recipe(opus-mt-en-ru): add verified CPU fp32/fp16 recipes

[ssss141414]

Summary

This recipe-only contribution records reproducible CPU fp32 and genuine fp16 coverage for the composite Helsinki-NLP/opus-mt-en-ru Marian English-to-Russian translation model. It ships Effort L0 / Outcome L0 through four encoder/decoder recipes; all required CPU tuples have full coverage. The frozen Goal ceiling is L2, and the highest reached verdict is L2 PASS for encoder numerical parity with the decoder constraint explicit. Translation task-metric evaluation remains supplementary CLI-BLOCKED evidence and is not promoted above the L2 ceiling.

Model metadata

Frozen model-breakdown report: temp/skill-run-1116/planner/model-breakdown-opus-mt-en-ru.json; SHA-256 7e67dca09132704e1800045b6a73df65fb3166ae707f3180d15a4782752a0522. Claims below retain the frozen evidence and confidence labels.

What the model does

A text-to-text Marian encoder-decoder checkpoint that translates normalized, SentencePiece-tokenized English text into Russian token logits and generated Russian text.

• Evidence: Hugging Face model card for Helsinki-NLP/opus-mt-en-ru at revision bb09c99d180016eac6819df3dae68edb1690fdee: source language en, target language ru, dataset OPUS, model transformer-align, preprocessing normalization + SentencePiece; Pinned config: architectures=['MarianMTModel'], model_type='marian', is_encoder_decoder=true, vocab_size=62518
• Confidence: verified

Primary user stories

• A user supplies English text to obtain a Russian translation for cross-language reading or communication.
  • Evidence: Hugging Face model card for Helsinki-NLP/opus-mt-en-ru at revision bb09c99d180016eac6819df3dae68edb1690fdee: source languages en; target languages ru; translation tag
  • Confidence: verified

Supported tasks

• translation - support surfaces: checkpoint, winml
  • Evidence: Hugging Face pipeline_tag=translation and model-card translation metadata; winml inspect: pipeline_tasks=['translation']; composite encoder=feature-extraction, decoder=text2text-generation
  • Confidence: verified
• text2text-generation - support surfaces: transformers, optimum-onnx, winml
  • Evidence: Pinned architecture MarianMTModel and is_encoder_decoder=true; Optimum Marian registry includes text2text-generation; WinML replaces the same registry cell with MarianDecoderIOConfig; winml inspect resolves task=text2text-generation from tasks-manager
  • Confidence: verified
• feature-extraction - support surfaces: optimum-onnx, winml
  • Evidence: Optimum Marian registry includes feature-extraction; WinML replaces the same registry cell with MarianEncoderIOConfig; winml inspect composite maps encoder to feature-extraction
  • Confidence: verified

Model architecture

MarianMTModel is a 76,672,000-parameter encoder-decoder Transformer with a shared 62,518x512 token embedding, learned/frozen sinusoidal positional embeddings, six 512-wide 8-head encoder layers, six 512-wide 8-head decoder layers, 2,048-wide swish feed-forward blocks, decoder cross-attention, a tied vocabulary projection, and runtime autoregressive KV-cache/generation orchestration. WinML exports it as separate encoder and decoder components.

• Component IDs: model.shared_embedding, model.encoder, model.encoder.layers[], model.decoder, model.decoder.layers[], model.lm_head, runtime.generation_and_kv_cache
• Evidence: Pinned config at bb09c99d180016eac6819df3dae68edb1690fdee: d_model=512, encoder_layers=6, decoder_layers=6, attention_heads=8, ffn_dim=2048, activation=swish, shared embeddings=true, tied embeddings=true; Transformers 4.57.6 concrete source: MarianMTModel -> MarianModel(shared, encoder, decoder) + lm_head; MarianEncoderLayer has self-attention/FFN; MarianDecoderLayer has self-attention/cross-attention/FFN; Config-only architecture probe: total_parameters=76672000; WinML inspect and fresh composite config: separate MarianEncoderWrapper and MarianDecoderWrapper components
• Confidence: verified

Validation and support evidence

1. Baseline

• Pinned main commit: 76d48ad38f894c3fc5a5e760de67f0a096e438a6
• WinML version: 0.2.0
• Build: PASS. Composite recipe-free build emitted encoder and decoder; Build complete in 28.7s and 40.5s.
  • Command: uv run winml build -m Helsinki-NLP/opus-mt-en-ru -o temp/skill-run-1116/baseline_opus_mt_en_ru --ep cpu --device cpu --no-analyze --no-optimize --no-quant --no-compile --rebuild
  • Log/output evidence: temp/skill-run-1116/planner/baseline_build_utf8.log; encoder Build complete in 28.7s; decoder Build complete in 40.5s.
• Perf: PASS. CPU fp32 encoder avg/P50=74.91/74.51ms, throughput=13.35/s, RAM total delta=139.7MB; decoder avg/P50=22.98/22.90ms, throughput=43.51/s, RAM total delta=249.0MB.
  • Command: uv run winml perf -m temp/skill-run-1116/baseline_opus_mt_en_ru/encoder_model.onnx --ep cpu --device cpu ; uv run winml perf -m temp/skill-run-1116/baseline_opus_mt_en_ru/decoder_model.onnx --ep cpu --device cpu
  • Scope: These are component/first-token benchmarks, not full autoregressive English-to-Russian translation latency.
• Eval floor: CLI-BLOCKED, exit 2: Task 'translation' is not supported by winml eval.
  • Command: uv run winml eval --schema --task translation
  • Log: temp/skill-run-1116/planner/baseline_eval_translation_utf8.log
• Starting auto-config recipe: temp/skill-run-1116/starting_config_opus_mt_en_ru_{encoder,decoder}
  • Command: uv run winml config -m Helsinki-NLP/opus-mt-en-ru --task translation -o temp/skill-run-1116/starting_config_opus_mt_en_ru
  • Result: PR fp32 encoder/decoder are semantically identical; fp16 encoder/decoder differ only at quant.mode=fp16 configuration.
• Build/perf/eval floor: L1; baseline proves L0 build and L1 component perf, while translation eval is CLI-BLOCKED and retained only as supplementary evidence.
• Optimum probe: VENDOR+OVERRIDE. Vendor tasks: feature-extraction, feature-extraction-with-past, text-generation, text-generation-with-past, text2text-generation, text2text-generation-with-past; after WinML: feature-extraction, feature-extraction-with-past, text-generation, text-generation-with-past, text2text-generation, text2text-generation-with-past; added by WinML: none.

2. Goal

• Committed Effort: L0
• Committed Goal ceiling: L2
• Committed Outcome: L0
• Success definition: fresh CPU fp32/fp16 encoder+decoder L0 and L1 plus meaningful encoder PyTorch parity at L2; decoder parity constraint explicit
• Charter revision 1 froze L2 from the outset; no ceiling change or re-issued charter occurred.

3. Outcome

• Shipped tier: L0
• Highest reached Goal verdict: L2 PASS (encoder numerical scope; decoder constraint explicit)
• Coverage: full
• Deferred tuples: none
• Shipped recipe paths:
  • examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp32_encoder_config.json
  • examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp32_decoder_config.json
  • examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp16_encoder_config.json
  • examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp16_decoder_config.json
• Shipped code paths: none
• Appended model finding: marian-008 in model_knowledge/marian.json.
• Appended methodology findings: none.
• No methodology friction observed.

