaries and compared ORT CPU flash/non-flash GQA against the WebNN decomposition. Layer 0 and external rotary matched; the first failing boundary followed layer-1 GQA dynamic quantization and output projection. Native ORT first-logit output changed from `0.669124` to `1.066261` when only its GQA kernel changed, while WebNN remained `1.211192`. | Reclassified FastVLM prefill from generic N1 to unsupported U1. Coverage remains **48/52**; no tolerance was weakened. |
+| 2026-09-17 — onnx2webnn `6f6ad8a` plus current Cast/Slice/comparison worktree, RustNN `65e76e67` plus Slice-backend worktree | Normalized ONNX numeric-to-Bool Cast through comparison, preserved positive Slice strides with WebNN extent semantics, and selected q4/q8 comparison envelopes from `MatMulNBits.bits`. | DETR, Donut encoder, and Qwen passed; real coverage rose **45 → 48** on the same 52 cases. FastVLM prefill, Voxtral level 4, and the two invalid Chronos artifacts remain blocked. |
+| 2026-09-16 — onnx2webnn `24fdd50` plus current validator/manifest worktree, RustNN `7f07a5e1` plus packed-4-bit archive worktree | Measured complete q4 output error distributions and applied a `1e-3` absolute/relative Float32 envelope only to source graphs containing `MatMulNBits`. Classified Voxtral's `accuracy_level=4` Int8-activation execution mode as unsupported. | SmolLM2 prefill/decode and Janus passed; Voxtral remained Q1. Real passes rose **42 → 45** on the same 52 cases. |
+| 2026-09-16 — onnx2webnn `24fdd50` plus current manifest/tests worktree, RustNN `7f07a5e1` plus packed-4-bit archive worktree | Stored packed Int4/Uint4 constants as versioned U8 Safetensors payloads while preserving logical dtype and shape in `.webnn`, then restored and executed them on reload. Marked the identical invalid Chronos publisher artifacts as upstream-blocked. | E1 was eliminated from the real sweep: all four q4 cases reached comparison and exposed N1. The aggregate remained **42/52** because those cases do not yet match numerically. |
+| 2026-09-15 — onnx2webnn `ec5ba275` plus current validator worktree, RustNN `7f07a5e1` | Reconciled native ONNX interfaces with branch-specialized cached WebNN interfaces, dispatching only retained inputs and accepting omitted outputs only when native ORT proves they are empty. | Real passes rose **32 → 42** on the same 52 cases. All ten V1 cases passed numerically and no new blocker family appeared. |
+| 2026-09-14 — onnx2webnn `aaa33e2` plus current validator worktree, RustNN `7f07a5e1` | Made standard `token_type_ids` zero, `attention_mask` one, and preserved zero-element input buffers without treating scalars as empty. | Real passes rose **28 → 32** on the same 52 cases. I1-I3 were eliminated: four cases passed, five exposed V1, and two exposed N1. |
+| 2026-09-12 — onnx2webnn `4926c3e`, RustNN `7f07a5e1` | RustNN `2e22b3db` unified MLGraphBuilder recording and GraphJSON loader inference, replacing the loader-only string-based inference loop. Typed inference added the previously missing/rejected Resample2d, RoundEven/Clamp, Conv2d, LogicalNot, and LogicalAnd reload paths. | On the same 52-case manifest, real passes rose **6 → 28**. Former reload families R1–R5 were eliminated; newly reachable cases exposed N1 and V1 instead. |
+| 2026-09-12 — RustNN `7f07a5e1` | Made `[]` unambiguously scalar and required completed GraphInfo descriptors to have known shapes. | Hardened the shared inference path but did not itself add the dispatch that cleared R1–R5; no separate coverage gain is attributed without a bisect. |
+| 2026-09-09 — onnx2webnn `e5a5f88c`, RustNN `2783a191` | Rebased onto the newer upstream model manifest and converter work. The manifest changed from 63 to 52 cases: 21 retained, 31 added, 42 removed. | Retained cases preserved their outcomes. Real coverage changed to 6 passes because the population changed, not because comparable coverage improved. One real Voxtral download was skipped. |
+| 2026-09-08/09 — onnx2webnn `f1ca548`, RustNN `b64cf495` | First complete real-weight full-model baseline with cache-backed export, reload, deterministic execution, and comparison. | 7/63 real cases passed. Every case that reached comparison matched; failures established the original reload, serialization, and input blocker families. |
+
+### RustNN inference-refactor attribution
+
+RustNN `2e22b3db` removed the separate ten-pass GraphJSON shape-inference implementation. Both
+normal graph construction and GraphJSON loading now pass typed operations through
+`GraphRecorder::record_operation` and `infer_operation_descriptors`. This directly removed:
+
+- R1: missing Resample2d output-shape inference.
+- R2: inconsistent RoundEven spelling and missing Clamp inference.
+- R3: missing Conv2d output-shape inference.
+- R4: inconsistent LogicalNot spelling.
+- R5: inconsistent LogicalAnd spelling.
+
+The full manifest was tested at `7f07a5e1`, not bisected at `2e22b3db`; attribution is based on
+the code changes between the recorded revisions. A focused reload benchmark found the changed
+graph reconstruction/inference stage about 6% faster on seven cached graphs, while end-to-end
+weight-heavy loading was effectively unchanged.
+
+## Maintenance rule
+
+When a new complete sweep is recorded:
+
+1. Replace the latest metadata, summary, failure counts, timing, and the **single** case ledger.
+2. Add one concise history row only when a code/revision change explains coverage movement.
+3. Do not append another historical case ledger or duplicate obsolete failure-family tables.
+4. Distinguish manifest/population changes from improvements on comparable cases.
diff --git a/docs/operator-conversion-plan.md b/docs/operator-conversion-plan.md
deleted file mode 100644
index 704b23b..0000000
--- a/docs/operator-conversion-plan.md
+++ /dev/null
@@ -1,483 +0,0 @@
-# ONNX → WebNN Operator Conversion Plan
-
-Phased plan for **onnx2webnn** to convert every **ai.onnx** operator at its **latest schema
-version** within the current ONNX standard. Baseline: **opset 26** (ONNX package 1.21.0,
-`onnx.defs.onnx_opset_version()`).
-
-Each stage below is a **conversion category**, not a calendar milestone. Work within a stage can
-proceed in parallel once opset-26 schema audits and the per-op workflow (see [Implementation
-workflow](#implementation-workflow)) are in place.
-
----
-
-## Goals
-
-1. Accept models declaring **ai.onnx opset 26** (raise `MIN/MAX_SUPPORTED_OPSET` from 11–18).
-2. Convert every ONNX operator that has a WebNN target — directly, via decomposition, or via
- convert-time constant folding.
-3. Reject or pre-fold operators with no WebNN equivalent, with explicit error messages.
-4. Track progress against a regenerated opset inventory (`webnn_exporter_supported` column).
-
-## Non-goals
-
-- Supporting multiple ONNX opsets simultaneously (target opset 26 only after completion).
-- Executing converted graphs at runtime (conversion validates via rustnn ORT `build()` only).
-- Implementing ONNX control flow (`If`, `Loop`, `Scan`) as WebNN graphs.
-
----
-
-## Coverage baseline (opset 26)
-
-| Category | Count | Share |
-|----------|------:|------:|
-| Total `ai.onnx` operators at opset 26 | 198 | 100% |
-| **Stage 1 — Direct mappable** | ~95 | ~48% |
-| **Stage 2 — Complex mappable** | ~19 | ~10% |
-| **Stage 3 — Impossible to map** | ~84 | ~42% |
-
-| Implementation status (today) | Count |
-|-------------------------------|------:|
-| Handlers in `src/onnx/ops/*.rs` | 59 |
-| Stage 1 backlog (direct, not yet exported) | ~36 |
-| Stage 2 backlog (complex) | ~19 |
-
-Regenerate counts after each handler lands:
-
-```powershell
-.\.venv\Scripts\python.exe scripts\generate_onnx_opsets.py --min 26 --max 26 -o docs\onnx-opsets
-.\.venv\Scripts\python.exe -c "import csv; r=list(csv.DictReader(open('docs/onnx-opsets/opset-26.csv'))); print(sum(1 for x in r if x['webnn_exporter_supported']=='yes'))"
-```
-
-`generate_onnx_opsets.py` defaults to `../webnn-graph/docs/onnx-opsets/`; pass `-o docs\onnx-opsets` to keep
-inventory CSVs in this repo.
-
-**Related artifacts**
-
-| Artifact | Path |
-|----------|------|
-| Operator inventory (regenerate for opset 26) | `docs/onnx-opsets/opset-26.csv` (or `../webnn-graph/docs/onnx-opsets/`) |
-| Exporter manifest | `scripts/webnn_onnx_ops.py` |
-| Op handlers (Rust) | `src/onnx/ops/*.rs`, registry in `src/onnx/ops/mod.rs` |
-| Opset gate | `src/onnx/convert.rs` (`MIN/MAX_SUPPORTED_OPSET`) |
-| ONNX ↔ WebNN name map | `webnn-onnx-utils/src/operation_names.rs` |
-| Per-op conversion tests | `tests/onnx_ops/{category}/{op}.rs` (auto-generated) |
-| Test generator | `scripts/generate_rust_op_conversion_tests.py` |
-| Model audit CLI | `scripts/onnx_ops_to_csv.py --check-webnn` |
-| Opset upgrade (ad hoc) | `scripts/upgrade_onnx_opset.py` |
-
-**Python scripts (`scripts/`)**
-
-| Script | Role |
-|--------|------|
-| `webnn_onnx_ops.py` | Supported-op manifest; keep in sync with `src/onnx/ops/*.rs` |
-| `generate_rust_op_conversion_tests.py` | Regenerate `tests/onnx_ops/` and `tests/onnx_op_tests.rs` |
-| `onnx_fixture_builders.py`, `onnx_test_builders.py`, `rust_model_emitter.py` | Libraries used by the test generator |
-| `onnx_ops_to_csv.py` | Audit a model's operators and WebNN support |
-| `upgrade_onnx_opset.py` | Upgrade a model to a target `ai.onnx` opset |
-| `generate_onnx_opsets.py` | Per-opset operator inventory CSVs (`webnn_exporter_supported` column) |
-
-Install deps from repo root: `pip install -r requirements.txt` (needs `onnx`, `numpy`).
-
----
-
-## Three conversion stages
-
-```
-ONNX model (opset 26, latest schema per op)
- │
- ▼
-┌───────────────────────┐
-│ Opset gate │ convert.rs → opset 26 only
-└───────────┬───────────┘
- ▼
-┌───────────────────────┐
-│ Stage 1: Direct │ 1 ONNX node → 1 WebNN op (± trivial attrs)
-└───────────┬───────────┘
- ▼
-┌───────────────────────┐
-│ Stage 2: Complex │ fold, decompose, or multi-node subgraph
-└───────────┬───────────┘
- ▼
-┌───────────────────────┐
-│ Stage 3: Impossible │ reject (or fold-only for metadata ops)
-└───────────────────────┘
-```
-
-### Stage 1 — Direct mappable operations
-
-ONNX operators that map to a **single** [WebNN `MLGraphBuilder`
-method](https://www.w3.org/TR/webnn/#operators) with straightforward attribute and I/O translation.
-No multi-node decomposition required; optional schema-version audit for ops whose active schema
-changed at opset 22/25/26.
-
-#### 1a. Implemented today (59)
-
-| Group | Operators | Rust module |
-|-------|-----------|-------------|
-| Linear | MatMul, Gemm | `matmul.rs` |
-| Convolution | Conv, ConvTranspose | `conv.rs` |
-| Pooling | MaxPool, AveragePool, GlobalMaxPool, GlobalAveragePool | `pool.rs` |
-| Elementwise binary | Add, Sub, Mul, Div, Pow, Min, Max | `elementwise.rs` |
-| Comparison | Greater, Less, Equal, GreaterOrEqual, LessOrEqual | `comparison.rs` |
-| Conditional | Where | `conditional.rs` |
-| Normalization | LayerNormalization, Softmax | `normalization.rs` |
-| Shape | Reshape, Transpose, Concat, Split, Unsqueeze, Squeeze, Tile, Expand, Flatten | `reshape.rs` |
-| Conversion | Cast, Constant | `conversion.rs` |
-| Utility | Shape, Gather, Slice, ConstantOfShape, Range, Trilu, OneHot | `utility.rs` |
-| Reduction | ReduceMean, ReduceSum, ReduceMax, ReduceMin | `reduction.rs` |
-| Activation / unary | Relu, Gelu, Tanh, Sigmoid, Sqrt, Exp, Log, Abs, Neg, Erf, Cos, Sin, Identity | `activation.rs` |
-| Scatter | ScatterND | `scatter.rs` |
-| Pad | Pad | `pad.rs` |
-
-#### 1b. Direct backlog — not yet exported (~36)
-
-Implement as single-op handlers. Suggested Rust home in parentheses.
-
-| Group | ONNX → WebNN | Module |
-|-------|--------------|--------|
-| Normalization | BatchNormalization → `batchNormalization()` | `normalization.rs` |
-| Normalization | InstanceNormalization → `instanceNormalization()` | `normalization.rs` |
-| Gather | GatherND → `gatherND()` | `utility.rs` |
-| Gather | GatherElements → `gatherElements()` | `utility.rs` |
-| Scatter | ScatterElements → `scatterElements()` | `scatter.rs` |
-| Resample | Resize → `resample2d()` | new `resize.rs` |
-| Activation | Elu → `elu()`, LeakyRelu → `leakyRelu()`, PRelu → `prelu()` | `activation.rs` |
-| Activation | HardSigmoid → `hardSigmoid()`, HardSwish → `hardSwish()` | `activation.rs` |
-| Activation | Softplus → `softplus()`, Softsign → `softsign()` | `activation.rs` |
-| Unary math | Ceil → `ceil()`, Floor → `floor()`, Reciprocal → `reciprocal()` | `activation.rs` |
-| Unary math | Sign → `sign()`, Tan → `tan()`, Round → `roundEven()` | `activation.rs` |
-| Clamp | Clip → `clamp()` | `activation.rs` |
-| Logical | Not → `logicalNot()`, And → `logicalAnd()`, Or → `logicalOr()`, Xor → `logicalXor()` | `comparison.rs` |
-| Predicate | IsNaN → `isNaN()`, IsInf → `isInfinite()` | `comparison.rs` |
-| Reduction | ArgMin → `argMin()`, ArgMax → `argMax()` | `reduction.rs` |
-| Reduction | ReduceL1 → `reduceL1()`, ReduceL2 → `reduceL2()` | `reduction.rs` |
-| Reduction | ReduceLogSum → `reduceLogSum()`, ReduceLogSumExp → `reduceLogSumExp()` | `reduction.rs` |
-| Reduction | ReduceProd → `reduceProduct()`, ReduceSumSquare → `reduceSumSquare()` | `reduction.rs` |
-| Reduction | CumSum → `cumulativeSum()` | `reduction.rs` |
-| Pool | LpPool → `l2Pool2d()`, GlobalLpPool → `globalAveragePool` + L2 (verify spec) | `pool.rs` |
-| Sequence | ReverseSequence → `reverse()` | `utility.rs` |
-
-**Stage 1 exit criteria:** ~95 operators with `webnn_exporter_supported=yes`; all direct-backlog
-fixtures convert; Birds EfficientNet and a small transformer ONNX convert without unsupported-op
-errors for Stage 1 ops.
-
-**Suggested implementation order (by model impact):**
-
-1. BatchNormalization, InstanceNormalization, GatherND, GatherElements, ScatterElements, Resize
-2. Remaining activations and unary math (Elu, LeakyRelu, PRelu, Clip, Ceil, Floor, …)
-3. Reduction family completion (ArgMin/Max, ReduceL1/L2/…, CumSum)
-4. Logical ops (Not, And, Or, Xor, IsNaN, IsInf)
-5. LpPool, ReverseSequence
-
----
-
-### Stage 2 — Complex mappable operations
-
-Operators **can** be lowered to WebNN, but conversion requires one or more of:
-
-- **Convert-time constant folding** (no runtime graph node)
-- **Multi-node decomposition** into Stage 1 ops
-- **Non-trivial attribute / mode mapping** (document limitations and reject unsupported modes early)
-- **Schema-version-specific inputs** (e.g. opset 25 `pads` tensor on Pad)
-
-#### 2a. Convert-time fold only (no runtime WebNN op)
-
-These appear in graphs but become constants or metadata during `--optimize` / conversion — not
-`MLGraphBuilder` nodes at inference time.
-
-| Operator | Strategy | Notes |
-|----------|----------|-------|
-| Constant | Inline tensor data into graph const pool | Already handled; audit opset 25 schema |
-| ConstantOfShape | Fold when shape is static | Audit opset 25 `value` attribute |
-| Range | Fold when start/limit/delta are constant | |
-| Shape | Fold when input shape is known | |
-| Size | Fold to scalar product of static dimensions | Reject if shape is dynamic |
-
-#### 2b. Decompose into Stage 1 subgraphs
-
-| Operator | Decomposition sketch | Priority |
-|----------|---------------------|----------|
-| CastLike | Type-infer `to` from second input → `Cast` | P1 |
-| Mean | `ReduceMean` over all axes (or specified) | P2 |
-| Sum | `ReduceSum` over all axes (or specified) | P2 |
-| LogSoftmax | `Softmax` + `Log` | P2 |
-| Hardmax | `Softmax` + one-hot style mask (or reject axis edge cases) | P3 |
-| Swish | `Sigmoid` × input (`Mul`) | P2 |
-| Celu | `Elu` variant with alpha (after Elu lands) | P3 |
-| Selu | scale × ELU-like piecewise | P3 |
-| Mish | `Softplus` + `Tanh` + `Mul` (after Softplus lands) | P3 |
-| ThresholdedRelu | `Relu` + comparison + `Where` | P3 |
-| GroupNormalization | Mean/var over group dims + `Sub`/`Div`/`Mul`/`Add` | P1 — common in vision |
-| RMSNormalization | L2 norm over axes + scale (new in opset 23) | P1 — common in LLMs |
-| SimplifiedLayerNormalization | Alias of RMSNormalization | Implemented — standard/exporter fusion |
-| SkipSimplifiedLayerNormalization | Residual/bias add + RMSNormalization | Implemented — com.microsoft fusion |
-| RotaryEmbedding | Gather caches + rotate pairs + concatenate | Implemented — standard and com.microsoft |
-| MatMulNBits | uint4/uint8 reinterpret + dequantizeLinear + transpose + matmul | Implemented — bits=4/8; reject g_idx |
-| MatMulInteger | centered float `matmul` + `int32` cast (mirrors ConvInteger) | Implemented — scalar or 1-D zero points |
-| Einsum | reduceSum + transpose/reshape + batched `matmul` | Implemented — ≤2 inputs; reject ellipsis/diagonals |
-| MatMulBnb4 | packed uint8 constant, nibbles unpacked with int32 div/sub, codebook `gather`, per-block absmax `mul`, `matmul` | Implemented; weight stays 4-bit in the graph, dense fallback when N*K is not a whole number of blocks |
-| GroupQueryAttention | reshape/transpose + KV-cache concat + scaled matmul + static causal mask + softmax | Implemented — unpadded batch assumption; reject packed QKV/do_rotary/softcap/local window |
-| MoE | dense expert evaluation (batched matmul) + iterative top-k mask + masked softmax blend | Implemented — fused interleaved swiglu/relu/gelu/sigmoid; reject fc3/sparse mixer; num_experts/k× FLOP overhead |
-| QMoE | uint4/uint8 expert constants with blockwise `dequantizeLinear`, then the MoE lowering | Implemented; same restrictions as MoE, dense fallback when `in` is not a whole number of blocks |
-| GatherBlockQuantized | `gather` packed rows, scales and zero points, unpack nibbles, `dequantizeLinear` on the gathered slice | Implemented; 2-D tables along axis 0 (uint8, INT4, UINT4), other layouts dequantize at conversion time |
-| SplitToSequence + SequenceAt | per-element slices registered as pseudo-operands, SequenceAt aliases by constant index | Implemented — torch split() export pattern; sequence graph outputs rejected |
-| If (constant condition) | branch inlined into the outer graph before conversion; a boolean graph input gate (optimum's `use_cache_branch`) becomes constant via `--pin-input` | Implemented — data-dependent conditions remain unsupported |
-| LSTM/GRU (direction=bidirectional/reverse) | WebNN native direction support + [dirs,...] bias split and Y layout | Implemented — LSTM initial_h/initial_c wired; sequence_lens/peephole rejected |
-| CumProd | `cumulativeSum` on log + `Exp`, or iterative multiply subgraph | P3 — new in opset 26 |
-
-#### 2c. Advanced single-entry WebNN ops (high attribute complexity)
-
-Treat as Stage 2 because spec coverage and testing cost exceed typical Stage 1 ops.
-
-| Operator | WebNN target | Key risks |
-|----------|--------------|-----------|
-| GRU | `gru()` / `gruCell()` | Layout, direction, hidden size, optional inputs |
-| LSTM | `lstm()` / `lstmCell()` | Same as GRU; bidirectional |
-| RNN | Partial via GRU/LSTM or reject | No dedicated WebNN `rnn()` |
-| QuantizeLinear | `quantizeLinear()` | Scale/zero-point types, axis |
-| DequantizeLinear | `dequantizeLinear()` | Pair with quant models |
-| Pad | `pad()` | Opset 25: `pads` as input tensor; verify `pad.rs` |
-| Resize | `resample2d()` | Coordinate transformation modes, cubic exclusion |
-| Gemm | `gemm()` | Already exported; audit transA/transB, alpha/beta at opset 13+ |
-| Conv / Pool | `conv2d()` / pools | Audit opset 22 dilations, auto_pad, ceil_mode |
-
-**Stage 2 exit criteria:** All fold/decompose paths documented in lowering notes; GroupNormalization
-and RMSNormalization decompositions tested on at least one real model; GRU/LSTM/quant handlers
-pass unit tests with documented limitations.
-
----
-
-### Stage 3 — Impossible to map operations
-
-No WebNN `MLGraphBuilder` equivalent exists. **Do not** add graph handlers. Strategy: **reject**
-at conversion with a clear message, unless the op can be eliminated earlier (optimizer, training
-vs inference export).
-
-#### 3a. Control flow and graph structure
-
-| Operators | Reason |
-|-----------|--------|
-| If, Loop, Scan | WebNN has no conditional or recurrent graph execution |
-
-#### 3b. Sequence and optional types
-
-| Operators | Reason |
-|-----------|--------|
-| SequenceAt, SequenceConstruct, SequenceEmpty, SequenceErase, SequenceInsert, SequenceLength, SequenceMap, SplitToSequence, ConcatFromSequence | WebNN tensors only; no sequence type |
-| Optional, OptionalGetElement, OptionalHasElement | No optional type in WebNN |
-
-#### 3c. Strings and NLP utilities
-
-| Operators | Reason |
-|-----------|--------|
-| StringConcat, StringSplit, StringNormalizer, RegexFullMatch, TfIdfVectorizer | No string tensors in WebNN |
-
-#### 3d. Random and stochastic
-
-| Operators | Reason |
-|-----------|--------|
-| Bernoulli, Multinomial, RandomNormal, RandomNormalLike, RandomUniform, RandomUniformLike | Non-deterministic; no WebNN RNG ops |
-
-#### 3e. Training, loss, and dropout
-
-| Operators | Reason |
-|-----------|--------|
-| Dropout, NegativeLogLikelihoodLoss, SoftmaxCrossEntropyLoss | Training-only; inference graphs should fold or omit |
-
-#### 3f. Detection, vision, and spatial extras
-
-| Operators | Reason |
-|-----------|--------|
-| NonMaxSuppression, RoiAlign, DeformConv, GridSample, ImageDecoder, MaxRoiPool, MaxUnpool, AffineGrid, CenterCropPad, Col2Im, DepthToSpace, SpaceToDepth | No WebNN ops; post-processing or custom CUDA territory |
-
-#### 3g. Signal processing and windows
-
-| Operators | Reason |
-|-----------|--------|
-| DFT, STFT, MelWeightMatrix, BlackmanWindow, HammingWindow, HannWindow | No WebNN signal ops |
-
-#### 3h. Quantized conv / matmul (integer)
-
-| Operators | Reason |
-|-----------|--------|
-| QLinearConv, QLinearMatMul | Fused quantized ops not in WebNN; use explicit float + quantization decomposition. `MatMulInteger` is now implemented via centered float matmul |
-
-`ConvInteger` is supported through centered float `conv2d` followed by an `int32` cast,
-`DynamicQuantizeLinear` is decomposed into reductions, scale/zero-point arithmetic, and WebNN
-`quantizeLinear`, and `MatMulNBits` (bits=4) follows ORT's WebNN EP path:
-`dequantizeLinear` → reshape → transpose → `matmul` (+ optional bias).
