recipe(opus-mt-fr-en): add CPU fp32 and fp16 recipes

[ssss141414]

Summary

This recipe-only contribution adds reviewed CPU fp32 and genuine fp16 coverage for the Helsinki-NLP/opus-mt-fr-en Marian French-to-English translation composite. The model does work by default on current main; the fp32 recipes are semantically identical to current auto-config and the fp16 recipes add an explicit, verified precision policy. Effort/Outcome are L0, the highest reached Goal is L2 PASS on the encoder, and L3 is CLI-BLOCKED because winml eval does not register translation.

Model metadata

Model-breakdown: temp/skill-run-1115/model-breakdown/opus-mt-fr-en.model-breakdown.json; SHA-256 3c15048e55223ca7aab6121ca5acdf534b16bb2837a9cd099eb813aec8eaafd9; checkpoint revision c4aed37b318c763fd177aa449b44e3b783cc6c02.

What the model does

A text-to-text Marian encoder-decoder checkpoint that translates French input text into English output text. The encoder produces contextual hidden states and the autoregressive decoder produces vocabulary logits plus updated key/value cache tensors.
Evidence: Hugging Face model identity opus-mt-fr-en; winml inspect: pipeline_tasks=[translation], architectures=[MarianMTModel]; exported encoder and decoder graph interfaces. Confidence: verified.

Primary user stories

• A user supplies French text to obtain an English translation for cross-language communication or downstream processing. Evidence: checkpoint identity and translation pipeline task. Confidence: verified.

Supported tasks

• translation — surfaces: checkpoint, transformers, winml; evidence: winml inspect pipeline_tasks; composite config generation; confidence: verified.
• text2text-generation — surfaces: transformers, optimum-onnx, winml; evidence: decoder task and Optimum Marian registry; confidence: verified.
• feature-extraction — surfaces: optimum-onnx, winml; evidence: encoder component task and build config; confidence: verified.

These support surfaces describe model loading/export/inference routing. They do not imply task-metric coverage in winml eval; the fresh L3 probe below confirms that its eval registry currently excludes translation.

Model architecture

A 6-layer Marian transformer encoder and 6-layer autoregressive Marian decoder, hidden width 512 with 8 attention heads. Each encoder layer contains self-attention and an MLP; each decoder layer contains self-attention, encoder cross-attention, an MLP, normalization/residual paths, and a KV-cache interface. Encoder and decoder export as separate ONNX graphs.
Components: encoder, encoder.embeddings_input, encoder.layers[], encoder.layers[].self_attention, encoder.layers[].mlp, encoder.layers[].normalization_residual, decoder, decoder.embeddings_input, decoder.layers[], decoder.layers[].self_attention, decoder.layers[].cross_attention, decoder.layers[].mlp, decoder.layers[].normalization_residual, decoder.kv_cache, decoder.lm_head. Evidence: HTP traced module hierarchies; ONNX node scopes and graph interfaces; winml inspect io_config. Confidence: mapped.

Validation and support evidence

1. Baseline

• Current main: 76d48ad38f894c3fc5a5e760de67f0a096e438a6; WinML 0.2.0.
• Command: winml build -m Helsinki-NLP/opus-mt-fr-en -o temp/skill-run-1115/baseline/build --ep cpu --device cpu --no-analyze --no-optimize --no-quant --no-compile --rebuild
• Command: winml perf -m temp/skill-run-1115/baseline/build/encoder_model.onnx --task feature-extraction --ep cpu --device cpu --iterations 10 --warmup 3 --format json
• Command: winml perf -m temp/skill-run-1115/baseline/build/decoder_model.onnx --task text2text-generation --ep cpu --device cpu --iterations 10 --warmup 3 --format json
• Command: winml eval --schema --task translation
• Default/current-main build PASS: Build complete in 37.9s; encoder_model.onnx and decoder_model.onnx emitted.
• Default encoder PASS: mean 71.773 ms, p50 71.038 ms, 14.03 samples/s, RSS delta 140.71 MB.
• Default decoder PASS: mean 21.897 ms, p50 22.274 ms, 45.67 samples/s, RSS delta 243.21 MB.
• Default composite perf PASS: encoder mean/p50 78.319/78.513 ms; decoder mean/p50 24.327/24.283 ms.
• Translation eval CLI-BLOCKED because task translation is absent from the eval registry.
• Optimum probe: vendor tasks are feature-extraction, feature-extraction-with-past, text-generation, text-generation-with-past, text2text-generation, and text2text-generation-with-past; WinML adds no new task key. Starting config is the two-file composite output from winml config --task translation.