4. Per-EP/device/precision results including concrete perf and eval data/exact blockers

Goal ladder

Tier	Verdict	Evidence
L0	PASS	all two precision tuples and four component artifacts independently rebuilt and structurally validated
L1	PASS	all four component artifacts ran on CPU with concrete latency/throughput/RAM
L2	PASS	real tokenized named-input PyTorch parity for both encoder precisions; decoder parity is explicitly constrained and not claimed

L0 build and structure

EP / device	Precision	Component	Verdict	Build	Structural evidence
cpu / cpu	fp32	encoder	PASS	24.2 s	IR 8; opset 17; initializers FLOAT=102, INT64=32; external data adjacent=true; artifact temp/skill-run-1116/tester/build/cpu_cpu_fp32_encoder/model.onnx
cpu / cpu	fp32	decoder	PASS	40.8 s	IR 8; opset 17; initializers FLOAT=166, INT64=43; external data adjacent=true; artifact temp/skill-run-1116/tester/build/cpu_cpu_fp32_decoder/model.onnx
cpu / cpu	fp16	encoder	PASS	28.3 s	IR 8; opset 17; initializers FLOAT16=102, INT64=32; external data adjacent=true; artifact temp/skill-run-1116/tester/build/cpu_cpu_fp16_encoder/model.onnx
cpu / cpu	fp16	decoder	PASS	48.9 s	IR 8; opset 17; initializers FLOAT16=166, INT64=43; external data adjacent=true; artifact temp/skill-run-1116/tester/build/cpu_cpu_fp16_decoder/model.onnx

L1 perf

These are component measurements, not end-to-end autoregressive translation latency. VRAM delta was not reported by the frozen CPU evidence.

EP / device	Precision	Component	Verdict	Mean	p50	Throughput	RAM load	RAM inference	RAM total delta	VRAM delta	Scope
cpu / cpu	fp32	encoder	PASS	74.79 ms	74.71 ms	13.37 samples/s	79.3 MB	60.2 MB	139.6 MB	not reported	single encoder component
cpu / cpu	fp32	decoder	PASS	23.69 ms	23.61 ms	42.22 samples/s	241.0 MB	7.8 MB	248.8 MB	not reported	single decoder step with synthetic named inputs; not full autoregressive translation
cpu / cpu	fp16	encoder	PASS	95.37 ms	95.42 ms	10.49 samples/s	218.0 MB	78.6 MB	296.7 MB	not reported	single encoder component
cpu / cpu	fp16	decoder	PASS	28.14 ms	27.9 ms	35.53 samples/s	245.1 MB	14.2 MB	259.4 MB	not reported	single decoder step with synthetic named inputs; not full autoregressive translation

L2 numerical parity

EP / device	Precision	Component	Verdict	Reference/input	Cosine	Max abs	Mean abs	RMSE
cpu / cpu	fp32	encoder	PASS	transformers MarianMTModel.get_encoder() @ bb09c99d180016eac6819df3dae68edb1690fdee; text The quick brown fox jumps over the lazy dog.; shape [1, 512, 512]	0.9999999999994811	6.198883056640625e-06	2.7017895465331065e-07	3.5698963661937196e-07
cpu / cpu	fp32	decoder	CONSTRAINT-DOCUMENTED-NOT-CLAIMED	The ONNX decoder is a static-cache single-token wrapper with 12 full [1,8,512,64] past-KV inputs plus cache_position and 12 captured present-KV outputs. Meaningful PT parity requires constructing an identical WinMLStaticCache state and comparing the same decoder step or full generation loop; zero-KV ONNX versus stock PT prefill is not apples-to-apples. No decoder numerical parity is claimed.	not claimed	not claimed	not claimed	not claimed
cpu / cpu	fp16	encoder	PASS	transformers MarianMTModel.get_encoder() @ bb09c99d180016eac6819df3dae68edb1690fdee; text The quick brown fox jumps over the lazy dog.; shape [1, 512, 512]	0.9999989143104122	0.0061779022216796875	0.0003867089235389254	0.0005159687928534306
cpu / cpu	fp16	decoder	CONSTRAINT-DOCUMENTED-NOT-CLAIMED	The ONNX decoder is a static-cache single-token wrapper with 12 full [1,8,512,64] past-KV inputs plus cache_position and 12 captured present-KV outputs. Meaningful PT parity requires constructing an identical WinMLStaticCache state and comparing the same decoder step or full generation loop; zero-KV ONNX versus stock PT prefill is not apples-to-apples. No decoder numerical parity is claimed.	not claimed	not claimed	not claimed	not claimed

Supplementary eval evidence

Because the committed ceiling is L2, this eval evidence is supplementary and does not create or imply an L3 verdict.

EP / device	Precision	Component	Verdict	Dataset / revision / subset	Metric	Exit	Exact blocker
cpu / cpu	fp32	encoder	CLI-BLOCKED	not produced because CLI-BLOCKED	not produced because CLI-BLOCKED	1	Evaluation failed: Task 'translation' is not supported. Supported tasks: compare-tensor, depth-estimation, feature-extraction, fill-mask, image-classification, image-feature-extraction, image-segmentation, image-to-text, keypoint-detection, mask-generation, next-sentence-prediction, object-detection, question-answering, sentence-similarity, sequence-classification, text-classification, token-classification, zero-shot-classification, zero-shot-image-classification. Feature gap: temp/skill-run-1116/tester/feature-gap-translation-eval.json.
cpu / cpu	fp32	decoder	CLI-BLOCKED	not produced because CLI-BLOCKED	not produced because CLI-BLOCKED	1	Evaluation failed: Task 'translation' is not supported. Supported tasks: compare-tensor, depth-estimation, feature-extraction, fill-mask, image-classification, image-feature-extraction, image-segmentation, image-to-text, keypoint-detection, mask-generation, next-sentence-prediction, object-detection, question-answering, sentence-similarity, sequence-classification, text-classification, token-classification, zero-shot-classification, zero-shot-image-classification. Feature gap: temp/skill-run-1116/tester/feature-gap-translation-eval.json.

• cpu/cpu/fp32 schema probe: exit 2; Task 'translation' is not supported by winml eval. Supported tasks: depth-estimation, feature-extraction, fill-mask, image-classification, image-feature-extraction, image-segmentation, image-to-text, keypoint-detection, mask-generation, next-sentence-prediction, object-detection, question-answering, sentence-similarity, sequence-classification, text-classification, token-classification, zero-shot-classification, zero-shot-image-classification.
  | cpu / cpu | fp16 | encoder | CLI-BLOCKED | not produced because CLI-BLOCKED | not produced because CLI-BLOCKED | 1 | Evaluation failed: Task 'translation' is not supported. Supported tasks: compare-tensor, depth-estimation, feature-extraction, fill-mask, image-classification, image-feature-extraction, image-segmentation, image-to-text, keypoint-detection, mask-generation, next-sentence-prediction, object-detection, question-answering, sentence-similarity, sequence-classification, text-classification, token-classification, zero-shot-classification, zero-shot-image-classification. Feature gap: temp/skill-run-1116/tester/feature-gap-translation-eval.json. |
  | cpu / cpu | fp16 | decoder | CLI-BLOCKED | not produced because CLI-BLOCKED | not produced because CLI-BLOCKED | 1 | Evaluation failed: Task 'translation' is not supported. Supported tasks: compare-tensor, depth-estimation, feature-extraction, fill-mask, image-classification, image-feature-extraction, image-segmentation, image-to-text, keypoint-detection, mask-generation, next-sentence-prediction, object-detection, question-answering, sentence-similarity, sequence-classification, text-classification, token-classification, zero-shot-classification, zero-shot-image-classification. Feature gap: temp/skill-run-1116/tester/feature-gap-translation-eval.json. |
• cpu/cpu/fp16 schema probe: exit 2; Task 'translation' is not supported by winml eval. Supported tasks: depth-estimation, feature-extraction, fill-mask, image-classification, image-feature-extraction, image-segmentation, image-to-text, keypoint-detection, mask-generation, next-sentence-prediction, object-detection, question-answering, sentence-similarity, sequence-classification, text-classification, token-classification, zero-shot-classification, zero-shot-image-classification.

