Add audeering wav2vec2 dimensional emotion (speech regression) support

[DingmaomaoBJTU]

This PR adds support for audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim, routed as audio-classification with an emotion regression head (wav2vec2_emotion_regression). The shipped outcome is L1 on the CPU target EP, and the produced ONNX artifact is fp32 with a matching fp32 recipe name. The highest Goal verdict honestly reached by the tester is L2 PASS, with cosine 0.9999998807907104 and max_abs 1.6689300537109375e-06.

1. Recipe path(s)

   • examples/recipes/audeering_wav2vec2-large-robust-12-ft-emotion-msp-dim/audio-classification_fp32_config.json

2. README row

   • true

3. Build output dir

   • temp\fix_build\

4. Build log

   • ✅ Build complete in 256.5s

5. Appended findings

   • model_knowledge/wav2vec2.json findings wav2vec2-001..wav2vec2-005 were appended on the skill repo Lane A, not in this model PR.
   • Captured: custom-head requirement + routing; L2 numeric parity cosine 0.9999998807907104 / max-abs 1.6689e-6; fp32 artifact; op-coverage 410 ops/14 unique; CPU fp32 perf representative median avg 381.46 ms across four runs (range 377.25-401.01 ms), throughput 2.49-2.65 samples/sec, RAM delta +690.7 to +697.0 MB.
   • Model ~165,336,707 params; trace coverage 71/231 modules; autoconf filled optim.gelu_fusion=true, optim.matmul_add_fusion=true, quant stayed null.

6. Optimum-coverage probe

   • VENDOR-ONLY
   • added_by_winml=[]

7. Claimed (Effort, Goal, Outcome)

   • Effort: L1
   • Goal ceiling: L2
   • Outcome: L1
   • Target EPs: ["cpu"]
   • Catalog gate: baseline_build=FAIL, verdict=real-engineering, origin_main_commit=130acfe42523b8aa553b1dd10eecd7a1328b832e
   • Baseline command (no recipe):

winml build -m audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim -o temp\review_baseline_fp32\

• Baseline HEAD: 130acfe42523b8aa553b1dd10eecd7a1328b832e (current origin/main at rebase time)
• winml --version: winml, version 0.2.0
• Baseline fail output:

[WinML] Installing Execution Provider: QNNExecutionProvider

Downloading... ░░░░░░░░░░   0%
Downloading... ██████████ 100%
Downloading... ██████████ 100%
QNNExecutionProvider EP installed successfully.
- Version: 2.2451.48.0
- Package Family Name: MicrosoftCorporationII.WinML.Qualcomm.QNN.EP.2_8wekyb3d8bbwe
Usage: winml build [OPTIONS]
Try 'winml build --help' for help.

Error: wav2vec2 doesn't support task text-classification for the onnx backend. Supported tasks are: feature-extraction, automatic-speech-recognition, audio-classification, audio-frame-classification, audio-xvector.

• Recipe-build delta: the same model succeeds with the checked-in recipe command winml build -c examples\recipes\audeering_wav2vec2-large-robust-12-ft-emotion-msp-dim\audio-classification_fp32_config.json -m audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim -o temp\fix_build\; build exit 0; ✅ Build complete in 256.5s; ONNX input input_values[1,16000], outputs hidden_states[1,1024] and logits[1,3]; initializer dtypes {FLOAT:233, INT64:10} with FLOAT16_COUNT 0, correctly fp32.