2. Goal

• Frozen charter revision 2: Effort L0, Goal ceiling L2, Outcome L0.
• Success definition: L0 builds and validates CPU fp32/fp16 encoder+decoder artifacts; L1 records component perf; L2 compares encoder ONNX to PyTorch by name at both precisions. Ceiling-independent eval support still requires a task metric or exact independent blocker for every tuple.
• Revision 2 supersedes stale pre-rebase evidence at 74342698e416bd9bc4e5f9b534f943425ce81c62; no later ceiling change was made.

3. Outcome

• Shipped tier: L0; highest reached Goal: L2 PASS; coverage: full; deferred tuples: none.
• Shipped four recipes under examples/recipes/Helsinki-NLP_opus-mt-fr-en/cpu/cpu/; no source code and no production recipe README change.
• Appended model finding marian-007 and methodology finding _meta-080 in ModelKitArtifacts Lane A commit be732458. Methodology friction observed: source-tree analyze reused a central rule tree only after selecting the implemented WINMLCLI_RULES_DIR override.

Goal ladder

Tier	Verdict	Evidence
L0	PASS	All four CPU/precision/component builds passed and were structurally validated.
L1	PASS	All four component artifacts completed CPU perf with concrete latency, throughput, and memory data.
L2	PASS	Encoder named-input PyTorch parity: fp32 cosine=0.9999999999995187, max_abs=3.337860107421875e-06; fp16 cosine=0.999999184734255, max_abs=0.004638791084289551. Decoder parity not claimed: identical DynamicCache/past-KV reconstruction is not provided by the current harness.

4. Per-EP/device/precision results — including perf and eval data

Perf

Ten measured iterations after three warmups, batch size 1. Component latency is not full autoregressive translation latency.

Component	EP / Device	Precision	Verdict	Mean	p50	Throughput	RAM delta	VRAM delta
encoder	CPUExecutionProvider / cpu	fp32	PASS	70.032 ms	70.415 ms	14.28 samples/s	140.57 MB	0.0 MB
decoder	CPUExecutionProvider / cpu	fp32	PASS	23.607 ms	23.677 ms	42.36 samples/s	243.33 MB	0.0 MB
encoder	CPUExecutionProvider / cpu	fp16	PASS	92.095 ms	92.482 ms	10.86 samples/s	291.08 MB	0.0 MB
decoder	CPUExecutionProvider / cpu	fp16	PASS	26.441 ms	26.45 ms	37.82 samples/s	252.65 MB	0.0 MB

Default composite-path perf also passed at fp32: encoder mean/p50 78.319/78.513 ms, decoder 24.327/24.283 ms. This confirms the model works by default independently of the checked-in recipes.