-
-#### 3i. Bitwise and type reinterpretation
-
-| Operators | Reason |
-|-----------|--------|
-| BitwiseAnd, BitwiseOr, BitwiseXor, BitwiseNot, BitShift, **BitCast** (new opset 26) | No bitwise ops in WebNN |
-
-#### 3j. Linear algebra and einsum
-
-| Operators | Reason |
-|-----------|--------|
-| Det | No determinant in WebNN. `Einsum` is now implemented for one/two-input equations without ellipsis or repeated per-term labels (lowered to reduceSum/transpose/reshape/matmul) |
-
-#### 3k. Indexing and sorting extras
-
-| Operators | Reason |
-|-----------|--------|
-| TopK, Unique, Compress, NonZero, OneHot, EyeLike | No WebNN equivalents |
-| **TensorScatter** (opset 24) | No WebNN op; use ScatterND/ScatterElements subset only |
-
-#### 3l. Attention and transformer extras (recent ONNX)
-
-| Operators | Reason |
-|-----------|--------|
-| **Attention** (opset 24) | Composite op; no single WebNN op — could be Stage 2 decomposition in future, but out of scope until spec stabilizes |
-| **MoE** (com.microsoft) | Requires TopK expert routing and efficient sparse dispatch, neither available in WebNN |
-
-#### 3m. Miscellaneous
-
-| Operators | Reason |
-|-----------|--------|
-| Acos, Acosh, Asin, Asinh, Atan, Atanh, Cosh, Sinh | No inverse/hyperbolic unary in WebNN |
-| LRN, LpNormalization, MeanVarianceNormalization | Legacy normalization; decompose or reject |
-| Mod, Shrink | No direct WebNN op |
-| Mod, Mish (as dedicated op) | Prefer Stage 2 decomposition via Swish/Mish subgraph if needed |
-
-**Stage 3 handling**
-
-| Strategy | When | Example |
-|----------|------|---------|
-| Reject | No WebNN op and not foldable | If, Loop, Attention |
-| Pre-fold / strip | Training artifact in inference export | Dropout → Identity |
-| Document workaround | Rare; manual graph rewrite | Replace TopK with external post-process |
-
----
-
-## Opset 26 schema revisions
-
-Operators whose **active schema version** changed at or after opset 21 need a fixture pass even when
-already exported. Priority: schema v25 and v26.
-
-| Operator | Active schema | Exporter today | Action |
-|----------|--------------:|----------------|--------|
-| Cast | 25 | yes | Audit `to` / saturate rules |
-| Constant | 25 | yes | Verify sparse / value attributes |
-| ConstantOfShape | 25 | yes | Verify optional `value` |
-| Flatten | 25 | yes | Verify `axis` default |
-| Identity | 25 | yes | Smoke test |
-| Reshape | 25 | yes | Verify `allowzero` |
-| Shape | 25 | yes | Constant-fold path |
-| Squeeze | 25 | yes | Verify axes input |
-| Transpose | 25 | yes | Verify optional `perm` |
-| Unsqueeze | 25 | yes | Verify axes input |
-| Pad | 25 | yes | Confirm `pads` tensor input |
-| If, Loop, Scan | 25 | no | Reject |
-| CastLike | 25 | no | Stage 2 decompose |
-| QuantizeLinear, DequantizeLinear | 25 | no | Stage 2 |
-| Conv, ConvTranspose, MaxPool, AveragePool, … | 22 | partial | Audit dilations, auto_pad |
-| **BitCast** | 26 | no | Stage 3 reject |
-| **CumProd** | 26 | no | Stage 2 decompose |
-
-Regenerate per-op Rust conversion tests (fixture opset 26, converter gate opset 18 today):
-
-```powershell
-.\.venv\Scripts\python.exe scripts\generate_rust_op_conversion_tests.py --fixture-opset 26 --test-opset 18
-cargo test --test onnx_op_tests
-```
-
----
-
-## Implementation workflow
-
-Use this checklist for **every** operator added in Stages 1–2:
-
-- [ ] **Audit real models** — `scripts/onnx_ops_to_csv.py model.onnx --check-webnn`
-- [ ] **Read ONNX schema** — `onnx.defs.get_schema(op, 26)` or opset-26 sidecar fixture
-- [ ] **Read WebNN spec** — method signature, options, `opSupportLimits`
-- [ ] **Implement handler** — `supports()` + `convert()` in `src/onnx/ops/*.rs`
-- [ ] **Shape inference** — update `src/onnx/shape_inference.rs` if needed
-- [ ] **Register** — add to `OpRegistry` in `src/onnx/ops/mod.rs`
-- [ ] **Sync manifest** — add to `scripts/webnn_onnx_ops.py`
-- [ ] **Update name map** — `webnn-onnx-utils/src/operation_names.rs` if new mapping
-- [ ] **Regenerate inventory** — `scripts/generate_onnx_opsets.py --min 26 --max 26 -o docs/onnx-opsets`
-- [ ] **Regenerate tests** — `scripts/generate_rust_op_conversion_tests.py --fixture-opset 26 --test-opset 18`
-- [ ] **Unit tests** — handler tests; `cargo test --test onnx_op_tests {op_snake}::opset26`
-- [ ] **Convert smoke test** — `cargo run -- convert --input model.onnx --optimize` on a real model using the op
-
----
-
-## Rollout phases
-
-Cross-cutting milestones that span the three stages:
-
-### Phase 0 — Foundation
-
-| Task | Files |
-|------|-------|
-| Regenerate opset-26 inventory | `scripts/generate_onnx_opsets.py` → `docs/onnx-opsets/` |
-| Set opset gate to 26 only | `src/onnx/convert.rs`, `scripts/webnn_onnx_ops.py` |
-| Opset 25/26 schema audit for 59 implemented ops | handlers + fixtures |
-| Sync `webnn_onnx_ops.py` with Rust registry | `scripts/webnn_onnx_ops.py` |
-
-**Exit:** Converter accepts opset 26, rejects 18; 59 ops pass v26 fixture conversion.
-
-### Phase 1 — Stage 1 high-impact gaps
-
-BatchNormalization, InstanceNormalization, GatherND, GatherElements, ScatterElements, Resize, Pad
-v25 verification.
-
-**Exit:** EfficientNet-class vision models convert; inventory ≥ 66 supported ops.
-
-### Phase 2 — Stage 1 activations, unary, logical, reductions
-
-Complete Stage 1 backlog table in [§1b](#1b-direct-backlog--not-yet-exported-36).
-
-**Exit:** ~95 supported ops; Stage 1 complete.
-
-### Phase 3 — Stage 2 decomposition and advanced ops
-
-GroupNormalization, RMSNormalization, CastLike, Mean/Sum/LogSoftmax, Swish; then GRU/LSTM/quant.
-
-**Exit:** Documented decomposition paths; optional advanced coverage for seq2seq and int8.
-
-### Phase 4 — Documentation and CI
-
-- Maintain supported/reject lists in lowering docs
-- CI: `scripts/onnx_ops_to_csv.py model.onnx --check-webnn` on checked-in ONNX models
-- Optional: generate `webnn_onnx_ops.py` from Rust `supports()` lists
-
----
-
-## Progress tracking
-
-| Milestone | Ops exported | Opset gate | Notes |
-|-----------|-------------:|------------|-------|
-| **Today** | 59 | 11–18 | ~30% of opset 26 |
-| Phase 0 done | 59 | 26 only | v25/v26 schema audit |
-| Phase 1 done | ~66 | 26 only | Vision models |
-| Phase 2 done | ~95 | 26 only | Stage 1 complete |
-| Phase 3 done | ~95+ | 26 only | Stage 2 decompositions |
-| Stage 3 | — | 26 only | 84 ops permanently rejected |
-
----
-
-## Risk register
-
-| Risk | Mitigation |
-|------|------------|
-| Schema v26 differs from v18 for same op name | Mandatory opset-26 fixture per op; read `schema_since_version` |
-| WebNN `resample2d` ≠ ONNX `Resize` | Document mode/coordinate mapping; reject unsupported modes |
-| `webnn_onnx_ops.py` drifts from Rust | Regenerate CSV after every op; optional codegen from `supports()` |
-| GroupNormalization / RMSNormalization in new models | Stage 2 decomposition before reject |
-| Attention / RotaryEmbedding in LLM exports | Stage 3 reject until decomposition designed |
-| Dynamic shapes | `--optimize` constant folding; static shapes at runtime |
-| MLGraphBuilder migration | Align handlers with [mlgraphbuilder-migration-plan.md](./mlgraphbuilder-migration-plan.md) |
-
----
-
-## References
-
-- [W3C WebNN specification](https://www.w3.org/TR/webnn/)
-- [ONNX Operator Schemas](https://github.com/onnx/onnx/blob/main/docs/Operators.md)
-- [webnn-onnx-utils operation_names.rs](../../webnn-onnx-utils/src/operation_names.rs)
-- [webnn-graph opset 21 plan (historical)](../../webnn-graph/docs/onnx-opsets/opset21-implementation-plan.md)
-- [onnx2webnn AGENTS.md](../AGENTS.md)
diff --git a/docs/operator-conversion-status.md b/docs/operator-conversion-status.md
new file mode 100644
index 0000000..a6ade0f
--- /dev/null
+++ b/docs/operator-conversion-status.md
@@ -0,0 +1,302 @@
+# ONNX → WebNN operator conversion status
+
+> **Last edited:** `2026-09-18T13:48:47Z`
+> **Checkout:** `fkrall/cache-backed-validation` at `558f641`
+>
+> **Freshness:** Use this document only when this provenance is recent relative to the relevant
+> code and commits; otherwise verify the implementation, tests, and Git history before relying
+> on it.
+
+This document is a capability inventory for the current **onnx2webnn** branch. It contains no
+implementation phases or rollout plan. The baseline is the 198 non-deprecated standard
+`ai.onnx` operator names active at opset 26.
+
+An operator marked implemented has at least one executable form. It does not imply that every
+schema revision, attribute combination, data type, rank, optional input/output, or dynamic-shape
+form is supported. The scope column records known material restrictions.
+
+## Status categories
+
+| Category | Meaning |
+|----------|---------|
+| **Direct WebNN** | At least one supported form primarily maps to a standardized WebNN builder primitive. |
+| **Decomposed** | Executes through multiple WebNN primitives because no matching native primitive exists. |
+| **Folded/static** | Resolved or materialized while converting. |
+| **Pattern-only** | Executes only when a specific static export pattern can be eliminated or inlined. |
+| **Unsupported** | No executable lowering exists on this branch. |
+
+Direct WebNN is the spec-level subset of executable support. Decomposed, folded, and pattern-only
+operators execute, but are not native WebNN operator mappings.
+
+### Estimated performance impact
+
+The performance-impact column is populated only for **Decomposed** operators. These estimates are
+derived from the current emitted WebNN graph, not from cross-backend benchmarks; actual cost depends
+on tensor shapes, backend fusion, memory movement, and hardware.
+
+| Tier | Meaning |
+|------|---------|
+| **minimal** | A small number of cheap elementwise, cast, or shape operations; backend fusion can often remove most overhead. |
+| **moderate** | Multiple full-tensor passes, reductions, or materialized intermediates are expected to add noticeable overhead. |
+| **high** | The lowering loses a fused, quantized, integer-accumulation, attention, or sparse-dispatch path and can materially increase compute or memory traffic. |
+| **not inferred** | Impact varies too strongly with the accepted equation, input count, shape, or backend to estimate from the lowering alone. |
+
+## Coverage
+
+| Opset-26 standard ONNX coverage | Count | Share |
+|---------------------------------|------:|------:|
+| Total non-deprecated `ai.onnx` names | 198 | 100% |
+| **Executable for at least one supported form** | **129** | **65.2%** |
+| └ Direct WebNN/spec-level subset | **96** | **48.5%** |
+| └ Decomposed, folded, or pattern-only | **33** | **16.7%** |
+| Unsupported | 69 | 34.8% |
+
+The 129 executable names comprise 127 standard names advertised by
+`scripts/webnn_onnx_ops.py`, plus the narrow registry-only `SplitToSequence` and `SequenceAt`
+pattern. The 96 direct names comprise 86 mappings in the pinned `webnn-onnx-utils` name map plus
+ten direct handlers not yet mirrored there: `CumSum`, `GatherElements`, `GatherND`, `IsInf`,
+`IsNaN`, `LpPool`, `ReverseSequence`, `Round`, `ScatterElements`, and `ScatterND`.
+
+## Complete opset-26 status matrix
+
+| ONNX operator | Status | Execution scope and WebNN limitation | Estimated perf impact |
+|---------------|--------|--------------------------------------|-----------------------|
+| `Abs` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Acos` | Unsupported | No current executable lowering. |
+| `Acosh` | Decomposed | Elementwise decomposition; WebNN has no native acosh. | moderate |
+| `Add` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `AffineGrid` | Unsupported | No current lowering; the required vision/spatial primitive is absent or not implemented. |
+| `And` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `ArgMax` | Direct WebNN | Direct WebNN argMax; select_last_index is not implemented. |
+| `ArgMin` | Direct WebNN | Direct WebNN argMin; select_last_index is not implemented. |
+| `Asin` | Unsupported | No current executable lowering. |
+| `Asinh` | Decomposed | Elementwise decomposition; WebNN has no native asinh. | moderate |
+| `Atan` | Unsupported | No current executable lowering. |
+| `Atanh` | Decomposed | Elementwise decomposition; WebNN has no native atanh. | moderate |
+| `Attention` | Unsupported | No native WebNN attention; no standard Attention decomposition is implemented. |
+| `AveragePool` | Direct WebNN | Direct WebNN pool; 1-D is emulated, higher spatial ranks rejected, and some ONNX padding modes are restricted. |
+| `BatchNormalization` | Direct WebNN | Direct WebNN primitive; inference mode only. |
+| `Bernoulli` | Unsupported | WebNN has no graph random-number operators. |
+| `BitCast` | Unsupported | WebNN has no corresponding bitwise/reinterpretation primitive. |
+| `BitShift` | Unsupported | WebNN has no corresponding bitwise/reinterpretation primitive. |
+| `BitwiseAnd` | Unsupported | WebNN has no corresponding bitwise/reinterpretation primitive. |
+| `BitwiseNot` | Unsupported | WebNN has no corresponding bitwise/reinterpretation primitive. |
+| `BitwiseOr` | Unsupported | WebNN has no corresponding bitwise/reinterpretation primitive. |
+| `BitwiseXor` | Unsupported | WebNN has no corresponding bitwise/reinterpretation primitive. |
+| `BlackmanWindow` | Unsupported | WebNN has no corresponding signal-processing primitive. |
+| `Cast` | Direct WebNN | Direct WebNN cast where the target type supports it; numeric-to-Boolean casts normalize nonzero values through a WebNN comparison. Unsupported ONNX/string element types are rejected. |
+| `CastLike` | Decomposed | Infers the second input's type at conversion time, then emits WebNN cast. | minimal |
+| `Ceil` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Celu` | Decomposed | Elementwise decomposition; no native WebNN celu. | moderate |
+| `CenterCropPad` | Unsupported | No current lowering; the required vision/spatial primitive is absent or not implemented. |
+| `Clip` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Col2Im` | Unsupported | No current lowering; the required vision/spatial primitive is absent or not implemented. |
+| `Compress` | Unsupported | No current lowering; missing primitive and/or data-dependent output shape. |
+| `Concat` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `ConcatFromSequence` | Unsupported | WebNN has no sequence operand type. |
+| `Constant` | Folded/static | Creates an inline WebNN constant; no runtime ONNX operation remains. |
+| `ConstantOfShape` | Folded/static | Materialized or expanded from statically resolvable shape/value data. |
+| `Conv` | Direct WebNN | Direct WebNN conv2d; 1-D is emulated and 3-D is unsupported. |
+| `ConvInteger` | Decomposed | Centers values in float, executes conv2d, then casts to int32; WebNN lacks integer accumulation, so large sums may lose exactness. | high |
+| `ConvTranspose` | Direct WebNN | Direct WebNN convTranspose2d; 1-D is emulated and 3-D is unsupported. |
+| `Cos` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Cosh` | Decomposed | Elementwise decomposition; WebNN has no native cosh. | moderate |
+| `CumProd` | Decomposed | exp(cumulativeSum(log(x))); valid only for positive floating-point inputs. Zero and negative inputs are not generally correct. | moderate |
+| `CumSum` | Direct WebNN | Direct WebNN cumulativeSum; axis must be constant. |
+| `DFT` | Unsupported | WebNN has no corresponding signal-processing primitive. |
+| `DeformConv` | Unsupported | No current lowering; the required vision/spatial primitive is absent or not implemented. |
+| `DepthToSpace` | Unsupported | No current lowering; the required vision/spatial primitive is absent or not implemented. |
+| `DequantizeLinear` | Direct WebNN | Direct WebNN primitive for supported scalar/effectively-scalar or full-rank block parameters; axis must be 1. |
+| `Det` | Unsupported | No current executable lowering. |
+| `Div` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Dropout` | Unsupported | No current inference lowering; exporters may remove applicable training artifacts. |
+| `DynamicQuantizeLinear` | Decomposed | Reduction/arithmetic decomposition followed by WebNN quantizeLinear. | moderate |
+| `Einsum` | Decomposed | Static reduce/transpose/reshape/matmul decomposition; at most two inputs, no ellipsis or repeated labels within a term. | not inferred |
+| `Elu` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Equal` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Erf` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Exp` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Expand` | Direct WebNN | Direct WebNN expand; target shape must be build-time resolvable. |
+| `EyeLike` | Unsupported | No current lowering; missing primitive and/or data-dependent output shape. |
+| `Flatten` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Floor` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `GRU` | Direct WebNN | Direct WebNN recurrent primitive with layout adaptation; sequence_lens and unsupported optional inputs are rejected. |
+| `Gather` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `GatherElements` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `GatherND` | Direct WebNN | Direct WebNN gatherND; ONNX batch_dims is currently ignored. |
+| `Gelu` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Gemm` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `GlobalAveragePool` | Direct WebNN | Direct WebNN pool; supported for the implemented 1-D/2-D spatial forms. |
+| `GlobalLpPool` | Unsupported | No current executable lowering. |
+| `GlobalMaxPool` | Direct WebNN | Direct WebNN pool; supported for the implemented 1-D/2-D spatial forms. |
+| `Greater` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `GreaterOrEqual` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `GridSample` | Unsupported | No current lowering; the required vision/spatial primitive is absent or not implemented. |
+| `GroupNormalization` | Decomposed | Reduction and elementwise decomposition; relevant ranks and dimensions must be known. | moderate |
+| `HammingWindow` | Unsupported | WebNN has no corresponding signal-processing primitive. |
+| `HannWindow` | Unsupported | WebNN has no corresponding signal-processing primitive. |
+| `HardSigmoid` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `HardSwish` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Hardmax` | Decomposed | Comparison/mask decomposition; requires known input shape. | moderate |
+| `Identity` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `If` | Folded/static | Only constant or pinned-constant conditions; the selected branch is resolved and inlined during conversion because WebNN has no control flow. Runtime-selectable models therefore require a separate `.webnn`/Safetensors pair per branch. Shared weights are not deduplicated, so keeping both compiled graphs resident may duplicate them in RAM and VRAM. |
+| `ImageDecoder` | Unsupported | No current lowering; the required vision/spatial primitive is absent or not implemented. |
+| `InstanceNormalization` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `IsInf` | Direct WebNN | Direct WebNN isInfinite; non-default ONNX detect_negative/detect_positive filtering is not implemented. |
+| `IsNaN` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `LRN` | Unsupported | No current executable lowering. |
+| `LSTM` | Direct WebNN | Direct WebNN recurrent primitive with layout adaptation; sequence_lens and peepholes are rejected. |
+| `LayerNormalization` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `LeakyRelu` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Less` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `LessOrEqual` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Log` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `LogSoftmax` | Decomposed | Softmax/log decomposition; no native WebNN logSoftmax. | moderate |
+| `Loop` | Unsupported | No WebNN runtime loop/control-flow representation. |
+| `LpNormalization` | Unsupported | No current executable lowering. |
+| `LpPool` | Direct WebNN | Direct WebNN l2Pool2d only when p=2; 1-D is emulated and higher spatial ranks rejected. |
+| `MatMul` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `MatMulInteger` | Decomposed | Centers values in float, executes matmul, then casts to int32; WebNN lacks integer accumulation, so large sums may lose exactness. | high |
+| `Max` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `MaxPool` | Direct WebNN | Direct WebNN pool; indices output is unsupported, 1-D is emulated, and higher spatial ranks are rejected. |
+| `MaxRoiPool` | Unsupported | No current lowering; the required vision/spatial primitive is absent or not implemented. |
+| `MaxUnpool` | Unsupported | No current lowering; the required vision/spatial primitive is absent or not implemented. |
+| `Mean` | Decomposed | Variadic binary fold; operation order and floating-point rounding can differ. | not inferred |
+| `MeanVarianceNormalization` | Unsupported | No current executable lowering. |
+| `MelWeightMatrix` | Unsupported | WebNN has no corresponding signal-processing primitive. |
+| `Min` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Mish` | Decomposed | Softplus/tanh/multiply decomposition; no native WebNN mish. | moderate |
+| `Mod` | Decomposed | A - B * q(A/B) decomposition for both fmod modes; no WebNN remainder primitive. | moderate |
+| `Mul` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Multinomial` | Unsupported | WebNN has no graph random-number operators. |
+| `Neg` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `NegativeLogLikelihoodLoss` | Unsupported | No current inference lowering; exporters may remove applicable training artifacts. |
+| `NonMaxSuppression` | Unsupported | No current lowering; the required vision/spatial primitive is absent or not implemented. |
+| `NonZero` | Unsupported | No current lowering; missing primitive and/or data-dependent output shape. |
+| `Not` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `OneHot` | Decomposed | Static comparison/mask decomposition; depth and values must be constant and indices shape known. | moderate |
+| `Optional` | Unsupported | WebNN has no optional operand type. |
+| `OptionalGetElement` | Unsupported | WebNN has no optional operand type. |
+| `OptionalHasElement` | Unsupported | WebNN has no optional operand type. |
+| `Or` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `PRelu` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Pad` | Direct WebNN | Direct WebNN pad; rank and non-negative pads must be known at build time. |
+| `Pow` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `QLinearConv` | Unsupported | No fused WebNN quantized linear-algebra primitive; float decomposition is not implemented. |
+| `QLinearMatMul` | Unsupported | No fused WebNN quantized linear-algebra primitive; float decomposition is not implemented. |
+| `QuantizeLinear` | Direct WebNN | Direct WebNN primitive for supported scalar/effectively-scalar or full-rank block parameters; axis must be 1. |
+| `RMSNormalization` | Decomposed | Reduction and elementwise decomposition; relevant ranks and dimensions must be known. | moderate |
+| `RNN` | Unsupported | No current executable lowering. |
+| `RandomNormal` | Unsupported | WebNN has no graph random-number operators. |
+| `RandomNormalLike` | Unsupported | WebNN has no graph random-number operators. |
+| `RandomUniform` | Unsupported | WebNN has no graph random-number operators. |
+| `RandomUniformLike` | Unsupported | WebNN has no graph random-number operators. |
+| `Range` | Folded/static | Implemented for constant scalar or supported bounded-symbolic forms. |
+| `Reciprocal` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `ReduceL1` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `ReduceL2` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `ReduceLogSum` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `ReduceLogSumExp` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `ReduceMax` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `ReduceMean` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `ReduceMin` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `ReduceProd` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `ReduceSum` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `ReduceSumSquare` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `RegexFullMatch` | Unsupported | WebNN has no string tensor/operation support. |
+| `Relu` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Reshape` | Direct WebNN | Direct WebNN reshape; target shape must be build-time resolvable. |
+| `Resize` | Direct WebNN | Maps to WebNN resample2d; nearest/linear supported forms only, with 1-D emulation. Cubic and unsupported coordinate/axis modes are rejected. |
+| `ReverseSequence` | Direct WebNN | Partial mapping to WebNN reverse: sequence_lens and batch_axis semantics are ignored, so only full-axis reversal is equivalent. |
+| `RoiAlign` | Unsupported | No current lowering; the required vision/spatial primitive is absent or not implemented. |
+| `RotaryEmbedding` | Decomposed | Gather/rotate/concat decomposition; concrete dimensions and compatible cache shapes required. | moderate |
+| `Round` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `STFT` | Unsupported | WebNN has no corresponding signal-processing primitive. |
+| `Scan` | Unsupported | No WebNN runtime loop/control-flow representation. |
+| `ScatterElements` | Direct WebNN | Direct WebNN scatterElements; only reduction=none. |
+| `ScatterND` | Direct WebNN | Direct WebNN scatterND; only reduction=none. |
+| `Selu` | Decomposed | Elementwise decomposition; no native WebNN selu. | moderate |
+| `SequenceAt` | Pattern-only | Only a constant index into the converter's SplitToSequence pseudo-sequence. |
+| `SequenceConstruct` | Unsupported | WebNN has no sequence operand type. |
+| `SequenceEmpty` | Unsupported | WebNN has no sequence operand type. |
+| `SequenceErase` | Unsupported | WebNN has no sequence operand type. |
+| `SequenceInsert` | Unsupported | WebNN has no sequence operand type. |
+| `SequenceLength` | Unsupported | WebNN has no sequence operand type. |
+| `SequenceMap` | Unsupported | WebNN has no sequence operand type. |
+| `Shape` | Folded/static | Resolved from known graph shape metadata. |
+| `Shrink` | Decomposed | Elementwise comparison/select decomposition; no native WebNN shrink. | moderate |
+| `Sigmoid` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Sign` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Sin` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Sinh` | Decomposed | Elementwise decomposition; WebNN has no native sinh. | moderate |
+| `Size` | Unsupported | No current executable lowering. |
+| `Slice` | Direct WebNN | Direct WebNN slice/reverse; starts, sizes, axes, and steps must resolve at build time. Positive steps are preserved as strides; negative steps support only the full-axis reverse pattern. |
+| `Softmax` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `SoftmaxCrossEntropyLoss` | Unsupported | No current inference lowering; exporters may remove applicable training artifacts. |
+| `Softplus` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Softsign` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `SpaceToDepth` | Unsupported | No current lowering; the required vision/spatial primitive is absent or not implemented. |
+| `Split` | Direct WebNN | Direct WebNN split; split sizes or the split dimension must be resolvable. |
+| `SplitToSequence` | Pattern-only | Only the static split pattern consumed by constant-index SequenceAt; no sequence graph output. |
+| `Sqrt` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Squeeze` | Direct WebNN | Direct WebNN squeeze; axes must be resolvable. |
+| `StringConcat` | Unsupported | WebNN has no string tensor/operation support. |
+| `StringNormalizer` | Unsupported | WebNN has no string tensor/operation support. |
+| `StringSplit` | Unsupported | WebNN has no string tensor/operation support. |
+| `Sub` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Sum` | Decomposed | Variadic binary fold; operation order and floating-point rounding can differ. | not inferred |
+| `Swish` | Decomposed | Sigmoid/multiply decomposition; no native WebNN swish. | minimal |
+| `Tan` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Tanh` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `TensorScatter` | Unsupported | No current lowering; missing primitive and/or data-dependent output shape. |
+| `TfIdfVectorizer` | Unsupported | WebNN has no string tensor/operation support. |
+| `ThresholdedRelu` | Decomposed | Comparison/select decomposition; no native WebNN thresholdedRelu. | minimal |
+| `Tile` | Direct WebNN | Direct WebNN tile; repeats must be constant. |
+| `TopK` | Unsupported | No current lowering; missing primitive and/or data-dependent output shape. |
+| `Transpose` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Trilu` | Direct WebNN | Direct WebNN triangular; k must be resolvable. |
+| `Unique` | Unsupported | No current lowering; missing primitive and/or data-dependent output shape. |
+| `Unsqueeze` | Direct WebNN | Direct WebNN unsqueeze; axes must be resolvable. |
+| `Where` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+| `Xor` | Direct WebNN | Direct standardized WebNN primitive; handler schema checks and WebNN operand-type limits apply. |
+
+## Additional contrib and fused operators
+
+These eight advertised names are executable but are not standard `ai.onnx` opset-26 names, so
+they are excluded from the 198-op percentages.