8. Goal-ladder verdict table

   Tier	Verdict	Command	Evidence
   L0	PASS	winml build -c examples\recipes\audeering_wav2vec2-large-robust-12-ft-emotion-msp-dim\audio-classification_fp32_config.json -m audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim -o temp\fix_build\	Build exit 0; ✅ Build complete in 256.5s; artifact temp\fix_build\model.onnx; IN [('input_values',[1,16000])]; OUT [('hidden_states',[1,1024]),('logits',[1,3])]; initializer dtypes {FLOAT:233, INT64:10}; FLOAT16_COUNT 0; actual precision fp32, matching recipe filename.
   L1	PASS	winml perf -m temp\fix_build\model.onnx --device cpu --ep cpu	Four CPU fp32 runs: Avg ms [377.25, 380.48, 382.43, 401.01], representative median avg 381.46 ms; P50 range 347.27-375.69 ms; throughput 2.49-2.65 samples/sec; RAM total delta +690.7 to +697.0 MB. Run-to-run variance includes a run-4 max-latency outlier (1278.88 ms), so the table reports median representative plus range rather than a single best run.
   L2	PASS	python temp\audeering_l2_parity.py	anchor_onnx_logits [[0.5460754632949829,0.60622638463974,0.4043162763118744]]; anchor_reference_logits [[0.5460754036903381,0.6062266230583191,0.40431657433509827]]; anchor_max_abs 2.980232238769531e-07; random_seed 0; onnx_logits [[0.6837583780288696,0.649890661239624,0.5071567296981812]]; pytorch_logits [[0.6837599873542786,0.6498923301696777,0.507158100605011]]; cosine 0.9999998807907104; max_abs 1.6689300537109375e-06

9. Methodology-evolution declaration

   • _meta-056 effort-mis-estimate (planner Optimum-probe edit), _meta-057 goal-ceiling mis-estimate (planner Goal-axis edit), and _meta-058 doc-code-drift user_skill→dev_skill (explainer/reviewer path fix) were filed on the SKILL repo Lane A branch.
   • The paired skill edits were committed on the skill worktree at commit 9d380bae and are intentionally not included in this model PR to keep lanes unpolluted.

10. Perf & eval data

EP / Device	Precision	Verdict	Mean	p50	Throughput	RAM Δ	Eval
CPUExecutionProvider / cpu	fp32	PASS	381.46 ms representative median avg (runs 377.25, 380.48, 382.43, 401.01 ms)	349.58 ms representative median p50 (run range 347.27-375.69 ms)	2.62 samples/sec representative (range 2.49-2.65)	+693.8 MB representative total delta (range +690.7 to +697.0 MB)	N/A — ceiling is L2

11. Component / op-level data

• Command: winml analyze --model temp\fix_build\model.onnx --ep all --format json
• Rules parquet count: 1746
• Command exit code: 1
• Total operators: 410
• Unique operator types: 14
• Per-EP classification: QNNExecutionProvider(NPU): 410/0/0/0 runtime_support=True has_errors=False has_warnings=False. OpenVINOExecutionProvider(NPU): 409/1/0/0 runtime_support=False has_warnings=True partial: OP/ai.onnx/Slice. VitisAIExecutionProvider(NPU): 0/0/0/410 unknown. CPUExecutionProvider: no rule data (all-unknown, expected).
• Artifact: temp\fix_analyze_output.txt

12. Reproducible commands

Set-Location C:\Users\qiowu\source\repos\winml-cli
winml build -m audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim -o temp\review_baseline_fp32\
winml build -c examples\recipes\audeering_wav2vec2-large-robust-12-ft-emotion-msp-dim\audio-classification_fp32_config.json -m audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim -o temp\fix_build\
winml perf -m temp\fix_build\model.onnx --device cpu --ep cpu
gh release download v0.2.0 --repo microsoft/winml-cli --pattern "rules-v*.zip" --dir temp; Expand-Archive .\temp\rules-v*.zip -DestinationPath src\winml\modelkit\analyze\rules\runtime_check_rules -Force
winml analyze --model temp\fix_build\model.onnx --ep all --format json
python temp\audeering_l2_parity.py

[DingmaomaoBJTU]

REVIEWER verdict: REQUEST_CHANGES

I independently re-ran the reviewer checklist from PR head da4d514a8d4238394a74ad028d2a4c7d8613016c. This is fixable, but I cannot approve from current evidence.

Blocking items (producer action required)

