aether-kit-phi2 / verification /qualification.json
zakkeown's picture
Publish qualified Phi-2 macOS Core AI bundle for Aether Kit
986b7a8 verified
Raw History Blame Contribute Delete
3.93 kB
{
"schema": 1,
"recordedAt": "2026-11-02T18:43:08.188400+00:00",
"status": "macOS spike qualified; completion API only",
"source": {
"repository": "microsoft/phi-2",
"revision": "9a5117d15e407c48fa40f8adbac1244abae52d6f",
"license": "mit"
},
"modelSHA256": "8b2a3f6388406f183163440449aced556008f05e6b7a0308b1d423bc2ce79085",
"architectureAdditions": [
"partial rotary dimensions and shared prefix rotation",
"optional output-head bias",
"parallel residual family block",
"noninterleaved strided KV updates and growth"
],
"arithmeticPolicy": {
"residuals": "float32",
"projections": "float32",
"normalization": "float32",
"weights": "float16",
"cache": "float16",
"logits": "float16",
"rotaryFrequencyInitialization": "float32"
},
"swiftRegression": {
"reportedTests": 714,
"testExecutables": 18,
"configuration": "Release, arm64-apple-macosx27.0",
"passed": true,
"logSHA256": "e5709cb19629857c6858980ecfff094986687a5dba4eeda21ac06dcdd31ad04a"
},
"pythonRegression": {
"testMethods": 10,
"passed": true,
"families": [
"qwen3",
"olmo2",
"gpt2",
"phi"
]
},
"coreAICPU": {
"passed": true,
"minimumCosine": 0.99298,
"exactArgmax": true,
"lostCacheRows": 0
},
"cases": {
"short": {
"promptTokens": 7,
"eagerPortVsReference": {
"running": {
"minCosine": 0.99999994,
"maxAbs": 0.003997,
"argmaxEqualRows": 64,
"rows": 64,
"withinFloat32Tolerance": true
},
"bucketed": {
"minCosine": 0.99999996,
"maxAbs": 0.003875,
"argmaxEqualRows": 64,
"rows": 64,
"withinFloat32Tolerance": true
}
},
"nativeMinimumCosine": 0.9984728627398196,
"nativeForcedRows": 64,
"nativeArgmaxMismatches": 0,
"nativeGreedyIdentical": true,
"cacheGrowths": 0,
"allocatedCapacity": 256,
"passed": true
},
"long": {
"promptTokens": 280,
"eagerPortVsReference": {
"running": {
"minCosine": 0.99999995,
"maxAbs": 0.008103,
"argmaxEqualRows": 64,
"rows": 64,
"withinFloat32Tolerance": true
},
"bucketed": {
"minCosine": 0.99999995,
"maxAbs": 0.008518,
"argmaxEqualRows": 64,
"rows": 64,
"withinFloat32Tolerance": true
}
},
"nativeMinimumCosine": 0.9991725048462431,
"nativeForcedRows": 64,
"nativeArgmaxMismatches": 0,
"nativeGreedyIdentical": true,
"cacheGrowths": 0,
"allocatedCapacity": 512,
"passed": true
},
"growth": {
"promptTokens": 224,
"eagerPortVsReference": {
"running": {
"minCosine": 0.99999995,
"maxAbs": 0.005407,
"argmaxEqualRows": 64,
"rows": 64,
"withinFloat32Tolerance": true
},
"bucketed": {
"minCosine": 0.99999996,
"maxAbs": 0.006926,
"argmaxEqualRows": 64,
"rows": 64,
"withinFloat32Tolerance": true
}
},
"nativeMinimumCosine": 0.9990648427426364,
"nativeForcedRows": 64,
"nativeArgmaxMismatches": 0,
"nativeGreedyIdentical": true,
"cacheGrowths": 3,
"allocatedCapacity": 512,
"passed": true
}
},
"float32Tolerance": {
"phiAtol": 0.01,
"rtol": 0,
"minimumCosine": 0.999999,
"argmax": "all rows equal",
"upstreamOnlyMaxAbs": 0.008141040802001953
},
"contiguousCacheNativeRegression": {
"family": "gpt2",
"passed": true
},
"localBundle": ".local/phi2-qualified/bundle",
"limitations": [
"no iOS qualification",
"base completion model; no chat template",
"no hosted bundle",
"interleaved KV storage remains unsupported"
]
}