SqueezeNet1.1 compatibility on the Wave

The complete trained SqueezeNet1.1 FP16 model passed Wave NPU inference on all four frozen images on 2026-10-14. A separate native profile repairs the rejected GlobalAveragePool lowering with a fixed depthwise convolution. All trained coefficients remain unchanged. This qualifies the exact profile and tested model, with numerical tolerances; no performance benchmark ran.

Mean repair and full-model result — PoC21

The original classifier failed at its first size-only export call in vivante.nn.pooling.layer2, before inference. REA inspected the actual firmware/device libOpenVX.so, SHA f9cd917d4232316ab53a443ec93cd1217893f4de96fc113811b800b9ae8ca0bb. Its ordinary Pooling NN support predicate clears the support flag for modes other than MAX. The verification foot copies this flag without overriding it. AVG therefore fails that route; changing stride alone cannot make that predicate admit AVG. The precise remaining fallback branch that failed is unobserved. This does not establish a universal 13×13 hardware limitation.

The isolated fp16-squeezenet-mean-v1 profile preserves the original ONNX GlobalAveragePool, graph and learned bits; only profile metadata changes. It emits a depthwise Conv2d with native weights [13,13,C,1], multiplier 1, stride 1, dilation 1, zero padding and positive FP32 zero bias. Each weight has FP16 bits 0x1e0f, exactly 1551/262144. The 169-weight gain is 1 - 25/262144, a −0.0095367431640625% error before accumulation and rounding. This is a qualified approximation to the mean. The original independent mean oracle and all previously frozen tolerances remain authoritative.

Control	Actual NN result	Agreement against original independent FP16 reference
21a	Single-channel 13×13 mean: export and one Run passed	Byte-for-byte agreement, mean 31.359375
21b	Trained classifier with 1,000-channel mean: export and one Run passed	1,000 finite logits, zero policy mismatches; max error 0.015625
21c	Complete 65-node model: export and four Runs passed	4,000 finite logits, zero policy mismatches; max error 0.03125

The full NBG is 3,278,944 bytes, SHA a965e230a5fe0aac63f89ccb8ce6618a46113c4d0e8c9c97120bed984596ece5. Actual target logs select NN throughout. The first full Run qualified the package; the remaining three used strict qualified execution with an ORT session throw guard. No CPU neural fallback or extra warm/benchmark Run occurred.

Frozen image	Max error vs independent FP16	Max probability error vs original FP32	Top-1 class unchanged	Top-5 overlap
dog	0.015625	0.000264418819	258	5/5
dog1	0.015625	0.000419957272	273	5/5
dog2	0.03125	0.000189096144	235	5/5
grace_hopper	0.015625	0.001945208801	457	5/5

Every output also passes the original FP32 logit policy abs(error) <= 0.05 + 0.02 * abs(reference), maximum probability error 0.01, identical top-1 and top-5 overlap at least 4/5. Agreement on these four images does not measure ImageNet accuracy or arbitrary model compatibility.

See the PoC21 result ledger, notebook and frozen repair design.

PoC21 used 3 compiler processes,6 public export API calls and 6 NPU Runs within its 4/8/8 budget. All 44 staging roots are absent and final captured health passes; Wi-Fi is restored to the original saved network. A Wi-Fi driver warning during the initial disconnect remains a failed pre-native interval. A separately reviewed quiet interval admitted the first new phase. The final read-only check leaves literal phase state ACTIVE with 46 events; the budget is closed and no native action remains. 415 host tests and5 subtests pass. Separate pure qualified loads and copied-artifact refusals pass.

Original pooling experiment — PoC20

Control	Trained graph	Result	Independent reference agreement
20a	First Conv3/stride2 + Relu; non-square input [1,3,17,13]	NN export and one Run passed	3,072 finite elements, zero mismatches; maximum error 0.001953125
20b	MaxPool3/stride2, [1,64,15,19] → [1,64,7,9]	NN export and one Run passed	4,032 elements, byte-for-byte agreement
20c	First Fire block: squeeze Conv, parallel expand branches, Relu, channel Concat	NN export and one Run passed	8,064 finite elements, zero mismatches; maximum error 0.001953125
20d	Classifier Conv1, Relu, GlobalAveragePool, Flatten	Size-only export failed, remote exit5	No NBG or native output
20e	Complete trained 65-node graph, four frozen images	Skipped after20d	No full-model native parity or timing measurement

Every qualified element must satisfy the policy frozen before device work: abs(error) <= 0.03 + 0.015 * abs(reference), with finite outputs and NaN canaries. The authoritative reference is independent NumPy FP32 accumulation with declared FP16 tensor rounding. Actual host ONNX Runtime outputs passed the same policy separately; their bytes are retained independently.