Eval

EP / Device	Precision	Verdict	Dataset / revision / subset	Metric	Exact blocker
CPUExecutionProvider / cpu	fp32	CLI-BLOCKED	unavailable / unavailable / unavailable	unavailable	Evaluation failed: Task 'translation' is not supported. Supported tasks: audio-classification, automatic-speech-recognition, depth-estimation, feature-extraction, fill-mask, image-classification, image-feature-extraction, image-to-image, image-to-text, object-detection, question-answering, semantic-segmentation, text-classification, token-classification, zero-shot-classification, zero-shot-image-classification
CPUExecutionProvider / cpu	fp16	CLI-BLOCKED	unavailable / unavailable / unavailable	unavailable	Evaluation failed: Task 'translation' is not supported. Supported tasks: audio-classification, automatic-speech-recognition, depth-estimation, feature-extraction, fill-mask, image-classification, image-feature-extraction, image-to-image, image-to-text, object-detection, question-answering, semantic-segmentation, text-classification, token-classification, zero-shot-classification, zero-shot-image-classification

No dataset or task metric can be reported before the translation eval task and generation loop are registered. Feature gap: add translation dataset/generation-loop support to winml eval / TASK_REGISTRY.

5. Delta

• examples/recipes/Helsinki-NLP_opus-mt-fr-en/cpu/cpu/translation_fp32_encoder_config.json
• examples/recipes/Helsinki-NLP_opus-mt-fr-en/cpu/cpu/translation_fp32_decoder_config.json
• examples/recipes/Helsinki-NLP_opus-mt-fr-en/cpu/cpu/translation_fp16_encoder_config.json
• examples/recipes/Helsinki-NLP_opus-mt-fr-en/cpu/cpu/translation_fp16_decoder_config.json
• fp32 encoder and decoder: semantic JSON diff from current-main winml config output is identical (formatting only).
• fp16 encoder and decoder: JSON pointer /quant changes from null to the generated fp16 policy; /quant/mode = "fp16", component-specific task/model metadata, /quant/fp16_keep_io_types = true, and generated default calibration fields. No export, optimizer, compile, or loader delta exists relative to the matching fp32 config.
• Reducibility: Current metadata and existing Marian registrations already derive a working default composite. No class-wide source defect exists; recipe-only scope is consistent. The fp32 files are filed as explicit reviewed CPU coverage, and fp16 files select a concrete precision policy.
• Code paths/symbols/class-wide changes: none. Recipe-free acceptance: PASS. examples/recipes/README.md remains untouched.

6. Analyze results — component level and op level

Status: ANALYZE-PARTIAL-SUCCESS; exit code 1. winml analyze exited 1 after emitting complete parseable 11-EP JSON for each artifact because requested all-EP classifications include partial/unsupported/unknown results; exit 2 rule-discovery failures were repaired by setting WINMLCLI_RULES_DIR.

Component-level analysis

All ten semantic regions map by scoped ONNX node names with mapped confidence. The mapping report lists representative nodes and a full region count; the table uses full region node/operator counts.

