Summary

This adds CPU fp32 and fp16 reranking recipes for cross-encoder/mmarco-mMiniLMv2-L12-H384-v1, a multilingual cross-encoder that emits one raw relevance logit per query-passage pair. The shipped Effort and Outcome are L0, layered on the generic reranking capability owned by dependency PR #1322. Candidate evidence reaches L3 PASS with full CPU fp32/fp16 coverage; exact current-main hosted evidence establishes the recipe-free fp32 baseline while confirming that current-main reranking config and Eval are unsupported.

Model metadata

What the model does

A multilingual cross-encoder reranker that jointly tokenizes a query and candidate passage, then emits one raw relevance logit used to sort passages in descending order.

• Evidence: the pinned checkpoint model card documents paired-query usage and decreasing-score ordering; its configuration selects XLMRobertaForSequenceClassification with one label and Identity activation; Transformers v5.14.1 supplies the one-logit sequence-classification path. Confidence: verified.

Primary user stories

• A user supplies a search query and retrieved candidate passages to obtain relevance scores for second-stage passage reranking. Evidence: pinned model card Information Retrieval and retrieve-and-rerank usage. Confidence: verified.
• A user supplies multilingual query-passage pairs to rank candidates in any of the 14 mMARCO languages advertised by the checkpoint. Evidence: pinned model card training statement and checkpoint language tags. Confidence: verified.

Supported tasks

• reranking on the checkpoint and Transformers surfaces. Evidence: checkpoint pipeline_tag=text-ranking and the paired-input raw-logit model-card example. Confidence: verified.
• text-classification on the Transformers and WinML surfaces. Evidence: checkpoint Transformers metadata selects AutoModelForSequenceClassification; the existing exact-model text-classification fp16 recipe remains unchanged. Confidence: mapped.

Model architecture

XLMRobertaForSequenceClassification
|-- XLM-R embeddings (250002 vocab, 384 hidden, learned absolute positions)
|-- Encoder stack x 12
|   |-- Self-attention (12 heads x 32 dimensions)
|   |-- Feed-forward (384 -> 1536 -> 384, GELU)
|   `-- Residual connections + LayerNorm
`-- Classification head on <s> (384 -> 384, tanh -> 1 raw logit)

• Source/confidence: pinned checkpoint configuration and Transformers v5.14.1 XLMRobertaForSequenceClassification source (verified).

Validation and support evidence

1. Baseline

• Pinned main: 0876e5ae1c98a169a6137e092e0d7b30bf9cee33.
• WinML version: winml, version 0.3.0. Hosted recovery supplied an exact-lock current-main environment after the earlier local hydration attempt was blocked.
• Recipe-free build: PASS in 111.313 seconds with auto-config, CPU, no analyze/optimize/quantize/compile, producing fp32 opset 17 with input_ids and attention_mask INT32 [1,512], FLOAT logits [1,1], 435 nodes, and external data.
• CPU perf: PASS over 3 iterations after 1 warmup: mean 210.972 ms, p50 211.587 ms, throughput 4.74 samples/s, and RSS total delta +80.81 MB.
• Starting auto-config resolves AutoModelForSequenceClassification, xlm-roberta, and text-classification. Explicit text-ranking and reranking config requests both exit 2. Current-main reranking Eval is UNSUPPORTED-TASK, exits 1, and emits no metrics.
• Optimum statically advertises vendor text-classification, and WinML adds no task: VENDOR-ONLY. The exact-lock live probe could not import Optimum's ONNX model-config module because the locked Optimum/Transformers pair is incompatible (_CAN_RECORD_REGISTRY import failure), so its evidence remains STATIC-MAPPED; LIVE-PROBE-LOCK-INCOMPATIBILITY; this does not become a claim that WinML added reranking support.
• Goal floor: L0 is the first required tier. The hosted run is actual current-main baseline evidence, not candidate evidence.

2. Goal

• Committed Effort: L0.
• Committed Goal ceiling: L3.
• Committed Outcome: L0.
• Success definition: CPU fp32/fp16 reranking build, perf, multilingual raw-logit parity and ordering, and a pinned bounded smoke through the shipped reranking evaluator.
• No ceiling change or re-issued charter occurred. Candidate evidence reached L3 PASS with full coverage and no deferred tuples; this does not change the shipped L0 Outcome.