+
+| Operator | Status | Execution scope and WebNN limitation | Estimated perf impact |
+|----------|--------|--------------------------------------|-----------------------|
+| `GatherBlockQuantized` | Decomposed | Gathers packed rows/scales/zero points and dequantizes only the selected slice for supported 2-D axis-0 tables; other layouts fall back at conversion time. | moderate |
+| `GroupQueryAttention` | Decomposed | Static attention/cache subgraph. WebNN has no fused attention; runtime sequence metadata is ignored and softcap/local-window modes are rejected. Fused backend kernels and the portable decomposition need not be bit-identical, so an immediately following dynamic quantizer can amplify legal rounding differences. | high |
+| `MatMulBnb4` | Decomposed | Packed NF4/FP4 block path where representable; incompatible tails become dense constants, losing low-bit memory/performance benefits. | high |
+| `MatMulNBits` | Decomposed | Supported 4/8-bit constant layouts dequantize into a matmul path; g_idx is rejected. WebNN has no low-bit matmul kernel. | high |
+| `MoE` | Decomposed | All experts execute densely because WebNN has no TopK or sparse dispatch. Approximate compute overhead is num_experts/k; supported activation/fusion forms only. | high |
+| `QMoE` | Decomposed | Supported 4/8-bit expert weights feed the same dense all-expert lowering; block-layout fallbacks may dequantize at conversion time. | high |
+| `SimplifiedLayerNormalization` | Decomposed | RMS-style normalization decomposition; no matching WebNN fused primitive. | moderate |
+| `SkipSimplifiedLayerNormalization` | Decomposed | Residual/bias plus simplified normalization decomposition; no matching WebNN fused primitive. | moderate |
+
+### MoE interpretation
+
+The previous documentation called MoE impossible because **native sparse MoE execution** cannot be
+represented in standard WebNN. That limitation remains. This branch nevertheless executes the
+supported `MoE` and `QMoE` forms by evaluating every expert and blending the selected outputs.
+This is functional fallback support, not sparse-dispatch performance parity.
+
+## Reproducing the inventory
+
+```bash
+python scripts/generate_onnx_opsets.py --min 26 --max 26 -o docs/onnx-opsets
+python scripts/onnx_ops_to_csv.py model.onnx --check-webnn
+```
+
+Sources of truth: `scripts/webnn_onnx_ops.py`, the handler registry in
+`src/onnx/ops/mod.rs`, handler restrictions in `src/onnx/ops/*.rs`, and the opset gate in
+`src/onnx/convert.rs`. Numerical model results are tracked separately in
+[model-validation-status.md](./model-validation-status.md).
+
+## References
+
+- [W3C WebNN specification](https://www.w3.org/TR/webnn/)
+- [ONNX operator schemas](https://github.com/onnx/onnx/blob/main/docs/Operators.md)
+- [ONNX Runtime MoE/QMoE notes](https://github.com/microsoft/onnxruntime/blob/main/docs/contrib_ops/cuda/moe_qmoe.md)
diff --git a/docs/real-weight-validation-failures.md b/docs/real-weight-validation-failures.md
new file mode 100644
index 0000000..87501e5
--- /dev/null
+++ b/docs/real-weight-validation-failures.md
@@ -0,0 +1,118 @@
+# Real-weight validation failures
+
+> **Last edited:** `2026-09-17T16:52:20Z`
+> **Checkout:** `fkrall/cache-backed-validation` at `25ea94c`
+>
+> **Freshness:** Use this document only when this provenance is recent relative to the relevant
+> code and commits; otherwise verify the implementation, tests, and Git history before relying
+> on it.
+
+This document is the current triage record for failures from full-model validation with publisher
+weights. Generated-weight failures are intentionally out of scope. The summary and complete case
+ledger remain in [Full-model numerical validation status](model-validation-status.md).
+
+## Recorded run
+
+- Date: 2026-09-17
+- Tested executable: onnx2webnn `25ea94c`
+- RustNN: `28fb3bbe`
+- Manifest: `tests/models/manifest.json` (51 current cases; verified within a completed 52-case superset)
+- Runtime: CPU ONNX Runtime 1.29.0
+- Cache state: complete; no model downloads or skips
+- Disk guard: minimum observed free space 78.7 GiB; stop threshold 15 GiB was not approached
+- Result for the current manifest: 47 passed and 4 failed
+- Recorded 52-case superset wall time: 7m 2.3s, one validation worker; peak RSS 26.4 GiB
+
+```bash
+ORT_DYLIB_PATH=../rustnn/target/onnxruntime/onnxruntime-linux-x64-1.29.0/lib/libonnxruntime.so.1.29.0 \
+ target/release/onnx2webnn validate-models \
+ --selection all --weights real --jobs 1
+```
+
+The current manifest is the tested 52-case run minus the passing FP32 Tiny RoFormer case; every
+retained case was executed in that run.
+
+The recorded result for each case is its first blocker. A case that fails before comparison may
+contain further converter, serialization, reload, execution, or numerical issues that are not yet
+observable.
+
+## Failure groups
+
+| Code | Cases | Furthest stage | Classification | Current ownership |
+|------|------:|----------------|----------------|-------------------|
+| U1 | 1 | Output comparison | Unsupported strict fused/decomposed attention equivalence across dynamic quantization boundaries | FastVLM prefill remains blocked; its decode specialization passes. |
+| Q1 | 1 | Output comparison | Unsupported `MatMulNBits accuracy_level=4` execution semantics | Requires a WebNN/backend mechanism for Int8-quantized activations with packed q4 weights; do not relax tolerance to hide it. |
+| O1 | 2 | Native ORT model load | The publisher ONNX is rejected before conversion can be compared | Blocked upstream pending a corrected publisher artifact. |
+
+The real sweep has no deterministic-input or cached-interface failures. U1 and Q1 both reach
+comparison, but the source and reconstructed graphs do not promise the same numerical execution
+mode. O1 fails in the reference model itself.
+
+## U1: precision-sensitive fused attention followed by dynamic quantization
+
+| # | Case | Comparison diagnostics |
+|--:|------|------------------------|
+| 11 | `onnx-community--FastVLM-0.5B-ONNX :: decoder_model_merged_quantized.onnx` (`sequence=64`, `past=0`) | `9,702,133 / 9,705,344` logits exceed tolerance; maximum absolute error `10.0572`, mean absolute error `0.760946`, RMSE `1.05325`. |
+
+Boundary probes rule out a shape, rotary, cache-layout, or serialization defect. Layer 0 query
+projection, external `RotaryEmbedding`, decomposed `GroupQueryAttention`, normalization, and MLP
+all pass the normal Float32 envelope. Layer 1 GQA output also remains within that envelope, but
+the immediately following `DynamicQuantizeLinear` makes the small difference discrete: the output
+projection is the first failing observed boundary (`825 / 57,344` values, maximum absolute error
+`0.001003`). Repeating this pattern through 24 layers produces the final logit disagreement above.
+
+ORT 1.29 CPU GQA has multiple valid kernels. With its default tiled flash-attention kernel the
+first native logit is `0.669124`; setting `ORT_GQA_DISABLE_FLASH_ATTENTION=1` selects ORT non-flash
+fallback and changes that same native logit to `1.066261`. The decomposed WebNN result is stable at
+`1.211192`. The source backend therefore materially disagrees with itself after repeated
+quantization when only its legal attention kernel changes.
+
+WebNN has no fused GQA primitive, and a decomposition into `matmul`, mask, `softmax`, and `matmul`
+cannot guarantee bit-identical rounding to either backend-specific fused kernel. Weakening the
+final tolerance would hide different Uint8 quantization decisions rather than accommodate ordinary
+output accumulation error. This fixed prefill specialization is therefore recorded as unsupported
+for strict cross-backend numerical validation; the sequence-1 decode specialization passes.
+
+## Q1: unsupported `MatMulNBits accuracy_level=4`
+
+Voxtral's decoder sets `accuracy_level=4` on all 211 `MatMulNBits` nodes. Native ORT may quantize
+Float32 activations to Int8 internally in that mode. WebNN has no fused low-bit matmul operator, so
+onnx2webnn emits packed Uint4 weights followed by `dequantizeLinear` and ordinary Float32 matmul.
+The two paths therefore do not promise the same numerical algorithm.
+
+This is not a packed-weight serialization problem. All 211 source weight payloads matched their
+saved q4 tensors byte-for-byte by SHA-256; archive marker, U8 storage shape, nibble ordering, and
+generated `0x88` zero points were also correct. Direct output analysis found mean absolute error
+`0.0401`, maximum error `0.367`, and correlation `0.999502`. Changing only the source attributes
+to `accuracy_level=0` reduced those figures to `9.66e-6`, `9.32e-5`, and `0.99999999997`.
+
+This document records case 28 as Q1. Both `all` and `match` attempt it normally, so a future
+implementation becomes visible. It is not counted as numerically supported.
+
+## O1: source ONNX rejected by native ORT
+
+Both Chronos files fail during native ORT model loading. A `Gather` indices input is the float
+output of `ConstantOfShape`; Gather indices must be integer.
+
+| # | Case |
+|--:|------|
+| 45 | `kashif--chronos-2-onnx :: encoder_model.onnx` |
+| 46 | `kashif--chronos-2-onnx :: decoder_model_merged.onnx` |
+
+The two manifest paths resolve to the same invalid publisher artifact. This document records them
+as O1; `all` and `match` still attempt them normally, so a future publisher correction will be
+visible. Until the source model runs in native ORT, these cases cannot
+provide a numerical oracle and are not evidence for or against onnx2webnn correctness.
+
+## Repair order
+
+1. Track U1 until WebNN or a backend-specific fusion can provide comparable attention precision
+ across the dynamic-quantization boundaries.
+2. Track Q1 until WebNN or a backend-specific fusion can express Int8 activations with packed q4 weights.
+3. Monitor the upstream Chronos repository for corrected reference exports.
+
+## Maintenance rule
+
+This file describes only the latest complete real-weight run. Replace counts, cases, diagnoses,
+and exact errors when a new run changes first blockers; do not append obsolete full failure tables.
+Record revision-attributed coverage changes in the main status file's history section.
diff --git a/tests/models/UNSUPPORTED_OPS.md b/docs/transformersjs_excluded_models.md
similarity index 74%
rename from tests/models/UNSUPPORTED_OPS.md
rename to docs/transformersjs_excluded_models.md
index f25f0fa..f2aff0f 100644
--- a/tests/models/UNSUPPORTED_OPS.md
+++ b/docs/transformersjs_excluded_models.md
@@ -1,8 +1,27 @@
-# Unsupported ONNX Operators by Hugging Face Repository
-
-This document records operators currently unsupported by `onnx2webnn`, grouped by Hugging Face repository.
-
-`scripts/generate_manifest.py` also reads it when regenerating `manifest.json`: a `## [org/repo](https://huggingface.co/org/repo)` section keeps that repo out of the sweep, and backticked `onnx/*.onnx` references inside the section narrow the exclusion to those components (the rest of the repo stays in).
+# Transformers.js Excluded Models
+
+> **Last edited:** `2026-09-23T11:35:58Z`
+> **Checkout:** `fkrall/cache-backed-validation` at `ac6d5ca`
+>
+> **Freshness:** Use this document only when this provenance is recent relative to the relevant
+> code and commits; otherwise verify the implementation, tests, and Git history before relying
+> on it.
+
+This document records model exports currently excluded from onnx2webnn's generated sweep because
+of unsupported operators, unsupported operator forms, unresolved dimensions, or conversion defects.
+For the complete operator-level capability matrix, see
+[Operator conversion status](operator-conversion-status.md).
+
+The [`scripts/generate_manifest.py` generator](../scripts/generate_manifest.py) reads this document
+when regenerating [`tests/models/manifest.json`](../tests/models/manifest.json). A
+`## [org/repo](https://huggingface.co/org/repo)` section identifies a repository. Backticked
+`onnx/*.onnx` references exclude only those component stems; a section without any such
+reference excludes the whole repository. The explanatory operator and error text is not parsed.
+
+An exclusions-disabled skeleton audit on 23 September 2026 found no listed component that became
+fully supported on the current branch. Generated cases retained a blocker, except that NLLB's
+decode/cache branch passed while its prefill branch still failed. Entries requiring unresolved or
+model-specific realistic dimensions were not fully re-audited and retain their previous diagnosis.
## [onnx-community/dpt-dinov2-small-kitti](https://huggingface.co/onnx-community/dpt-dinov2-small-kitti)
@@ -87,4 +106,6 @@ This document records operators currently unsupported by `onnx2webnn`, grouped b
## [Xenova/nllb-200-distilled-600M](https://huggingface.co/Xenova/nllb-200-distilled-600M)
-- `onnx/decoder_model_merged*.onnx` — the shared output-embedding bias reshape mis-derives its target shape from `decoder_sequence_length` instead of the vocab size, producing an incompatible broadcast (`decoder_sequence_length` vs. 256206); open (encoder_model converts)
+- `onnx/decoder_model_merged*.onnx` — the shared output-embedding bias reshape mis-derives its target shape from `decoder_sequence_length` instead of the vocab size, producing an incompatible broadcast (`decoder_sequence_length` vs. 256206); open. The decode/cache
+ branch builds, but the prefill branch still fails, so the component remains excluded
+ (encoder_model converts).
diff --git a/scripts/generate_manifest.py b/scripts/generate_manifest.py
index db03698..cf73d47 100644
--- a/scripts/generate_manifest.py
+++ b/scripts/generate_manifest.py
@@ -9,7 +9,7 @@
`transformers.js` (--top N for more per task; the ranking is fetched from
the Hub and cached under target/, --refresh-ranking re-fetches), minus
repos and components documented as unconvertible in
- tests/models/UNSUPPORTED_OPS.md. Tasks with nothing to convert
+ docs/transformersjs_excluded_models.md. Tasks with nothing to convert
(`(tokenizer-only)`) are skipped; models without onnx/*.onnx files are
reported.
2. Which files: what transformers.js itself loads for that repo -- the
@@ -69,7 +69,7 @@
SCRIPT_DIR = Path(__file__).resolve().parent
REPO_ROOT = SCRIPT_DIR.parent
DEFAULT_MANIFEST = REPO_ROOT / "tests" / "models" / "manifest.json"
-DEFAULT_UNSUPPORTED_FILE = REPO_ROOT / "tests" / "models" / "UNSUPPORTED_OPS.md"
+DEFAULT_UNSUPPORTED_FILE = REPO_ROOT / "docs" / "transformersjs_excluded_models.md"
DEFAULT_OVERRIDES = REPO_ROOT / "tests" / "models" / "manifest-overrides.json"
# target/ is gitignored: the ranking snapshot and the review copy live there.
DEFAULT_MODELS_JSON = REPO_ROOT / "target" / "transformers-js-models.json"
@@ -277,7 +277,7 @@ def select_repos(models: list[dict], top: int, only: list[str]) -> list[dict]:
return picked
-# ---- UNSUPPORTED_OPS.md: repos and components known not to convert ----
+# ---- transformersjs_excluded_models.md: repos and components known not to convert ----
SECTION_RE = re.compile(r"^## \[[^\]]+\]\(https://huggingface\.co/([^)]+)\)\s*$", re.MULTILINE)
ONNX_REF_RE = re.compile(r"`onnx/([A-Za-z0-9_]+?)\*?(?:\.onnx)?`")
@@ -799,7 +799,7 @@ def __init__(self, existing: list[dict]):
for name, value in (e.get("override_dims") or {}).items():
if self._branch_invariant(name):
self.repo_dims.setdefault(repo, {}).setdefault(name, value)
- extra = {k: v for k, v in e.items() if k not in ("file", "heavy", "pin_inputs", "override_dims")}
+ extra = {k: v for k, v in e.items() if k not in ("file", "heavy", "pin_inputs", "override_dims", "validation")}
if extra:
self.manual_fields[key] = extra
diff --git a/scripts/test_generate_manifest.py b/scripts/test_generate_manifest.py
new file mode 100755
index 0000000..9123f26
--- /dev/null
+++ b/scripts/test_generate_manifest.py
@@ -0,0 +1,35 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+import sys
+import types
+import unittest
+
+# This focused test does not use ONNX.
+sys.modules.setdefault("onnx", types.ModuleType("onnx"))
+
+import generate_manifest as manifest
+
+
+class LegacyValidationMetadataTests(unittest.TestCase):
+ def test_validation_metadata_is_not_preserved(self):
+ file = "org--repo/onnx/model.onnx"
+ baselines = manifest.Baselines(
+ [
+ {
+ "file": file,
+ "validation": {"tier": "blocked", "reason": "legacy"},
+ "coreml_slow": "preserve unrelated metadata",
+ }
+ ]
+ )
+
+ self.assertEqual(
+ baselines.manual_fields[(file, ())],
+ {"coreml_slow": "preserve unrelated metadata"},
+ )
+
+
+if __name__ == "__main__":
+ unittest.main()
diff --git a/src/cache.rs b/src/cache.rs
new file mode 100644
index 0000000..72aeb80
--- /dev/null
+++ b/src/cache.rs
@@ -0,0 +1,91 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+//! Platform-appropriate cache locations shared by the CLI and model validator.
+
+use std::path::PathBuf;
+
+const APPLICATION_DIR: &str = "onnx2webnn";
+
+/// Return the directory for cached source ONNX models.
+///
+/// `O2W_ONNX_CACHE` overrides only this cache. `O2W_CACHE_DIR` relocates all
+/// onnx2webnn caches while retaining their `onnx`/`webnn` subdirectories.
+pub fn onnx_cache_dir() -> Result {
+ cache_dir("O2W_ONNX_CACHE", "onnx")
+}
+
+/// Return the directory for exported WebNN graphs and Safetensors archives.
+///
+/// `O2W_WEBNN_CACHE` overrides only this cache. `O2W_CACHE_DIR` relocates all
+/// onnx2webnn caches while retaining their `onnx`/`webnn` subdirectories.
+pub fn webnn_cache_dir() -> Result {
+ cache_dir("O2W_WEBNN_CACHE", "webnn")
+}
+
+fn cache_dir(specific_env: &str, kind: &str) -> Result {
+ resolve_cache_dir(
+ nonempty_env(specific_env),
+ nonempty_env("O2W_CACHE_DIR"),
+ dirs::cache_dir(),
+ kind,
+ )
+ .ok_or_else(|| {
+ format!(
+ "cannot determine the operating-system cache directory; set {specific_env} or O2W_CACHE_DIR"
+ )
+ })
+}
+
+fn nonempty_env(name: &str) -> Option {
+ std::env::var_os(name)
+ .filter(|value| !value.is_empty())
+ .map(PathBuf::from)
+}
+
+fn resolve_cache_dir(
+ specific: Option,
+ shared: Option,
+ os_cache: Option,
+ kind: &str,
+) -> Option {
+ specific.or_else(|| {
+ shared
+ .or_else(|| os_cache.map(|root| root.join(APPLICATION_DIR)))
+ .map(|root| root.join(kind))
+ })
+}
+
+#[cfg(test)]
+mod tests {
+ use super::*;
+ use std::path::Path;
+
+ #[test]
+ fn cache_path_precedence_is_specific_shared_then_os() {
+ let specific = PathBuf::from("/specific");
+ let shared = PathBuf::from("/shared");
+ let os = PathBuf::from("/os-cache");
+
+ assert_eq!(
+ resolve_cache_dir(
+ Some(specific.clone()),
+ Some(shared.clone()),
+ Some(os.clone()),
+ "onnx"
+ ),
+ Some(specific)
+ );
+ assert_eq!(
+ resolve_cache_dir(None, Some(shared), Some(os.clone()), "onnx"),
+ Some(Path::new("/shared/onnx").to_path_buf())
+ );
+ assert_eq!(
+ resolve_cache_dir(None, None, Some(os), "webnn"),
+ Some(Path::new("/os-cache/onnx2webnn/webnn").to_path_buf())
+ );
+ assert_eq!(resolve_cache_dir(None, None, None, "onnx"), None);
+ }
+}
diff --git a/src/lib.rs b/src/lib.rs
index 001302d..1571d5e 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -4,13 +4,20 @@
* SPDX-License-Identifier: Apache-2.0
*/
+pub mod cache;
pub mod debug;
+pub mod model_validation;
pub mod protos;
+pub mod validation;
pub mod onnx;
pub use onnx::convert::{
- convert_model_proto, convert_onnx, ConvertOptions, OnnxError, UnsupportedOpEntry,
- ValidatedGraph,
+ cache_onnx_model, convert_model_proto, convert_onnx, ConvertOptions, OnnxError,
+ UnsupportedOpEntry, ValidatedGraph,
};
pub use onnx::test_models;
+pub use validation::{
+ validate_cached_model, validate_cached_model_with_options,
+ validate_cached_model_with_overrides, ValidationSummary,
+};
diff --git a/src/main.rs b/src/main.rs
index bdcf2fe..40af7f9 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -18,10 +18,12 @@
use clap::{Parser, Subcommand};
use std::collections::HashMap;
-use std::path::Path;
+use std::path::{Path, PathBuf};
-use onnx2webnn::convert_onnx;
+use onnx2webnn::model_validation::manifest::Selection;
+use onnx2webnn::model_validation::{run_manifest_validation, RunOptions, WeightMode};
use onnx2webnn::ConvertOptions;
+use onnx2webnn::{cache_onnx_model, convert_onnx, validate_cached_model_with_overrides};
#[derive(Parser)]
#[command(name = "onnx2webnn")]
@@ -71,9 +73,66 @@ enum Command {
/// Preserve unresolved symbolic input dims as dynamic metadata (experimental)
#[arg(long)]
experimental_dynamic_inputs: bool,
+
+ /// Save the converted graph and self-contained ONNX model to the OS cache directory
+ #[arg(long)]
+ output: bool,
+
+ /// Reload cached WebNN artifacts, dispatch deterministic inputs, and compare with native ORT
+ #[arg(long, requires = "output")]
+ validate: bool,
+
+ /// Validate existing cache artifacts only; does not read or convert the source ONNX.
+ #[arg(long, conflicts_with_all = ["validate", "output"])]
+ validate_cached: bool,
+ },
+
+ /// Numerically validate manifest models against reloaded WebNN artifacts
+ ValidateModels {
+ /// Manifest selection: all or match=
+ #[arg(long)]
+ selection: String,
+
+ /// Weight source. Only publisher weights are currently supported.
+ #[arg(long, default_value_t = WeightMode::Real)]
+ weights: WeightMode,
+
+ /// Override tests/models/manifest.json
+ #[arg(long)]
+ manifest: Option,
+
+ /// Number of parallel light-model workers; heavy models remain sequential
+ #[arg(long, default_value_t = 1)]
+ jobs: usize,
},
}
+fn cache_paths(input: &Path) -> anyhow::Result<(PathBuf, PathBuf)> {
+ let canonical = input.canonicalize().unwrap_or_else(|_| input.to_path_buf());
+ let mut hash = 0xcbf29ce484222325u64;
+ for byte in canonical.to_string_lossy().bytes() {
+ hash ^= u64::from(byte);
+ hash = hash.wrapping_mul(0x100000001b3);
+ }
+ let stem = input
+ .file_stem()
+ .and_then(|value| value.to_str())
+ .unwrap_or("model");
+ let safe_stem: String = stem
+ .chars()
+ .map(|c| if c.is_ascii_alphanumeric() { c } else { '_' })
+ .collect();
+ let key = format!("{safe_stem}-{hash:016x}");
+ Ok((
+ onnx2webnn::cache::onnx_cache_dir()
+ .map_err(anyhow::Error::msg)?