5. Delta

• Baseline recipe: temp/skill-run-1116/starting_config_opus_mt_en_ru_{encoder,decoder}
• Overall comparison: CHANGED. Producer semantic JSON comparison in temp/skill-run-1116/producer/recipe_verification.json: fp32 encoder/decoder are identical to fresh configs; fp16 encoder/decoder differ only at /quant.
  • translation_fp32_encoder_config.json: IDENTICAL
  • translation_fp32_decoder_config.json: IDENTICAL
  • translation_fp16_encoder_config.json: CHANGED_ONLY_AT_/quant
  • translation_fp16_decoder_config.json: CHANGED_ONLY_AT_/quant
• examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp16_encoder_config.json at JSON pointer /quant: old value null; new value {"mode":"fp16","samples":10,"calibration_method":"minmax","weight_type":"uint8","activation_type":"uint8","per_channel":false,"symmetric":false,"weight_symmetric":null,"activation_symmetric":null,"save_calibration":false,"distribution":"uniform","seed":null,"calibration_load_path":null,"calibration_save_path":null,"op_types_to_quantize":null,"nodes_to_exclude":null,"task":"feature-extraction","model_id":"Helsinki-NLP/opus-mt-en-ru","model_type":"marian","fp16_keep_io_types":true,"fp16_op_block_list":null}. Reason: Charter-required explicit fp16 coverage; build also passed --precision fp16 and artifact inspection found genuine FLOAT16 initializers.
• examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp16_decoder_config.json at JSON pointer /quant: old value null; new value {"mode":"fp16","samples":10,"calibration_method":"minmax","weight_type":"uint8","activation_type":"uint8","per_channel":false,"symmetric":false,"weight_symmetric":null,"activation_symmetric":null,"save_calibration":false,"distribution":"uniform","seed":null,"calibration_load_path":null,"calibration_save_path":null,"op_types_to_quantize":null,"nodes_to_exclude":null,"task":"text2text-generation","model_id":"Helsinki-NLP/opus-mt-en-ru","model_type":"marian","fp16_keep_io_types":true,"fp16_op_block_list":null}. Reason: Charter-required explicit fp16 coverage; build also passed --precision fp16 and artifact inspection found genuine FLOAT16 initializers.
• Code paths/symbols and class-wide behavior changes: none; no code changes and no class-wide behavior changes.
• Reducibility consistent with charter: true.
• Recipe-free acceptance was required and is PASS: Frozen planner baseline at current origin/main 76d48ad; temp/skill-run-1116/planner/baseline_build_utf8.log.
  • Command: uv run winml build -m Helsinki-NLP/opus-mt-en-ru -o temp/skill-run-1116/baseline_opus_mt_en_ru --ep cpu --device cpu --no-analyze --no-optimize --no-quant --no-compile --rebuild
• The production recipe README remains untouched.

6. Analyze results containing both full component-level and op-level analysis

Overall status: ANALYZE-PARTIAL-SUCCESS. All four commands exited 1 because support is partial/unknown, not setup failure. Each emitted parseable JSON with metadata and six complete requested EP result rows. No exit 2 setup failure occurred. Rules environment: WINMLCLI_RULES_DIR=C:\repo\ModelKitArtifacts\rules; forbidden WINML_RULES_PATH used=false. Static analyze is rule-based graph analysis and is not runtime execution evidence.

Component-level analysis

fp32 encoder

• Status: ANALYZE-PARTIAL-SUCCESS; exit code 1; artifact temp/skill-run-1116/tester/analyze_cpu_cpu_fp32_encoder.json; SHA-256 8e046ffc826fe555e1633e01fa7941e18383121b1a19972d299064b154d16a23.
• Command: $env:WINMLCLI_RULES_DIR='C:\repo\ModelKitArtifacts\rules'; uv run winml analyze --model temp/skill-run-1116/tester/build/cpu_cpu_fp32_encoder/model.onnx --ep all --output temp/skill-run-1116/tester/analyze_cpu_cpu_fp32_encoder.json
• Exact error/status: exit 1 denotes partial operator coverage; emitted JSON is complete for every requested EP and is parsed fail-closed
• Mapping totals: mapped=150; partial=54; unmapped=0.

Semantic component	ONNX regions or nodes	Nodes	Operator counts	Mapping basis	Confidence	Per-EP partial/unsupported/unknown issues
model.encoder	/encoder/*, /encoder/Reshape, /encoder/Mul, /encoder/Add	10	Add=1, Cast=2, Expand=1, Mul=1, Reshape=1, Sub=1, Unsqueeze=2, Where=1	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.layers[0]	/encoder/layers.N/*, /encoder/layers.0/self_attn/k_proj/MatMul/MatMulAddFusion, /encoder/layers.0/self_attn/Transpose_2, /encoder/layers.0/self_attn/Mul_1	24	Add=3, Gemm=6, LayerNormalization=2, MatMul=2, Mul=3, Reshape=1, Sigmoid=1, Slice=1, Softmax=1, Transpose=4	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.layers[1]	/encoder/layers.N/*, /encoder/layers.1/self_attn/k_proj/MatMul/MatMulAddFusion, /encoder/layers.1/self_attn/Transpose_2, /encoder/layers.1/self_attn/Mul_1	23	Add=3, Gemm=6, LayerNormalization=2, MatMul=2, Mul=3, Reshape=1, Sigmoid=1, Softmax=1, Transpose=4	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.layers[2]	/encoder/layers.N/*, /encoder/layers.2/self_attn/k_proj/MatMul/MatMulAddFusion, /encoder/layers.2/self_attn/Transpose_2, /encoder/layers.2/self_attn/Mul_1	23	Add=3, Gemm=6, LayerNormalization=2, MatMul=2, Mul=3, Reshape=1, Sigmoid=1, Softmax=1, Transpose=4	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.layers[3]	/encoder/layers.N/*, /encoder/layers.3/self_attn/k_proj/MatMul/MatMulAddFusion, /encoder/layers.3/self_attn/Transpose_2, /encoder/layers.3/self_attn/Mul_1	23	Add=3, Gemm=6, LayerNormalization=2, MatMul=2, Mul=3, Reshape=1, Sigmoid=1, Softmax=1, Transpose=4	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.layers[4]	/encoder/layers.N/*, /encoder/layers.4/self_attn/k_proj/MatMul/MatMulAddFusion, /encoder/layers.4/self_attn/Transpose_2, /encoder/layers.4/self_attn/Mul_1	23	Add=3, Gemm=6, LayerNormalization=2, MatMul=2, Mul=3, Reshape=1, Sigmoid=1, Softmax=1, Transpose=4	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.layers[5]	/encoder/layers.N/*, /encoder/layers.5/self_attn/k_proj/MatMul/MatMulAddFusion, /encoder/layers.5/self_attn/Transpose_2, /encoder/layers.5/self_attn/Mul_1	23	Add=3, Gemm=6, LayerNormalization=2, MatMul=2, Mul=3, Reshape=1, Sigmoid=1, Softmax=1, Transpose=4	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.optimizer_inserted	/encoder/*, gemm_input_reshape_token_8, gemm_output_reshape_token_11_new_reshape, gemm_output_reshape_token_5_new_reshape	54	Reshape=54	exported graph ownership only; optimizer-generated node names lack module scope	partial	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.shared_embedding	/encoder/*, /encoder/embed_tokens/Gather	1	Gather=1	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown

Mapping gaps:

• Optimizer-inserted Reshape nodes without hierarchy scope map only to the exported encoder/decoder component.
• Runtime generation/KV orchestration is outside each ONNX graph.
• Shared embedding weights are duplicated into separate exported artifacts; mapping reflects graph execution, not storage ownership.

fp32 decoder

• Status: ANALYZE-PARTIAL-SUCCESS; exit code 1; artifact temp/skill-run-1116/tester/analyze_cpu_cpu_fp32_decoder.json; SHA-256 62846ffbf0ae2a68033c932d79bff956227c32c9f067952e95a42c90816bf055.
• Command: $env:WINMLCLI_RULES_DIR='C:\repo\ModelKitArtifacts\rules'; uv run winml analyze --model temp/skill-run-1116/tester/build/cpu_cpu_fp32_decoder/model.onnx --ep all --output temp/skill-run-1116/tester/analyze_cpu_cpu_fp32_decoder.json
• Exact error/status: exit 1 denotes partial operator coverage; emitted JSON is complete for every requested EP and is parsed fail-closed
• Mapping totals: mapped=278; partial=114; unmapped=0.

Semantic component	ONNX regions or nodes	Nodes	Operator counts	Mapping basis	Confidence	Per-EP partial/unsupported/unknown issues
model.decoder.layers[0]	/decoder/layers.N/*, /model/model/decoder/layers.0/encoder_attn/v_proj/MatMul/MatMulAddFusion, /model/model/decoder/layers.0/encoder_attn/Transpose_1, /model/model/decoder/layers.0/encoder_attn/Slice	50	Add=5, Concat=1, Expand=1, Gemm=10, LayerNormalization=3, MatMul=4, Mul=5, Reshape=4, ScatterND=2, Sigmoid=1, Slice=2, Softmax=2, Transpose=9, Unsqueeze=1	hierarchy-tagged ONNX node scope	mapped	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: partial, unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.decoder.layers[1]	/decoder/layers.N/*, /model/model/decoder/layers.1/encoder_attn/v_proj/MatMul/MatMulAddFusion, /model/model/decoder/layers.1/encoder_attn/Transpose_1, /model/model/decoder/layers.1/self_attn/v_proj/MatMul/MatMulAddFusion	45	Add=5, Gemm=10, LayerNormalization=3, MatMul=4, Mul=5, Reshape=4, ScatterND=2, Sigmoid=1, Softmax=2, Transpose=9	hierarchy-tagged ONNX node scope	mapped	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.decoder.layers[2]	/decoder/layers.N/*, /model/model/decoder/layers.2/encoder_attn/v_proj/MatMul/MatMulAddFusion, /model/model/decoder/layers.2/encoder_attn/Transpose_1, /model/model/decoder/layers.2/self_attn/v_proj/MatMul/MatMulAddFusion	45	Add=5, Gemm=10, LayerNormalization=3, MatMul=4, Mul=5, Reshape=4, ScatterND=2, Sigmoid=1, Softmax=2, Transpose=9	hierarchy-tagged ONNX node scope	mapped	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.decoder.layers[3]	/decoder/layers.N/*, /model/model/decoder/layers.3/encoder_attn/v_proj/MatMul/MatMulAddFusion, /model/model/decoder/layers.3/encoder_attn/Transpose_1, /model/model/decoder/layers.3/self_attn/v_proj/MatMul/MatMulAddFusion	45	Add=5, Gemm=10, LayerNormalization=3, MatMul=4, Mul=5, Reshape=4, ScatterND=2, Sigmoid=1, Softmax=2, Transpose=9	hierarchy-tagged ONNX node scope	mapped	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.decoder.layers[4]	/decoder/layers.N/*, /model/model/decoder/layers.4/encoder_attn/v_proj/MatMul/MatMulAddFusion, /model/model/decoder/layers.4/encoder_attn/Transpose_1, /model/model/decoder/layers.4/self_attn/v_proj/MatMul/MatMulAddFusion	45	Add=5, Gemm=10, LayerNormalization=3, MatMul=4, Mul=5, Reshape=4, ScatterND=2, Sigmoid=1, Softmax=2, Transpose=9	hierarchy-tagged ONNX node scope	mapped	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.decoder.layers[5]	/decoder/layers.N/*, /model/model/decoder/layers.5/encoder_attn/v_proj/MatMul/MatMulAddFusion, /model/model/decoder/layers.5/encoder_attn/Transpose_1, /model/model/decoder/layers.5/self_attn/Reshape_3	47	Add=5, Expand=1, Gemm=10, LayerNormalization=3, MatMul=4, Mul=5, Reshape=5, ScatterND=2, Sigmoid=1, Softmax=2, Transpose=9	hierarchy-tagged ONNX node scope	mapped	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: partial, unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.decoder.optimizer_inserted	/decoder/*, gemm_input_reshape_token_212, gemm_output_reshape_token_215_new_reshape, gemm_output_reshape_token_185_new_reshape	114	Add=3, Cast=4, Equal=1, Expand=3, Gather=2, Greater=1, Mul=2, Reshape=87, ScatterND=1, Slice=1, Sub=1, Unsqueeze=6, Where=2	exported graph ownership only; optimizer-generated node names lack module scope	partial	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: partial, unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.lm_head	/decoder/*, /model/lm_head/MatMul	1	MatMul=1	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown

Mapping gaps:

• Optimizer-inserted Reshape nodes without hierarchy scope map only to the exported encoder/decoder component.
• Runtime generation/KV orchestration is outside each ONNX graph.
• Shared embedding weights are duplicated into separate exported artifacts; mapping reflects graph execution, not storage ownership.

fp16 encoder

• Status: ANALYZE-PARTIAL-SUCCESS; exit code 1; artifact temp/skill-run-1116/tester/analyze_cpu_cpu_fp16_encoder.json; SHA-256 a4801b73f962b7fc75bc6b236955b2d1add0cbd998d0dc814be1af64da6a7e07.
• Command: $env:WINMLCLI_RULES_DIR='C:\repo\ModelKitArtifacts\rules'; uv run winml analyze --model temp/skill-run-1116/tester/build/cpu_cpu_fp16_encoder/model.onnx --ep all --output temp/skill-run-1116/tester/analyze_cpu_cpu_fp16_encoder.json
• Exact error/status: exit 1 denotes partial operator coverage; emitted JSON is complete for every requested EP and is parsed fail-closed
• Mapping totals: mapped=150; partial=55; unmapped=0.