Semantic component	ONNX region / representative nodes	Nodes	Operator counts	Basis / confidence	Relevant per-EP issues
encoder.embeddings_input — Encoder embedding/input preparation	65 nodes in encoder_model.onnx; first 20 listed; /encoder/Reshape, /encoder/embed_tokens/Gather	65	Reshape:55, Gather:1, Mul:1, Add:1, Unsqueeze:2, Expand:1, Cast:2, Sub:1, Where:1	scope-name / mapped	No partial/unsupported rule-backed op type
encoder.layers[].self_attention — Encoder self-attention	91 nodes in encoder_model.onnx; first 20 listed; /encoder/layers.0/self_attn/k_proj/MatMul/MatMulAddFusion, /encoder/layers.0/self_attn/Transpose_2	91	Gemm:24, Transpose:24, Mul:12, MatMul:12, Slice:1, Add:6, Softmax:6, Reshape:6	scope-name / mapped	No partial/unsupported rule-backed op type
encoder.layers[].normalization_residual — Encoder normalization/residual	24 nodes in encoder_model.onnx; first 20 listed; /encoder/layers.0/Add, /encoder/layers.0/self_attn_layer_norm/LayerNormalization	24	Add:12, LayerNormalization:12	scope-name / mapped	No partial/unsupported rule-backed op type
encoder.layers[].mlp — Encoder feed-forward network	24 nodes in encoder_model.onnx; first 20 listed; /encoder/layers.0/fc1/MatMul/MatMulAddFusion, /encoder/layers.0/activation_fn/Sigmoid	24	Gemm:12, Sigmoid:6, Mul:6	scope-name / mapped	No partial/unsupported rule-backed op type
decoder.embeddings_input — Decoder embedding/input preparation	114 nodes in decoder_model.onnx; first 20 listed; gemm_input_reshape_token_212, gemm_output_reshape_token_215_new_reshape	114	Reshape:87, Unsqueeze:6, Expand:3, Cast:4, Sub:1, Where:2, Gather:2, Mul:2, Add:3, Greater:1, Slice:1, Equal:1, ScatterND:1	scope-name / mapped	fp32 NvTensorRTRTXExecutionProvider/gpu: unknown=ScatterND; fp32 QNNExecutionProvider/npu: unknown=ScatterND; fp32 QNNExecutionProvider/gpu: partial=Cast,Expand, unknown=ScatterND; fp32 OpenVINOExecutionProvider/npu: unknown=ScatterND; fp32 OpenVINOExecutionProvider/gpu: unknown=ScatterND; fp32 OpenVINOExecutionProvider/cpu: unknown=ScatterND; fp16 NvTensorRTRTXExecutionProvider/gpu: unknown=ScatterND; fp16 QNNExecutionProvider/npu: partial=Expand, unknown=ScatterND; fp16 QNNExecutionProvider/gpu: partial=Expand, unsupported=Cast, unknown=ScatterND; fp16 OpenVINOExecutionProvider/npu: unknown=ScatterND; fp16 OpenVINOExecutionProvider/gpu: unknown=ScatterND; fp16 OpenVINOExecutionProvider/cpu: unknown=ScatterND
decoder.layers[].cross_attention — Decoder cross-attention	91 nodes in decoder_model.onnx; first 20 listed; /model/model/decoder/layers.5/encoder_attn/v_proj/MatMul/MatMulAddFusion, /model/model/decoder/layers.5/encoder_attn/Transpose_1	91	Gemm:24, Transpose:24, Slice:1, Mul:12, MatMul:12, Add:6, Softmax:6, Reshape:6	scope-name / mapped	No partial/unsupported rule-backed op type
decoder.layers[].self_attention — Decoder self-attention	126 nodes in decoder_model.onnx; first 20 listed; /model/model/decoder/layers.0/self_attn/v_proj/MatMul/MatMulAddFusion, /model/model/decoder/layers.0/self_attn/Transpose_2	126	Gemm:24, Transpose:30, Reshape:19, Expand:2, Unsqueeze:1, Concat:1, ScatterND:12, Slice:1, Mul:12, MatMul:12, Add:6, Softmax:6	scope-name / mapped	fp32 NvTensorRTRTXExecutionProvider/gpu: unknown=ScatterND; fp32 QNNExecutionProvider/npu: unknown=ScatterND; fp32 QNNExecutionProvider/gpu: partial=Concat,Expand, unknown=ScatterND; fp32 OpenVINOExecutionProvider/npu: unknown=ScatterND; fp32 OpenVINOExecutionProvider/gpu: unknown=ScatterND; fp32 OpenVINOExecutionProvider/cpu: unknown=ScatterND; fp16 NvTensorRTRTXExecutionProvider/gpu: unknown=ScatterND; fp16 QNNExecutionProvider/npu: partial=Expand, unknown=ScatterND; fp16 QNNExecutionProvider/gpu: partial=Concat,Expand, unknown=ScatterND; fp16 OpenVINOExecutionProvider/npu: unknown=ScatterND; fp16 OpenVINOExecutionProvider/gpu: unknown=ScatterND; fp16 OpenVINOExecutionProvider/cpu: unknown=ScatterND
decoder.layers[].normalization_residual — Decoder normalization/residual	36 nodes in decoder_model.onnx; first 20 listed; /model/model/decoder/layers.0/Add, /model/model/decoder/layers.0/self_attn_layer_norm/LayerNormalization	36	Add:18, LayerNormalization:18	scope-name / mapped	No partial/unsupported rule-backed op type
decoder.layers[].mlp — Decoder feed-forward network	24 nodes in decoder_model.onnx; first 20 listed; /model/model/decoder/layers.0/fc1/MatMul/MatMulAddFusion, /model/model/decoder/layers.0/activation_fn/Sigmoid	24	Gemm:12, Sigmoid:6, Mul:6	scope-name / mapped	No partial/unsupported rule-backed op type
decoder.lm_head — Language-model projection	1 nodes in decoder_model.onnx; first 20 listed; /model/lm_head/MatMul	1	MatMul:1	scope-name / mapped	No partial/unsupported rule-backed op type

