NagaYu/sludge-small

1{2  "model": "checkpoints/sludge-small",3  "preset": "small",4  "parameters": 12011719,5  "platform": "macOS-27.0-arm64-arm-64bit-Mach-O",6  "processor": "arm64",7  "torch_threads": 1,8  "mean_nodes_per_screen": 13.175,9  "formats": {10    "pytorch": {11      "ms_mean": 10.71451664007327,12      "ms_p50": 10.695083000427985,13      "ms_p95": 11.57509119984752,14      "size_mb": 48.046876,15      "available": true16    },17    "onnx": {18      "ms_mean": 33.21225328021683,19      "ms_p50": 33.24016600072355,20      "ms_p95": 33.34932440011471,21      "size_mb": 48.260867,22      "available": true,23      "max_abs_diff_vs_pytorch": 1.239776611328125e-05,24      "max_prob_diff_vs_pytorch": 6.6356733441352844e-09,25      "node_decision_agreement_vs_pytorch": 1.0,26      "screen_decision_agreement_vs_pytorch": 1.0,27      "max_screen_prob_diff_vs_pytorch": 2.2402343025085258e-07,28      "decision_agreement_vs_pytorch": 1.0,29      "path": "sludge-small.onnx",30      "screen_decision_agreement_over_corpus": 1.0,31      "screens_with_any_flip": 0,32      "corpus_screens": 4033    },34    "onnx_int8": {35      "ms_mean": 13.350311599970155,36      "ms_p50": 13.336750000235043,37      "ms_p95": 13.508375000674278,38      "size_mb": 12.41416,39      "available": true,40      "max_abs_diff_vs_pytorch": 10.417057991027832,41      "max_prob_diff_vs_pytorch": 0.9890071749687195,42      "node_decision_agreement_vs_pytorch": 0.9997106481481481,43      "screen_decision_agreement_vs_pytorch": 0.9629629629629629,44      "max_screen_prob_diff_vs_pytorch": 0.6227820437142244,45      "decision_agreement_vs_pytorch": 0.9629629629629629,46      "path": "sludge-small.int8.onnx",47      "note": "int8 weight quantisation shifts logits; the meaningful check is decision agreement at the detector's threshold, reported above.",48      "screen_decision_agreement_over_corpus": 0.9898148148148148,49      "screens_with_any_flip": 11,50      "corpus_screens": 4051    },52    "coreml": {53      "ms_mean": 2.3157748799712863,54      "ms_p50": 2.360292000048503,55      "ms_p95": 2.4256751998109394,56      "size_mb": 24.183565,57      "available": true,58      "path": "sludge-small.mlpackage",59      "max_abs_diff_vs_pytorch": 1.2060108184814453,60      "max_prob_diff_vs_pytorch": 0.0005344669334590435,61      "node_decision_agreement_vs_pytorch": 1.0,62      "screen_decision_agreement_vs_pytorch": 1.0,63      "max_screen_prob_diff_vs_pytorch": 0.01530037641045523,64      "decision_agreement_vs_pytorch": 1.0,65      "precision": "float16"66    },67    "gguf": {68      "size_mb": 24.030848,69      "available": true,70      "path": "sludge-small.f16.gguf",71      "dtype": "f16",72      "max_abs_diff_vs_pytorch": 0.00128173828125,73      "max_prob_diff_vs_pytorch": 2.1227169781923294e-06,74      "node_decision_agreement_vs_pytorch": 1.0,75      "screen_decision_agreement_vs_pytorch": 1.0,76      "max_screen_prob_diff_vs_pytorch": 2.3298150106043636e-05,77      "decision_agreement_vs_pytorch": 1.0,78      "runnable_by_llama_cpp": false,79      "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."80    }81  },82  "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.",83  "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.",84  "end_to_end": {85    "ms_mean": 4.9734986666483865,86    "ms_p50": 4.9054374999286665,87    "ms_p95": 5.962366649646356,88    "ms_p99": 6.116143729759642,89    "n": 120,90    "rss_delta_mb": 0.147456,91    "rss_mb": 1305.837568,92    "params": 1201171993  }94}