File size: 3,830 Bytes
76c6038 4947d56 76c6038 4947d56 76c6038 4947d56 76c6038 4947d56 76c6038 4947d56 76c6038 4947d56 76c6038 4947d56 76c6038 4947d56 76c6038 4947d56 76c6038 4947d56 76c6038 4947d56 76c6038 4947d56 76c6038 4947d56 76c6038 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 | {
"model": "checkpoints/sludge-small",
"preset": "small",
"parameters": 12011719,
"platform": "macOS-27.0-arm64-arm-64bit-Mach-O",
"processor": "arm64",
"torch_threads": 1,
"mean_nodes_per_screen": 13.175,
"formats": {
"pytorch": {
"ms_mean": 10.71451664007327,
"ms_p50": 10.695083000427985,
"ms_p95": 11.57509119984752,
"size_mb": 48.046876,
"available": true
},
"onnx": {
"ms_mean": 33.21225328021683,
"ms_p50": 33.24016600072355,
"ms_p95": 33.34932440011471,
"size_mb": 48.260867,
"available": true,
"max_abs_diff_vs_pytorch": 1.239776611328125e-05,
"max_prob_diff_vs_pytorch": 6.6356733441352844e-09,
"node_decision_agreement_vs_pytorch": 1.0,
"screen_decision_agreement_vs_pytorch": 1.0,
"max_screen_prob_diff_vs_pytorch": 2.2402343025085258e-07,
"decision_agreement_vs_pytorch": 1.0,
"path": "sludge-small.onnx",
"screen_decision_agreement_over_corpus": 1.0,
"screens_with_any_flip": 0,
"corpus_screens": 40
},
"onnx_int8": {
"ms_mean": 13.350311599970155,
"ms_p50": 13.336750000235043,
"ms_p95": 13.508375000674278,
"size_mb": 12.41416,
"available": true,
"max_abs_diff_vs_pytorch": 10.417057991027832,
"max_prob_diff_vs_pytorch": 0.9890071749687195,
"node_decision_agreement_vs_pytorch": 0.9997106481481481,
"screen_decision_agreement_vs_pytorch": 0.9629629629629629,
"max_screen_prob_diff_vs_pytorch": 0.6227820437142244,
"decision_agreement_vs_pytorch": 0.9629629629629629,
"path": "sludge-small.int8.onnx",
"note": "int8 weight quantisation shifts logits; the meaningful check is decision agreement at the detector's threshold, reported above.",
"screen_decision_agreement_over_corpus": 0.9898148148148148,
"screens_with_any_flip": 11,
"corpus_screens": 40
},
"coreml": {
"ms_mean": 2.3157748799712863,
"ms_p50": 2.360292000048503,
"ms_p95": 2.4256751998109394,
"size_mb": 24.183565,
"available": true,
"path": "sludge-small.mlpackage",
"max_abs_diff_vs_pytorch": 1.2060108184814453,
"max_prob_diff_vs_pytorch": 0.0005344669334590435,
"node_decision_agreement_vs_pytorch": 1.0,
"screen_decision_agreement_vs_pytorch": 1.0,
"max_screen_prob_diff_vs_pytorch": 0.01530037641045523,
"decision_agreement_vs_pytorch": 1.0,
"precision": "float16"
},
"gguf": {
"size_mb": 24.030848,
"available": true,
"path": "sludge-small.f16.gguf",
"dtype": "f16",
"max_abs_diff_vs_pytorch": 0.00128173828125,
"max_prob_diff_vs_pytorch": 2.1227169781923294e-06,
"node_decision_agreement_vs_pytorch": 1.0,
"screen_decision_agreement_vs_pytorch": 1.0,
"max_screen_prob_diff_vs_pytorch": 2.3298150106043636e-05,
"decision_agreement_vs_pytorch": 1.0,
"runnable_by_llama_cpp": false,
"note": "Valid GGUF v3 container and round-trips through sludge.gguf. llama.cpp cannot execute it because this is not one of its LLM graph architectures; the graph lives in sludge.model."
}
},
"note": "Per-screen figures are the model forward pass only, on this machine. The end-to-end number in the CLI's `sludge bench` additionally includes featurisation and decoding.",
"graph_shape": "static: the node axis is fixed at max_nodes and the caller pads and masks. A dynamic node axis bakes the traced node count into the attention reshapes and only runs for that one screen size.",
"end_to_end": {
"ms_mean": 4.9734986666483865,
"ms_p50": 4.9054374999286665,
"ms_p95": 5.962366649646356,
"ms_p99": 6.116143729759642,
"n": 120,
"rss_delta_mb": 0.147456,
"rss_mb": 1305.837568,
"params": 12011719
}
} |