Unmapped node count: 0. Mapping gaps: MLP and normalization/residual are functional regions inferred from scoped node names/topology rather than separate HTP module boundaries; the frozen component map uses current-main fp32 graphs, while fp16 topology/precision is validated by tester artifacts.

Op-level analysis

fp32_encoder — 204 operators, 16 unique types

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 / device	runtime_support	supported	partial	unsupported	unknown
NvTensorRTRTXExecutionProvider/gpu	true	16 types	none	none	none
CUDAExecutionProvider/gpu	false	0 types	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid
MIGraphXExecutionProvider/gpu	false	0 types	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid
QNNExecutionProvider/npu	true	16 types	none	none	none
QNNExecutionProvider/gpu	true	16 types	none	none	none
OpenVINOExecutionProvider/npu	true	16 types	none	none	none
OpenVINOExecutionProvider/gpu	true	16 types	none	none	none
OpenVINOExecutionProvider/cpu	true	16 types	none	none	none
DmlExecutionProvider/gpu	false	0 types	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid
CPUExecutionProvider/cpu	false	0 types	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid
VitisAIExecutionProvider/npu	false	0 types	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid

fp32_decoder — 392 operators, 20 unique types

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 / device	runtime_support	supported	partial	unsupported	unknown
NvTensorRTRTXExecutionProvider/gpu	false	20 types	none	none	ScatterND
CUDAExecutionProvider/gpu	false	0 types	none	none	Reshape, Gemm, Transpose, Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Concat, ScatterND, Greater, Equal, MatMul, Softmax, LayerNormalization, Sigmoid
MIGraphXExecutionProvider/gpu	false	0 types	none	none	Reshape, Gemm, Transpose, Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Concat, ScatterND, Greater, Equal, MatMul, Softmax, LayerNormalization, Sigmoid
QNNExecutionProvider/npu	false	20 types	none	none	ScatterND
QNNExecutionProvider/gpu	false	19 types	Expand, Cast, Concat	none	ScatterND
OpenVINOExecutionProvider/npu	false	20 types	none	none	ScatterND
OpenVINOExecutionProvider/gpu	false	20 types	none	none	ScatterND
OpenVINOExecutionProvider/cpu	false	20 types	none	none	ScatterND
DmlExecutionProvider/gpu	false	0 types	none	none	Reshape, Gemm, Transpose, Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Concat, ScatterND, Greater, Equal, MatMul, Softmax, LayerNormalization, Sigmoid
CPUExecutionProvider/cpu	false	0 types	none	none	Reshape, Gemm, Transpose, Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Concat, ScatterND, Greater, Equal, MatMul, Softmax, LayerNormalization, Sigmoid
VitisAIExecutionProvider/npu	false	0 types	none	none	Reshape, Gemm, Transpose, Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Concat, ScatterND, Greater, Equal, MatMul, Softmax, LayerNormalization, Sigmoid