1. Baseline freshness (_meta-052 / reviewer.md L100) — PR cites baseline origin_main_commit=670ad35e..., but after git fetch origin main, git rev-list --count 670ad35e..origin/main returned 1; current origin/main is 130acfe4. Fix: rebase onto current origin/main and re-run the no--c baseline; update PR item 7/body with the new baseline HEAD.
2. Baseline-build gate evidence incomplete (reviewer.md L99) — PR body item 7 names baseline_build=FAIL but does not quote the required no--c baseline command output plus winml --version and git rev-parse HEAD. Reviewer re-run: winml --version → 0.2.0; winml build -m audeering/... -o temp\review_baseline\ → Error: wav2vec2 doesn't support task text-classification ..., no ✅ Build complete. Fix: after rebase, paste the exact baseline command, fail output, version, and baseline HEAD into the PR body.
3. Goal-L0 precision mismatch (_meta-014 / reviewer.md L110) — recipe path is audio-classification_fp16_config.json, but recipe has quant: null (...\audio-classification_fp16_config.json:36) and reviewer initializer scan of temp\review_build\model.onnx returned DTYPES {1:233, 7:10} / FLOAT16_COUNT 0; winml perf also reports Model Precision: fp32. Fix: either rename the recipe to _fp32_config.json and update README/PR body references, or add a real fp16/quantization path that emits FLOAT16 initializers.
4. L1 perf number not independently verified (reviewer.md L111, fail-closed) — PR body claims CPU avg 161.38 ms, but reviewer re-run on the PR-built artifact returned Avg 416.38 ms, Throughput 2.40 samples/sec, RAM +697.4 MB. Fix: after rebase/precision fix, re-run winml perf -m <fresh artifact> --device cpu --ep cpu, paste the fresh log and update item 10 to match the actual run (or explain host variance with evidence).

Checked evidence

• PR is real/draft: gh pr view 1084 returned draft PR head da4d514a....
• Checkout/diff scope: git rev-parse HEAD = da4d514a...; git diff --name-only origin/main...HEAD returned exactly examples/recipes/README.md, the audeering recipe JSON, src/winml/modelkit/models/hf/__init__.py, src/winml/modelkit/models/hf/wav2vec2.py.
• Recipe sanity: input input_values float32 [1,16000], outputs hidden_states/logits, variant loader wav2vec2_emotion_regression (audio-classification_fp16_config.json:12-42).
• Build: winml build -c ... -m audeering/... -o temp\review_build\ produced ✅ Build complete in 263.8s.
• Structure: ONNX IR 8, opset 17, inputs [('input_values',[1,16000])], outputs [('hidden_states',[1,1024]),('logits',[1,3])].
• External data: temp\review_build contains model.onnx plus colocated model.onnx.data (661,342,208 bytes).
• L2 parity: reviewer script temp\review_l2_parity.py passed; zeros ONNX logits [[0.54607546,0.60622638,0.40431628]], reference max abs 2.98e-7; random seed 0 cosine 1.0, max abs 8.94e-7.
• Analyze: recursive parquet count 1746; winml analyze --ep all produced 410 ops / 14 unique; QNN 410/0/0/0, OpenVINO 409/1/0/0 partial Slice, VitisAI 0/0/0/410 unknown.
• Code/design: custom code is confined to per-arch src/winml/modelkit/models/hf/wav2vec2.py; @register_onnx_overwrite at wav2vec2.py:73; imported in __init__.py:98-103; mapping included at __init__.py:119-138; routing path threads loader.model_type through commands/build.py:1602-1608, loader/hf.py:227-237, and loader/resolution.py:452-453,515-517. No new shared hardcoded if model_type == branch was found.
• Tests: no consistent per-model pytest convention for all neighbouring models/hf additions was found (e.g. recent ViTPose added code without adjacent tests); existing relevant loader scope passed: uv run pytest tests\integration\loader\test_hf_model_class_mapping.py tests\integration\loader\test_load_hf_model.py -q → 24 passed.
• Knowledge capture: Lane-A model_knowledge\wav2vec2.json has populated scope.validated_on (lines 17-19, 43-45, etc.), analyze totals/op types (lines 89-91), export metadata 165,336,707 params and 71/231 trace (lines 15, 41, 120), and autoconf diff (lines 15, 41, 66).
• Optimum probe: exact reviewer probe with import optimum.exporters.onnx.model_configs returned vendor tasks ['audio-classification', 'audio-frame-classification', 'audio-xvector', 'automatic-speech-recognition', 'feature-extraction'], after_winml same, added_by_winml=[]; L1-with-code remains justified by checkpoint-specific head evidence in _meta-056 and baseline failure.
• Methodology declaration: PR item 9 declares _meta-056/_meta-057/_meta-058 and says skill edits are Lane A; per this review handoff, those edits are audited in the skill repo, not required in this Lane-B PR.

Reachable verified EP: CPUExecutionProvider/cpu. Deferred target EPs: none (target_eps is ["cpu"]).