Semantic component	ONNX regions or nodes	Nodes	Operator counts	Mapping basis	Confidence	Per-EP partial/unsupported/unknown issues
model.encoder	/encoder/*, /encoder/Reshape, /encoder/Mul, /encoder/Add	10	Add=1, Cast=2, Expand=1, Mul=1, Reshape=1, Sub=1, Unsqueeze=2, Where=1	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.layers[0]	/encoder/layers.N/*, /encoder/layers.0/self_attn/k_proj/MatMul/MatMulAddFusion, /encoder/layers.0/self_attn/Transpose_2, /encoder/layers.0/self_attn/Mul_1	24	Add=3, Gemm=6, LayerNormalization=2, MatMul=2, Mul=3, Reshape=1, Sigmoid=1, Slice=1, Softmax=1, Transpose=4	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.layers[1]	/encoder/layers.N/*, /encoder/layers.1/self_attn/k_proj/MatMul/MatMulAddFusion, /encoder/layers.1/self_attn/Transpose_2, /encoder/layers.1/self_attn/Mul_1	23	Add=3, Gemm=6, LayerNormalization=2, MatMul=2, Mul=3, Reshape=1, Sigmoid=1, Softmax=1, Transpose=4	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.layers[2]	/encoder/layers.N/*, /encoder/layers.2/self_attn/k_proj/MatMul/MatMulAddFusion, /encoder/layers.2/self_attn/Transpose_2, /encoder/layers.2/self_attn/Mul_1	23	Add=3, Gemm=6, LayerNormalization=2, MatMul=2, Mul=3, Reshape=1, Sigmoid=1, Softmax=1, Transpose=4	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.layers[3]	/encoder/layers.N/*, /encoder/layers.3/self_attn/k_proj/MatMul/MatMulAddFusion, /encoder/layers.3/self_attn/Transpose_2, /encoder/layers.3/self_attn/Mul_1	23	Add=3, Gemm=6, LayerNormalization=2, MatMul=2, Mul=3, Reshape=1, Sigmoid=1, Softmax=1, Transpose=4	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.layers[4]	/encoder/layers.N/*, /encoder/layers.4/self_attn/k_proj/MatMul/MatMulAddFusion, /encoder/layers.4/self_attn/Transpose_2, /encoder/layers.4/self_attn/Mul_1	23	Add=3, Gemm=6, LayerNormalization=2, MatMul=2, Mul=3, Reshape=1, Sigmoid=1, Softmax=1, Transpose=4	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.layers[5]	/encoder/layers.N/*, /encoder/layers.5/self_attn/k_proj/MatMul/MatMulAddFusion, /encoder/layers.5/self_attn/Transpose_2, /encoder/layers.5/self_attn/Mul_1	23	Add=3, Gemm=6, LayerNormalization=2, MatMul=2, Mul=3, Reshape=1, Sigmoid=1, Softmax=1, Transpose=4	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.encoder.optimizer_inserted	/encoder/*, gemm_input_reshape_token_8, gemm_output_reshape_token_11_new_reshape, gemm_output_reshape_token_5_new_reshape	55	Cast=1, Reshape=54	exported graph ownership only; optimizer-generated node names lack module scope	partial	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown
model.shared_embedding	/encoder/*, /encoder/embed_tokens/Gather	1	Gather=1	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown

Mapping gaps:

• Optimizer-inserted Reshape nodes without hierarchy scope map only to the exported encoder/decoder component.
• Runtime generation/KV orchestration is outside each ONNX graph.
• Shared embedding weights are duplicated into separate exported artifacts; mapping reflects graph execution, not storage ownership.

fp16 decoder

• Status: ANALYZE-PARTIAL-SUCCESS; exit code 1; artifact temp/skill-run-1116/tester/analyze_cpu_cpu_fp16_decoder.json; SHA-256 7e4c2cc7088bc463f01b9f3dd88570d4cfee9dfb996f27b39c3b6da8625a3398.
• Command: $env:WINMLCLI_RULES_DIR='C:\repo\ModelKitArtifacts\rules'; uv run winml analyze --model temp/skill-run-1116/tester/build/cpu_cpu_fp16_decoder/model.onnx --ep all --output temp/skill-run-1116/tester/analyze_cpu_cpu_fp16_decoder.json
• Exact error/status: exit 1 denotes partial operator coverage; emitted JSON is complete for every requested EP and is parsed fail-closed
• Mapping totals: mapped=278; partial=140; unmapped=0.

Semantic component	ONNX regions or nodes	Nodes	Operator counts	Mapping basis	Confidence	Per-EP partial/unsupported/unknown issues
model.decoder.layers[0]	/decoder/layers.N/*, /model/model/decoder/layers.0/encoder_attn/Slice, /model/model/decoder/layers.0/self_attn/v_proj/MatMul/MatMulAddFusion, /model/model/decoder/layers.0/self_attn/Transpose_2	50	Add=5, Concat=1, Expand=1, Gemm=10, LayerNormalization=3, MatMul=4, Mul=5, Reshape=4, ScatterND=2, Sigmoid=1, Slice=2, Softmax=2, Transpose=9, Unsqueeze=1	hierarchy-tagged ONNX node scope	mapped	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: partial, unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.decoder.layers[1]	/decoder/layers.N/*, /model/model/decoder/layers.1/encoder_attn/v_proj/MatMul/MatMulAddFusion, /model/model/decoder/layers.1/encoder_attn/Transpose_1, /model/model/decoder/layers.1/self_attn/v_proj/MatMul/MatMulAddFusion	45	Add=5, Gemm=10, LayerNormalization=3, MatMul=4, Mul=5, Reshape=4, ScatterND=2, Sigmoid=1, Softmax=2, Transpose=9	hierarchy-tagged ONNX node scope	mapped	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.decoder.layers[2]	/decoder/layers.N/*, /model/model/decoder/layers.2/encoder_attn/v_proj/MatMul/MatMulAddFusion, /model/model/decoder/layers.2/encoder_attn/Transpose_1, /model/model/decoder/layers.2/self_attn/v_proj/MatMul/MatMulAddFusion	45	Add=5, Gemm=10, LayerNormalization=3, MatMul=4, Mul=5, Reshape=4, ScatterND=2, Sigmoid=1, Softmax=2, Transpose=9	hierarchy-tagged ONNX node scope	mapped	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.decoder.layers[3]	/decoder/layers.N/*, /model/model/decoder/layers.3/encoder_attn/v_proj/MatMul/MatMulAddFusion, /model/model/decoder/layers.3/encoder_attn/Transpose_1, /model/model/decoder/layers.3/self_attn/v_proj/MatMul/MatMulAddFusion	45	Add=5, Gemm=10, LayerNormalization=3, MatMul=4, Mul=5, Reshape=4, ScatterND=2, Sigmoid=1, Softmax=2, Transpose=9	hierarchy-tagged ONNX node scope	mapped	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.decoder.layers[4]	/decoder/layers.N/*, /model/model/decoder/layers.4/encoder_attn/v_proj/MatMul/MatMulAddFusion, /model/model/decoder/layers.4/encoder_attn/Transpose_1, /model/model/decoder/layers.4/self_attn/v_proj/MatMul/MatMulAddFusion	45	Add=5, Gemm=10, LayerNormalization=3, MatMul=4, Mul=5, Reshape=4, ScatterND=2, Sigmoid=1, Softmax=2, Transpose=9	hierarchy-tagged ONNX node scope	mapped	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.decoder.layers[5]	/decoder/layers.N/*, /model/model/decoder/layers.5/self_attn/Reshape_3, /model/model/decoder/layers.5/self_attn/Expand_2, /model/model/decoder/layers.5/encoder_attn/v_proj/MatMul/MatMulAddFusion	47	Add=5, Expand=1, Gemm=10, LayerNormalization=3, MatMul=4, Mul=5, Reshape=5, ScatterND=2, Sigmoid=1, Softmax=2, Transpose=9	hierarchy-tagged ONNX node scope	mapped	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: partial, unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.decoder.optimizer_inserted	/decoder/*, /model/model/decoder/Unsqueeze_8, /model/model/decoder/Unsqueeze_9, /model/model/decoder/Expand_6	140	Add=3, Cast=30, Equal=1, Expand=3, Gather=2, Greater=1, Mul=2, Reshape=87, ScatterND=1, Slice=1, Sub=1, Unsqueeze=6, Where=2	exported graph ownership only; optimizer-generated node names lack module scope	partial	NvTensorRTRTXExecutionProvider: unknown; CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; QNNExecutionProvider: partial, unsupported, unknown; OpenVINOExecutionProvider: unknown; DmlExecutionProvider: unknown
model.lm_head	/decoder/*, /model/lm_head/MatMul	1	MatMul=1	hierarchy-tagged ONNX node scope	mapped	CUDAExecutionProvider: unknown; MIGraphXExecutionProvider: unknown; DmlExecutionProvider: unknown