+ .join(format!("{key}.onnx")),
+ onnx2webnn::cache::webnn_cache_dir()
+ .map_err(anyhow::Error::msg)?
+ .join(format!("{key}.webnn")),
+ ))
+}
+
fn main() -> anyhow::Result<()> {
let cli = Cli::parse();
if cli.debug {
@@ -89,6 +148,9 @@ fn main() -> anyhow::Result<()> {
allow_missing_external_data,
optimize,
experimental_dynamic_inputs,
+ output,
+ validate,
+ validate_cached,
} => {
let mut free_dim_overrides = if let Some(path) = override_dims_file {
let content = std::fs::read_to_string(&path)?;
@@ -144,18 +206,86 @@ fn main() -> anyhow::Result<()> {
pinned_inputs.insert(name, value);
}
+ let validation_overrides = free_dim_overrides.clone();
+ if validate_cached {
+ let (onnx_cache, webnn_cache) = cache_paths(input_path)?;
+ if !onnx_cache.exists() || !webnn_cache.exists() {
+ return Err(anyhow::anyhow!(
+ "cached validation requires {} and {}; run with --output first",
+ onnx_cache.display(),
+ webnn_cache.display()
+ ));
+ }
+ let summary = validate_cached_model_with_overrides(
+ &onnx_cache,
+ &webnn_cache,
+ &validation_overrides,
+ )
+ .map_err(|e| anyhow::anyhow!("{e}"))?;
+ println!(
+ "✓ cached deterministic validation passed ({} inputs, {} outputs)",
+ summary.input_count, summary.output_count
+ );
+ return Ok(());
+ }
+
+ let cache_paths = output.then(|| cache_paths(input_path)).transpose()?;
+ if let Some((onnx_cache, _)) = &cache_paths {
+ cache_onnx_model(input_path, onnx_cache, allow_missing_external_data)
+ .map_err(|e| anyhow::anyhow!("{e}"))?;
+ }
+
let options = ConvertOptions {
free_dim_overrides,
optimize,
experimental_dynamic_inputs,
pinned_inputs,
zero_fill_missing_external_data: allow_missing_external_data,
+ output_path: cache_paths.as_ref().map(|(_, webnn)| webnn.clone()),
};
let _validated = convert_onnx(input_path.to_str().unwrap(), options)
.map_err(|e| anyhow::anyhow!("{e}"))?;
// stdout (not stderr): PowerShell treats native stderr as NativeCommandError
println!("✓ ORT graph build succeeded for {}", input);
+ if let Some((onnx_cache, webnn_cache)) = &cache_paths {
+ println!("✓ cached ONNX at {}", onnx_cache.display());
+ println!("✓ cached WebNN at {}", webnn_cache.display());
+ if validate {
+ let summary = validate_cached_model_with_overrides(
+ onnx_cache,
+ webnn_cache,
+ &validation_overrides,
+ )
+ .map_err(|e| anyhow::anyhow!("{e}"))?;
+ println!(
+ "✓ deterministic validation passed ({} inputs, {} outputs)",
+ summary.input_count, summary.output_count
+ );
+ }
+ }
+ }
+ Command::ValidateModels {
+ selection,
+ weights,
+ manifest,
+ jobs,
+ } => {
+ let selection = Selection::parse(&selection).map_err(anyhow::Error::msg)?;
+ let mut options = RunOptions::new(selection, weights).map_err(anyhow::Error::msg)?;
+ options.jobs = jobs;
+ if let Some(manifest) = manifest {
+ options.manifest = manifest;
+ }
+ let summary = run_manifest_validation(options).map_err(anyhow::Error::msg)?;
+ println!("{summary}");
+ if summary.has_failures() {
+ return Err(anyhow::anyhow!(
+ "{} of {} model validations failed",
+ summary.failed.len(),
+ summary.selected
+ ));
+ }
}
}
diff --git a/src/model_validation/full_model.rs b/src/model_validation/full_model.rs
new file mode 100644
index 0000000..e8767c1
--- /dev/null
+++ b/src/model_validation/full_model.rs
@@ -0,0 +1,538 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+//! Full Hugging Face model downloads for the manual numerical sweep.
+
+use super::manifest::Entry;
+use crate::protos::onnx::{GraphProto, ModelProto};
+use huggingface_hub::{HFClient, HFClientSync, HFRepositorySync, RepoDownloadFileParams};
+use prost::Message;
+use serde::{Deserialize, Serialize};
+use sha2::{Digest, Sha256};
+use std::collections::BTreeSet;
+use std::fs;
+use std::io::Read;
+use std::path::{Component, Path, PathBuf};
+
+const CACHE_FORMAT: u32 = 3;
+const METADATA_DIR: &str = ".onnx2webnn-validation";
+
+#[derive(Debug, Deserialize, Serialize)]
+struct CacheMetadata {
+ format: u32,
+ repository: String,
+ revision: String,
+ sha256: Option,
+ primary: String,
+ files: Vec,
+}
+
+#[derive(Debug, Deserialize, Serialize)]
+struct CachedFile {
+ path: String,
+ length: u64,
+}
+
+pub fn cache_root() -> Result {
+ resolve_cache_root(
+ nonempty_env("O2W_ONNX_CACHE"),
+ nonempty_env("O2W_CACHE_DIR"),
+ nonempty_env("HF_HUB_CACHE"),
+ nonempty_env("HUGGINGFACE_HUB_CACHE"),
+ nonempty_env("HF_HOME"),
+ dirs::cache_dir(),
+ )
+ .ok_or_else(|| {
+ "cannot determine the Hugging Face cache directory; set O2W_ONNX_CACHE or HF_HUB_CACHE"
+ .to_string()
+ })
+}
+
+fn resolve_cache_root(
+ o2w_onnx: Option,
+ o2w_shared: Option,
+ hf_hub: Option,
+ legacy_hf_hub: Option,
+ hf_home: Option,
+ os_cache: Option,
+) -> Option {
+ o2w_onnx
+ .or_else(|| o2w_shared.map(|path| path.join("onnx")))
+ .or(hf_hub)
+ .or(legacy_hf_hub)
+ .or_else(|| hf_home.map(|path| path.join("hub")))
+ .or_else(|| os_cache.map(|path| path.join("huggingface").join("hub")))
+}
+
+fn nonempty_env(name: &str) -> Option {
+ std::env::var_os(name)
+ .filter(|value| !value.is_empty())
+ .map(PathBuf::from)
+}
+
+pub fn cache_full_model(entry: &Entry) -> Result {
+ let cache_root = cache_root()?;
+ fs::create_dir_all(&cache_root).map_err(|e| format!("create {}: {e}", cache_root.display()))?;
+ let revision = entry.revision();
+ let (repository, repository_path) = parse_manifest_file(&entry.file)?;
+ let (owner, name) = repository
+ .split_once('/')
+ .ok_or_else(|| format!("invalid Hugging Face repository {repository:?}"))?;
+ let client = HFClient::builder()
+ .cache_dir(&cache_root)
+ .build()
+ .map_err(|e| format!("create Hugging Face client: {e}"))?;
+ let client = HFClientSync::from_api(client)
+ .map_err(|e| format!("create blocking Hugging Face client: {e}"))?;
+ let repository_client = client.model(owner, name);
+ let metadata_path = metadata_path(&cache_root, entry);
+ let refresh = std::env::var_os("O2W_MODEL_CACHE_REFRESH").is_some();
+
+ if !refresh {
+ if let Some(model_path) = complete_cache(
+ &repository_client,
+ &metadata_path,
+ &repository,
+ revision,
+ entry.sha256.as_deref(),
+ )? {
+ return Ok(model_path);
+ }
+ }
+
+ if metadata_path.exists() {
+ fs::remove_file(&metadata_path)
+ .map_err(|e| format!("remove stale {}: {e}", metadata_path.display()))?;
+ }
+
+ let mut model_path = download_hub_file(
+ &repository_client,
+ &repository_path,
+ revision,
+ refresh,
+ false,
+ )?;
+ if let Some(expected) = entry.sha256.as_deref() {
+ if let Err(error) = verify_sha256(&model_path, expected) {
+ if refresh {
+ return Err(error);
+ }
+ model_path =
+ download_hub_file(&repository_client, &repository_path, revision, true, false)?;
+ verify_sha256(&model_path, expected)?;
+ }
+ }
+
+ let model_bytes =
+ fs::read(&model_path).map_err(|e| format!("read {}: {e}", model_path.display()))?;
+ let model = ModelProto::decode(model_bytes.as_slice())
+ .map_err(|e| format!("decode {}: {e}", model_path.display()))?;
+ let locations = external_locations(&model)?;
+ let repository_parent = Path::new(&repository_path)
+ .parent()
+ .unwrap_or_else(|| Path::new(""));
+ let mut files = vec![cached_file(&repository_path, &model_path)?];
+ for location in locations {
+ let repository_sidecar = repository_parent.join(&location);
+ let repository_sidecar = path_for_url(&repository_sidecar)?;
+ let sidecar_path = download_hub_file(
+ &repository_client,
+ &repository_sidecar,
+ revision,
+ refresh,
+ false,
+ )?;
+ files.push(cached_file(&repository_sidecar, &sidecar_path)?);
+ }
+
+ let metadata = CacheMetadata {
+ format: CACHE_FORMAT,
+ repository,
+ revision: revision.to_string(),
+ sha256: entry.sha256.clone(),
+ primary: repository_path,
+ files,
+ };
+ write_metadata(&metadata_path, &metadata)?;
+ Ok(model_path)
+}
+
+fn download_hub_file(
+ repository: &HFRepositorySync,
+ path: &str,
+ revision: &str,
+ force_download: bool,
+ local_files_only: bool,
+) -> Result {
+ repository
+ .download_file(
+ &RepoDownloadFileParams::builder()
+ .filename(path)
+ .revision(revision)
+ .force_download(force_download)
+ .local_files_only(local_files_only)
+ .build(),
+ )
+ .map_err(|e| format!("download {path} at revision {revision}: {e}"))
+}
+
+fn cached_file(repository_path: &str, local_path: &Path) -> Result {
+ let length = fs::metadata(local_path)
+ .map_err(|e| format!("inspect {}: {e}", local_path.display()))?
+ .len();
+ Ok(CachedFile {
+ path: repository_path.to_string(),
+ length,
+ })
+}
+
+fn parse_manifest_file(file: &str) -> Result<(String, String), String> {
+ let (org_repo, relative) = file
+ .split_once('/')
+ .ok_or_else(|| format!("{file}: expected --/"))?;
+ let (org, repo) = org_repo
+ .split_once("--")
+ .ok_or_else(|| format!("{file}: expected --/"))?;
+ if org.is_empty() || repo.is_empty() {
+ return Err(format!(
+ "{file}: empty Hugging Face organization or repository"
+ ));
+ }
+ let relative = path_for_url(&safe_relative_path(relative)?)?;
+ Ok((format!("{org}/{repo}"), relative))
+}
+
+fn safe_relative_path(path: impl AsRef) -> Result {
+ let path = path.as_ref();
+ if path.as_os_str().is_empty() || path.is_absolute() {
+ return Err(format!("unsafe external-data path '{}'", path.display()));
+ }
+ let mut clean = PathBuf::new();
+ for component in path.components() {
+ match component {
+ Component::Normal(part) => clean.push(part),
+ _ => return Err(format!("unsafe external-data path '{}'", path.display())),
+ }
+ }
+ Ok(clean)
+}
+
+fn path_for_url(path: &Path) -> Result {
+ let safe = safe_relative_path(path)?;
+ Ok(safe
+ .components()
+ .map(|component| component.as_os_str().to_string_lossy())
+ .collect::>()
+ .join("/"))
+}
+
+fn metadata_path(cache_root: &Path, entry: &Entry) -> PathBuf {
+ let digest = Sha256::digest(entry.source_key().as_bytes());
+ let key = digest
+ .iter()
+ .map(|byte| format!("{byte:02x}"))
+ .collect::();
+ cache_root
+ .join(METADATA_DIR)
+ .join(format!("{key}.complete.json"))
+}
+
+fn complete_cache(
+ repository_client: &HFRepositorySync,
+ metadata_path: &Path,
+ repository: &str,
+ revision: &str,
+ expected_sha256: Option<&str>,
+) -> Result, String> {
+ let Ok(bytes) = fs::read(metadata_path) else {
+ return Ok(None);
+ };
+ let Ok(metadata) = serde_json::from_slice::(&bytes) else {
+ return Ok(None);
+ };
+ if !metadata_matches(&metadata, repository, revision, expected_sha256) {
+ return Ok(None);
+ }
+
+ let mut primary = None;
+ for file in &metadata.files {
+ let path = match path_for_url(Path::new(&file.path)) {
+ Ok(path) => path,
+ Err(_) => return Ok(None),
+ };
+ let local_path = match download_hub_file(repository_client, &path, revision, false, true) {
+ Ok(path) => path,
+ Err(_) => return Ok(None),
+ };
+ let valid_length = fs::metadata(&local_path)
+ .map(|value| value.is_file() && value.len() == file.length)
+ .unwrap_or(false);
+ if !valid_length {
+ return Ok(None);
+ }
+ if path == metadata.primary {
+ primary = Some(local_path);
+ }
+ }
+
+ let Some(primary) = primary else {
+ return Ok(None);
+ };
+ if expected_sha256.is_some_and(|expected| verify_sha256(&primary, expected).is_err()) {
+ return Ok(None);
+ }
+ Ok(Some(primary))
+}
+
+fn metadata_matches(
+ metadata: &CacheMetadata,
+ repository: &str,
+ revision: &str,
+ expected_sha256: Option<&str>,
+) -> bool {
+ metadata.format == CACHE_FORMAT
+ && metadata.repository == repository
+ && metadata.revision == revision
+ && metadata.sha256.as_deref() == expected_sha256
+ && !metadata.primary.is_empty()
+}
+
+fn sha256_file(path: &Path) -> Result {
+ let mut file = fs::File::open(path).map_err(|e| format!("open {}: {e}", path.display()))?;
+ let mut digest = Sha256::new();
+ let mut buffer = [0u8; 1 << 20];
+ loop {
+ let read = file
+ .read(&mut buffer)
+ .map_err(|e| format!("hash {}: {e}", path.display()))?;
+ if read == 0 {
+ break;
+ }
+ digest.update(&buffer[..read]);
+ }
+ Ok(digest
+ .finalize()
+ .iter()
+ .map(|byte| format!("{byte:02x}"))
+ .collect())
+}
+
+fn verify_sha256(path: &Path, expected: &str) -> Result<(), String> {
+ let actual = sha256_file(path)?;
+ if actual == expected {
+ Ok(())
+ } else {
+ Err(format!(
+ "SHA-256 mismatch for {}: expected {expected}, got {actual}",
+ path.display()
+ ))
+ }
+}
+
+fn write_metadata(path: &Path, metadata: &CacheMetadata) -> Result<(), String> {
+ if let Some(parent) = path.parent() {
+ fs::create_dir_all(parent).map_err(|e| format!("create {}: {e}", parent.display()))?;
+ }
+ let file_name = path
+ .file_name()
+ .and_then(|name| name.to_str())
+ .unwrap_or("complete.json");
+ let part = path.with_file_name(format!("{file_name}.part"));
+ let bytes = serde_json::to_vec_pretty(metadata).map_err(|e| e.to_string())?;
+ fs::write(&part, bytes).map_err(|e| format!("write {}: {e}", part.display()))?;
+ if path.exists() {
+ fs::remove_file(path).map_err(|e| format!("replace {}: {e}", path.display()))?;
+ }
+ fs::rename(&part, path)
+ .map_err(|e| format!("move {} to {}: {e}", part.display(), path.display()))
+}
+
+fn external_locations(model: &ModelProto) -> Result, String> {
+ fn walk(graph: &GraphProto, locations: &mut BTreeSet) -> Result<(), String> {
+ for tensor in &graph.initializer {
+ if tensor.data_location != 1 {
+ continue;
+ }
+ let location = tensor
+ .external_data
+ .iter()
+ .find(|entry| entry.key == "location")
+ .map(|entry| entry.value.as_str())
+ .ok_or_else(|| format!("external tensor '{}' has no location", tensor.name))?;
+ locations.insert(safe_relative_path(location)?);
+ }
+ for node in &graph.node {
+ for attribute in &node.attribute {
+ if let Some(graph) = &attribute.g {
+ walk(graph, locations)?;
+ }
+ for graph in &attribute.graphs {
+ walk(graph, locations)?;
+ }
+ }
+ }
+ Ok(())
+ }
+
+ let mut locations = BTreeSet::new();
+ if let Some(graph) = &model.graph {
+ walk(graph, &mut locations)?;
+ }
+ Ok(locations.into_iter().collect())
+}
+
+#[cfg(test)]
+mod tests {
+ use super::*;
+ use crate::protos::onnx::{StringStringEntryProto, TensorProto};
+
+ #[test]
+ fn model_cache_precedence_prefers_o2w_then_hugging_face() {
+ let path = resolve_cache_root(
+ Some("/o2w-models".into()),
+ Some("/o2w".into()),
+ Some("/hf-hub".into()),
+ Some("/legacy-hf-hub".into()),
+ Some("/hf-home".into()),
+ Some("/os-cache".into()),
+ );
+ assert_eq!(path, Some(PathBuf::from("/o2w-models")));
+
+ let path = resolve_cache_root(
+ None,
+ Some("/o2w".into()),
+ Some("/hf-hub".into()),
+ None,
+ None,
+ None,
+ );
+ assert_eq!(path, Some(PathBuf::from("/o2w/onnx")));
+
+ let path = resolve_cache_root(
+ None,
+ None,
+ Some("/hf-hub".into()),
+ None,
+ Some("/hf-home".into()),
+ Some("/os-cache".into()),
+ );
+ assert_eq!(path, Some(PathBuf::from("/hf-hub")));
+
+ let path = resolve_cache_root(None, None, None, None, None, Some("/os-cache".into()));
+ assert_eq!(path, Some(PathBuf::from("/os-cache/huggingface/hub")));
+ }
+
+ #[test]
+ fn cache_metadata_requires_current_format_revision_and_digest() {
+ let metadata = CacheMetadata {
+ format: CACHE_FORMAT,
+ repository: "org/repo".into(),
+ revision: "0123456789abcdef0123456789abcdef01234567".into(),
+ sha256: Some("0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef".into()),
+ primary: "onnx/model.onnx".into(),
+ files: Vec::new(),
+ };
+ let digest = metadata.sha256.as_deref();
+ assert!(metadata_matches(
+ &metadata,
+ &metadata.repository,
+ &metadata.revision,
+ digest
+ ));
+ assert!(!metadata_matches(
+ &metadata,
+ &metadata.repository,
+ "other-revision",
+ digest
+ ));
+ assert!(!metadata_matches(
+ &metadata,
+ &metadata.repository,
+ &metadata.revision,
+ None
+ ));
+
+ let mut old = metadata;
+ old.format -= 1;
+ assert!(!metadata_matches(
+ &old,
+ &old.repository,
+ &old.revision,
+ old.sha256.as_deref()
+ ));
+ }
+
+ #[test]
+ fn computes_sha256_for_download_verification() {
+ let directory = tempfile::tempdir().unwrap();
+ let path = directory.path().join("model.onnx");
+ fs::write(&path, b"abc").unwrap();
+ let digest = "ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad";
+ assert_eq!(sha256_file(&path).unwrap(), digest);
+ verify_sha256(&path, digest).unwrap();
+ let error = verify_sha256(
+ &path,
+ "0000000000000000000000000000000000000000000000000000000000000000",
+ )
+ .unwrap_err();
+ assert!(error.contains("SHA-256 mismatch"));
+ }
+
+ #[test]
+ fn metadata_path_is_stable_for_the_source_identity() {
+ let entry = Entry {
+ file: "org--repo/onnx/model.onnx".into(),
+ revision: Some("deadbeef".into()),
+ sha256: None,
+ heavy: false,
+ coreml_unsupported: None,
+ coreml_slow: None,
+ override_dims: Default::default(),
+ pin_inputs: Default::default(),
+ };
+ let root = Path::new("/cache");
+ let first = metadata_path(root, &entry);
+ assert_eq!(first, metadata_path(root, &entry));
+ assert!(first.starts_with(root.join(METADATA_DIR)));
+ assert_eq!(
+ first.extension().and_then(|value| value.to_str()),
+ Some("json")
+ );
+ }
+
+ #[test]
+ fn rejects_external_paths_that_escape_the_model_directory() {
+ assert!(safe_relative_path("../weights.bin").is_err());
+ assert!(safe_relative_path("/weights.bin").is_err());
+ assert_eq!(
+ safe_relative_path("weights/model.data").unwrap(),
+ Path::new("weights/model.data")
+ );
+ }
+
+ #[test]
+ fn finds_unique_external_data_locations() {
+ let tensor = TensorProto {
+ name: "weight".into(),
+ data_location: 1,
+ external_data: vec![StringStringEntryProto {
+ key: "location".into(),
+ value: "model.onnx_data".into(),
+ }],
+ ..Default::default()
+ };
+ let model = ModelProto {
+ graph: Some(GraphProto {
+ initializer: vec![tensor.clone(), tensor],
+ ..Default::default()
+ }),
+ ..Default::default()
+ };
+ assert_eq!(
+ external_locations(&model).unwrap(),
+ vec![PathBuf::from("model.onnx_data")]
+ );
+ }
+}
diff --git a/src/model_validation/manifest.rs b/src/model_validation/manifest.rs
new file mode 100644
index 0000000..6250743
--- /dev/null
+++ b/src/model_validation/manifest.rs
@@ -0,0 +1,285 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+//! Shared parsing and selection for the transformers.js model manifest.
+
+use serde::Deserialize;
+use std::collections::HashMap;
+use std::path::{Path, PathBuf};
+
+#[derive(Clone, Debug, Deserialize, PartialEq, Eq)]
+#[serde(deny_unknown_fields)]
+pub struct Entry {
+ pub file: String,
+ /// Immutable Hugging Face revision for reproducible downloads.
+ #[serde(default)]
+ pub revision: Option,
+ /// Expected SHA-256 of the primary ONNX file.
+ #[serde(default)]
+ pub sha256: Option,
+ #[serde(default)]
+ pub heavy: bool,
+ /// Reason this model cannot build on the CoreML backend.
+ #[serde(default)]
+ pub coreml_unsupported: Option,
+ /// Reason this model is prohibitively slow to build on the CoreML backend.
+ #[serde(default)]
+ pub coreml_slow: Option,
+ #[serde(default)]
+ pub override_dims: HashMap,
+ #[serde(default)]
+ pub pin_inputs: HashMap,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub enum Selection {
+ All,
+ Match(String),
+}
+
+impl Selection {
+ pub fn parse(value: &str) -> Result {
+ match value {
+ "all" => Ok(Self::All),
+ value if value.starts_with("match=") && value.len() > "match=".len() => {
+ Ok(Self::Match(value["match=".len()..].to_string()))
+ }
+ _ => Err(format!(
+ "invalid selection {value:?}: expected all or match="
+ )),
+ }
+ }
+
+ pub fn includes(&self, index: usize, entry: &Entry) -> bool {
+ match self {
+ Self::All => true,
+ Self::Match(needle) => entry.label(index).contains(needle),
+ }
+ }
+}
+
+impl Entry {
+ pub fn revision(&self) -> &str {
+ self.revision.as_deref().unwrap_or("main")
+ }
+
+ pub fn source_key(&self) -> String {
+ format!(
+ "{}@{}#{}",
+ self.file,
+ self.revision(),
+ self.sha256.as_deref().unwrap_or("")
+ )
+ }
+
+ pub fn label(&self, index: usize) -> String {
+ format!(
+ "#{index} {} dims={:?} pins={:?}",
+ self.file, self.override_dims, self.pin_inputs
+ )
+ }
+
+ /// Stable cache identity. Map entries are sorted before hashing, so two
+ /// cases using the same file but different overrides or pins cannot clash.