[DingmaomaoBJTU]

Addressed all 4 items in f3b531a: rebased onto origin/main 130acfe, renamed recipe fp16→fp32 (0 FLOAT16 initializers confirmed), added baseline-FAIL gate evidence to item 7, re-ran CPU perf (representative median avg 381.46 ms across runs 377.25, 380.48, 382.43, 401.01 ms; variance noted). Ready for re-review.

[DingmaomaoBJTU]

APPROVE

Verified PR #1084 from fresh PR head f3b531a6ab1237af1a4135ec6f1c3b2b175b9c84 on branch add-audeering-wav2vec2-emotion.

Checkout / scope

• Branch/head: git rev-parse --abbrev-ref HEAD => add-audeering-wav2vec2-emotion; git rev-parse HEAD => f3b531a6ab1237af1a4135ec6f1c3b2b175b9c84.
• Rebase/base freshness: git merge-base HEAD origin/main => 130acfe42523b8aa553b1dd10eecd7a1328b832e, equal to git rev-parse origin/main; HEAD is 0 commits behind current main.
• GitHub mergeability: gh pr view 1084 --json headRefOid,headRefName,mergeable,isDraft => head f3b531a6..., mergeable: MERGEABLE, draft PR.
• Lane-B diff scope (git diff --name-only origin/main...HEAD) is exactly:
  • examples/recipes/README.md
  • examples/recipes/audeering_wav2vec2-large-robust-12-ft-emotion-msp-dim/audio-classification_fp32_config.json
  • src/winml/modelkit/models/hf/__init__.py
  • src/winml/modelkit/models/hf/wav2vec2.py
• No copilot-skills/, SKILL.md, agents/*.md, tests/, or unrelated src/ churn in the PR diff.

Engineering/code review

• src/winml/modelkit/models/hf/wav2vec2.py:28-63 transcribes the model-card RegressionHead + EmotionModel exactly: wav2vec2 backbone, mean pooling over sequence, dense/tanh/dropout/out_proj regression head, outputs (hidden_states, logits).
• src/winml/modelkit/models/hf/wav2vec2.py:73-93 registers an ONNX config for wav2vec2_emotion_regression / audio-classification with raw input_values dynamic axes and hidden_states + logits batch axes; this matches the recipe and emitted ONNX.
• src/winml/modelkit/models/hf/wav2vec2.py:96-104 uses the established custom-wrapper pattern: hyphenated lookup key in MODEL_CLASS_MAPPING, plus register_specialization(..., WinMLModelForGenericTask).
• src/winml/modelkit/models/hf/__init__.py:98-103,119-138 imports the new module and merges its mapping, so decorators and class lookup run at import time.
• examples/recipes/.../audio-classification_fp32_config.json:12-42 declares input_values float32 [1,16000], value_range [-1,1], outputs hidden_states/logits, quant: null, and loader override model_type=wav2vec2_emotion_regression, model_class=EmotionModel. No hardcoded model branching was added in shared paths (rg 'if model_type ==' src\winml\modelkit found only pre-existing utility/catalog entries).

PR body / report audit

• The PR body contains all 12 Step-6 sections: recipe path, README row, build dir/log, appended findings, Optimum probe, claimed tiers + baseline gate, Goal ladder, methodology declaration, perf/eval table, component/analyze data, reproducible commands (temp\pr1084_body_recheck.md:3-93).
• Baseline-gate evidence is present in body item 7 with the no-recipe command and exact error string (temp\pr1084_body_recheck.md:29-53).
• Methodology-evolution declaration is present (temp\pr1084_body_recheck.md:65-67); skill changes are explicitly kept out of this Lane-B model PR, consistent with the observed 4-file Lane-B scope.
• Perf/eval item 10 is per target EP/device with CPU runtime row only; target EPs are ['cpu'], so no DML/QNN/OpenVINO runtime PASS is fabricated (temp\pr1084_body_recheck.md:69-73). Static analyze notes for other EPs are in the component section, not claimed as runtime L1.
• The body no longer has 161 as a headline or claim; it reports representative CPU avg 381.46 ms with run range (Select-String '161' no matches; temp\pr1084_body_recheck.md:17,62,73).