fp16_encoder — 205 operators, 16 unique types

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 / device	runtime_support	supported	partial	unsupported	unknown
NvTensorRTRTXExecutionProvider/gpu	true	16 types	none	none	none
CUDAExecutionProvider/gpu	false	0 types	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid
MIGraphXExecutionProvider/gpu	false	0 types	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid
QNNExecutionProvider/npu	true	16 types	none	none	none
QNNExecutionProvider/gpu	true	16 types	none	none	none
OpenVINOExecutionProvider/npu	true	16 types	none	none	none
OpenVINOExecutionProvider/gpu	true	16 types	none	none	none
OpenVINOExecutionProvider/cpu	true	16 types	none	none	none
DmlExecutionProvider/gpu	false	0 types	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid
CPUExecutionProvider/cpu	false	0 types	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid
VitisAIExecutionProvider/npu	false	0 types	none	none	Reshape, Gather, Mul, Add, Gemm, Transpose, MatMul, Unsqueeze, Expand, Cast, Sub, Where, Slice, Softmax, LayerNormalization, Sigmoid

fp16_decoder — 418 operators, 20 unique types

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 / device	runtime_support	supported	partial	unsupported	unknown
NvTensorRTRTXExecutionProvider/gpu	false	20 types	none	none	ScatterND
CUDAExecutionProvider/gpu	false	0 types	none	none	Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Reshape, Gemm, Transpose, Concat, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid
MIGraphXExecutionProvider/gpu	false	0 types	none	none	Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Reshape, Gemm, Transpose, Concat, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid
QNNExecutionProvider/npu	false	20 types	Expand	none	ScatterND
QNNExecutionProvider/gpu	false	19 types	Expand, Concat	Cast	ScatterND
OpenVINOExecutionProvider/npu	false	20 types	none	none	ScatterND
OpenVINOExecutionProvider/gpu	false	20 types	none	none	ScatterND
OpenVINOExecutionProvider/cpu	false	20 types	none	none	ScatterND
DmlExecutionProvider/gpu	false	0 types	none	none	Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Reshape, Gemm, Transpose, Concat, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid
CPUExecutionProvider/cpu	false	0 types	none	none	Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Reshape, Gemm, Transpose, Concat, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid
VitisAIExecutionProvider/npu	false	0 types	none	none	Unsqueeze, Expand, Cast, Sub, Where, Slice, Gather, Mul, Add, Reshape, Gemm, Transpose, Concat, Greater, Equal, ScatterND, MatMul, Softmax, LayerNormalization, Sigmoid

Artifacts: temp/skill-run-1115/tester/analyze_fp32_encoder.json, temp/skill-run-1115/tester/analyze_fp32_decoder.json, temp/skill-run-1115/tester/analyze_fp16_encoder.json, temp/skill-run-1115/tester/analyze_fp16_decoder.json.

