recipe(wav2vec2): add jakeBland/wav2vec-vm-finetune (audio-classification)

[codykk]

Recipe-only contribution (Effort L0). Adds fp32 and fp16 recipe configs for
jakeBland/wav2vec-vm-finetune (Wav2Vec2ForSequenceClassification, task audio-classification)
on two verified EP/device combinations: CPU and DML (GPU). Goal L1 (perf) PASS on all 4 configs.

---

1. Recipe path(s)

• examples/recipes/jakeBland_wav2vec-vm-finetune/cpu/cpu/audio-classification_fp32_config.json
• examples/recipes/jakeBland_wav2vec-vm-finetune/cpu/cpu/audio-classification_fp16_config.json
• examples/recipes/jakeBland_wav2vec-vm-finetune/dml/gpu/audio-classification_fp32_config.json
• examples/recipes/jakeBland_wav2vec-vm-finetune/dml/gpu/audio-classification_fp16_config.json

2. README row

N/A — README not modified in this PR. The model has not yet passed fp16 eval on all 10 (EP, device) buckets; only cpu/cpu and dml/gpu are verified.

3. Build output dir

• temp/wav2vec_cpu_fp32/ (cpu/cpu fp32)
• temp/wav2vec_cpu_fp16/ (cpu/cpu fp16)
• temp/wav2vec_dml_fp32/ (dml/gpu fp32)
• temp/wav2vec_dml_fp16/ (dml/gpu fp16)

4. Build log

All 4 configs completed successfully:

• cpu/cpu fp32: ✅ Build complete in 56.2s (Export 29.7s, Optimize 25.5s)
• cpu/cpu fp16: ✅ Build complete in 64.0s (Export 28.8s, Optimize 24.3s, FP16 10.3s)
• dml/gpu fp32: ✅ Build complete in 53.2s (Export 28.2s, Optimize 24.2s)
• dml/gpu fp16: ✅ Build complete in 54.0s (Export 28.5s, Optimize 24.7s)

5. Appended findings

N/A — no model_knowledge/ or skill_meta/ entries added (recipe-only L0 contribution; skill repo is separate from this working repo).

6. Optimum-coverage probe

N/A — wav2vec2 architecture is already supported by Optimum's Wav2Vec2OnnxConfig. No custom OnnxConfig was needed.

7. Claimed (Effort, Goal, Outcome)

Axis	Tier
Effort	L0 (recipe-only, no per-architecture code)
Goal	L1 (build + perf)
Outcome	L0 (recipe + report)

8. Goal-ladder verdict table

Tier	Verdict	Evidence
L0 (build)	PASS	All 4 configs (cpu/cpu fp32, cpu/cpu fp16, dml/gpu fp32, dml/gpu fp16) build successfully with winml build
L1 (perf)	PASS	All 4 configs produce valid latency/throughput numbers via winml perf (see item 10)
L2 (numeric vs PyTorch)	N/A	Not attempted (Goal ceiling = L1)
L3 (task metric)	N/A	Not attempted (Goal ceiling = L1)

9. Methodology-evolution declaration

No methodology friction observed during this contribution.

10. Perf & eval data

EP / Device	Precision	Verdict	Mean	p50	p90	Throughput	RAM Δ	VRAM Δ (local)
OpenVINOExecutionProvider / cpu	fp32	PASS	128.478 ms	127.733 ms	137.807 ms	7.78 samples/s	+2511.28 MB	—
OpenVINOExecutionProvider / cpu	fp16	PASS	128.124 ms	127.399 ms	134.016 ms	7.80 samples/s	+2514.16 MB	—
DmlExecutionProvider / gpu	fp32	PASS	35.032 ms	34.790 ms	36.966 ms	28.54 samples/s	+1437.18 MB	+1273.35 MB
DmlExecutionProvider / gpu	fp16	PASS	35.047 ms	35.078 ms	36.145 ms	28.53 samples/s	+1420.45 MB	+1273.35 MB
QNNExecutionProvider / npu	*	HOST-BLOCKED	—	—	—	—	—	No NPU hardware available on test host

Model size: fp32 = 1204.57 MB (export) → 1.2 GB (optimized); fp16 = 602.4 MB (quantized, 50% size reduction).

11. Component / op-level data

From winml analyze (post-optimization):

• Total operators: 771 (after fusion)
• Unique operator types: 13
• Operator breakdown: Reshape (244), Gemm (147), Transpose (111), LayerNormalization (57), Add (49), Mul (48), MatMul (48), Gelu (32), Softmax (24), Conv (8), Unsqueeze (1), Slice (1), ReduceMean (1)
• Fusion patterns applied: gelu_fusion, matmul_add_fusion (autoconf converged in 2 iterations, 897 → 771 nodes)
• Per-EP op classification: all 13 op types classified as unknown (no per-EP support matrix available in current analyze data)
• Artifact: temp/wav2vec_cpu_fp32/analyze_result.json

12. Reproducible commands

# Build (cpu/cpu fp32)
winml build -c examples/recipes/jakeBland_wav2vec-vm-finetune/cpu/cpu/audio-classification_fp32_config.json -m jakeBland/wav2vec-vm-finetune -o temp/wav2vec_cpu_fp32

# Build (cpu/cpu fp16)
winml build -c examples/recipes/jakeBland_wav2vec-vm-finetune/cpu/cpu/audio-classification_fp16_config.json -m jakeBland/wav2vec-vm-finetune -o temp/wav2vec_cpu_fp16

# Build (dml/gpu fp32)
winml build -c examples/recipes/jakeBland_wav2vec-vm-finetune/dml/gpu/audio-classification_fp32_config.json -m jakeBland/wav2vec-vm-finetune -o temp/wav2vec_dml_fp32 --ep dml --device gpu

# Build (dml/gpu fp16)
winml build -c examples/recipes/jakeBland_wav2vec-vm-finetune/dml/gpu/audio-classification_fp16_config.json -m jakeBland/wav2vec-vm-finetune -o temp/wav2vec_dml_fp16 --ep dml --device gpu

# Perf (cpu fp32)
winml perf -m temp/wav2vec_cpu_fp32/model.onnx --device cpu --iterations 10

# Perf (cpu fp16)
winml perf -m temp/wav2vec_cpu_fp16/model.onnx --device cpu --iterations 10

# Perf (dml/gpu fp32)
winml perf -m temp/wav2vec_dml_fp32/model.onnx --device gpu --ep dml --iterations 10

# Perf (dml/gpu fp16)
winml perf -m temp/wav2vec_dml_fp16/model.onnx --device gpu --ep dml --iterations 10