+ pub fn cache_key(&self) -> String {
+ let mut canonical = format!("webnn-cache-v1\nfile={}\n", self.file);
+ if let Some(revision) = &self.revision {
+ canonical.push_str(&format!("revision={revision}\n"));
+ }
+ if let Some(sha256) = &self.sha256 {
+ canonical.push_str(&format!("sha256={sha256}\n"));
+ }
+ let mut dims: Vec<_> = self.override_dims.iter().collect();
+ dims.sort_unstable_by_key(|(name, _)| *name);
+ for (name, value) in dims {
+ canonical.push_str(&format!("dim={name}:{value}\n"));
+ }
+ let mut pins: Vec<_> = self.pin_inputs.iter().collect();
+ pins.sort_unstable_by_key(|(name, _)| *name);
+ for (name, value) in pins {
+ canonical.push_str(&format!("pin={name}:{value}\n"));
+ }
+ let hash = canonical.bytes().fold(0xcbf29ce484222325u64, |hash, byte| {
+ (hash ^ u64::from(byte)).wrapping_mul(0x100000001b3)
+ });
+ let stem = Path::new(&self.file)
+ .file_stem()
+ .and_then(|s| s.to_str())
+ .unwrap_or("model")
+ .chars()
+ .map(|c| if c.is_ascii_alphanumeric() { c } else { '-' })
+ .collect::();
+ format!("{stem}-{hash:016x}")
+ }
+}
+
+pub fn manifest_path() -> PathBuf {
+ std::env::var_os("O2W_MANIFEST")
+ .filter(|v| !v.is_empty())
+ .map(PathBuf::from)
+ .unwrap_or_else(|| Path::new(env!("CARGO_MANIFEST_DIR")).join("tests/models/manifest.json"))
+}
+
+pub fn load_manifest() -> Result, String> {
+ load_manifest_from(&manifest_path())
+}
+
+pub fn load_manifest_from(path: &Path) -> Result, String> {
+ let text =
+ std::fs::read_to_string(path).map_err(|e| format!("read {}: {e}", path.display()))?;
+ parse_manifest(&text).map_err(|e| format!("parse {}: {e}", path.display()))
+}
+pub fn parse_manifest(text: &str) -> Result, String> {
+ let entries: Vec = serde_json::from_str(text).map_err(|e| e.to_string())?;
+ for (index, entry) in entries.iter().enumerate() {
+ if entry.file.is_empty() {
+ return Err(format!("manifest entry #{index} has an empty file"));
+ }
+ if let Some(revision) = &entry.revision {
+ if revision.is_empty()
+ || !revision
+ .bytes()
+ .all(|byte| byte.is_ascii_alphanumeric() || matches!(byte, b'.' | b'_' | b'-'))
+ {
+ return Err(format!(
+ "manifest entry #{index} has an invalid revision {revision:?}"
+ ));
+ }
+ }
+ if let Some(sha256) = &entry.sha256 {
+ if entry.revision.is_none() {
+ return Err(format!(
+ "manifest entry #{index} has sha256 without an immutable revision"
+ ));
+ }
+ if !valid_sha256(sha256) {
+ return Err(format!(
+ "manifest entry #{index} has an invalid lowercase sha256"
+ ));
+ }
+ }
+ }
+ Ok(entries)
+}
+
+fn valid_sha256(value: &str) -> bool {
+ value.len() == 64
+ && value
+ .bytes()
+ .all(|byte| byte.is_ascii_digit() || matches!(byte, b'a'..=b'f'))
+}
+
+#[cfg(test)]
+mod tests {
+ use super::*;
+ use std::collections::HashSet;
+
+ fn entry() -> Entry {
+ serde_json::from_str(
+ r#"{"file":"org--repo/onnx/model.onnx","override_dims":{"b":1},"pin_inputs":{"branch":0}}"#,
+ )
+ .unwrap()
+ }
+
+ #[test]
+ fn parses_and_selects_manifest_cases() {
+ let entry = entry();
+ assert!(Selection::All.includes(0, &entry));
+ assert!(Selection::Match("model.onnx".into()).includes(0, &entry));
+ assert!(!Selection::Match("missing".into()).includes(0, &entry));
+ }
+
+ #[test]
+ fn cache_key_is_stable_and_configuration_specific() {
+ let first = entry();
+ let reordered: Entry = serde_json::from_str(
+ r#"{"pin_inputs":{"branch":0},"override_dims":{"b":1},"file":"org--repo/onnx/model.onnx"}"#,
+ )
+ .unwrap();
+ assert_eq!(first.cache_key(), reordered.cache_key());
+
+ let mut changed = first.clone();
+ changed.override_dims.insert("b".into(), 2);
+ assert_ne!(first.cache_key(), changed.cache_key());
+ changed = first.clone();
+ changed.pin_inputs.insert("branch".into(), 1);
+ assert_ne!(first.cache_key(), changed.cache_key());
+
+ changed = first.clone();
+ changed.revision = Some("0123456789abcdef0123456789abcdef01234567".into());
+ changed.sha256 =
+ Some("0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef".into());
+ assert_ne!(first.cache_key(), changed.cache_key());
+ }
+
+ #[test]
+ fn validates_revision_and_digest() {
+ assert!(
+ parse_manifest(
+ r#"[{"file":"org--repo/model.onnx","sha256":"0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef"}]"#
+ )
+ .unwrap_err()
+ .contains("without an immutable revision")
+ );
+ assert!(
+ parse_manifest(r#"[{"file":"org--repo/model.onnx","revision":"../main"}]"#)
+ .unwrap_err()
+ .contains("invalid revision")
+ );
+ assert!(
+ parse_manifest(
+ r#"[{"file":"org--repo/model.onnx","revision":"0123456789abcdef0123456789abcdef01234567","sha256":"ABCDEF"}]"#
+ )
+ .unwrap_err()
+ .contains("invalid lowercase sha256")
+ );
+ }
+
+ #[test]
+ fn rejects_removed_validation_metadata() {
+ let error =
+ parse_manifest(r#"[{"file":"org--repo/model.onnx","validation":{"tier":"blocked"}}]"#)
+ .unwrap_err();
+ assert!(error.contains("unknown field"));
+ }
+
+ #[test]
+ fn curated_ci_manifest_is_small_pinned_and_unique() {
+ let path = Path::new(env!("CARGO_MANIFEST_DIR")).join("tests/models/ci-validation.json");
+ let entries = load_manifest_from(&path).expect("load curated CI manifest");
+ assert!(!entries.is_empty(), "CI manifest must not be empty");
+ let mut cases = HashSet::new();
+ for entry in entries {
+ assert!(!entry.heavy, "CI model {} must not be heavy", entry.file);
+ let revision = entry.revision.as_deref().expect("CI model revision");
+ assert!(
+ revision.len() == 40
+ && revision
+ .bytes()
+ .all(|byte| byte.is_ascii_digit() || matches!(byte, b'a'..=b'f')),
+ "CI model {} must use a full immutable commit SHA",
+ entry.file
+ );
+ assert!(
+ entry.sha256.as_deref().is_some_and(valid_sha256),
+ "CI model {} must declare its SHA-256",
+ entry.file
+ );
+ assert!(
+ cases.insert(entry.cache_key()),
+ "duplicate CI model case for {}",
+ entry.file
+ );
+ }
+ }
+
+ #[test]
+ fn selector_rejects_tiers_and_malformed_values() {
+ assert!(Selection::parse("smoke").is_err());
+ assert!(Selection::parse("extended").is_err());
+ assert!(Selection::parse("quick").is_err());
+ assert!(Selection::parse("match=").is_err());
+ }
+}
diff --git a/src/model_validation/mod.rs b/src/model_validation/mod.rs
new file mode 100644
index 0000000..2f461d3
--- /dev/null
+++ b/src/model_validation/mod.rs
@@ -0,0 +1,11 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+pub mod full_model;
+pub mod manifest;
+pub mod runner;
+pub mod skeleton;
+
+pub use runner::{run_manifest_validation, ModelFailure, RunOptions, RunSummary, WeightMode};
diff --git a/src/model_validation/runner.rs b/src/model_validation/runner.rs
new file mode 100644
index 0000000..f6a7893
--- /dev/null
+++ b/src/model_validation/runner.rs
@@ -0,0 +1,314 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+use super::full_model::cache_full_model;
+use super::manifest::{load_manifest_from, Entry, Selection};
+use crate::{convert_onnx, validate_cached_model_with_options, ConvertOptions};
+use std::collections::HashMap;
+use std::fmt;
+use std::path::{Path, PathBuf};
+use std::str::FromStr;
+use std::sync::{Arc, Mutex, OnceLock};
+
+type ModelCell = Arc>>;
+
+#[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
+pub enum WeightMode {
+ #[default]
+ Real,
+}
+
+impl FromStr for WeightMode {
+ type Err = String;
+ fn from_str(value: &str) -> Result {
+ match value {
+ "real" => Ok(Self::Real),
+ _ => Err(format!("invalid weight mode {value:?}; expected real")),
+ }
+ }
+}
+
+impl fmt::Display for WeightMode {
+ fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+ f.write_str(match self {
+ Self::Real => "real",
+ })
+ }
+}
+
+#[derive(Clone, Debug)]
+pub struct RunOptions {
+ pub selection: Selection,
+ pub weights: WeightMode,
+ pub manifest: PathBuf,
+ pub jobs: usize,
+ pub webnn_cache: PathBuf,
+}
+
+impl RunOptions {
+ pub fn new(selection: Selection, weights: WeightMode) -> Result {
+ let root = Path::new(env!("CARGO_MANIFEST_DIR"));
+ Ok(Self {
+ selection,
+ weights,
+ manifest: std::env::var_os("O2W_MANIFEST")
+ .filter(|v| !v.is_empty())
+ .map(PathBuf::from)
+ .unwrap_or_else(|| root.join("tests/models/manifest.json")),
+ jobs: 1,
+ webnn_cache: crate::cache::webnn_cache_dir()?,
+ })
+ }
+}
+
+#[derive(Clone, Debug, Default, Eq, PartialEq)]
+pub struct RunSummary {
+ pub selected: usize,
+ pub passed: usize,
+ pub succeeded: Vec,
+ pub failed: Vec,
+}
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub struct ModelFailure {
+ pub label: String,
+ pub error: String,
+}
+
+impl RunSummary {
+ pub fn pass_percentage(&self) -> f64 {
+ if self.selected == 0 {
+ 0.0
+ } else {
+ self.passed as f64 * 100.0 / self.selected as f64
+ }
+ }
+
+ pub fn has_failures(&self) -> bool {
+ !self.failed.is_empty()
+ }
+}
+
+impl fmt::Display for RunSummary {
+ fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+ writeln!(f, "Model validation summary")?;
+ writeln!(f, "Succeeded ({}):", self.succeeded.len())?;
+ if self.succeeded.is_empty() {
+ writeln!(f, " (none)")?;
+ } else {
+ for label in &self.succeeded {
+ writeln!(f, " PASS {label}")?;
+ }
+ }
+ writeln!(f, "Failed ({}):", self.failed.len())?;
+ if self.failed.is_empty() {
+ writeln!(f, " (none)")?;
+ } else {
+ for failure in &self.failed {
+ writeln!(f, " FAIL {}: {}", failure.label, failure.error)?;
+ }
+ }
+ write!(
+ f,
+ "Overall: {}/{} passed ({:.1}%)",
+ self.passed,
+ self.selected,
+ self.pass_percentage()
+ )
+ }
+}
+
+pub fn run_manifest_validation(options: RunOptions) -> Result {
+ if options.jobs == 0 {
+ return Err("validation jobs must be at least 1".to_string());
+ }
+ let entries = load_manifest_from(&options.manifest)?;
+ let selected: Vec<_> = entries
+ .into_iter()
+ .enumerate()
+ .filter(|(index, entry)| options.selection.includes(*index, entry))
+ .collect();
+ if selected.is_empty() {
+ return Err("validation selection matched no models".to_string());
+ }
+ let selected_count = selected.len();
+ let (heavy, light): (Vec<_>, Vec<_>) = selected.into_iter().partition(|(_, entry)| entry.heavy);
+ let sweep = Sweep {
+ options,
+ models: Mutex::new(HashMap::new()),
+ results: Mutex::new(Vec::new()),
+ };
+ sweep.run(light, sweep.options.jobs);
+ sweep.run(heavy, 1);
+ let mut results = sweep.results.into_inner().unwrap();
+ results.sort_unstable_by_key(|result| result.index);
+ let mut succeeded = Vec::new();
+ let mut failed = Vec::new();
+ for result in results {
+ if let Some(error) = result.error {
+ failed.push(ModelFailure {
+ label: result.label,
+ error,
+ });
+ } else {
+ succeeded.push(result.label);
+ }
+ }
+ let passed = succeeded.len();
+ Ok(RunSummary {
+ selected: selected_count,
+ passed,
+ succeeded,
+ failed,
+ })
+}
+
+struct ModelResult {
+ index: usize,
+ label: String,
+ error: Option,
+}
+
+struct Sweep {
+ options: RunOptions,
+ models: Mutex>,
+ results: Mutex>,
+}
+
+impl Sweep {
+ fn model(&self, entry: &Entry) -> Result {
+ let key = format!("{}:{}", self.options.weights, entry.source_key());
+ let cell = self.models.lock().unwrap().entry(key).or_default().clone();
+ cell.get_or_init(|| cache_full_model(entry)).clone()
+ }
+
+ fn validate(&self, index: usize, entry: &Entry) {
+ let label = entry.label(index);
+ match self.validate_inner(entry) {
+ Ok((inputs, pins, outputs)) => {
+ eprintln!(
+ "ok {label}\n {inputs} inputs + {pins} pinned, {outputs} outputs ({})",
+ self.options.weights
+ );
+ self.results.lock().unwrap().push(ModelResult {
+ index,
+ label,
+ error: None,
+ });
+ }
+ Err(error) => {
+ eprintln!("FAIL {label}\n {error}");
+ self.results.lock().unwrap().push(ModelResult {
+ index,
+ label,
+ error: Some(error),
+ });
+ }
+ }
+ }
+
+ fn validate_inner(&self, entry: &Entry) -> Result<(usize, usize, usize), String> {
+ let onnx_path = self
+ .model(entry)
+ .map_err(|e| format!("model preparation: {e}"))?;
+ let webnn_path = self.options.webnn_cache.join(format!(
+ "{}-{}.webnn",
+ entry.cache_key(),
+ self.options.weights
+ ));
+ std::fs::create_dir_all(webnn_path.parent().expect("cache parent"))
+ .map_err(|e| format!("export: {e}"))?;
+ convert_onnx(
+ &onnx_path,
+ ConvertOptions {
+ free_dim_overrides: entry.override_dims.clone(),
+ optimize: true,
+ experimental_dynamic_inputs: false,
+ pinned_inputs: entry.pin_inputs.clone(),
+ zero_fill_missing_external_data: false,
+ output_path: Some(webnn_path.clone()),
+ },
+ )
+ .map_err(|e| format!("conversion/export: {e}"))?;
+ if !webnn_path.exists() || !webnn_path.with_extension("safetensors").exists() {
+ return Err("export did not produce .webnn and .safetensors".to_string());
+ }
+ let summary = validate_cached_model_with_options(
+ &onnx_path,
+ &webnn_path,
+ &entry.override_dims,
+ &entry.pin_inputs,
+ )
+ .map_err(|e| classify(&e.to_string()))?;
+ Ok((
+ summary.input_count,
+ summary.pinned_input_count,
+ summary.output_count,
+ ))
+ }
+
+ fn run(&self, entries: Vec<(usize, Entry)>, workers: usize) {
+ let queue = Mutex::new(entries);
+ std::thread::scope(|scope| {
+ for _ in 0..workers {
+ scope.spawn(|| loop {
+ let Some((index, entry)) = queue.lock().unwrap().pop() else {
+ break;
+ };
+ self.validate(index, &entry);
+ });
+ }
+ });
+ }
+}
+
+fn classify(error: &str) -> String {
+ let stage = if error.contains("native ORT") {
+ "native ORT"
+ } else if error.contains("reload WebNN") || error.contains("graph build") {
+ "reload"
+ } else if error.contains("dispatch") || error.contains("write") || error.contains("read output")
+ {
+ "dispatch"
+ } else if error.contains("mismatch") || error.contains("did not produce") {
+ "comparison"
+ } else {
+ "validation"
+ };
+ format!("{stage}: {error}")
+}
+
+#[cfg(test)]
+mod tests {
+ use super::*;
+ #[test]
+ fn parses_weight_modes() {
+ assert_eq!("real".parse(), Ok(WeightMode::Real));
+ assert!("generated".parse::().is_err());
+ assert!("random".parse::().is_err());
+ }
+ #[test]
+ fn real_is_default() {
+ assert_eq!(WeightMode::default(), WeightMode::Real);
+ }
+
+ #[test]
+ fn summary_lists_all_results_and_percentage() {
+ let summary = RunSummary {
+ selected: 3,
+ passed: 2,
+ succeeded: vec!["#0 first".into(), "#2 third".into()],
+ failed: vec![ModelFailure {
+ label: "#1 second".into(),
+ error: "comparison: mismatch".into(),
+ }],
+ };
+ let report = summary.to_string();
+ assert!(report.contains("Succeeded (2):\n PASS #0 first\n PASS #2 third"));
+ assert!(report.contains("Failed (1):\n FAIL #1 second: comparison: mismatch"));
+ assert!(report.contains("Overall: 2/3 passed (66.7%)"));
+ assert!(summary.has_failures());
+ }
+}
diff --git a/tests/common/skeleton.rs b/src/model_validation/skeleton.rs
similarity index 91%
rename from tests/common/skeleton.rs
rename to src/model_validation/skeleton.rs
index b8e9742..3618d88 100644
--- a/tests/common/skeleton.rs
+++ b/src/model_validation/skeleton.rs
@@ -51,6 +51,9 @@ const TENSOR_DATA_LOCATION: u32 = 14;
/// Random-access bytes: a local file or a remote file read by ranges.
pub trait ByteSource {
fn len(&self) -> u64;
+ fn is_empty(&self) -> bool {
+ self.len() == 0
+ }
fn read_at(&mut self, start: u64, len: usize) -> Result, String>;
}
@@ -88,16 +91,21 @@ pub struct HubSource {
agent: ureq::Agent,
url: String,
len: u64,
+ revision: Option,
}
impl HubSource {
/// `--/onnx/.onnx` -> `https://huggingface.co///resolve/main/onnx/.onnx`.
pub fn open(file: &str) -> Result {
+ Self::open_revision(file, "main")
+ }
+
+ pub fn open_revision(file: &str, requested_revision: &str) -> Result {
let (org_repo, rel) = file
.split_once('/')
.ok_or_else(|| format!("{file}: expected --/"))?;
let repo = org_repo.replacen("--", "/", 1);
- let url = format!("https://huggingface.co/{repo}/resolve/main/{rel}");
+ let url = format!("https://huggingface.co/{repo}/resolve/{requested_revision}/{rel}");
let agent = ureq::AgentBuilder::new()
.timeout(Duration::from_secs(120))
.build();
@@ -110,6 +118,10 @@ impl HubSource {
}
let response = with_retries(|| request.clone().call().map_err(|e| e.to_string()))
.map_err(|e| format!("resolve {url}: {e}"))?;
+ let revision = response
+ .header("x-repo-commit")
+ .map(str::to_string)
+ .or_else(|| (requested_revision != "main").then(|| requested_revision.to_string()));
let len = response
.header("Content-Length")
.and_then(|v| v.parse().ok())
@@ -118,8 +130,17 @@ impl HubSource {
agent,
url: response.get_url().to_string(),
len,
+ revision,
})
}
+
+ pub fn revision(&self) -> Option<&str> {
+ self.revision.as_deref()
+ }
+
+ pub fn resolved_url(&self) -> &str {
+ &self.url
+ }
}
impl ByteSource for HubSource {
@@ -365,6 +386,8 @@ fn strip_tensor(
let mut saw_data = false;
let mut dims = Vec::new();
let mut data_type = 0u64;
+ let mut source_offset = None;
+ let mut source_encoding = None;
while r.pos < end {
let (field, wt) = r.read_tag()?;
if TENSOR_DATA_FIELDS.contains(&field) {
@@ -372,6 +395,14 @@ fn strip_tensor(
if wt == LENGTH_DELIMITED {
let n = r.read_varint()?;
data_bytes += n;
+ if n as usize > keep_bytes {
+ source_offset = Some(r.pos);
+ source_encoding = Some(if field == 9 {
+ "raw_data"
+ } else {
+ "protobuf_data"
+ });
+ }
if n as usize <= keep_bytes {
kept.extend(encode_len_field(field, &r.read(n as usize)?));
} else {
@@ -426,6 +457,16 @@ fn strip_tensor(
let mut entry = encode_len_field(1, b"length");
entry.extend(encode_len_field(2, length.to_string().as_bytes()));
kept.extend(encode_len_field(TENSOR_EXTERNAL_DATA, &entry));
+ if let Some(offset) = source_offset {
+ let mut source = encode_len_field(1, b"o2w_source_offset");
+ source.extend(encode_len_field(2, offset.to_string().as_bytes()));
+ kept.extend(encode_len_field(TENSOR_EXTERNAL_DATA, &source));
+ }
+ if let Some(encoding) = source_encoding {
+ let mut source = encode_len_field(1, b"o2w_source_encoding");
+ source.extend(encode_len_field(2, encoding.as_bytes()));
+ kept.extend(encode_len_field(TENSOR_EXTERNAL_DATA, &source));
+ }
kept.extend(encode_tag(TENSOR_DATA_LOCATION, VARINT));
kept.extend(encode_varint(1));
Ok(kept)
diff --git a/src/onnx/builder_helpers.rs b/src/onnx/builder_helpers.rs
index 523b5b3..eb5f943 100644
--- a/src/onnx/builder_helpers.rs
+++ b/src/onnx/builder_helpers.rs
@@ -206,6 +206,21 @@ pub fn slice_with_params(
map_op_result(b.builder.slice_with_options(input, starts, sizes, opts))
}
+pub fn slice_with_params_and_strides(
+ b: &mut OnnxBuilder<'_, '_, '_>,
+ input: MLOperand,
+ label: &str,
+ starts: &[u32],
+ sizes: &[MLDimension],
+ strides: &[u32],
+) -> Result {
+ let opts = MLSliceOptions {
+ label: label.to_string(),
+ strides: strides.to_vec(),
+ };
+ map_op_result(b.builder.slice_with_options(input, starts, sizes, opts))
+}
+
/// First scalar element of an ONNX tensor as WebNN `MLNumber` (`serde_json::Value`).
pub fn ml_number_from_tensor(t: &TensorProto) -> Result {
match t.data_type {
diff --git a/src/onnx/convert.rs b/src/onnx/convert.rs
index 8f8d6d2..f0f968c 100644
--- a/src/onnx/convert.rs
+++ b/src/onnx/convert.rs
@@ -30,9 +30,9 @@ use rustnn::mlcontext::{
};
use rustnn::DataType;
use serde_json::Value as JsonValue;
-use std::collections::{HashMap, HashSet};
+use std::collections::{BTreeMap, HashMap, HashSet};
use std::fs;
-use std::path::Path;
+use std::path::{Path, PathBuf};
use thiserror::Error;
use webnn_onnx_utils::{data_types as utils_data_types, identifiers};
@@ -101,6 +101,9 @@ pub enum OnnxError {
#[error("shape inference failed for node: {0}")]
ShapeInference(String),
+
+ #[error("validation failed: {0}")]
+ Validation(String),
}
impl OnnxError {
@@ -184,6 +187,9 @@ pub struct ConvertOptions {
/// weight-stripped "skeleton" model exercise the full conversion and ORT
/// graph build (values never matter for graph structure).
pub zero_fill_missing_external_data: bool,
+ /// Write the lowered graph as a rustnn `.webnn` file plus sibling
+ /// `.safetensors` constants before building it for structural validation.
+ pub output_path: Option,
}
/// Parse a `--pin-input NAME=VALUE` argument (`true`/`false` or an integer).
@@ -706,10 +712,6 @@ Provide --override-dim {}= or enable --experimental-dynamic-inputs.",
)));
};
- if shape.is_empty() {
- continue;
- }
-
b.register_input(&raw_name, data_type, &shape)?;
value_name_map.insert(raw_name.clone(), name.clone());
@@ -1664,6 +1666,28 @@ pub fn convert_onnx>(
convert_model(model, &options)
}
+/// Cache an ONNX model with all external tensor data embedded in the protobuf.
+///
+/// Keeping the cache self-contained means native ORT validation can reopen the
+/// cached `.onnx` without relying on a source-side `.onnx.data` file.
+pub fn cache_onnx_model, Q: AsRef>(
+ source_path: P,
+ cache_path: Q,
+ zero_fill_missing_external_data: bool,
+) -> Result<(), OnnxError> {
+ let source_path = source_path.as_ref();
+ let bytes = fs::read(source_path)?;
+ let mut model = ModelProto::decode(bytes.as_slice())
+ .map_err(|e| OnnxError::ProtobufError(e.to_string()))?;
+ load_external_tensor_data(&mut model, source_path, zero_fill_missing_external_data)?;
+ let cache_path = cache_path.as_ref();
+ if let Some(parent) = cache_path.parent() {
+ fs::create_dir_all(parent)?;
+ }
+ fs::write(cache_path, model.encode_to_vec())?;
+ Ok(())
+}
+
/// Inline `If` nodes whose condition folds to a constant (e.g. pyannote's
/// shape-derived `Equal(Gather(Shape(x)), k)` gate). The chosen branch's
/// nodes and initializers are spliced into the outer graph with internal
@@ -1944,7 +1968,7 @@ pub(crate) fn convert_model(
.as_ref()
.ok_or_else(|| OnnxError::ProtobufError("Missing graph in model".to_string()))?;
- let mut outputs: HashMap = HashMap::new();
+ let mut outputs: BTreeMap = BTreeMap::new();
for output in onnx_graph.output.as_slice() {
// Sequences only exist as lowered elements; one escaping to a graph
// output has no WebNN representation.