20a proves the tested NCHW/WHCN mapping using distinct non-square dimensions. 20c preserves both expand branches’ order when ONNX channel Concat axis1 becomes native axis2. All successful export logs select NN, with no CPU/SH neural fallback. The pool-only control could admit TP, but its actual route was NN. These results qualify the exact signatures, not arbitrary geometries.

The selected ceiling-mode MaxPool geometries give the same output as floor mode. This experiment does not qualify general ceiling behavior. The classifier lowers GlobalAveragePool to public TIM-VX AVG Pool2d, kernel and stride 13×13, zero padding, FLOOR, followed by an explicit reshape. Its input is FP16 [13,13,1000,1]; pooling output is [1,1,1000,1].

The raw20d log reports initialization failure for vivante.nn.pooling.layer2, status−1, then a size-only compile failure. The first CompileToBinary(nullptr, &size) returned false, so the data-export call never occurred. Vendor teardown also emitted two force-release reference diagnostics. No NBG was produced, no inference ran, captured health stayed unchanged, and cleanup removed the staging root. This does not identify a universal average-pooling limit or prove which size, stride, channel count or backend parameter caused the rejection.

See the result ledger, experiment notebook and frozen design.

Trained model and reproducibility

The source is torchvision0.20.1 SqueezeNet1.1, with its published ImageNet checkpoint. All52 state tensors and1,235,496 parameters match the exported original FP32 bits; FP16 conversion is explicit. The eval graph has26 Conv,26 Relu,8 Concat,3 MaxPool,1 GlobalAveragePool and1 Flatten, with no Dropout. Batch1 input is [1,3,224,224]; output is [1,1000]. Inline FP16 parameters occupy2,470,992 bytes.

Four real images and preprocessing are frozen: RGB, resize256 with bilinear antialiasing, center crop224, ImageNet normalization, Pillow11.3. Original FP32, actual ORT FP16 and independent NumPy outputs pass the predeclared logit/probability/top-k policy on all four images on the host. PoC20 did not run these full inputs on the NPU; PoC21 runs the same frozen inputs through the repaired native profile, as reported above.

Artifact	SHA-256
Published checkpoint	b8a52dc049b60e4b6ab68ad0df457362afab8b6304b2febdc1650a5dab4d7e7b
Full FP16 ONNX	f5ac6b08b8304372c06fb897c2cad63e0c30191e9bc866680b1777e5bb62e081
Original FP32 ONNX	76d825183eba140d2462759837ccb7c3802b9da807399ba7138cdad90c3af541
Frozen32-file preparation	848c42d1c0bfd704d4f19ab40d573ea04379f26b7fc73ba5c3a532c474367f3c

The original bounded importer is fp16-squeezenet-v1. It rejects unsupported operators, attributes, shapes and unconsumed coefficients. Importability does not confer native qualification. The retained20d and20e packages refuse strict native execution and uncompiled qualification locally, before phase activation, a CPU neural session or device contact.20a–20c reload as qualified packages without CPU sessions or ADB; twelve copied-artifact tamper cases refuse. Original packages and earlier source bytes remain intact.

Original PoC20 continuity and closure

The new budget was6 compiler processes,12 public export API calls and30 NPU Runs. Actual use was 4 processes,7 API calls and3 Runs, producing3 NBGs. There was no native retry. Full-model and benchmark allocations were unused; unfinished conditional benchmark source was removed. No system installation update or production integration was required.

A boot change and Wi-Fi warning during host preparation occurred before any SqueezeNet vendor call. That failed health comparison remains retained; the cause of the boot change is unknown. A separately reviewed206.882-second disconnected quiet interval admitted the current boot’s phase. Every staging, immediate preflight, native operation and CLEAN cleanup has continuous diagnostic evidence. The final read-only phase check added event36 and left its literal state ACTIVE; no native operation remained and the experiment closed without silently reseeding or advancing it.

The first final raw comparison flagged different temporary-root CLC-absence keys between the seed and final snapshots. Its FAIL remains preserved. A separate assessment compares the identical three system CLC paths and checks temporary roots independently: all38 roots are absent, pinned hashes, boot, services and parameters match, no lab processes/galcore holders remain, and seed-to-final/post-restoration diagnostics have no new alerts.

Wi-Fi is verified restored to the original saved network, COMPLETED/adminUP with an assigned IP; no configuration was saved.381 host tests and 5 subtests passed. Raw artifacts remain under artifacts/npu/poc20/ and artifacts/npu/runs/rea-poc20*.

PoC21 uses its own bounded experiment, roots and phase; it does not retry or modify the closed PoC20 experiment. For audio, the speech-enhancement shortlist selects GTCRN as the next device baseline. UL-UNAS has stronger matched published quality but an unresolved identical-feed/reset discrepancy in the Mac host runtime. DPDFNet remains the stronger full-band candidate, and LocalVQE/DTLN-AEC the echo-cancellation candidates. Their Wave performance and native compatibility remain unmeasured. VAD integration stays deferred under root backlog C20.