Independent rebuild and artifact verification

• Rebuilt from PR-head recipe using the body command after deleting stale temp\fix_build: winml build -c examples\recipes\audeering_wav2vec2-large-robust-12-ft-emotion-msp-dim\audio-classification_fp32_config.json -m audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim -o temp\fix_build\.
• Build result: exit 0, stdout contains ✅ Build complete in 240.1s; log saved at temp\review1084_build.log; final artifact temp\fix_build\model.onnx.
• L0 structural ONNX check: IR 8, opset 17; input input_values float32 [1,16000]; outputs hidden_states float32 [1,1024], logits float32 [1,3].
• Precision honesty: initializer dtype counts {FLOAT: 233, INT64: 10}, FLOAT16_COUNT=0; recipe filename is now _fp32_ and winml_build_config.json has quant: null.
• External-data layout is valid: model.onnx and model.onnx.data are both in temp\fix_build\.
• Build artifacts read directly: winml_build_config.json loader is {task: audio-classification, model_class: EmotionModel, model_type: wav2vec2_emotion_regression}, optim autoconf filled gelu_fusion=true, matmul_add_fusion=true, quant null; export_htp_metadata.json class EmotionModel, 165,336,707 parameters, trace 71/231; analyze_result.json has total_operators=410, unique_operator_types=14, top op counts include Reshape=122, Gemm=75, Transpose=63, and QNN static-rule row is supported.

Goal ladder re-verification

• L0 PASS: build + structural check above.
• L1 PASS on target CPU: three independent winml perf -m temp\fix_build\model.onnx --device cpu --ep cpu runs yielded avg latencies 376.84 ms, 381.93 ms, 433.75 ms; throughputs 2.65, 2.62, 2.31 samples/sec; total RAM deltas +696.1 MB, +696.5 MB, +688.8 MB. Logs: temp\review1084_perf1.log..temp\review1084_perf3.log. These are in honest variance of the PR's representative 381.46 ms / 2.62 samples/sec; no threshold gate applies.
• L2 PASS: direct ONNX Runtime CPU inference on zeros float32[1,16000] produced logits [[0.5460754633, 0.6062263846, 0.4043162763]]; reference is [[0.5460754, 0.6062266, 0.40431657]]; max-abs 2.98e-07 (<1e-4). This proves the trained regression head exported, not a random stock head.
• Optimum coverage probe re-run for native wav2vec2: vendor and after-winml tasks both ['audio-classification','audio-frame-classification','audio-xvector','automatic-speech-recognition','feature-extraction'], added_by_winml=[]; the L1 code remains justified by this model's non-loadable custom regression head.

Baseline gate

• Re-ran no-recipe baseline: winml build -m audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim -o temp\review_baseline_fp32\.
• Result: exit code 2 with Error: wav2vec2 doesn't support task text-classification for the onnx backend. Supported tasks are: feature-extraction, automatic-speech-recognition, audio-classification, audio-frame-classification, audio-xvector. Log: temp\review1084_baseline.log.
• Therefore the recipe + variant routing are load-bearing and the baseline fail -> recipe success delta is real on current origin/main.

Prior REQUEST_CHANGES items

1. Stale rebase base: FIXED. Merge-base equals current origin/main (130acfe42523b8aa553b1dd10eecd7a1328b832e), and PR is MERGEABLE.
2. Baseline-gate evidence missing from PR body: FIXED. Body item 7 includes the failing command and exact error; independent baseline re-run reproduced it.
3. fp16 filename / fp32 artifact mismatch: FIXED. Diff contains only audio-classification_fp32_config.json; no _fp16_ audeering recipe remains; quant:null; rebuilt artifact has 0 FLOAT16 initializers.
4. Perf honesty: FIXED. Body no longer headlines 161ms; reports representative 381.46 ms with range. My three-run perf (376.84, 381.93, 433.75 ms) confirms that is honest host-load variance.

EP coverage annotation

• charter.target_eps = [cpu].
• Reachable verified: CPUExecutionProvider / cpu completed L0, L1, and L2.
• Deferred target EPs: none. DML/QNN/OpenVINO/VitisAI are not target runtime EPs for this charter; any NPU/GPU entries are static analyze/rules-only, not runtime PASS.
• Coverage: full for the only target EP (cpu).