@@ -1961,6 +1985,17 @@ pub(crate) fn convert_model(
let output_key = onnx_builder.build_output_key(output.name.as_str());
outputs.insert(output_key, op);
}
+ if let Some(path) = &options.output_path {
+ if let Some(parent) = path.parent() {
+ fs::create_dir_all(parent)?;
+ }
+ let output_refs: BTreeMap<&str, MLOperand> =
+ outputs.iter().map(|(k, v)| (k.as_str(), *v)).collect();
+ onnx_builder
+ .builder
+ .rustnn_save_webnn(&output_refs, path)
+ .map_err(map_rustnn_error)?;
+ }
let output_refs: HashMap<&str, MLOperand> =
outputs.iter().map(|(k, v)| (k.as_str(), *v)).collect();
diff --git a/src/onnx/ops/conversion.rs b/src/onnx/ops/conversion.rs
index 83dfaab..cba4cee 100644
--- a/src/onnx/ops/conversion.rs
+++ b/src/onnx/ops/conversion.rs
@@ -101,8 +101,36 @@ impl ConversionHandler {
let output_name = output_label(node, node_name);
let input = b.resolve_operand(&inputs[0])?;
+ let requested_type = to_type.unwrap() as i32;
+ if requested_type == crate::protos::onnx::TensorProto_DataType::Bool as i32 {
+ // ONNX numeric-to-bool casts normalize zero to false and every nonzero
+ // value (including NaN) to true. WebNN represents logical tensors as
+ // uint8, so a numeric cast alone would incorrectly preserve values.
+ let input_type = resolve_value_type(context, &inputs[0]).ok_or_else(|| {
+ OnnxError::InvalidShape(format!(
+ "Cast could not infer input type from '{}'",
+ inputs[0]
+ ))
+ })?;
+ let zero_name = format!("{output_name}__zero");
+ let zero_len = input_type.storage_byte_length(1).ok_or_else(|| {
+ OnnxError::InvalidShape("Cast boolean zero size overflow".to_string())
+ })?;
+ b.register_constant_from_bytes(&zero_name, input_type, &[], vec![0; zero_len])?;
+ let zero = b.resolve_operand(&zero_name)?;
+ let out = b
+ .builder
+ .not_equal_with_options(input, zero, OnnxBuilder::labeled_options(&output_name))
+ .map_err(map_op_error)?;
+ if let Some(onnx_out) = node.output.first() {
+ record_node_output(b, onnx_out, &output_name, out);
+ } else {
+ b.record_operand(&[&output_name], out);
+ }
+ return Ok(ConversionResult::default());
+ }
// float64 is lowered to float32 (WebNN has no float64).
- let mut to_type = to_type.unwrap() as i32;
+ let mut to_type = requested_type;
if to_type == crate::protos::onnx::TensorProto_DataType::Double as i32 {
to_type = crate::protos::onnx::TensorProto_DataType::Float as i32;
}
diff --git a/src/onnx/ops/utility.rs b/src/onnx/ops/utility.rs
index 13365a3..b95bdb6 100644
--- a/src/onnx/ops/utility.rs
+++ b/src/onnx/ops/utility.rs
@@ -21,7 +21,7 @@
use crate::onnx::builder::{map_op_error, OnnxBuilder};
use crate::onnx::builder_helpers::{
ast_dims_to_mldim, expand_with_shape, i64_starts_as_u32, output_label, record_node_output,
- slice_sizes_from_i64, slice_with_params,
+ slice_sizes_from_i64, slice_with_params, slice_with_params_and_strides,
};
use crate::onnx::convert::{sanitize_identifier, OnnxError};
use crate::onnx::ops::conv::lookup_shape;
@@ -1681,7 +1681,7 @@ impl UtilityHandler {
}
let output_name = output_label(node, node_name);
- let mut slice_params: Option<(Vec, Vec)> = None;
+ let mut slice_params: Option<(Vec, Vec, Vec)> = None;
// Axes handled as full reverses (negative-step slices spanning the dim).
let mut reverse_axes: Vec = Vec::new();
@@ -1904,11 +1904,9 @@ impl UtilityHandler {
start = start.clamp(0, dim);
end = end.clamp(0, dim);
- let size = if end <= start {
- 0
- } else {
- (end - start + step - 1) / step
- };
+ // WebNN `sizes` are input extents. The output dimension is
+ // ceil(size / stride), which RustNN infers from the options.
+ let size = (end - start).max(0);
// If this end value came from a dynamic dimension, mark the size as dynamic
if let Some(dims) = ends_dims {
@@ -1928,6 +1926,16 @@ impl UtilityHandler {
slice_params = Some((
i64_starts_as_u32(&dense_starts)?,
slice_sizes_from_i64(&dense_sizes, &dynamic_size_info)?,
+ dense_strides
+ .into_iter()
+ .map(|stride| {
+ u32::try_from(stride).map_err(|_| {
+ OnnxError::InvalidShape(format!(
+ "Slice '{node_name}' stride {stride} is out of range"
+ ))
+ })
+ })
+ .collect::, _>>()?,
));
} else {
return Err(OnnxError::InvalidShape(
@@ -2025,11 +2033,8 @@ impl UtilityHandler {
start = start.clamp(0, dim);
end = end.clamp(0, dim);
- let size = if end <= start {
- 0
- } else {
- (end - start + step - 1) / step
- };
+ // WebNN `sizes` are input extents, not output element counts.
+ let size = (end - start).max(0);
dense_starts[axis] = start;
dense_sizes[axis] = size;
@@ -2039,11 +2044,21 @@ impl UtilityHandler {
slice_params = Some((
i64_starts_as_u32(&dense_starts)?,
slice_sizes_from_i64(&dense_sizes, &vec![None; rank])?,
+ dense_strides
+ .into_iter()
+ .map(|stride| {
+ u32::try_from(stride).map_err(|_| {
+ OnnxError::InvalidShape(format!(
+ "Slice '{node_name}' stride {stride} is out of range"
+ ))
+ })
+ })
+ .collect::, _>>()?,
));
}
}
- let (starts, sizes) = slice_params.ok_or_else(|| {
+ let (starts, sizes, strides) = slice_params.ok_or_else(|| {
OnnxError::InvalidShape(
"Slice requires static starts/sizes for MLGraphBuilder".to_string(),
)
@@ -2064,7 +2079,8 @@ impl UtilityHandler {
} else {
format!("{output_name}__pre_reverse")
};
- let mut out = slice_with_params(b, input, &slice_label, &starts, &sizes)?;
+ let mut out =
+ slice_with_params_and_strides(b, input, &slice_label, &starts, &sizes, &strides)?;
if !reverse_axes.is_empty() {
out = b
.builder
diff --git a/src/validation.rs b/src/validation.rs
new file mode 100644
index 0000000..3e6a355
--- /dev/null
+++ b/src/validation.rs
@@ -0,0 +1,890 @@
+//! Cache reload and deterministic native-ORT validation for exported WebNN graphs.
+use crate::onnx::builder::OnnxBuilder;
+use crate::onnx::convert::{OnnxError, ValidatedGraph};
+use crate::protos::onnx::{
+ tensor_shape_proto::dimension::Value as DimensionValue, type_proto::Value as TypeProtoValue,
+ ModelProto, TensorProto_DataType, ValueInfoProto,
+};
+use half::f16;
+use prost::Message;
+use rustnn::graph::OperandDescriptor;
+use rustnn::mlcontext::{
+ MLContext, MLContextOptions, MLPowerPreference, MLTensor, MLTensorDescriptor,
+};
+use rustnn::operator_enums::MLOperandDataType;
+use rustnn::{load_graph_from_path, run_onnx_path_with_inputs, OnnxInput, TensorData};
+use std::collections::{BTreeMap, HashMap, HashSet};
+use std::fs;
+use std::path::Path;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct ValidationSummary {
+ pub input_count: usize,
+ pub pinned_input_count: usize,
+ pub output_count: usize,
+}
+
+/// Reload saved WebNN artifacts, dispatch deterministic fixed-shape inputs,
+/// and compare results with native ONNX Runtime on the cached ONNX model.
+pub fn validate_cached_model(
+ onnx_path: impl AsRef,
+ webnn_path: impl AsRef,
+) -> Result {
+ validate_cached_model_with_options(onnx_path, webnn_path, &HashMap::new(), &HashMap::new())
+}
+
+/// Validate cached artifacts using explicit bindings for symbolic ONNX input dimensions.
+pub fn validate_cached_model_with_overrides(
+ onnx_path: impl AsRef,
+ webnn_path: impl AsRef,
+ free_dim_overrides: &HashMap,
+) -> Result {
+ validate_cached_model_with_options(onnx_path, webnn_path, free_dim_overrides, &HashMap::new())
+}
+
+/// Validate cached artifacts with symbolic dimension bindings and converter-pinned inputs.
+pub fn validate_cached_model_with_options(
+ onnx_path: impl AsRef,
+ webnn_path: impl AsRef,
+ free_dim_overrides: &HashMap,
+ pinned_inputs: &HashMap,
+) -> Result {
+ let onnx_bytes = fs::read(onnx_path.as_ref())?;
+ let model = ModelProto::decode(onnx_bytes.as_slice())
+ .map_err(|e| OnnxError::ProtobufError(e.to_string()))?;
+ let inputs = deterministic_inputs(&model, free_dim_overrides, pinned_inputs)?;
+ let reference = run_onnx_path_with_inputs(onnx_path.as_ref(), clone_inputs(&inputs))
+ .map_err(|e| OnnxError::Validation(format!("native ORT run failed: {e}")))?;
+ let graph_info = load_graph_from_path(webnn_path.as_ref())
+ .map_err(|e| OnnxError::Validation(format!("failed to reload WebNN cache: {e}")))?;
+ let mut context = MLContext::create(&MLContextOptions::new(MLPowerPreference::Default, false))
+ .map_err(|e| OnnxError::Validation(format!("MLContext::create failed: {e}")))?;
+ let graph = context
+ .rustnn_build_graph(graph_info)
+ .map_err(|e| OnnxError::Validation(format!("cached graph build failed: {e}")))?;
+ let mut validated = ValidatedGraph { context, graph };
+ let input_count = validated.graph.input_descriptors.len();
+ let actual = dispatch_and_collect(&mut validated, &model, &inputs, pinned_inputs)?;
+ compare_outputs(&model, &reference, &actual)?;
+ Ok(ValidationSummary {
+ input_count,
+ pinned_input_count: pinned_inputs.len(),
+ output_count: reference.len(),
+ })
+}
+
+fn graph(model: &ModelProto) -> Result<&crate::protos::onnx::GraphProto, OnnxError> {
+ model
+ .graph
+ .as_ref()
+ .ok_or_else(|| OnnxError::ProtobufError("Missing graph in model".to_string()))
+}
+
+fn tensor_dims(
+ vi: &ValueInfoProto,
+ free_dim_overrides: &HashMap,
+) -> Result<(i32, Vec), OnnxError> {
+ let ty = vi
+ .r#type
+ .as_ref()
+ .and_then(|ty| ty.value.as_ref())
+ .ok_or_else(|| OnnxError::Validation(format!("missing tensor type for {}", vi.name)))?;
+ let tensor = match ty {
+ TypeProtoValue::TensorType(tensor) => tensor,
+ _ => {
+ return Err(OnnxError::Validation(format!(
+ "non-tensor input {}",
+ vi.name
+ )))
+ }
+ };
+ let shape = tensor
+ .shape
+ .as_ref()
+ .ok_or_else(|| OnnxError::Validation(format!("missing shape for {}", vi.name)))?;
+ let dims = shape
+ .dim
+ .iter()
+ .map(|dim| match dim.value.as_ref() {
+ Some(DimensionValue::DimValue(value)) if *value >= 0 => Ok(*value as usize),
+ Some(DimensionValue::DimParam(name)) => free_dim_overrides
+ .get(name)
+ .map(|value| *value as usize)
+ .ok_or_else(|| {
+ OnnxError::Validation(format!(
+ "dynamic input dimension {name} in {}; provide an override for validation",
+ vi.name
+ ))
+ }),
+ _ => Err(OnnxError::Validation(format!(
+ "dynamic input dimension in {}; use fixed shapes for validation",
+ vi.name
+ ))),
+ })
+ .collect::, _>>()?;
+ Ok((tensor.elem_type, dims))
+}
+
+fn feedable_inputs(model: &ModelProto) -> Result, OnnxError> {
+ let graph = graph(model)?;
+ let initializer_names: HashSet<&str> =
+ graph.initializer.iter().map(|t| t.name.as_str()).collect();
+ Ok(graph
+ .input
+ .iter()
+ .filter(|input| !initializer_names.contains(input.name.as_str()))
+ .collect())
+}
+
+fn checked_element_count(name: &str, shape: &[usize]) -> Result {
+ if shape.contains(&0) {
+ return Ok(0);
+ }
+ shape.iter().try_fold(1usize, |count, &dimension| {
+ count.checked_mul(dimension).ok_or_else(|| {
+ OnnxError::Validation(format!(
+ "input {name} shape {shape:?} overflows element count"
+ ))
+ })
+ })
+}
+
+fn semantic_input_value(name: &str) -> Option {
+ match name {
+ "token_type_ids" => Some(0),
+ "attention_mask" => Some(1),
+ _ => None,
+ }
+}
+
+fn repeated_data(
+ elem_type: i32,
+ count: usize,
+ value: i64,
+ name: &str,
+) -> Result {
+ macro_rules! checked {
+ ($variant:ident, $type:ty) => {
+ TensorData::$variant(vec![
+ <$type>::try_from(value).map_err(|_| {
+ OnnxError::Validation(format!("value {value} is out of range for {name}"))
+ })?;
+ count
+ ])
+ };
+ }
+ Ok(match elem_type {
+ x if x == TensorProto_DataType::Float as i32 => {
+ TensorData::Float32(vec![value as f32; count])
+ }
+ x if x == TensorProto_DataType::Float16 as i32 => {
+ TensorData::Float16(vec![f16::from_f32(value as f32).to_bits(); count])
+ }
+ x if x == TensorProto_DataType::Int8 as i32 => checked!(Int8, i8),
+ x if x == TensorProto_DataType::Uint8 as i32 => checked!(Uint8, u8),
+ x if x == TensorProto_DataType::Int32 as i32 => checked!(Int32, i32),
+ x if x == TensorProto_DataType::Uint32 as i32 => checked!(Uint32, u32),
+ x if x == TensorProto_DataType::Int64 as i32 => TensorData::Int64(vec![value; count]),
+ x if x == TensorProto_DataType::Uint64 as i32 => checked!(Uint64, u64),
+ x if x == TensorProto_DataType::Bool as i32 && matches!(value, 0 | 1) => {
+ TensorData::Uint8(vec![value as u8; count])
+ }
+ x if x == TensorProto_DataType::Bool as i32 => {
+ return Err(OnnxError::Validation(format!(
+ "bool input {name} must be 0 or 1, got {value}"
+ )))
+ }
+ other => {
+ return Err(OnnxError::Validation(format!(
+ "unsupported deterministic input dtype {other} for {name}"
+ )))
+ }
+ })
+}
+
+fn deterministic_data(elem_type: i32, count: usize, name: &str) -> Result {
+ Ok(match elem_type {
+ x if x == TensorProto_DataType::Float as i32 => {
+ TensorData::Float32((0..count).map(|i| ((i % 17) as f32 - 8.0) / 16.0).collect())
+ }
+ x if x == TensorProto_DataType::Float16 as i32 => TensorData::Float16(
+ (0..count)
+ .map(|i| f16::from_f32(((i % 17) as f32 - 8.0) / 16.0).to_bits())
+ .collect(),
+ ),
+ x if x == TensorProto_DataType::Int8 as i32 => {
+ TensorData::Int8((0..count).map(|i| (i % 7) as i8).collect())
+ }
+ x if x == TensorProto_DataType::Uint8 as i32 => {
+ TensorData::Uint8((0..count).map(|i| (i % 7) as u8).collect())
+ }
+ x if x == TensorProto_DataType::Int32 as i32 => {
+ TensorData::Int32((0..count).map(|i| (i % 7) as i32).collect())
+ }
+ x if x == TensorProto_DataType::Uint32 as i32 => {
+ TensorData::Uint32((0..count).map(|i| (i % 7) as u32).collect())
+ }
+ x if x == TensorProto_DataType::Int64 as i32 => {
+ TensorData::Int64((0..count).map(|i| (i % 7) as i64).collect())
+ }
+ x if x == TensorProto_DataType::Uint64 as i32 => {
+ TensorData::Uint64((0..count).map(|i| (i % 7) as u64).collect())
+ }
+ x if x == TensorProto_DataType::Bool as i32 => {
+ TensorData::Uint8((0..count).map(|i| u8::from(i % 2 == 0)).collect())
+ }
+ other => {
+ return Err(OnnxError::Validation(format!(
+ "unsupported deterministic input dtype {other} for {name}"
+ )))
+ }
+ })
+}
+
+fn input_data(
+ elem_type: i32,
+ count: usize,
+ name: &str,
+ pinned_value: Option,
+) -> Result {
+ if let Some(value) = pinned_value.or_else(|| semantic_input_value(name)) {
+ repeated_data(elem_type, count, value, name)
+ } else {
+ deterministic_data(elem_type, count, name)
+ }
+}
+
+fn deterministic_inputs(
+ model: &ModelProto,
+ free_dim_overrides: &HashMap,
+ pinned_inputs: &HashMap,
+) -> Result, OnnxError> {
+ let feedable = feedable_inputs(model)?;
+ let feedable_names: HashSet<&str> = feedable.iter().map(|input| input.name.as_str()).collect();
+ for name in pinned_inputs.keys() {
+ if !feedable_names.contains(name.as_str()) {
+ return Err(OnnxError::Validation(format!(
+ "pinned input {name} is not a feedable graph input"
+ )));
+ }
+ }
+ feedable
+ .into_iter()
+ .map(|input| {
+ let (elem_type, shape) = tensor_dims(input, free_dim_overrides)?;
+ let count = checked_element_count(&input.name, &shape)?;
+ let data = input_data(
+ elem_type,
+ count,
+ &input.name,
+ pinned_inputs.get(&input.name).copied(),
+ )?;
+ Ok(OnnxInput {
+ name: input.name.clone(),
+ shape,
+ data,
+ })
+ })
+ .collect()
+}
+
+fn clone_inputs(inputs: &[OnnxInput]) -> Vec {
+ inputs
+ .iter()
+ .map(|input| OnnxInput {
+ name: input.name.clone(),
+ shape: input.shape.clone(),
+ data: match &input.data {
+ TensorData::Float32(v) => TensorData::Float32(v.clone()),
+ TensorData::Float16(v) => TensorData::Float16(v.clone()),
+ TensorData::Int8(v) => TensorData::Int8(v.clone()),
+ TensorData::Uint8(v) => TensorData::Uint8(v.clone()),
+ TensorData::Int32(v) => TensorData::Int32(v.clone()),
+ TensorData::Uint32(v) => TensorData::Uint32(v.clone()),
+ TensorData::Int64(v) => TensorData::Int64(v.clone()),
+ TensorData::Uint64(v) => TensorData::Uint64(v.clone()),
+ },
+ })
+ .collect()
+}
+
+fn tensor_descriptor(desc: &OperandDescriptor) -> MLTensorDescriptor {
+ let data_type = MLOperandDataType::try_from(desc.data_type).expect("WebNN operand type");
+ let mut tensor = MLTensorDescriptor::new(
+ data_type,
+ desc.static_or_max_shape()
+ .into_iter()
+ .map(u64::from)
+ .collect(),
+ );
+ tensor.set_readable(true);
+ tensor.set_writable(true);
+ tensor
+}
+
+fn write_input(
+ context: &mut MLContext,
+ tensor: &MLTensor,
+ input: &OnnxInput,
+) -> Result<(), OnnxError> {
+ if checked_element_count(&input.name, &input.shape)? == 0 {
+ return Ok(());
+ }
+ let result = match &input.data {
+ TensorData::Float32(data) => context.write_tensor(tensor, data),
+ TensorData::Float16(data) => context.write_tensor(tensor, data),
+ TensorData::Int8(data) => context.write_tensor(tensor, data),
+ TensorData::Uint32(data) => context.write_tensor(tensor, data),
+ TensorData::Uint64(data) => context.write_tensor(tensor, data),
+ TensorData::Int32(data) => context.write_tensor(tensor, data),
+ TensorData::Int64(data) => context.write_tensor(tensor, data),
+ TensorData::Uint8(data) => context.write_tensor(tensor, data),
+ };
+ result.map_err(|e| OnnxError::Validation(format!("failed to write {}: {e}", input.name)))
+}
+
+enum CollectedOutput {
+ Numeric(Vec),
+ Int64(Vec),
+ Uint64(Vec),
+}
+
+impl CollectedOutput {
+ fn len(&self) -> usize {
+ match self {
+ Self::Numeric(data) => data.len(),
+ Self::Int64(data) => data.len(),
+ Self::Uint64(data) => data.len(),
+ }
+ }
+}
+
+fn read_output(
+ context: &mut MLContext,
+ tensor: &MLTensor,
+ desc: &OperandDescriptor,
+) -> Result {
+ let count = desc
+ .element_count()
+ .ok_or_else(|| OnnxError::Validation("output shape overflows element count".to_string()))?;
+ macro_rules! read_numeric {
+ ($type:ty) => {{
+ let mut data = vec![<$type>::default(); count];
+ if count != 0 {
+ context
+ .read_tensor(tensor, &mut data)
+ .map_err(|e| OnnxError::Validation(format!("failed to read output: {e}")))?;
+ }
+ CollectedOutput::Numeric(data.into_iter().map(|v| v as f64).collect())
+ }};
+ }
+ Ok(match desc.data_type {
+ rustnn::DataType::Float32 => read_numeric!(f32),
+ rustnn::DataType::Float16 => {
+ let mut data = vec![0u16; count];
+ if count != 0 {
+ context
+ .read_tensor(tensor, &mut data)
+ .map_err(|e| OnnxError::Validation(format!("failed to read output: {e}")))?;
+ }
+ CollectedOutput::Numeric(
+ data.into_iter()
+ .map(|v| f64::from(f16::from_bits(v).to_f32()))
+ .collect(),
+ )
+ }
+ rustnn::DataType::Int8 => read_numeric!(i8),
+ rustnn::DataType::Int32 => read_numeric!(i32),
+ rustnn::DataType::Int64 => {
+ let mut data = vec![0i64; count];
+ if count != 0 {
+ context
+ .read_tensor(tensor, &mut data)
+ .map_err(|e| OnnxError::Validation(format!("failed to read output: {e}")))?;
+ }
+ CollectedOutput::Int64(data)
+ }
+ rustnn::DataType::Uint8 => read_numeric!(u8),
+ rustnn::DataType::Uint32 => read_numeric!(u32),
+ rustnn::DataType::Uint64 => {
+ let mut data = vec![0u64; count];
+ if count != 0 {
+ context
+ .read_tensor(tensor, &mut data)
+ .map_err(|e| OnnxError::Validation(format!("failed to read output: {e}")))?;
+ }
+ CollectedOutput::Uint64(data)
+ }
+ other => {
+ return Err(OnnxError::Validation(format!(
+ "unsupported output dtype {other:?}"
+ )))
+ }
+ })
+}
+
+fn dispatch_and_collect(
+ validated: &mut ValidatedGraph,
+ model: &ModelProto,
+ inputs: &[OnnxInput],
+ pinned_inputs: &HashMap,
+) -> Result, OnnxError> {
+ let graph_proto = graph(model)?;
+ let mut source_inputs = HashMap::new();
+ for input in inputs
+ .iter()
+ .filter(|input| !pinned_inputs.contains_key(&input.name))
+ {
+ let key = OnnxBuilder::webnn_id(&input.name);
+ if let Some(previous) = source_inputs.insert(key.clone(), input) {
+ return Err(OnnxError::Validation(format!(
+ "ONNX inputs {} and {} both map to cached input key {key}",
+ previous.name, input.name
+ )));
+ }
+ }
+
+ // Pinning a merged model's branch selector specializes the converted graph. Inputs used only
+ // by the other branch are deliberately absent from the cached interface, while native ORT
+ // still needs the complete source-model feed above. Drive WebNN from its actual interface and
+ // require every retained cached input to map back to the source model.