3. Outcome

The shipped Outcome is L0: exactly two model-specific CPU reranking recipes declare verified fp32 and fp16 coverage while generic reranking remains owned by dependency PR #1322. The highest reached Goal verdict is L3 PASS on candidate bc7377782d96a4788e565ff3acf0a6c2f06ae56d, with full required-tuple coverage and no deferred tuples.

Learner findings xlm-roberta-005 through xlm-roberta-009 retain the resolved autoconf/HTP facts, complete static analysis, fp32/fp16 structure and bounded perf, raw-logit parity/order, and pinned Chinese functional-smoke evidence. They are separate Lane A knowledge changes, as is methodology finding _meta-113 (Lane A commit 113413ae, compare); none is included in this model PR.

Exact-candidate quality: Ruff passed, mypy reported no issues in 438 source files, and the five candidate test partitions passed an aggregate 8420 tests: models 1534 passed / 6 skipped / 2 xfailed; optim 848 passed / 16 skipped / 1 xfailed; commands 3641 passed / 9 skipped / 1 warning; analyze 1526 passed / 45 skipped; remaining 871 passed / 2 skipped / 1 deselected / 1 warning. Aggregate non-pass accounting was 78 skipped, 3 xfailed, and 1 deselected.

The original exact-head validation workflow concluded failure: its model job stopped when winml analyze returned exit 1 for partial/unknown static findings, so the wrapper skipped its summarizer even though complete JSON had already been emitted and was independently validated. This was a harness false negative and a non-candidate blocker, not a candidate change or a claim that the original workflow was green.

Methodology finding _meta-113 is now mechanism confirmed by validation-only run 32660811255 on unchanged candidate bc7377782d96a4788e565ff3acf0a6c2f06ae56d, through closed, unmerged draft PR #1341. The repaired control flow captured analyze exit 1, continued to the summarizer, validated the complete ANALYZE-PARTIAL-SUCCESS result with 387 operators, 20 unique types, and all 7 requested EP rows, then returned wrapper exit 0 with both validator and final job successful. This exercises the repair without changing the candidate.

4. Per-EP/device/precision results and Functional smoke Eval

• L0 fp32: input_ids and attention_mask INT32 [1,512]; FLOAT logits [1,1]; 204 FLOAT initializers; artifact size 470724506 bytes.
• L0 fp16: input_ids and attention_mask INT32 [1,512]; FLOAT logits [1,1]; 204 FLOAT16 initializers; artifact size 235435418 bytes.
• L2 fp32: cosine 0.9999999999999137, max absolute difference 0.0000057220458984375; reference and ONNX descending order [0,1,2].
• L2 fp16: cosine 0.9999999857145679, max absolute difference 0.0023097991943359375; reference and ONNX descending order [0,1,2].
• L2 scope: exactly three pairs at sequence length 512, with English, German, and unrelated Chinese passages. Scores were Identity-activated raw scalar logits with no postprocessing.

Functional smoke Eval: L3 PASS on final candidate bc7377782d96a4788e565ff3acf0a6c2f06ae56d, FP32 CPU, using C-MTEB/Mmarco-reranking, configuration default, split dev, pinned revision 8e0c766dbe9e16e1d221116a3f36795fbade07f6. Deterministic first-N selection with no shuffle retained source candidate order and processed/scored 2/2 groups, 20/20 expanded pairs, with zero skipped groups and zero groups without a positive. Caps were 2 groups, 10 candidates per group, 20 total pairs, and sequence length 512. Schema (query, positive, negative), positive/negative relevance labels, and descending raw-logit prediction semantics were verified. MRR@10 = 0.333333, Recall@1 = 0.0, and Recall@10 = 1.0. This Chinese run is functional smoke only: it proves end-to-end evaluator operability, not benchmark accuracy or multilingual quality. No fp16 or accelerator Eval claim is made. The former blocker was the absence of a WinML reranking evaluator; dependency PR #1322 supplies generic task resolution, paired-input inference, raw-logit evaluation, grouped relevance metrics, and dataset adaptation.