7. Reproduce commands

git rev-parse HEAD; winml --version
winml inspect -m Helsinki-NLP/opus-mt-fr-en --format json --output temp/skill-run-1115/baseline/inspect.json
winml config -m Helsinki-NLP/opus-mt-fr-en --task translation -o temp/skill-run-1115/baseline/config --overwrite
winml build -m Helsinki-NLP/opus-mt-fr-en -c examples/recipes/Helsinki-NLP_opus-mt-fr-en/cpu/cpu/translation_fp32_encoder_config.json -o temp/skill-run-1115/tester/translation_fp32_encoder_config --ep cpu --device cpu --precision fp32 --no-analyze --no-compile --rebuild
winml build -m Helsinki-NLP/opus-mt-fr-en -c examples/recipes/Helsinki-NLP_opus-mt-fr-en/cpu/cpu/translation_fp32_decoder_config.json -o temp/skill-run-1115/tester/translation_fp32_decoder_config --ep cpu --device cpu --precision fp32 --no-analyze --no-compile --rebuild
winml build -m Helsinki-NLP/opus-mt-fr-en -c examples/recipes/Helsinki-NLP_opus-mt-fr-en/cpu/cpu/translation_fp16_encoder_config.json -o temp/skill-run-1115/tester/translation_fp16_encoder_config --ep cpu --device cpu --precision fp16 --no-analyze --no-compile --rebuild
winml build -m Helsinki-NLP/opus-mt-fr-en -c examples/recipes/Helsinki-NLP_opus-mt-fr-en/cpu/cpu/translation_fp16_decoder_config.json -o temp/skill-run-1115/tester/translation_fp16_decoder_config --ep cpu --device cpu --precision fp16 --no-analyze --no-compile --rebuild
winml perf -m temp/skill-run-1115/tester/translation_fp32_encoder_config/model.onnx --task feature-extraction --ep cpu --device cpu --iterations 10 --warmup 3 --format json --output temp/skill-run-1115/tester/perf_fp32_encoder.json --overwrite
winml perf -m temp/skill-run-1115/tester/translation_fp32_decoder_config/model.onnx --task text2text-generation --ep cpu --device cpu --iterations 10 --warmup 3 --format json --output temp/skill-run-1115/tester/perf_fp32_decoder.json --overwrite
winml perf -m temp/skill-run-1115/tester/translation_fp16_encoder_config/model.onnx --task feature-extraction --ep cpu --device cpu --iterations 10 --warmup 3 --format json --output temp/skill-run-1115/tester/perf_fp16_encoder.json --overwrite
winml perf -m temp/skill-run-1115/tester/translation_fp16_decoder_config/model.onnx --task text2text-generation --ep cpu --device cpu --iterations 10 --warmup 3 --format json --output temp/skill-run-1115/tester/perf_fp16_decoder.json --overwrite
python temp/skill-run-1115/tester/compare_encoder.py temp/skill-run-1115/tester/translation_fp32_encoder_config/model.onnx --output temp/skill-run-1115/tester/l2_fp32_encoder.json
python temp/skill-run-1115/tester/compare_encoder.py temp/skill-run-1115/tester/translation_fp16_encoder_config/model.onnx --output temp/skill-run-1115/tester/l2_fp16_encoder.json
winml eval --schema --task translation
$env:WINMLCLI_RULES_DIR='C:\repo\ModelKitArtifacts\rules'; winml analyze -m temp/skill-run-1115/tester/translation_fp32_encoder_config/model.onnx --ep all --device all --format json --output temp/skill-run-1115/tester/analyze_fp32_encoder.json --overwrite
$env:WINMLCLI_RULES_DIR='C:\repo\ModelKitArtifacts\rules'; winml analyze -m temp/skill-run-1115/tester/translation_fp32_decoder_config/model.onnx --ep all --device all --format json --output temp/skill-run-1115/tester/analyze_fp32_decoder.json --overwrite
$env:WINMLCLI_RULES_DIR='C:\repo\ModelKitArtifacts\rules'; winml analyze -m temp/skill-run-1115/tester/translation_fp16_encoder_config/model.onnx --ep all --device all --format json --output temp/skill-run-1115/tester/analyze_fp16_encoder.json --overwrite
$env:WINMLCLI_RULES_DIR='C:\repo\ModelKitArtifacts\rules'; winml analyze -m temp/skill-run-1115/tester/translation_fp16_decoder_config/model.onnx --ep all --device all --format json --output temp/skill-run-1115/tester/analyze_fp16_decoder.json --overwrite
winml perf -m Helsinki-NLP/opus-mt-fr-en --task translation --ep cpu --device cpu --precision fp32 --iterations 10 --warmup 3 --format json --output temp/skill-run-1115/tester/perf_composite_default_fp32.json --overwrite

[xieofxie]

same question for all your prs: does the model work by default?
if not - why
if so - what's the difference between default config and ones in pr?

Thanks

[ssss141414]

Thanks — yes, this model works by default on current main (76d48ad38f894c3fc5a5e760de67f0a096e438a6). A recipe-free composite build passed and emitted both encoder and decoder; default composite CPU perf also passed (encoder mean/p50 78.319/78.513 ms, decoder 24.327/24.283 ms).