Sign-off

• Verdict: APPROVE.
• Verified commit: f3b531a6ab1237af1a4135ec6f1c3b2b175b9c84.
• Build output dir: C:\Users\qiowu\source\repos\winml-cli\temp\fix_build\.
• Logs/artifacts produced: temp\review1084_build.log, temp\review1084_perf1.log, temp\review1084_perf2.log, temp\review1084_perf3.log, temp\review1084_baseline.log, temp\pr1084_body_recheck.md.

[xieofxie]

Nice, clean addition — the routing follows house style well: hyphenated MODEL_CLASS_MAPPING keys that rely on the resolver's _→- normalization (consistent with the qwen3-* variants), register_specialization(..., "WinMLModelForGenericTask"), and the model_type_override plumbing via the recipe loader block. Lint passes and the RegressionHead/mean-pool forward match the audeering reference, which lines up with the L2 cosine ≈1.0 you reported.

My main ask is test coverage — see the inline note on wav2vec2.py. Every comparable model-class/ONNX-config addition ships a small unit test, and this one doesn't. The rest are minor nits.

[xieofxie]

No unit test ships with this module. Every comparable MODEL_CLASS_MAPPING / OnnxConfig addition has one under tests/unit/models/ — e.g. test_vitpose_mapping.py, segformer/test_onnx_config.py, sam2/test_onnx_config.py, and the qwen3-* modeling tests. A tiny test here would lock in the behavior that actually matters:

• ("wav2vec2-emotion-regression", "audio-classification") is present in the aggregated MODEL_CLASS_MAPPING and resolves to EmotionModel;
• resolve_task(cfg, task="audio-classification", model_class="EmotionModel", model_type_override="wav2vec2_emotion_regression") returns EmotionModel (guards the _→- normalization contract this whole file depends on);
• the OnnxConfig exposes the expected inputs/outputs axes.

Without it, a rename of EMOTION_REGRESSION_MODEL_TYPE or a resolver change to the normalization could silently break routing with nothing failing in CI.

[DingmaomaoBJTU]

Added tests/unit/models/wav2vec2/test_onnx_config.py (commit 7110135). It locks in all three behaviors:

• ("wav2vec2-emotion-regression", "audio-classification") is present in the aggregated MODEL_CLASS_MAPPING and resolves to EmotionModel (plus the module mapping ⊆ aggregate check);
• resolve_task(cfg, task="audio-classification", model_class="EmotionModel", model_type_override="wav2vec2_emotion_regression") returns EmotionModel, guarding the _→- normalization contract;
• Wav2Vec2EmotionRegressionIOConfig is registered with TasksManager for audio-classification and exposes the expected inputs/outputs axes.

8 tests, all passing.

[xieofxie]

This side-effect import is missing the # triggers registration trailing comment that every other *IOConfig import in this file carries (bart, blip, segformer, qwen3-*, zoedepth, …). Worth adding for consistency so it's clear the import exists for the @register_onnx_overwrite side effect, not for use. It can also be folded into the same from .wav2vec2 import (...) block as the mapping import above rather than a second from .wav2vec2 statement.

[DingmaomaoBJTU]

Added the # triggers registration comment and collapsed the mapping import to a single statement (commit 7110135).

I kept it as two from .wav2vec2 statements rather than one block, though: lint.isort.combine-as-imports is false in pyproject.toml, so ruff splits aliased imports back onto separate statements (same reason bart/t5/siglip each get their own line). The comment sits on its own line inside the parens, matching the existing qwen3 long-name imports.

[xieofxie]

This rewords the shared header from "pass fp16 eval on all 10 (EP, device) buckets" to "with curated model recipes" for all 76 rows. This model is legitimately fp32/CPU-only so it can't meet the old bar — but relaxing the sentence for the whole table quietly drops the fp16/all-buckets guarantee that the other 75 entries do satisfy. Consider keeping the stronger wording and noting the fp32/CPU exception, rather than weakening the guarantee for every row.

[xieofxie]

by design?

[DingmaomaoBJTU]

Good catch — not by design. Reworded in commit 7110135 to keep the stronger guarantee instead of relaxing it for every row:

Total: 76 (model, task) tuples with curated recipes. 75 pass fp16 eval on all 10 (EP, device) buckets; audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim (audio-classification) is fp32/CPU-only.

So the fp16/all-buckets bar still stands for the other 75 and the fp32/CPU exception is called out explicitly.