+ let mut input_storage = Vec::new();
+ let mut input_keys = Vec::new();
+ for (key, desc) in &validated.graph.input_descriptors {
+ let input = source_inputs.get(key).ok_or_else(|| {
+ OnnxError::Validation(format!(
+ "cached graph input {key} has no matching unpinned ONNX input"
+ ))
+ })?;
+ let tensor = validated
+ .context
+ .create_tensor(&tensor_descriptor(desc))
+ .map_err(|e| OnnxError::Validation(format!("failed to create {key}: {e}")))?;
+ write_input(&mut validated.context, &tensor, input)?;
+ input_keys.push(key.clone());
+ input_storage.push(tensor);
+ }
+ let input_bindings: BTreeMap<&str, &MLTensor> = input_keys
+ .iter()
+ .zip(input_storage.iter())
+ .map(|(name, tensor)| (name.as_str(), tensor))
+ .collect();
+ let mut output_storage = Vec::new();
+ let mut output_keys = Vec::new();
+ let mut output_map = HashMap::new();
+ let cached_input_names: HashSet =
+ validated.graph.input_descriptors.keys().cloned().collect();
+ let mut source_outputs = HashMap::new();
+ for output in &graph_proto.output {
+ let key = OnnxBuilder::output_key_for(&output.name, &cached_input_names);
+ if let Some(previous) = source_outputs.insert(key.clone(), output.name.as_str()) {
+ return Err(OnnxError::Validation(format!(
+ "ONNX outputs {previous} and {} both map to cached output key {key}",
+ output.name
+ )));
+ }
+ }
+ for (key, desc) in &validated.graph.output_descriptors {
+ let onnx_name = source_outputs.get(key).ok_or_else(|| {
+ OnnxError::Validation(format!(
+ "cached graph output {key} has no matching ONNX output"
+ ))
+ })?;
+ let tensor = validated
+ .context
+ .create_tensor(&tensor_descriptor(desc))
+ .map_err(|e| OnnxError::Validation(format!("failed to create {key}: {e}")))?;
+ output_keys.push(key.clone());
+ output_storage.push(tensor);
+ output_map.insert((*onnx_name).to_string(), key.clone());
+ }
+ let output_bindings: BTreeMap<&str, &MLTensor> = output_keys
+ .iter()
+ .zip(output_storage.iter())
+ .map(|(name, tensor)| (name.as_str(), tensor))
+ .collect();
+ validated
+ .context
+ .dispatch(&mut validated.graph, &input_bindings, &output_bindings)
+ .map_err(|e| OnnxError::Validation(format!("cached graph dispatch failed: {e}")))?;
+ output_map
+ .into_iter()
+ .map(|(onnx_name, key)| {
+ let desc = validated
+ .graph
+ .output_descriptors
+ .get(&key)
+ .expect("validated above");
+ let tensor = output_bindings.get(key.as_str()).expect("bound above");
+ read_output(&mut validated.context, tensor, desc).map(|values| (onnx_name, values))
+ })
+ .collect()
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum FloatComparisonProfile {
+ Standard,
+ MatMulNBits4,
+ MatMulNBits8,
+ UnsupportedMatMulNBitsAccuracy(i64),
+}
+
+fn float_comparison_profile(model: &ModelProto) -> Result {
+ let mut profile = FloatComparisonProfile::Standard;
+ for node in graph(model)?
+ .node
+ .iter()
+ .filter(|node| node.op_type == "MatMulNBits")
+ {
+ let bits = node
+ .attribute
+ .iter()
+ .find(|attr| attr.name == "bits")
+ .map(|attr| attr.i)
+ .unwrap_or(4);
+ let accuracy_level = node
+ .attribute
+ .iter()
+ .find(|attr| attr.name == "accuracy_level")
+ .map(|attr| attr.i)
+ .unwrap_or(0);
+ if accuracy_level != 0 {
+ return Ok(FloatComparisonProfile::UnsupportedMatMulNBitsAccuracy(
+ accuracy_level,
+ ));
+ }
+ profile = match (profile, bits) {
+ (_, 8) => FloatComparisonProfile::MatMulNBits8,
+ (FloatComparisonProfile::Standard, 4) => FloatComparisonProfile::MatMulNBits4,
+ (current, 4) => current,
+ (current, _) => current,
+ };
+ }
+ Ok(profile)
+}
+
+fn float_tolerance(elem_type: Option, expected: f64, profile: FloatComparisonProfile) -> f64 {
+ match elem_type {
+ Some(x) if x == TensorProto_DataType::Float16 as i32 => 1e-3 + expected.abs() * 1e-2,
+ Some(x)
+ if x == TensorProto_DataType::Float as i32
+ && profile == FloatComparisonProfile::MatMulNBits8 =>
+ {
+ 2e-3 + expected.abs() * 2e-3
+ }
+ Some(x)
+ if x == TensorProto_DataType::Float as i32
+ && profile == FloatComparisonProfile::MatMulNBits4 =>
+ {
+ 1e-3 + expected.abs() * 1e-3
+ }
+ Some(x) if x == TensorProto_DataType::Float as i32 => 1e-5 + expected.abs() * 1e-4,
+ _ => 0.0,
+ }
+}
+
+fn compare_outputs(
+ model: &ModelProto,
+ reference: &[rustnn::OnnxOutputWithData],
+ actual: &HashMap,
+) -> Result<(), OnnxError> {
+ let outputs = &graph(model)?.output;
+ let comparison_profile = float_comparison_profile(model)?;
+ if let FloatComparisonProfile::UnsupportedMatMulNBitsAccuracy(level) = comparison_profile {
+ return Err(OnnxError::Validation(format!(
+ "MatMulNBits accuracy_level={level} is unsupported for numerical comparison"
+ )));
+ }
+ if outputs.len() != reference.len() {
+ return Err(OnnxError::Validation(
+ "native ORT output count mismatch".to_string(),
+ ));
+ }
+ let mut reference_by_name = HashMap::new();
+ for expected in reference {
+ if reference_by_name
+ .insert(expected.name.as_str(), expected)
+ .is_some()
+ {
+ return Err(OnnxError::Validation(format!(
+ "native ORT returned duplicate output {}",
+ expected.name
+ )));
+ }
+ }
+ for output in outputs {
+ let expected = reference_by_name.get(output.name.as_str()).ok_or_else(|| {
+ OnnxError::Validation(format!(
+ "native ORT did not return declared output {}",
+ output.name
+ ))
+ })?;
+ let Some(got) = actual.get(&output.name) else {
+ if expected.data.is_empty() {
+ continue;
+ }
+ return Err(OnnxError::Validation(format!(
+ "cached graph omitted non-empty ONNX output {} ({} elements)",
+ output.name,
+ expected.data.len()
+ )));
+ };
+ let elem_type = output
+ .r#type
+ .as_ref()
+ .and_then(|ty| ty.value.as_ref())
+ .and_then(|value| match value {
+ TypeProtoValue::TensorType(tensor) => Some(tensor.elem_type),
+ _ => None,
+ });
+ if expected.data.len() != got.len() {
+ return Err(OnnxError::Validation(format!(
+ "{} length mismatch: ORT={}, WebNN={}",
+ output.name,
+ expected.data.len(),
+ got.len()
+ )));
+ }
+ match got {
+ CollectedOutput::Int64(actual) => {
+ let expected = expected.int64_data.as_ref().ok_or_else(|| {
+ OnnxError::Validation(format!(
+ "native ORT did not return typed int64 data for {}",
+ output.name
+ ))
+ })?;
+ if let Some(index) = expected.iter().zip(actual).position(|(a, b)| a != b) {
+ return Err(OnnxError::Validation(format!(
+ "{}[{index}] mismatch: ORT={}, WebNN={}",
+ output.name, expected[index], actual[index]
+ )));
+ }
+ }
+ CollectedOutput::Uint64(actual) => {
+ let expected = expected.uint64_data.as_ref().ok_or_else(|| {
+ OnnxError::Validation(format!(
+ "native ORT did not return typed uint64 data for {}",
+ output.name
+ ))
+ })?;
+ if let Some(index) = expected.iter().zip(actual).position(|(a, b)| a != b) {
+ return Err(OnnxError::Validation(format!(
+ "{}[{index}] mismatch: ORT={}, WebNN={}",
+ output.name, expected[index], actual[index]
+ )));
+ }
+ }
+ CollectedOutput::Numeric(actual) => {
+ let mut first_mismatch = None;
+ let mut failure_count = 0usize;
+ let mut max_abs_error = 0.0f64;
+ let mut sum_abs_error = 0.0f64;
+ let mut sum_squared_error = 0.0f64;
+ let mut max_normalized_error = 0.0f64;
+ for (index, (expected, actual)) in expected.data.iter().zip(actual).enumerate() {
+ if expected.is_nan() && actual.is_nan() {
+ continue;
+ }
+ let tolerance = float_tolerance(elem_type, *expected, comparison_profile);
+ let abs_error = if expected == actual {
+ 0.0
+ } else if !expected.is_finite() || !actual.is_finite() {
+ f64::INFINITY
+ } else {
+ (expected - actual).abs()
+ };
+ max_abs_error = max_abs_error.max(abs_error);
+ sum_abs_error += abs_error;
+ sum_squared_error += abs_error * abs_error;
+ let normalized_error = if tolerance == 0.0 {
+ if abs_error == 0.0 {
+ 0.0
+ } else {
+ f64::INFINITY
+ }
+ } else {
+ abs_error / tolerance
+ };
+ max_normalized_error = max_normalized_error.max(normalized_error);
+ if abs_error > tolerance {
+ failure_count += 1;
+ first_mismatch.get_or_insert((index, *expected, *actual, tolerance));
+ }
+ }
+ if let Some((index, expected, observed, tolerance)) = first_mismatch {
+ let count = actual.len().max(1) as f64;
+ return Err(OnnxError::Validation(format!(
+ "{}[{index}] mismatch: ORT={expected}, WebNN={observed}, tolerance={tolerance}; failures={failure_count}/{}, max_abs_error={max_abs_error}, mean_abs_error={}, rmse={}, max_normalized_error={max_normalized_error}",
+ output.name,
+ actual.len(),
+ sum_abs_error / count,
+ (sum_squared_error / count).sqrt(),
+ )));
+ }
+ }
+ }
+ }
+ Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+ use super::*;
+
+ fn comparison_model(op_type: &str) -> ModelProto {
+ ModelProto {
+ graph: Some(crate::protos::onnx::GraphProto {
+ node: vec![crate::protos::onnx::NodeProto {
+ op_type: op_type.to_string(),
+ ..Default::default()
+ }],
+ ..Default::default()
+ }),
+ ..Default::default()
+ }
+ }
+
+ fn matmul_nbits_model(bits: i64, accuracy_level: i64) -> ModelProto {
+ let mut model = comparison_model("MatMulNBits");
+ let node = &mut model.graph.as_mut().unwrap().node[0];
+ node.attribute.push(crate::protos::onnx::AttributeProto {
+ name: "bits".to_string(),
+ i: bits,
+ ..Default::default()
+ });
+ if accuracy_level != 0 {
+ node.attribute.push(crate::protos::onnx::AttributeProto {
+ name: "accuracy_level".to_string(),
+ i: accuracy_level,
+ ..Default::default()
+ });
+ }
+ model
+ }
+
+ #[test]
+ fn matmul_nbits_models_use_attribute_derived_tolerances() {
+ let q4 = matmul_nbits_model(4, 0);
+ let q8 = matmul_nbits_model(8, 0);
+ let unsupported = matmul_nbits_model(4, 4);
+ assert_eq!(
+ float_comparison_profile(&q4).unwrap(),
+ FloatComparisonProfile::MatMulNBits4
+ );
+ assert_eq!(
+ float_comparison_profile(&q8).unwrap(),
+ FloatComparisonProfile::MatMulNBits8
+ );
+ assert_eq!(
+ float_comparison_profile(&unsupported).unwrap(),
+ FloatComparisonProfile::UnsupportedMatMulNBitsAccuracy(4)
+ );
+
+ let expected = 0.009632587432861328f64;
+ let ordinary = float_tolerance(
+ Some(TensorProto_DataType::Float as i32),
+ expected,
+ FloatComparisonProfile::Standard,
+ );
+ let q4_tolerance = float_tolerance(
+ Some(TensorProto_DataType::Float as i32),
+ expected,
+ FloatComparisonProfile::MatMulNBits4,
+ );
+ let q8_tolerance = float_tolerance(
+ Some(TensorProto_DataType::Float as i32),
+ expected,
+ FloatComparisonProfile::MatMulNBits8,
+ );
+ assert!(ordinary < q4_tolerance);
+ assert!(q4_tolerance < q8_tolerance);
+
+ let qwen_expected = -0.12115895748138428f64;
+ let qwen_observed = -0.11994504928588867f64;
+ let qwen_difference = (qwen_expected - qwen_observed).abs();
+ assert!(
+ qwen_difference
+ > float_tolerance(
+ Some(TensorProto_DataType::Float as i32),
+ qwen_expected,
+ FloatComparisonProfile::MatMulNBits4,
+ )
+ );
+ assert!(
+ qwen_difference
+ <= float_tolerance(
+ Some(TensorProto_DataType::Float as i32),
+ qwen_expected,
+ FloatComparisonProfile::MatMulNBits8,
+ )
+ );
+ assert_eq!(
+ float_comparison_profile(&comparison_model("MatMul")).unwrap(),
+ FloatComparisonProfile::Standard
+ );
+ }
+
+ #[test]
+ fn element_count_distinguishes_scalars_zero_dimensions_and_overflow() {
+ assert_eq!(checked_element_count("scalar", &[]).unwrap(), 1);
+ assert_eq!(checked_element_count("empty", &[1, 2, 0, 64]).unwrap(), 0);
+ assert_eq!(
+ checked_element_count("empty", &[usize::MAX, 2, 0]).unwrap(),
+ 0
+ );
+ let TensorData::Float32(empty) =
+ input_data(TensorProto_DataType::Float as i32, 0, "past_key", None).unwrap()
+ else {
+ panic!("expected float32 empty input");
+ };
+ assert!(empty.is_empty());
+ assert!(checked_element_count("huge", &[usize::MAX, 2]).is_err());
+ }
+
+ #[test]
+ fn semantic_inputs_use_valid_deterministic_values() {
+ let TensorData::Int64(token_types) = input_data(
+ TensorProto_DataType::Int64 as i32,
+ 4,
+ "token_type_ids",
+ None,
+ )
+ .unwrap() else {
+ panic!("expected int64 token types");
+ };
+ assert_eq!(token_types, vec![0; 4]);
+
+ let TensorData::Int64(mask) = input_data(
+ TensorProto_DataType::Int64 as i32,
+ 65,
+ "attention_mask",
+ None,
+ )
+ .unwrap() else {
+ panic!("expected int64 attention mask");
+ };
+ assert_eq!(mask, vec![1; 65]);
+ assert_eq!(mask.iter().sum::() - 1, 64);
+ }
+
+ #[test]
+ fn generic_pattern_and_pinned_precedence_are_preserved() {
+ let TensorData::Int64(generic) =
+ input_data(TensorProto_DataType::Int64 as i32, 9, "input_ids", None).unwrap()
+ else {
+ panic!("expected int64 generic input");
+ };
+ assert_eq!(generic, vec![0, 1, 2, 3, 4, 5, 6, 0, 1]);
+
+ let TensorData::Int64(pinned) = input_data(
+ TensorProto_DataType::Int64 as i32,
+ 3,
+ "attention_mask",
+ Some(0),
+ )
+ .unwrap() else {
+ panic!("expected int64 pinned input");
+ };
+ assert_eq!(pinned, vec![0; 3]);
+ }
+}
diff --git a/tests/cache_validation.rs b/tests/cache_validation.rs
new file mode 100644
index 0000000..09b5cdd
--- /dev/null
+++ b/tests/cache_validation.rs
@@ -0,0 +1,795 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+//! End-to-end cache export, reload, dispatch, and native ORT comparison.
+
+use std::fs;
+
+use onnx2webnn::onnx::test_models::prelude::*;
+use onnx2webnn::{
+ cache_onnx_model, convert_onnx, validate_cached_model, validate_cached_model_with_options,
+ ConvertOptions,
+};
+use prost::Message;
+
+#[test]
+fn saved_add_graph_reloads_and_matches_native_ort() {
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join("source.onnx");
+ let cached_onnx = dir.path().join("cached.onnx");
+ let cached_webnn = dir.path().join("cached.webnn");
+ let model = model(
+ 17,
+ graph(
+ "add",
+ vec![f32_input("x", &[1, 2])],
+ vec![f32_output("y", &[1, 2])],
+ vec![node("Add", "add", &["x", "w"], &["y"], &[])],
+ vec![f32_init("w", &[1, 2], &[0.5, 1.0])],
+ ),
+ );
+ fs::write(&source, model.encode_to_vec()).expect("write source model");
+ cache_onnx_model(&source, &cached_onnx, false).expect("cache ONNX");
+ convert_onnx(
+ &cached_onnx,
+ ConvertOptions {
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert and export");
+ assert!(cached_webnn.with_extension("safetensors").exists());
+ assert_eq!(
+ validate_cached_model(&cached_onnx, &cached_webnn)
+ .expect("validate")
+ .output_count,
+ 1
+ );
+}
+
+#[test]
+fn packed_uint4_matmul_round_trips_and_matches_native_ort() {
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join("matmul-nbits.onnx");
+ let cached_webnn = dir.path().join("matmul-nbits.webnn");
+ let (k, n, block_size) = (32i64, 2i64, 32i64);
+ let packed_weights = (0..n * block_size / 2)
+ .map(|index| ((index * 13 + 7) % 256) as u8)
+ .collect::>();
+ let mut matmul = node(
+ "MatMulNBits",
+ "matmul_nbits",
+ &["A", "B", "scales"],
+ &["Y"],
+ &[
+ attr_int("K", k),
+ attr_int("N", n),
+ attr_int("bits", 4),
+ attr_int("block_size", block_size),
+ ],
+ );
+ matmul.domain = "com.microsoft".to_string();
+ let mut fixture = model(
+ 17,
+ graph(
+ "packed_uint4_matmul",
+ vec![f32_input("A", &[1, k])],
+ vec![f32_output("Y", &[1, n])],
+ vec![matmul],
+ vec![
+ u8_init("B", &[n, 1, block_size / 2], &packed_weights),
+ f32_init("scales", &[n], &[0.05, 0.08]),
+ ],
+ ),
+ );
+ fixture
+ .opset_import
+ .push(onnx2webnn::protos::onnx::OperatorSetIdProto {
+ domain: "com.microsoft".to_string(),
+ version: 1,
+ });
+ fs::write(&source, fixture.encode_to_vec()).expect("write packed fixture");
+
+ convert_onnx(
+ &source,
+ ConvertOptions {
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert and export packed fixture");
+
+ assert!(cached_webnn.with_extension("safetensors").exists());
+
+ let summary = validate_cached_model(&source, &cached_webnn).expect("validate packed fixture");
+ assert_eq!(summary.input_count, 1);
+ assert_eq!(summary.output_count, 1);
+}
+
+fn integer_dtype_round_trip(label: &str, elem_type: i32) -> Result<(), String> {
+ let dir = tempfile::tempdir().map_err(|error| format!("temporary cache: {error}"))?;
+ let source = dir.path().join(format!("{label}.onnx"));
+ let cached_webnn = dir.path().join(format!("{label}.webnn"));
+ let model = model(
+ 17,
+ graph(
+ label,
+ vec![tensor_input("x", elem_type, &[4])],
+ vec![tensor_output("y", elem_type, &[4])],
+ vec![node("Identity", "identity", &["x"], &["y"], &[])],
+ vec![],
+ ),
+ );
+ fs::write(&source, model.encode_to_vec())
+ .map_err(|error| format!("write {label} dtype model: {error}"))?;
+ convert_onnx(
+ &source,
+ ConvertOptions {
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .map_err(|error| format!("convert {label}: {error}"))?;
+ let summary = validate_cached_model(&source, &cached_webnn)
+ .map_err(|error| format!("validate {label}: {error}"))?;
+ if summary.input_count != 1 || summary.output_count != 1 {
+ return Err(format!(
+ "validate {label}: expected one input and output, got {} and {}",
+ summary.input_count, summary.output_count
+ ));
+ }
+ Ok(())
+}
+
+fn expect_integer_dtype_round_trip(label: &str, elem_type: i32) {
+ integer_dtype_round_trip(label, elem_type)
+ .unwrap_or_else(|error| panic!("{label} should round-trip exactly: {error}"));
+}
+
+// WebNN permits all operand data types for Identity, but only requires float32,
+// float16, and int32. CoreML therefore legitimately omits optional int8 graph
+// boundary support. Keep executing it as a strict XFAIL instead of hiding it:
+// a different error is a regression, and unexpected success is an XPASS.
