recipe(segformer): add nvidia b4 ADE recipe coverage

[kujin66]

Add recipe/catalog coverage for nvidia/segformer-b4-finetuned-ade-512-512 image segmentation.

This is a recipe/catalog contribution. Current origin/main already auto-builds the model, so the value of this PR is discoverable recipe/catalog/eval coverage plus a checked-in w8a16 quantized variant. The measured delta is footprint, not CPU latency: w8a16 reduces artifact footprint by 73.6%, while CPU latency is 2.06x slower on this host.

Perf And Footprint Summary

Artifact	Precision	Size	Mean Latency	P50 Latency	Throughput	RSS Delta	Compared With Baseline
temp/baseline_segformer_b4/model.onnx	fp32	245.3 MB	707.060 ms	709.386 ms	1.41 samples/s	1202.64 MB	baseline
temp/segformer-b4-w8a16/model.onnx	w8a16	64.7 MB	1456.313 ms	1462.053 ms	0.69 samples/s	1247.51 MB	-73.6% size, 2.06x slower CPU latency

Conclusion: the w8a16 recipe is a quantized footprint variant for deployment/catalog coverage. It should not be read as a CPU performance optimization.

1. Recipe path(s)

• examples/recipes/nvidia_segformer-b4-finetuned-ade-512-512/image-segmentation_fp16_config.json
• examples/recipes/nvidia_segformer-b4-finetuned-ade-512-512/image-segmentation_w8a16_config.json

2. README row

• examples/recipes/README.md contains nvidia/segformer-b4-finetuned-ade-512-512 | image-segmentation; total updated to 77 on current origin/main.

3. Build output dir

• Baseline auto-config: temp/baseline_segformer_b4
• Recipe fp16/unquantized: temp/segformer-b4-fp16
• Recipe w8a16: temp/segformer-b4-w8a16

4. Build and perf evidence

Run	Command	Result
Baseline build	./.venv/Scripts/winml.exe build -m nvidia/segformer-b4-finetuned-ade-512-512 -o temp/baseline_segformer_b4 --ep cpu --device cpu --no-analyze --no-optimize --no-quant --no-compile --rebuild	Build complete in 80.5s; final artifact temp/baseline_segformer_b4/model.onnx
Recipe fp16/unquantized build	./.venv/Scripts/winml.exe build -c examples/recipes/nvidia_segformer-b4-finetuned-ade-512-512/image-segmentation_fp16_config.json -m nvidia/segformer-b4-finetuned-ade-512-512 -o temp/segformer-b4-fp16 --no-analyze --no-optimize --no-quant --no-compile --rebuild	Build complete in 97.7s; final artifact temp/segformer-b4-fp16/model.onnx
Recipe w8a16 build	./.venv/Scripts/winml.exe build -c examples/recipes/nvidia_segformer-b4-finetuned-ade-512-512/image-segmentation_w8a16_config.json -m nvidia/segformer-b4-finetuned-ade-512-512 -o temp/segformer-b4-w8a16 --no-analyze --no-optimize --no-compile --rebuild	Build complete in 218.8s; final artifact temp/segformer-b4-w8a16/model.onnx
Baseline CPU perf	./.venv/Scripts/winml.exe perf -m temp/baseline_segformer_b4/model.onnx --device cpu --ep cpu --iterations 10 --warmup 2 --format json -o temp/baseline_segformer_b4/perf_cpu.json --overwrite	mean 707.060 ms; throughput 1.41 samples/s
W8A16 CPU perf	./.venv/Scripts/winml.exe perf -m temp/segformer-b4-w8a16/model.onnx --device cpu --ep cpu --iterations 10 --warmup 2 --format json -o temp/segformer-b4-w8a16/perf_cpu.json --overwrite	mean 1456.313 ms; throughput 0.69 samples/s

5. Appended findings

• C:/Users/jinkun/.codex/skills/adding-model-support/model_knowledge/segformer.json, finding segformer-001.
• No model-repo model_knowledge file is included in the Lane B diff.

6. Optimum-coverage probe

• Probe output for model_type=segformer: {'vendor': [], 'after_winml': ['feature-extraction', 'image-classification', 'image-segmentation', 'semantic-segmentation'], 'added_by_winml': ['feature-extraction', 'image-classification', 'image-segmentation', 'semantic-segmentation']}.
• Verdict: WINML-ONLY; no new source code needed for this checkpoint.

7. Claimed (Effort, Goal, Outcome)

• Effort: L0, recipe/catalog/index only.
• Goal ceiling: L1, because baseline L0 is already covered by current origin/main; contribution adds discoverable recipe variants and w8a16 footprint delta evidence. CPU perf is reported as a tradeoff, not an improvement.
• Outcome: L0 for model repo files; learner finding recorded in the local skill repo Lane A.
• Target EPs: [cpu].

