Feature gap
ONNX Runtime exposes exact node_block_list support for FP16 conversion, but WinML needs shared, opt-in plumbing for graph-derived node names and a bounded strict-type validation gate.
Required shared contract
- Add optional
WinMLQuantizationConfig.fp16_nodes_to_exclude: list[str] | None with stable declaration-order de-duplication and non-empty-string validation.
- Serialize the field only for FP16 mode and include distinct non-empty lists in cache identity; omitted and null preserve existing behavior.
- Forward the list unchanged through
FP16Pass.run as ORT node_block_list.
- Forward the same list in both normal and
disable_shape_infer=True convert_to_fp16 calls.
- Before conversion, run checker, strict shape/type inference, and checker on an isolated input model. If that succeeds, require the identical sequence on the converted model and raise
RuntimeError('FP16 conversion produced incompatible FP16 types.') on failure.
- Only
AttributeError, google.protobuf.message.EncodeError, checker.ValidationError, and shape_inference.InferenceError from the input capability preflight may skip the generic converted-model strict sequence. Existing specialized safeguards remain mandatory.
Exact graph-derived boundary names
The recipe must use exactly these 74 ONNX node.name values, in this order, as fp16_nodes_to_exclude:
InsertedPrecisionFreeCast_/deberta/embeddings/LayerNorm/LayerNormalization_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.0/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.0/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.0/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.0/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.0/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.0/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.1/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.1/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.1/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.1/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.1/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.1/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.2/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.2/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.2/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.2/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.2/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.2/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.3/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.3/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.3/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.3/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.3/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.3/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.4/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.4/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.4/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.4/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.4/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.4/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.5/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.5/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.5/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.5/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.5/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.5/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.6/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.6/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.6/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.6/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.6/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.6/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.7/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.7/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.7/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.7/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.7/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.7/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.8/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.8/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.8/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.8/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.8/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.8/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.9/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.9/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.9/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.9/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.9/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.9/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.10/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.10/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.10/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.10/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.10/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.10/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.11/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.11/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.11/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.11/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.11/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.11/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/pooler/Gather_output_0
InsertedPrecisionFreeCast_/deberta/embeddings/Cast_output_0 is deliberately absent because it is INT32-to-FLOAT, not FP16-to-FLOAT, and must convert to INT32-to-FLOAT16 under ordinary semantics.
Acceptance and evidence
- Exact-name tests prove all 74 names are unique and ordered, exist once in the fresh optimized graph, remain FLOAT boundaries in FP16, and the excluded embeddings Cast converts normally.
- Config/pass/converter tests cover absent, null, non-empty, duplicate, invalid, serialization, cache, normal/fallback forwarding, unknown-name, strict-capability, and specialized-regression behavior.
- Fresh quality evidence
producer-v3-quality-20260824T130316Z-7e682b5e is PASS: 152 focused converter tests, 426 focused config/pass tests, 1 focused recipe test, and 858 affected optimization tests passed (16 skipped, 1 xfailed), with format, lint, type, diff, and license gates complete.
- Fresh build attempt
producer-v3-artifacts-20260824T130610Z-971c84dd recorded a validation-script failure and was recovered idempotently by producer-v3-artifacts-20260824T131754Z-0d34abcb, whose model validation and artifact completion are PASS with 74 selected Casts and bounded CPU FP16 performance evidence.
Explicitly prohibited
No broad Cast op blocking, prefix/wildcard/substring matching, automatic graph heuristics, runtime synthesis or fallback exclusions, model/checkpoint conditionals, arbitrary exception catch, or suppression of specialized validation safeguards.
Planning reference: sealed planner-charter.v4.json; authorized base 3708969b731425b0c6d4b97920d1b5e6519bb013; dependency #1322. Invalid candidate 4cc9e8ee56848ffea02642294b27647e3eaeadb0 is not reusable.
Feature gap
ONNX Runtime exposes exact
node_block_listsupport for FP16 conversion, but WinML needs shared, opt-in plumbing for graph-derived node names and a bounded strict-type validation gate.Required shared contract
WinMLQuantizationConfig.fp16_nodes_to_exclude: list[str] | Nonewith stable declaration-order de-duplication and non-empty-string validation.FP16Pass.runas ORTnode_block_list.disable_shape_infer=Trueconvert_to_fp16calls.RuntimeError('FP16 conversion produced incompatible FP16 types.')on failure.AttributeError,google.protobuf.message.EncodeError,checker.ValidationError, andshape_inference.InferenceErrorfrom the input capability preflight may skip the generic converted-model strict sequence. Existing specialized safeguards remain mandatory.Exact graph-derived boundary names
The recipe must use exactly these 74 ONNX
node.namevalues, in this order, asfp16_nodes_to_exclude:InsertedPrecisionFreeCast_/deberta/embeddings/LayerNorm/LayerNormalization_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.0/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.0/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.0/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.0/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.0/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.0/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.1/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.1/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.1/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.1/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.1/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.1/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.2/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.2/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.2/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.2/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.2/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.2/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.3/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.3/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.3/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.3/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.3/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.3/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.4/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.4/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.4/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.4/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.4/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.4/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.5/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.5/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.5/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.5/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.5/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.5/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.6/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.6/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.6/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.6/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.6/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.6/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.7/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.7/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.7/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.7/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.7/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.7/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.8/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.8/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.8/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.8/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.8/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.8/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.9/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.9/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.9/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.9/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.9/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.9/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.10/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.10/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.10/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.10/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.10/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.10/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.11/attention/self/Transpose_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.11/attention/self/Reshape_1_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.11/attention/self/Reshape_3_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.11/attention/self/Transpose_8_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.11/attention/self/Reshape_5_output_0InsertedPrecisionFreeCast_/deberta/encoder/layer.11/attention/self/Reshape_13_output_0InsertedPrecisionFreeCast_/pooler/Gather_output_0InsertedPrecisionFreeCast_/deberta/embeddings/Cast_output_0is deliberately absent because it is INT32-to-FLOAT, not FP16-to-FLOAT, and must convert to INT32-to-FLOAT16 under ordinary semantics.Acceptance and evidence
producer-v3-quality-20260824T130316Z-7e682b5eis PASS: 152 focused converter tests, 426 focused config/pass tests, 1 focused recipe test, and 858 affected optimization tests passed (16 skipped, 1 xfailed), with format, lint, type, diff, and license gates complete.producer-v3-artifacts-20260824T130610Z-971c84ddrecorded a validation-script failure and was recovered idempotently byproducer-v3-artifacts-20260824T131754Z-0d34abcb, whose model validation and artifact completion are PASS with 74 selected Casts and bounded CPU FP16 performance evidence.Explicitly prohibited
No broad Cast op blocking, prefix/wildcard/substring matching, automatic graph heuristics, runtime synthesis or fallback exclusions, model/checkpoint conditionals, arbitrary exception catch, or suppression of specialized validation safeguards.
Planning reference: sealed
planner-charter.v4.json; authorized base3708969b731425b0c6d4b97920d1b5e6519bb013; dependency #1322. Invalid candidate4cc9e8ee56848ffea02642294b27647e3eaeadb0is not reusable.