5. Delta

The candidate diff contains exactly these two model-specific recipes:

• examples/recipes/cross-encoder_mmarco-mMiniLMv2-L12-H384-v1/cpu/cpu/reranking_fp32_config.json
• examples/recipes/cross-encoder_mmarco-mMiniLMv2-L12-H384-v1/cpu/cpu/reranking_fp16_config.json

The current-main generated auto-config was recovered and compared field by field after stripping note-only fields:

• The 14 differences for each recipe are complete: 12 shipped /eval leaves absent from auto-config (task; dataset path/name/revision/split/samples/shuffle/streaming; query/positive/negative/max-candidates mappings), /loader/task changes text-classification to reranking, and auto-config-only /export/compatibility/transformers_attention=eager is omitted. The eval block encodes the Producer-owned pinned bounded smoke plan; the loader delta selects dependency PR Add WinML reranking support for cross-encoder/ms-marco-MiniLM-L6-v2 #1322's paired-input reranking semantics; omitting the compatibility hint preserves the existing exact-model recipe contract.
• The common identical fields cover /compile; /optim/clamp_constant_values; loader model class/type; export batch, cleanup, folding, dynamo, hierarchy, parameter, verbosity and opset settings; both named INT32 input shapes and value ranges; and the logits output name. fp32 also has identical /quant=null.
• The 22 fp16 precision-specific pointers are complete: /quant itself plus /quant/{activation_symmetric,activation_type,calibration_load_path,calibration_method,calibration_save_path,distribution,fp16_keep_io_types,fp16_op_block_list,mode,model_id,model_type,nodes_to_exclude,op_types_to_quantize,per_channel,samples,save_calibration,seed,symmetric,task,weight_symmetric,weight_type}. They have no like-for-like value because current-main auto-config was generated without a precision flag; this is a precision realization, not an auto-config regression.

Relative to the existing exact-model legacy recipe, export, optimization, and compile configuration are identical; the recipes preserve AutoModelForSequenceClassification, xlm-roberta, opset 17, INT32 [1,512] named inputs, no token_type_ids, and one logits output. The existing text-classification recipe is unchanged. No source, tests, workflow files, or production recipe README are changed. Current-main recipe-free fp32 acceptance passes, while current-main reranking config/Eval remain unsupported; generic reranking behavior and acceptance remain dependency PR #1322 ownership. The two-recipe delta is reducibility-consistent with the charter and contains no checkpoint-specific shared-code behavior.

6. Analyze summary - component level and op level

Static rule analysis completed as ANALYZE-PARTIAL-SUCCESS with exit code 1; complete JSON was emitted and independently validated. This is static compatibility analysis, not accelerator runtime execution.

Component-level summary

The unresolved mapping gap is 127 optimizer-generated/helper nodes in fp32; repeated encoder layers are collapsed above.

Op-level summary

No unsupported operator type was reported. CUDA, MIGraphX, TensorRT, and DML have no rule classifications. These are complete per-EP static findings, not runtime support claims.

7. Reproduce commands

$OUT='temp/mmarco-mminilmv2-l12-h384-v1-test'
uv run winml build -c examples/recipes/cross-encoder_mmarco-mMiniLMv2-L12-H384-v1/cpu/cpu/reranking_fp32_config.json -m cross-encoder/mmarco-mMiniLMv2-L12-H384-v1 -o $OUT/fp32
uv run winml build -c examples/recipes/cross-encoder_mmarco-mMiniLMv2-L12-H384-v1/cpu/cpu/reranking_fp16_config.json -m cross-encoder/mmarco-mMiniLMv2-L12-H384-v1 -o $OUT/fp16 --precision fp16
uv run winml analyze --model $OUT/fp32/model.onnx --ep all --output $OUT/analyze-all.json
uv run winml perf -m $OUT/fp32/model.onnx --device cpu --ep cpu --iterations 3 --warmup 1
uv run winml eval -m $OUT/fp32/model.onnx --model-id cross-encoder/mmarco-mMiniLMv2-L12-H384-v1 --task reranking --dataset C-MTEB/Mmarco-reranking --dataset-name default --dataset-revision 8e0c766dbe9e16e1d221116a3f36795fbade07f6 --split dev --samples 2 --no-shuffle --streaming --column query_column=query --column positive_column=positive --column negative_column=negative --column max_candidates=10 --ep cpu --device cpu