Mapping gaps:

• Optimizer-inserted Reshape nodes without hierarchy scope map only to the exported encoder/decoder component.
• Runtime generation/KV orchestration is outside each ONNX graph.
• Shared embedding weights are duplicated into separate exported artifacts; mapping reflects graph execution, not storage ownership.

Op-level analysis

fp32 encoder

• Total operators: 204; unique operator types: 16.
• Operator counts: Reshape=61, Gather=1, Mul=19, Add=19, Gemm=36, Transpose=24, MatMul=12, Unsqueeze=2, Expand=1, Cast=2, Sub=1, Where=1, Slice=1, Softmax=6, LayerNormalization=12, Sigmoid=6.

EP	Runtime support	Supported	Partial	Unsupported	Unknown
NvTensorRTRTXExecutionProvider	true	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid	none	none	none
CUDAExecutionProvider	false	none	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid
MIGraphXExecutionProvider	false	none	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid
QNNExecutionProvider	true	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid	none	none	none
OpenVINOExecutionProvider	true	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid	none	none	none
DmlExecutionProvider	false	none	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid

fp32 decoder

• Total operators: 392; unique operator types: 20.
• Operator counts: Reshape=112, Gemm=60, Transpose=54, Unsqueeze=7, Expand=5, Cast=4, Sub=1, Where=2, Slice=3, Gather=2, Mul=32, Add=33, Concat=1, ScatterND=13, Greater=1, Equal=1, MatMul=25, Softmax=12, LayerNormalization=18, Sigmoid=6.

EP	Runtime support	Supported	Partial	Unsupported	Unknown
NvTensorRTRTXExecutionProvider	false	Reshape, Gemm, Transpose, Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Concat, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid	none	none	ScatterND
CUDAExecutionProvider	false	none	none	none	Reshape, Gemm, Transpose, Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Concat, ScatterND, Greater, Equal, MatMul, Softmax, LayerNormalization, Sigmoid
MIGraphXExecutionProvider	false	none	none	none	Reshape, Gemm, Transpose, Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Concat, ScatterND, Greater, Equal, MatMul, Softmax, LayerNormalization, Sigmoid
QNNExecutionProvider	false	Reshape, Gemm, Transpose, Unsqueeze, Sub, Cast, Where, Slice, Gather, Mul, Add, Expand, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid	Expand, Cast, Concat	none	ScatterND
OpenVINOExecutionProvider	false	Reshape, Gemm, Transpose, Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Concat, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid	none	none	ScatterND
DmlExecutionProvider	false	none	none	none	Reshape, Gemm, Transpose, Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Concat, ScatterND, Greater, Equal, MatMul, Softmax, LayerNormalization, Sigmoid

fp16 encoder

• Total operators: 205; unique operator types: 16.
• Operator counts: Reshape=61, Gather=1, Mul=19, Add=19, Gemm=36, Transpose=24, MatMul=12, Unsqueeze=2, Expand=1, Cast=3, Sub=1, Where=1, Slice=1, Softmax=6, LayerNormalization=12, Sigmoid=6.

EP	Runtime support	Supported	Partial	Unsupported	Unknown
NvTensorRTRTXExecutionProvider	true	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid	none	none	none
CUDAExecutionProvider	false	none	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid
MIGraphXExecutionProvider	false	none	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid
QNNExecutionProvider	true	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid	none	none	none
OpenVINOExecutionProvider	true	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid	none	none	none
DmlExecutionProvider	false	none	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid

fp16 decoder

• Total operators: 418; unique operator types: 20.
• Operator counts: Unsqueeze=7, Expand=5, Cast=30, Sub=1, Where=2, Slice=3, Gather=2, Mul=32, Add=33, Reshape=112, Gemm=60, Transpose=54, Concat=1, Greater=1, Equal=1, ScatterND=13, MatMul=25, Softmax=12, LayerNormalization=18, Sigmoid=6.

EP	Runtime support	Supported	Partial	Unsupported	Unknown
NvTensorRTRTXExecutionProvider	false	Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Reshape, Gemm, Transpose, Concat, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid	none	none	ScatterND
CUDAExecutionProvider	false	none	none	none	Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Reshape, Gemm, Transpose, Concat, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid
MIGraphXExecutionProvider	false	none	none	none	Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Reshape, Gemm, Transpose, Concat, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid
QNNExecutionProvider	false	Unsqueeze, Sub, Cast, Where, Slice, Gather, Mul, Add, Reshape, Gemm, Transpose, Expand, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid	Expand, Concat	Cast	ScatterND
OpenVINOExecutionProvider	false	Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Reshape, Gemm, Transpose, Concat, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid	none	none	ScatterND
DmlExecutionProvider	false	none	none	none	Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Reshape, Gemm, Transpose, Concat, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid

7. Reproduce commands

uv run winml build -c examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp32_encoder_config.json -m Helsinki-NLP/opus-mt-en-ru -o temp/skill-run-1116/tester/build/cpu_cpu_fp32_encoder --ep cpu --device cpu --precision fp32 --no-analyze --no-compile --rebuild
uv run winml build -c examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp32_decoder_config.json -m Helsinki-NLP/opus-mt-en-ru -o temp/skill-run-1116/tester/build/cpu_cpu_fp32_decoder --ep cpu --device cpu --precision fp32 --no-analyze --no-compile --rebuild
uv run winml build -c examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp16_encoder_config.json -m Helsinki-NLP/opus-mt-en-ru -o temp/skill-run-1116/tester/build/cpu_cpu_fp16_encoder --ep cpu --device cpu --precision fp16 --no-analyze --no-compile --rebuild
uv run winml build -c examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp16_decoder_config.json -m Helsinki-NLP/opus-mt-en-ru -o temp/skill-run-1116/tester/build/cpu_cpu_fp16_decoder --ep cpu --device cpu --precision fp16 --no-analyze --no-compile --rebuild
uv run winml perf -m temp/skill-run-1116/tester/build/cpu_cpu_fp32_encoder/model.onnx --ep cpu --device cpu
uv run winml perf -m temp/skill-run-1116/tester/build/cpu_cpu_fp32_decoder/model.onnx --ep cpu --device cpu
uv run winml perf -m temp/skill-run-1116/tester/build/cpu_cpu_fp16_encoder/model.onnx --ep cpu --device cpu
uv run winml perf -m temp/skill-run-1116/tester/build/cpu_cpu_fp16_decoder/model.onnx --ep cpu --device cpu
uv run python temp/skill-run-1116/tester/l2_encoder_parity.py
uv run winml eval --schema --task translation
uv run winml eval -m temp/skill-run-1116/tester/build/cpu_cpu_fp32_encoder/model.onnx --model-id Helsinki-NLP/opus-mt-en-ru --task translation --ep cpu --device cpu
uv run winml eval -m temp/skill-run-1116/tester/build/cpu_cpu_fp32_decoder/model.onnx --model-id Helsinki-NLP/opus-mt-en-ru --task translation --ep cpu --device cpu
uv run winml eval -m temp/skill-run-1116/tester/build/cpu_cpu_fp16_encoder/model.onnx --model-id Helsinki-NLP/opus-mt-en-ru --task translation --ep cpu --device cpu
uv run winml eval -m temp/skill-run-1116/tester/build/cpu_cpu_fp16_decoder/model.onnx --model-id Helsinki-NLP/opus-mt-en-ru --task translation --ep cpu --device cpu
$env:WINMLCLI_RULES_DIR='C:\repo\ModelKitArtifacts\rules'; uv run winml analyze --model temp/skill-run-1116/tester/build/cpu_cpu_fp32_encoder/model.onnx --ep all --output temp/skill-run-1116/tester/analyze_cpu_cpu_fp32_encoder.json
$env:WINMLCLI_RULES_DIR='C:\repo\ModelKitArtifacts\rules'; uv run winml analyze --model temp/skill-run-1116/tester/build/cpu_cpu_fp32_decoder/model.onnx --ep all --output temp/skill-run-1116/tester/analyze_cpu_cpu_fp32_decoder.json
$env:WINMLCLI_RULES_DIR='C:\repo\ModelKitArtifacts\rules'; uv run winml analyze --model temp/skill-run-1116/tester/build/cpu_cpu_fp16_encoder/model.onnx --ep all --output temp/skill-run-1116/tester/analyze_cpu_cpu_fp16_encoder.json
$env:WINMLCLI_RULES_DIR='C:\repo\ModelKitArtifacts\rules'; uv run winml analyze --model temp/skill-run-1116/tester/build/cpu_cpu_fp16_decoder/model.onnx --ep all --output temp/skill-run-1116/tester/analyze_cpu_cpu_fp16_decoder.json

[ssss141414]

Independent reviewer verdict: APPROVE

Reviewed PR head 4479a5d9 against current origin/main 74342698e416bd9bc4e5f9b534f943425ce81c62.

• Scope: exactly four nested recipe JSON files under examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/; production examples/recipes/README.md is unchanged.
• Threads: 0 review threads (hasNextPage=false), 0 review comments, and 0 issue comments.
• CPU coverage freshly verified: fp32/fp16 encoder and decoder all build and perf successfully. Host providers are ['AzureExecutionProvider', 'CPUExecutionProvider']; no accelerator runtime claim is made.
• Precision honesty: fp16 commands include explicit --precision fp16; encoder/decoder contain 102/166 FLOAT16 initializers and 0 FLOAT32 initializers; perf reports Model Precision: fp16.
• Goal ladder: L0 PASS, L1 PASS, L2 encoder PASS (fp32 cosine 0.9999999403953552; fp16 cosine 0.9999976754188538), L3 CLI-BLOCKED with the exact unsupported-translation error.
• Mandatory all-EP static analysis is present for all four artifacts. Static findings are correctly not presented as accelerator execution.
• Recipe-vs-current-auto-config delta is zero. Under the current skill contract this is a valid verified CPU coverage fixture; there is no metadata-derived generalized code fix.
• Scoped lint reports no Python files and passes. Full-tree ruff check reports five pre-existing findings in scripts/sam3_reference_check.py, which is unchanged by this recipe-only PR.

No actionable issue found. Coverage annotation: fresh CPU fp32/fp16 verification; no deferred tuple in this CPU-only precision plan.

[ssss141414]

Refreshed evidence supersedes the earlier terminal verdict: frozen Lane A commit bc39a4e2 now grades L0 PASS, L1 PASS, and L2 PASS for encoder numerical parity at fp32/fp16, with the decoder static-cache constraint explicitly not claimed. Translation eval remains supplementary CLI-BLOCKED evidence because the charter ceiling is L2. The PR body now contains the complete frozen baseline, per-component perf/eval evidence, component/op analysis, delta, and reproduction commands.

[ssss141414]

Independent reviewer verdict: REQUEST_CHANGES

Reviewed live #1116 independently at PR head 224fdec14928fd1b613decc091e9fa3bf50ae0f5 against current origin/main / PR base 76d48ad38f894c3fc5a5e760de67f0a096e438a6. The old approval comment was treated as stale and was not inherited.

Owner-routed blocking issues

Explainer — restore the frozen PR body without encoding corruption

The live body is not byte/text-equivalent to the frozen explainer handoff temp/skill-run-1116/PR_1116.md (live normalized SHA-256 737e46acc1b4276cee9d75eadf74b5cfbdee86df55bc710cb98082e7c6762ff0; frozen normalized SHA-256 dc3a0d636fa870ef68ffdc123860e1093941f360e5c886dcd578ae7ddfc792fe). Three task bullets render an em dash as ΓÇö, and the architecture renders 62,518×512 as 62,518×512. Re-publish the UTF-8 frozen body without shell transcoding, then verify the live body matches the frozen file exactly. This is a PR-body communication defect, not a recipe defect.

Explainer — add the mandatory methodology-evolution declaration

The live Outcome/PR body contains Appended methodology findings: none, but contains neither exact allowed declaration: No methodology friction observed nor Methodology friction observed: _meta-NNN..NNN added. The reviewer contract's methodology-evolution audit explicitly makes silence REQUEST_CHANGES. Add No methodology friction observed because the learner handoff records no_friction: true and meta_findings: [], then retain the supporting Lane A knowledge citation.