8. Goal-ladder verdict table

Tier	Verdict	Evidence
L0	PASS	Baseline auto-config build passed on current origin/main; recipe fp16 and w8a16 builds passed. Structural validation for temp/segformer-b4-w8a16/model.onnx: IR 8, opset 17, input pixel_values [1,3,512,512], output logits [1,150,128,128], 926 uint8 initializers. External data stayed next to ONNX artifacts.
L1	PASS on CPU	Providers: ['DmlExecutionProvider', 'CPUExecutionProvider']. Baseline fp32 CPU mean 707.060 ms; w8a16 CPU mean 1456.313 ms. Artifact footprint changed from 245.3 MB to 64.7 MB (-73.6%); CPU latency is 2.06x slower.

9. Methodology-evolution declaration

• Methodology friction observed and handled in local Lane A skill updates, not mixed into this model PR: _meta-055 documents that winml perf --json is invalid and the current CLI uses --format json; _meta-056 documents that winml analyze cannot run when src/winml/modelkit/analyze/rules/runtime_check_rules contains only README.md and no parquet rule files.
• Local skill files updated: agents/tester.md, agents/learner.md, agents/reviewer.md, skill_meta/findings.json, and model_knowledge/segformer.json.
• The skill directory is not a git repo on this machine, so no separate Lane A PR could be opened from it in this session.
• Draft PR created: recipe(segformer): add nvidia b4 ADE recipe coverage #1123. GitHub CLI auth was refreshed before shipment.

[kujin66]

Producer-side reviewer sanity check for the SegFormer B4 recipe PR.

Verdict: APPROVE-PENDING-CI (comment-only; GitHub does not allow approving my own PR).

Evidence checked:

• Draft PR exists and body contains all 9 hand-off items: recipe(segformer): add nvidia b4 ADE recipe coverage #1123
• Diff scope from origin/main...HEAD is limited to recipe/catalog/doc/eval-index files for nvidia/segformer-b4.
• Recipe JSON and touched JSON indexes parse with python -m json.tool.
• Baseline auto-config build on current origin/main 904c201 passed: Build complete in 80.5s.
• Recipe builds passed: fp16/unquantized Build complete in 97.7s; w8a16 Build complete in 218.8s.
• Structural validation passed for temp/segformer-b4-w8a16/model.onnx: input pixel_values [1,3,512,512], output logits [1,150,128,128], 926 uint8 initializers.
• L1 CPU perf passed: mean 1456.313 ms, throughput 0.69 samples/s.

Open items before ready-for-merge:

• Wait for pending CI/checks to finish.
• PR is intentionally draft until maintainer/reviewer approval.

[kujin66]

Follow-up evidence added after baseline-delta review:

• Baseline auto-config CPU perf: temp/baseline_segformer_b4/perf_cpu.json -> mean 707.060 ms, throughput 1.41 samples/s, RSS total delta 1202.64 MB, artifact footprint 245.3 MB including external data.
• W8A16 recipe CPU perf: temp/segformer-b4-w8a16/perf_cpu.json -> mean 1456.313 ms, throughput 0.69 samples/s, RSS total delta 1247.51 MB, artifact footprint 64.7 MB.
• Delta is therefore footprint-focused, not CPU-latency-positive: 245.3 MB -> 64.7 MB (-73.6%); CPU latency is 2.06x slower on this CPU host.

Lane A methodology follow-up was also completed locally (kept out of this Lane B model PR):

• _meta-055: current perf JSON flag is --format json, not --json; agents/tester.md updated.
• _meta-056: when runtime rule parquet files are absent, cite exact winml analyze failure and rules-dir listing instead of requiring fabricated analyze_result.json; agents/learner.md and agents/reviewer.md updated.
• model_knowledge/segformer.json updated with the baseline-vs-w8a16 footprint/latency tradeoff.

The local skill directory is not itself a git repo, so I could not open a separate Lane A skill PR from this machine in this session.

[kujin66]

@kujin66 please read the following Contributor License Agreement(CLA). If you agree with the CLA, please reply with the following information.

@microsoft-github-policy-service agree [company="{your company}"]

Options:

• (default - no company specified) I have sole ownership of intellectual property rights to my Submissions and I am not making Submissions in the course of work for my employer.

@microsoft-github-policy-service agree

• (when company given) I am making Submissions in the course of work for my employer (or my employer has intellectual property rights in my Submissions by contract or applicable law). I have permission from my employer to make Submissions and enter into this Agreement on behalf of my employer. By signing below, the defined term “You” includes me and my employer.

@microsoft-github-policy-service agree company="Microsoft"

Contributor License Agreement

[kujin66]

@microsoft-github-policy-service agree [company="{your company}"]

[kujin66]

@microsoft-github-policy-service agree

[kujin66]

Perf evidence has been reorganized into a compact table in the PR body.

Artifact	Precision	Size	Mean Latency	Throughput	RSS Delta	Interpretation
Baseline auto-config	fp32	245.3 MB	707.060 ms	1.41 samples/s	1202.64 MB	baseline
Recipe variant	w8a16	64.7 MB	1456.313 ms	0.69 samples/s	1247.51 MB	-73.6% size, 2.06x slower CPU latency

So the recipe delta is now stated as footprint/catalog coverage, not a CPU speedup claim.

[kujin66]

@microsoft-github-policy-service agree company="Microsoft"