+// https://github.com/webmachinelearning/webnn/blob/main/index.bs#tensor-limits-elementwise-unary-identity
+#[cfg(all(target_os = "macos", feature = "coreml"))]
+fn expect_coreml_identity_boundary_xfail(label: &str, elem_type: i32) {
+ let error = match integer_dtype_round_trip(label, elem_type) {
+ Ok(()) => panic!("XPASS: CoreML now supports {label} Identity graph I/O"),
+ Err(error) => error,
+ };
+ let normalized = error.to_ascii_lowercase();
+ assert!(
+ normalized.contains("coreml") && normalized.contains("identity"),
+ "unexpected failure for CoreML {label} Identity graph I/O: {error}"
+ );
+ eprintln!("XFAIL: CoreML {label} Identity graph I/O: {error}");
+}
+
+#[cfg(not(all(target_os = "macos", feature = "coreml")))]
+#[test]
+fn accepted_int8_dtype_round_trips_exactly() {
+ expect_integer_dtype_round_trip(
+ "int8",
+ onnx2webnn::protos::onnx::TensorProto_DataType::Int8 as i32,
+ );
+}
+
+#[test]
+fn accepted_uint32_dtype_round_trips_exactly() {
+ expect_integer_dtype_round_trip(
+ "uint32",
+ onnx2webnn::protos::onnx::TensorProto_DataType::Uint32 as i32,
+ );
+}
+
+#[test]
+fn accepted_uint64_dtype_round_trips_exactly() {
+ expect_integer_dtype_round_trip(
+ "uint64",
+ onnx2webnn::protos::onnx::TensorProto_DataType::Uint64 as i32,
+ );
+}
+
+#[cfg(all(target_os = "macos", feature = "coreml"))]
+#[test]
+fn coreml_xfail_int8_identity_boundary() {
+ expect_coreml_identity_boundary_xfail(
+ "int8",
+ onnx2webnn::protos::onnx::TensorProto_DataType::Int8 as i32,
+ );
+}
+
+#[test]
+fn pinned_input_is_used_by_native_ort_but_not_dispatched() {
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join("pinned.onnx");
+ let cached_webnn = dir.path().join("pinned.webnn");
+ let model = model(
+ 17,
+ graph(
+ "pinned",
+ vec![f32_input("x", &[2]), f32_input("scale", &[])],
+ vec![f32_output("y", &[2])],
+ vec![node("Mul", "mul", &["x", "scale"], &["y"], &[])],
+ vec![],
+ ),
+ );
+ fs::write(&source, model.encode_to_vec()).expect("write pinned model");
+ let pins = std::collections::HashMap::from([("scale".to_string(), 2)]);
+ convert_onnx(
+ &source,
+ ConvertOptions {
+ pinned_inputs: pins.clone(),
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert pinned model");
+ let summary = validate_cached_model_with_options(
+ &source,
+ &cached_webnn,
+ &std::collections::HashMap::new(),
+ &pins,
+ )
+ .expect("validate pinned model");
+ assert_eq!(summary.input_count, 1);
+ assert_eq!(summary.pinned_input_count, 1);
+ assert_eq!(summary.output_count, 1);
+}
+
+#[test]
+fn pinned_input_name_reused_as_output_uses_the_converted_output_key() {
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join("pinned-output-collision.onnx");
+ let cached_webnn = dir.path().join("pinned-output-collision.webnn");
+ let model = model(
+ 17,
+ graph(
+ "pinned-output-collision",
+ vec![f32_input("x", &[2]), f32_input("scale", &[])],
+ vec![f32_output("y", &[2]), f32_output("scale", &[])],
+ vec![node("Identity", "identity", &["x"], &["y"], &[])],
+ vec![],
+ ),
+ );
+ fs::write(&source, model.encode_to_vec()).expect("write pinned-output collision model");
+ let pins = std::collections::HashMap::from([("scale".to_string(), 2)]);
+ convert_onnx(
+ &source,
+ ConvertOptions {
+ pinned_inputs: pins.clone(),
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert pinned-output collision model");
+
+ let summary = validate_cached_model_with_options(
+ &source,
+ &cached_webnn,
+ &std::collections::HashMap::new(),
+ &pins,
+ )
+ .expect("validate output whose name matches a pinned input");
+ assert_eq!(summary.input_count, 1);
+ assert_eq!(summary.pinned_input_count, 1);
+ assert_eq!(summary.output_count, 2);
+}
+
+#[test]
+fn pinned_if_branches_validate_with_their_specialized_input_interfaces() {
+ use onnx2webnn::protos::onnx::{AttributeProto, GraphProto};
+
+ let branch = |name: &str, input: &str| GraphProto {
+ name: name.to_string(),
+ node: vec![node(
+ "Identity",
+ &format!("{name}_identity"),
+ &[input],
+ &["branch_output"],
+ &[],
+ )],
+ output: vec![f32_output("branch_output", &[2])],
+ ..Default::default()
+ };
+ let mut if_node = node("If", "gate", &["use_cache_branch"], &["y"], &[]);
+ if_node.attribute = vec![
+ AttributeProto {
+ name: "then_branch".to_string(),
+ r#type: 5,
+ g: Some(branch("then", "then_input")),
+ ..Default::default()
+ },
+ AttributeProto {
+ name: "else_branch".to_string(),
+ r#type: 5,
+ g: Some(branch("else", "else_input")),
+ ..Default::default()
+ },
+ ];
+ let model = model(
+ 17,
+ graph(
+ "merged",
+ vec![
+ f32_input("then_input", &[2]),
+ f32_input("else_input", &[2]),
+ bool_input("use_cache_branch", &[1]),
+ ],
+ vec![f32_output("y", &[2])],
+ vec![if_node],
+ vec![],
+ ),
+ );
+
+ for branch_value in [0, 1] {
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join("merged.onnx");
+ let cached_webnn = dir.path().join("merged.webnn");
+ fs::write(&source, model.encode_to_vec()).expect("write merged model");
+ let pins =
+ std::collections::HashMap::from([("use_cache_branch".to_string(), branch_value)]);
+ convert_onnx(
+ &source,
+ ConvertOptions {
+ pinned_inputs: pins.clone(),
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert selected branch");
+
+ let summary = validate_cached_model_with_options(
+ &source,
+ &cached_webnn,
+ &std::collections::HashMap::new(),
+ &pins,
+ )
+ .expect("validate selected branch");
+ assert_eq!(summary.input_count, 1);
+ assert_eq!(summary.output_count, 1);
+ }
+}
+
+#[test]
+fn specialized_cache_accepts_pruned_inputs_and_empty_outputs() {
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join("specialized.onnx");
+ let cached_webnn = dir.path().join("specialized.webnn");
+ let model = model(
+ 17,
+ graph(
+ "specialized",
+ vec![
+ f32_input("x", &[2]),
+ f32_input("dead_branch_input", &[2]),
+ bool_input("use_cache_branch", &[]),
+ ],
+ vec![f32_output("y", &[2]), f32_output("empty_cache", &[0])],
+ vec![node("Identity", "identity", &["x"], &["y"], &[])],
+ vec![f32_init("empty_cache", &[0], &[])],
+ ),
+ );
+ fs::write(&source, model.encode_to_vec()).expect("write specialized model");
+ let pins = std::collections::HashMap::from([("use_cache_branch".to_string(), 0)]);
+ convert_onnx(
+ &source,
+ ConvertOptions {
+ pinned_inputs: pins.clone(),
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert specialized model");
+
+ let summary = validate_cached_model_with_options(
+ &source,
+ &cached_webnn,
+ &std::collections::HashMap::new(),
+ &pins,
+ )
+ .expect("validate specialized interface");
+ assert_eq!(summary.input_count, 1);
+ assert_eq!(summary.output_count, 2);
+}
+
+#[test]
+fn specialized_cache_rejects_an_omitted_nonempty_output() {
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join("source.onnx");
+ let cache_source = dir.path().join("cache-source.onnx");
+ let cached_webnn = dir.path().join("cached.webnn");
+ let source_model = model(
+ 17,
+ graph(
+ "source",
+ vec![f32_input("x", &[2])],
+ vec![f32_output("y", &[2]), f32_output("z", &[2])],
+ vec![
+ node("Identity", "y", &["x"], &["y"], &[]),
+ node("Identity", "z", &["x"], &["z"], &[]),
+ ],
+ vec![],
+ ),
+ );
+ let cache_model = model(
+ 17,
+ graph(
+ "cache",
+ vec![f32_input("x", &[2])],
+ vec![f32_output("y", &[2])],
+ vec![node("Identity", "y", &["x"], &["y"], &[])],
+ vec![],
+ ),
+ );
+ fs::write(&source, source_model.encode_to_vec()).expect("write source model");
+ fs::write(&cache_source, cache_model.encode_to_vec()).expect("write cache source");
+ convert_onnx(
+ &cache_source,
+ ConvertOptions {
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert cache source");
+
+ let error = validate_cached_model(&source, &cached_webnn)
+ .expect_err("non-empty source output must not be silently omitted");
+ assert!(error
+ .to_string()
+ .contains("cached graph omitted non-empty ONNX output z"));
+}
+
+#[test]
+fn specialized_cache_rejects_extra_inputs_and_outputs() {
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join("source.onnx");
+ let extra_input_source = dir.path().join("extra-input.onnx");
+ let extra_output_source = dir.path().join("extra-output.onnx");
+ let extra_input_webnn = dir.path().join("extra-input.webnn");
+ let extra_output_webnn = dir.path().join("extra-output.webnn");
+ let source_model = model(
+ 17,
+ graph(
+ "source",
+ vec![f32_input("x", &[2])],
+ vec![f32_output("y", &[2])],
+ vec![node("Identity", "y", &["x"], &["y"], &[])],
+ vec![],
+ ),
+ );
+ let extra_input_model = model(
+ 17,
+ graph(
+ "extra-input",
+ vec![f32_input("x", &[2]), f32_input("extra", &[2])],
+ vec![f32_output("y", &[2])],
+ vec![node("Add", "y", &["x", "extra"], &["y"], &[])],
+ vec![],
+ ),
+ );
+ let extra_output_model = model(
+ 17,
+ graph(
+ "extra-output",
+ vec![f32_input("x", &[2])],
+ vec![f32_output("y", &[2]), f32_output("extra", &[2])],
+ vec![
+ node("Identity", "y", &["x"], &["y"], &[]),
+ node("Identity", "extra", &["x"], &["extra"], &[]),
+ ],
+ vec![],
+ ),
+ );
+ fs::write(&source, source_model.encode_to_vec()).expect("write source model");
+ fs::write(&extra_input_source, extra_input_model.encode_to_vec())
+ .expect("write extra-input model");
+ fs::write(&extra_output_source, extra_output_model.encode_to_vec())
+ .expect("write extra-output model");
+
+ convert_onnx(
+ &extra_input_source,
+ ConvertOptions {
+ output_path: Some(extra_input_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert extra-input model");
+ let error = validate_cached_model(&source, &extra_input_webnn)
+ .expect_err("cached-only input must be rejected");
+ assert!(error
+ .to_string()
+ .contains("cached graph input extra has no matching unpinned ONNX input"));
+
+ convert_onnx(
+ &extra_output_source,
+ ConvertOptions {
+ output_path: Some(extra_output_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert extra-output model");
+ let error = validate_cached_model(&source, &extra_output_webnn)
+ .expect_err("cached-only output must be rejected");
+ assert!(error
+ .to_string()
+ .contains("cached graph output extra has no matching ONNX output"));
+}
+
+#[test]
+fn sanitized_source_input_collisions_are_rejected() {
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join("source.onnx");
+ let cache_source = dir.path().join("cache-source.onnx");
+ let cached_webnn = dir.path().join("cached.webnn");
+ let source_model = model(
+ 17,
+ graph(
+ "source",
+ vec![f32_input("a/b", &[2]), f32_input("a_b", &[2])],
+ vec![f32_output("y", &[2])],
+ vec![node("Identity", "y", &["a/b"], &["y"], &[])],
+ vec![],
+ ),
+ );
+ let cache_model = model(
+ 17,
+ graph(
+ "cache",
+ vec![f32_input("a_b", &[2])],
+ vec![f32_output("y", &[2])],
+ vec![node("Identity", "y", &["a_b"], &["y"], &[])],
+ vec![],
+ ),
+ );
+ fs::write(&source, source_model.encode_to_vec()).expect("write source model");
+ fs::write(&cache_source, cache_model.encode_to_vec()).expect("write cache source");
+ convert_onnx(
+ &cache_source,
+ ConvertOptions {
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert cache source");
+
+ let error = validate_cached_model(&source, &cached_webnn)
+ .expect_err("ambiguous source input mapping must be rejected");
+ assert!(error
+ .to_string()
+ .contains("both map to cached input key a_b"));
+}
+
+#[test]
+fn external_data_model_round_trips_without_embedding_weights() {
+ use onnx2webnn::protos::onnx::StringStringEntryProto;
+
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join("external.onnx");
+ let weights = dir.path().join("weights.bin");
+ let cached_webnn = dir.path().join("external.webnn");
+ let mut model = model(
+ 17,
+ graph(
+ "external",
+ vec![f32_input("x", &[2])],
+ vec![f32_output("y", &[2])],
+ vec![node("Add", "add", &["x", "w"], &["y"], &[])],
+ vec![f32_init("w", &[2], &[0.5, 1.0])],
+ ),
+ );
+ let weight = &mut model.graph.as_mut().unwrap().initializer[0];
+ weight.float_data.clear();
+ weight.data_location = 1;
+ weight.external_data = vec![
+ StringStringEntryProto {
+ key: "location".into(),
+ value: "weights.bin".into(),
+ },
+ StringStringEntryProto {
+ key: "offset".into(),
+ value: "0".into(),
+ },
+ StringStringEntryProto {
+ key: "length".into(),
+ value: "8".into(),
+ },
+ ];
+ fs::write(&source, model.encode_to_vec()).expect("write external-data model");
+ fs::write(
+ &weights,
+ [0.5f32, 1.0f32]
+ .into_iter()
+ .flat_map(f32::to_le_bytes)
+ .collect::>(),
+ )
+ .expect("write external weights");
+
+ convert_onnx(
+ &source,
+ ConvertOptions {
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert external-data model");
+ validate_cached_model(&source, &cached_webnn)
+ .expect("path-based native ORT resolves external weights");
+}
+
+fn validate_semantic_index_input(input_name: &str, length: i64) {
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join(format!("{input_name}.onnx"));
+ let cached_webnn = dir.path().join(format!("{input_name}.webnn"));
+ let model = model(
+ 17,
+ graph(
+ input_name,
+ vec![i64_input(input_name, &[length])],
+ vec![f32_output("y", &[length, 1])],
+ vec![node(
+ "Gather",
+ "gather",
+ &["table", input_name],
+ &["y"],
+ &[],
+ )],
+ vec![f32_init("table", &[2, 1], &[0.25, 0.75])],
+ ),
+ );
+ fs::write(&source, model.encode_to_vec()).expect("write semantic-input model");
+ convert_onnx(
+ &source,
+ ConvertOptions {
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert semantic-input model");
+ validate_cached_model(&source, &cached_webnn).expect("semantic input stays in range");
+}
+
+#[test]
+fn token_type_ids_stay_within_a_two_row_embedding() {
+ validate_semantic_index_input("token_type_ids", 4);
+}
+
+#[test]
+fn attention_mask_is_binary_and_valid_for_indexing() {
+ validate_semantic_index_input("attention_mask", 65);
+}
+
+#[test]
+fn zero_element_input_round_trips_without_becoming_a_scalar() {
+ use onnx2webnn::protos::onnx::{
+ tensor_shape_proto::dimension::Value as DimensionValue, type_proto::Value as TypeValue,
+ };
+
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join("zero-element-input.onnx");
+ let cached_webnn = dir.path().join("zero-element-input.webnn");
+ let mut empty_cache = f32_input("past_key", &[1, 2, 1, 4]);
+ let Some(TypeValue::TensorType(tensor)) =
+ empty_cache.r#type.as_mut().and_then(|ty| ty.value.as_mut())
+ else {
+ panic!("tensor input");
+ };
+ tensor.shape.as_mut().unwrap().dim[2].value =
+ Some(DimensionValue::DimParam("past_sequence_length".into()));
+ let model = model(
+ 17,
+ graph(
+ "zero-element-input",
+ vec![empty_cache, f32_input("x", &[1])],
+ vec![f32_output("y", &[1])],
+ vec![node("Identity", "identity", &["x"], &["y"], &[])],
+ vec![],
+ ),
+ );
+ fs::write(&source, model.encode_to_vec()).expect("write zero-element-input model");
+ let overrides = std::collections::HashMap::from([("past_sequence_length".to_string(), 0)]);
+ convert_onnx(
+ &source,
+ ConvertOptions {
+ free_dim_overrides: overrides.clone(),
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert zero-element-input model");
+ let summary = validate_cached_model_with_options(
+ &source,
+ &cached_webnn,
+ &overrides,
+ &std::collections::HashMap::new(),
+ )
+ .expect("zero-element input survives native and cached execution");
+ assert_eq!(summary.input_count, 2);
+ assert_eq!(summary.output_count, 1);
+}
+
+#[test]
+fn numeric_cast_to_bool_normalizes_nonzero_values() {
+ use onnx2webnn::protos::onnx::TensorProto_DataType;
+
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join("cast-bool.onnx");
+ let cached_webnn = dir.path().join("cast-bool.webnn");
+ let fixture = model(
+ 17,
+ graph(
+ "cast-bool",
+ vec![],
+ vec![f32_output("y", &[5])],
+ vec![
+ node(
+ "Cast",
+ "to_bool",
+ &["x"],
+ &["as_bool"],
+ &[attr_int("to", TensorProto_DataType::Bool as i64)],
+ ),
+ node(
+ "Cast",
+ "to_float",
+ &["as_bool"],
+ &["y"],
+ &[attr_int("to", TensorProto_DataType::Float as i64)],
+ ),
+ ],
+ vec![f32_init("x", &[5], &[0.0, 1.0, 2.0, -1.0, f32::NAN])],
+ ),
+ );
+ fs::write(&source, fixture.encode_to_vec()).expect("write cast-bool fixture");
+ convert_onnx(
+ &source,
+ ConvertOptions {
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert cast-bool fixture");
+ validate_cached_model(&source, &cached_webnn).expect("validate cast-bool fixture");
+}
+
+#[test]
+fn positive_step_slice_round_trips_with_extent_semantics() {
+ let dir = tempfile::tempdir().expect("temporary cache");
+ let source = dir.path().join("strided-slice.onnx");
+ let cached_webnn = dir.path().join("strided-slice.webnn");
+ let fixture = model(
+ 17,
+ graph(
+ "strided-slice",
+ vec![f32_input("x", &[10])],
+ vec![f32_output("y", &[4])],
+ vec![node(
+ "Slice",
+ "slice",
+ &["x", "starts", "ends", "axes", "steps"],
+ &["y"],
+ &[],
+ )],
+ vec![
+ i64_init("starts", &[1], &[1]),
+ i64_init("ends", &[1], &[9]),
+ i64_init("axes", &[1], &[0]),
+ i64_init("steps", &[1], &[2]),
+ ],
+ ),
+ );
+ fs::write(&source, fixture.encode_to_vec()).expect("write strided-slice fixture");
+ convert_onnx(
+ &source,
+ ConvertOptions {
+ output_path: Some(cached_webnn.clone()),
+ ..ConvertOptions::default()
+ },
+ )
+ .expect("convert strided-slice fixture");
+ let serialized = fs::read_to_string(&cached_webnn).expect("read serialized graph");
+ assert!(serialized.contains("sizes=[8]"));
+ assert!(serialized.contains("strides=[2]"));
+ validate_cached_model(&source, &cached_webnn).expect("validate strided-slice fixture");
+}
diff --git a/tests/common/mod.rs b/tests/common/mod.rs
index 93d9588..4f8ef1e 100644
--- a/tests/common/mod.rs
+++ b/tests/common/mod.rs
@@ -5,8 +5,10 @@
//! Shared helpers for ONNX op conversion integration tests.
-mod runner;
-pub mod skeleton;
+mod runner;
+
+#[allow(unused_imports)]
+pub use onnx2webnn::model_validation::{full_model, manifest, skeleton};
// Not every test crate uses every helper.
#[allow(unused_imports)]
diff --git a/tests/model_skeletons.rs b/tests/model_skeletons.rs
index 8852d09..484bba6 100644
--- a/tests/model_skeletons.rs
+++ b/tests/model_skeletons.rs
@@ -33,36 +33,11 @@ use std::path::{Path, PathBuf};
use std::sync::atomic::{AtomicUsize, Ordering};
use std::sync::{Arc, Mutex, OnceLock};
+use common::manifest::{load_manifest, Entry};
use common::skeleton::{strip_model, FileSource, HubSource, KEEP_BYTES};
use onnx2webnn::protos::onnx::ModelProto;
use onnx2webnn::{convert_model_proto, convert_onnx, ConvertOptions};
use prost::Message;
-use serde::Deserialize;
-
-/// Large models recurse deeply in shape inference; the default 2 MB thread
-/// stack is not enough.
-const WORKER_STACK_BYTES: usize = 256 << 20;
-
-#[derive(Deserialize)]
-struct Entry {
- file: String,
- #[serde(default)]
- heavy: bool,
- /// Reason this model cannot build on the CoreML backend (e.g. a hard
- /// backend limit like max tensor rank 5). Skipped when built with
- /// `--features coreml`; still exercised on the ORT backend.
- #[serde(default)]
- coreml_unsupported: Option,
- /// Reason this model, while it does build on CoreML, takes so long there
- /// that it would dominate the sweep (e.g. pathological Espresso compile
- /// times). Skipped like `coreml_unsupported`.
- #[serde(default)]
- coreml_slow: Option,
- #[serde(default)]
- override_dims: HashMap,
- #[serde(default)]
- pin_inputs: HashMap,
-}
enum Source {
Hub,
@@ -107,23 +82,32 @@ struct Sweep {
}
impl Sweep {
- fn skeleton(&self, file: &str) -> Skeleton {
+ fn skeleton(&self, entry: &Entry) -> Skeleton {
+ let file = entry.file.as_str();
let cell = self
.skeletons
.lock()
.unwrap()
- .entry(file.to_string())
+ .entry(entry.source_key())
.or_default()
.clone();
cell.get_or_init(|| {
let started = std::time::Instant::now();
- let cached = matches!(self.source, Source::Hub).then(|| cache_dir().join(file));
+ let cached = matches!(self.source, Source::Hub).then(|| {
+ entry
+ .revision
+ .as_deref()
+ .map(|revision| cache_dir().join("revisions").join(revision).join(file))
+ .unwrap_or_else(|| cache_dir().join(file))
+ });
if let Some(bytes) = cached.as_ref().and_then(|path| std::fs::read(path).ok()) {
let note = format!("skeleton {:.2} MB (cached)", bytes.len() as f64 / 1e6);
return Ok((Arc::new(bytes), note));
}
let (bytes, stats) = match &self.source {
- Source::Hub => strip_model(HubSource::open(file)?, KEEP_BYTES)?,
+ Source::Hub => {
+ strip_model(HubSource::open_revision(file, entry.revision())?, KEEP_BYTES)?
+ }
Source::StripDir(dir) => strip_model(FileSource::open(&dir.join(file))?, KEEP_BYTES)?,
Source::Dir(_) => unreachable!("full models are converted from disk"),
};
@@ -152,31 +136,27 @@ impl Sweep {
}
fn convert(&self, idx: usize, entry: &Entry) {
- let label = format!(
- "#{idx} {} dims={:?} pins={:?}",
- entry.file, entry.override_dims, entry.pin_inputs
- );
+ let label = entry.label(idx);
let options = ConvertOptions {
free_dim_overrides: entry.override_dims.clone(),
optimize: true,
experimental_dynamic_inputs: false,
pinned_inputs: entry.pin_inputs.clone(),
zero_fill_missing_external_data: true,
+ output_path: None,
};
let started = std::time::Instant::now();
let result = match &self.source {
Source::Dir(dir) => convert_onnx(dir.join(&entry.file), options)
.map(|_| String::new())
.map_err(|e| e.to_string()),
- Source::Hub | Source::StripDir(_) => {
- self.skeleton(&entry.file).and_then(|(bytes, note)| {
- let model = ModelProto::decode(&bytes[..])
- .map_err(|e| format!("decode skeleton: {e}"))?;
- convert_model_proto(model, &options)
- .map(|_| note)
- .map_err(|e| e.to_string())
- })
- }
+ Source::Hub | Source::StripDir(_) => self.skeleton(entry).and_then(|(bytes, note)| {
+ let model =
+ ModelProto::decode(&bytes[..]).map_err(|e| format!("decode skeleton: {e}"))?;
+ convert_model_proto(model, &options)
+ .map(|_| note)
+ .map_err(|e| e.to_string())
+ }),
};
match result {
Ok(note) => {
@@ -205,17 +185,17 @@ impl Sweep {
if matches!(self.source, Source::Dir(_)) {
return;
}
- let mut files: Vec<&str> = entries.iter().map(|(_, e)| e.file.as_str()).collect();
- files.sort_unstable();
- files.dedup();
- let queue = Mutex::new(files);
+ let mut sources: Vec<&Entry> = entries.iter().map(|(_, entry)| *entry).collect();
+ sources.sort_unstable_by_key(|entry| entry.source_key());
+ sources.dedup_by(|left, right| left.source_key() == right.source_key());
+ let queue = Mutex::new(sources);
std::thread::scope(|scope| {
for _ in 0..workers {
scope.spawn(|| loop {
- let Some(file) = queue.lock().unwrap().pop() else {
+ let Some(entry) = queue.lock().unwrap().pop() else {
break;
};
- let _ = self.skeleton(file);
+ let _ = self.skeleton(entry);
});
}
});
@@ -226,15 +206,12 @@ impl Sweep {
let queue = Mutex::new(entries);
std::thread::scope(|scope| {
for _ in 0..workers {
- std::thread::Builder::new()
- .stack_size(WORKER_STACK_BYTES)
- .spawn_scoped(scope, || loop {
- let Some((idx, entry)) = queue.lock().unwrap().pop() else {
- break;
- };
- self.convert(idx, entry);
- })
- .expect("spawn sweep worker");
+ scope.spawn(|| loop {
+ let Some((idx, entry)) = queue.lock().unwrap().pop() else {
+ break;
+ };
+ self.convert(idx, entry);
+ });
}
});
}
@@ -244,16 +221,7 @@ impl Sweep {
fn manifest_models_convert_and_build() {
// Backend build errors are only visible through `log` (RUST_LOG=error).
let _ = pretty_env_logger::try_init();
- // O2W_MANIFEST points at another manifest (e.g. candidates under evaluation).
- let manifest_path = std::env::var_os("O2W_MANIFEST")
- .filter(|v| !v.is_empty())
- .map(PathBuf::from)
- .unwrap_or_else(|| {
- Path::new(env!("CARGO_MANIFEST_DIR")).join("tests/models/manifest.json")
- });
- let manifest = std::fs::read_to_string(&manifest_path)
- .unwrap_or_else(|e| panic!("read {}: {e}", manifest_path.display()));
- let entries: Vec = serde_json::from_str(&manifest).expect("parse manifest");
+ let entries = load_manifest().unwrap_or_else(|error| panic!("{error}"));
let Some(source) = source() else {
eprintln!("skipping model sweep: set O2W_MODELS=hub, dir= or strip=");
return;
diff --git a/tests/models/ci-validation.json b/tests/models/ci-validation.json
new file mode 100644
index 0000000..3642528
--- /dev/null
+++ b/tests/models/ci-validation.json
@@ -0,0 +1,12 @@
+[
+ {
+ "file": "Xenova--tiny-random-RoFormerForMultipleChoice/onnx/model.onnx",
+ "revision": "3db6c9837f894f35ff0c26afe547bcbf24c97de0",
+ "sha256": "62128b34993971fcb7ded053c9af6932f2490724247315a112317fe398c97572",
+ "override_dims": {
+ "batch_size": 1,
+ "num_choices": 2,
+ "sequence_length": 16
+ }
+ }
+]
diff --git a/tests/models/manifest.json b/tests/models/manifest.json
index e73ae2a..64a31dd 100644
--- a/tests/models/manifest.json
+++ b/tests/models/manifest.json
@@ -27,7 +27,7 @@
{"file": "Xenova--musicgen-small/onnx/decoder_model_merged_quantized.onnx", "pin_inputs": {"use_cache_branch": 1}, "override_dims": {"batch_size": 1, "encoder_sequence_length": 16, "total_batch_size_x_num_codebooks": 8, "decoder_sequence_length": 1, "total_batch_size": 2, "past_decoder_sequence_length": 8, "encoder_sequence_length_out": 16}},
{"file": "Mozilla--distilvit/onnx/encoder_model_quantized.onnx", "override_dims": {"batch_size": 1, "num_channels": 3, "height": 224, "width": 224}},
{"file": "onnx-community--Voxtral-Mini-3B-2507-ONNX/onnx/embed_tokens_fp16.onnx", "override_dims": {"batch_size": 1, "sequence_length": 16}},
- {"file": "onnx-community--Voxtral-Mini-3B-2507-ONNX/onnx/decoder_model_merged_q4.onnx", "heavy": true, "override_dims": {"batch_size": 1, "sequence_length": 16, "total_sequence_length": 16, "past_sequence_length": 0}},
+ {"file": "onnx-community--Voxtral-Mini-3B-2507-ONNX/onnx/decoder_model_merged_q4.onnx", "heavy": true, "override_dims": {"batch_size": 1, "sequence_length": 16, "total_sequence_length": 16, "past_sequence_length": 0}, "coreml_unsupported": "CoreML MLProgram compilation overflows a native worker thread stack (SIGABRT)"},
{"file": "onnx-community--Voxtral-Mini-3B-2507-ONNX/onnx/audio_encoder_quantized.onnx", "override_dims": {"batch_size": 1}},
{"file": "Xenova--LaMini-Flan-T5-783M/onnx/encoder_model_quantized.onnx", "override_dims": {"batch_size": 1, "encoder_sequence_length": 128}},
{"file": "Xenova--LaMini-Flan-T5-783M/onnx/decoder_model_merged_quantized.onnx", "pin_inputs": {"use_cache_branch": 0}, "override_dims": {"batch_size": 1, "encoder_sequence_length": 128, "decoder_sequence_length": 16, "past_decoder_sequence_length": 0, "encoder_sequence_length_out": 128}},
diff --git a/tests/scalar_io.rs b/tests/scalar_io.rs
new file mode 100644
index 0000000..f322746
--- /dev/null
+++ b/tests/scalar_io.rs
@@ -0,0 +1,34 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+//! Regression coverage for rank-0 ONNX graph inputs and outputs.
+
+#[allow(dead_code)]
+mod common;
+
+use common::{assert_op_matches_ort, ExpectConvertOp};
+use onnx2webnn::test_models::prelude::*;
+
+#[test]
+fn scalar_input_and_output_match_ort() {
+ let model = model(
+ 17,
+ graph(
+ "scalar_identity",
+ vec![f32_input("input", &[])],
+ vec![f32_output("output", &[])],
+ vec![node(
+ "Identity",
+ "scalar_identity",
+ &["input"],
+ &["output"],
+ &[],
+ )],
+ vec![],
+ ),
+ );
+
+ assert_op_matches_ort(model, ExpectConvertOp::Success, 17);
+}