Verified evidence

• Live metadata: OPEN, DRAFT=true, author ssss141414, label model-scale-by-skill, base main, head 224fdec14928fd1b613decc091e9fa3bf50ae0f5. Draft was not changed.
• Conversation gate: paginated REST enumeration found 2 issue comments, 0 line comments, and 0 reviews. Cursor-paginated GraphQL enumeration found 0 review threads and therefore 0 open threads. The refresh comment at recipe(opus-mt-en-ru): add verified CPU fp32/fp16 recipes #1116 (comment) explicitly supersedes the stale earlier verdict.
• Body hierarchy: top-level order is Summary, Model metadata, Validation and support evidence; validation contains Baseline, Goal, Outcome, per-tuple results, Delta, both Analyze levels, and Reproduce commands. Model metadata matches the frozen model-breakdown profile SHA-256 7e67dca09132704e1800045b6a73df65fb3166ae707f3180d15a4782752a0522.
• Freshness and scope: current head, tester git_head.txt, and live PR head all equal 224fdec14928fd1b613decc091e9fa3bf50ae0f5; current main, tester origin_main.txt, and PR base all equal 76d48ad38f894c3fc5a5e760de67f0a096e438a6. Diff is exactly four added nested recipe JSON files. No Python file changed; examples/recipes/README.md diff is empty.
• Frozen handoffs: learner references to charter, deliverable, and tester verdict all hash-match their files. Charter is Effort L0 / Goal L2 / Outcome L0 with required CPU fp32 and fp16 tuples. Lane A commit bc39a4e2 exists and contains marian-008 with validated_on entries, parameter/module/trace counts, optimization choices, analyze totals, fp16 evidence, parity, and exact eval blockers.
• Recipe versus auto-config: producer semantic comparison confirms fp32 encoder/decoder are identical to fresh current-main config; fp16 differs only at /quant for explicit fp16. This is a valid recipe-only verified-coverage fixture and not a disguised generalized code fix.
• L0 / fp16 / artifact loads: independently rebuilt current-head fp16 encoder with explicit --precision fp16 and no --no-quant; exit 0 and stdout contained Build complete in 29.3s. Direct ONNX load found IR 8, opset 17, 102 FLOAT16 initializers, correct named inputs [input_ids, attention_mask], output encoder_hidden_states [1,512,512], and adjacent external data where applicable. Direct loads of frozen fp32/fp16 decoder artifacts found 166 FLOAT/166 FLOAT16 initializers respectively and the complete correctly named/static-cache-shaped input/output contract.
• L1: independent fp16 encoder CPU perf rerun reported Model Precision: fp16, mean 93.97 ms, p50 93.53 ms, throughput 10.64 samples/s, and RAM total delta 296.2 MB, consistent with frozen mean 95.37 ms within 20%. Frozen evidence supplies concrete mean/p50/throughput/RAM for all four component artifacts and clearly does not claim end-to-end autoregressive latency.
• L2 Goal logic: independently reran temp/skill-run-1116/tester/l2_encoder_parity.py; exit 0. fp32 cosine 0.9999999999994811, max abs 6.198883056640625e-06; fp16 cosine 0.9999989143104122, max abs 0.0061779022216796875. Decoder parity is explicitly not claimed because equivalent static-cache state is unavailable; no higher-tier PASS is fabricated. L0→L1→L2 is complete at the frozen ceiling.
• Eval: exact per-component/per-precision blocker is present. Independent winml eval --schema --task translation rerun exited 2 with Task 'translation' is not supported; supplementary evidence is not promoted to L3.
• Analyze: all four JSON artifacts parse as six requested-EP rows each. Totals are fp32 encoder 204/16 unique, fp32 decoder 392/20, fp16 encoder 205/16, fp16 decoder 418/20. The body includes semantic component-to-node mappings, mapped/partial/unmapped counts, and complete per-EP supported/partial/unsupported/unknown op classifications; exit 1 is honestly reported as ANALYZE-PARTIAL-SUCCESS, not runtime accelerator evidence.
• Lint classification: uv run ruff check reports 6 errors on both PR head and a clean current-origin/main worktree; complete logs are byte-identical. With zero changed Python files, these are pre-existing repository failures and not a PR regression.

Required re-entry evidence

1. Publish the corrected UTF-8 body and prove live normalized body equals temp/skill-run-1116/PR_1116.md.
2. Include the explicit No methodology friction observed declaration in that frozen/live body.
3. Re-enumerate all comments and review threads after the body update; retain DRAFT and model-scale-by-skill.

No producer, tester, learner, or planner recipe/evidence correction is requested. The technical deliverable is otherwise verified; approval is withheld solely for the two explainer-owned contract failures above.

[ssss141414]

Explainer re-entry response (superseding the REQUEST_CHANGES comment for the two explainer-owned blockers): the PR body is now regenerated with ASCII-safe wording, contains the exact declaration No methodology friction observed., and has no known mojibake or replacement corruption. The local frozen body and live PR body were verified byte-for-byte identical. Draft state, label, technical facts, verdicts, recipes, commits, and Goal scope remain unchanged.

[ssss141414]

Independent reviewer verdict (re-review iteration 2): APPROVE

Independently re-reviewed live microsoft/winml-cli#1116 after the explainer fixes. The prior REQUEST_CHANGES verdict was used only as the blocker inventory; approval was not inherited.

Re-entry blockers

• PASS - frozen/live body parity: temp/skill-run-1116/PR_1116.md and the live PR body are byte-for-byte and text-identical: 52,030 UTF-8 bytes, SHA-256 a83f32b1de8ae429524576e7a9e215b3a1bddc7d8c00e4e58ecbda0698a896bc.
• PASS - corruption removed: independent scans found zero occurrences of the prior mojibake byte patterns, the Unicode replacement character, or placeholder sequences (?? / ???). Both bodies contain zero non-ASCII code points, so no hidden transcoding difference remains.
• PASS - methodology declaration: the exact sentence No methodology friction observed. occurs exactly once in the frozen body and exactly once in the live body.

Live conversation gate

Fresh paginated enumeration after the body update found:

• issue comments: 4;
• line comments: 0;
• formal reviews: 0;
• cursor-paginated GraphQL review threads: 0 total, 0 open, with hasNextPage=false.

The latest explainer response is #1116 (comment). No unresolved external thread or newly asserted technical failure exists.

Metadata, freshness, and scope

• PR state: OPEN, DRAFT=true; author ssss141414; label set exactly includes model-scale-by-skill. Draft state was not changed.
• Live/fetched/local head: 224fdec14928fd1b613decc091e9fa3bf50ae0f5 in all three places.
• Live base/current origin/main/merge base: 76d48ad38f894c3fc5a5e760de67f0a096e438a6 in all three places; PR is one commit ahead and zero commits behind.
• Scope remains exactly four added files (214 additions, 0 deletions), all nested recipe JSONs:
  • examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp16_decoder_config.json
  • examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp16_encoder_config.json
  • examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp32_decoder_config.json
  • examples/recipes/Helsinki-NLP_opus-mt-en-ru/cpu/cpu/translation_fp32_encoder_config.json
• examples/recipes/README.md remains unchanged; no Python, source, test, or skill file entered Lane B.

Frozen-evidence reuse decision

Reuse is allowed because the PR head, current base, merge base, four-file scope, and frozen handoffs did not move. Handoff hashes independently recomputed and matched their cross-references:

• charter: 6e5ed3d4a2a83c12994d2336dd1627e3e92429fb2ead2ca4a8b34dd448b0c77c;
• deliverable: c1002d70802d54649448f5330ae37e93ec7507756dff9076357a7aa19b27c315;
• tester verdict: d22132622477bb9c131da11a4938e0daf5094fdbc5d5af8e0af02f15627c6b13;
• learner findings: fb03861238552f567a4a33f1b2c72528c6e6085e8b61686b006563aa6a90b04c;
• model profile: 7e67dca09132704e1800045b6a73df65fb3166ae707f3180d15a4782752a0522;
• independent L2 script: 5ccaa50ab6710ed947d16855d5cbd99411e63abaefef91a379dc86530e57860a.

Accordingly, the prior independent current-head evidence remains valid: fp16 encoder rebuild and structural load passed; decoder artifacts loaded with the declared static-cache contract; CPU perf was within tolerance; fp32/fp16 encoder parity passed; translation eval remained exact CLI-BLOCKED; all four analyze outputs parsed with complete requested-EP rows; and full-tree ruff failures were identical to clean current main and outside this recipe-only diff.

Final disposition

APPROVE. Both explainer-owned blockers are fixed, no state or evidence moved, no conversation blocker exists, and the technical deliverable remains independently verified. Coverage is fresh CPU fp32/fp16 for encoder and decoder; there are no deferred tuples in the frozen CPU-only precision plan.

The PR must remain Draft with model-scale-by-skill; this approval does not authorize changing Draft state.