The current-main winml config --task translation output is semantically identical to the two fp32 recipes in this PR. Those fp32 files therefore add an explicit, reviewed CPU coverage fixture rather than fixing a default-build failure. The fp16 pair differs at /quant: it adds the generated mode: fp16 policy, and the refreshed test used explicit --precision fp16; the outputs contain 102/166 FLOAT16 initializers and winml perf reports fp16.

I rebased onto current main, reran the four builds, perf, encoder PyTorch parity, and both component-level and op-level analysis, and folded the exact evidence into the PR description. Translation task-metric eval remains CLI-BLOCKED because translation is not registered by winml eval; that is independent of the recipes.

[ssss141414]

APPROVE

Independent post-PR review passed after the requested charter correction.

• Reviewed PR head 2b2c13df2c4ab33abdf32680d3481babf11c0015 against current origin/main 76d48ad38f894c3fc5a5e760de67f0a096e438a6; diff is exactly four CPU recipe JSON files and leaves examples/recipes/README.md unchanged.
• Enumerated 0 line-review threads and 0 open GraphQL review threads. Two issue comments are present: the maintainer question and the direct refreshed answer. No unresolved review thread remains.
• GitHub metadata verified: Draft is true; label model-scale-by-skill is present.
• Required PR-body hierarchy and all seven validation subsections are present. The frozen model-breakdown hash is 3c15048e55223ca7aab6121ca5acdf534b16bb2837a9cd099eb813aec8eaafd9.
• Charter revision 2 now commits Goal ceiling L2. Tester marches L0/L1/L2 to PASS; fp32/fp16 eval remains complete ceiling-independent support evidence with the exact translation CLI blocker.
• Re-ran winml --version (0.2.0) and independently loaded all four artifacts with ONNX: IR 8, opset 17, expected encoder/decoder named inputs. fp16 initializer/perf evidence and all four perf rows match the frozen tester artifacts.
• uv run ruff check reports six diagnostics, reproduced identically on a clean current-main worktree. They are outside this JSON-only diff, so this PR introduces no lint regression.
• Component-level mappings cover ten semantic regions with zero unmapped nodes; op-level analysis contains complete 11-EP classifications for all four artifacts. All analyze JSON files are parseable; exit 1 is preserved as ANALYZE-PARTIAL-SUCCESS.
• The human question is answered directly: default build and default composite perf pass; fp32 recipes are semantically identical to auto-config, while fp16 adds and verifies the explicit /quant/mode=fp16 policy.
• Learner artifacts are shipped separately in ModelKitArtifacts Lane A commit be732458 (marian-007, _meta-080).

Approval does not change shipment state. PR #1115 must remain Draft.

[xieofxie]

Thanks — yes, this model works by default on current main (76d48ad38f894c3fc5a5e760de67f0a096e438a6). A recipe-free composite build passed and emitted both encoder and decoder; default composite CPU perf also passed (encoder mean/p50 78.319/78.513 ms, decoder 24.327/24.283 ms).

The current-main winml config --task translation output is semantically identical to the two fp32 recipes in this PR. Those fp32 files therefore add an explicit, reviewed CPU coverage fixture rather than fixing a default-build failure. The fp16 pair differs at /quant: it adds the generated mode: fp16 policy, and the refreshed test used explicit --precision fp16; the outputs contain 102/166 FLOAT16 initializers and winml perf reports fp16.

I rebased onto current main, reran the four builds, perf, encoder PyTorch parity, and both component-level and op-level analysis, and folded the exact evidence into the PR description. Translation task-metric eval remains CLI-BLOCKED because translation is not registered by winml eval; that is independent of the recipes.

The fp16 pair differs at /quant: it adds the generated mode: fp16 policy, and the refreshed test used explicit --precision fp16

It means it is identical to the one with winml config --precision fp16?