File size: 1,498 Bytes
c69aaec | 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 | {
"created_utc": "2026-10-06T16:01:38.503748+00:00",
"status": "FP4 W4A16; full Decision Index and native JEV inference validated",
"model_path": "Maincode/matilda-jev-fp4",
"source_checkpoint": "/shared/yue/jev-best-tritask-20261006/runs/balanced/checkpoints/step-00800",
"source_decision_index": 62.43,
"quantized_decision_index": 61.77,
"format": "FP4 E2M1 packed weights; FP8 E4M3 scale per16weights; FP32global scale; BF16activations and GEMM",
"weight_tensor_GB": 17.20065232,
"original_weight_file_GB": 52.170802736,
"native_fp4_tensor_core_GEMM": false,
"readout_and_tokenizer_identical": true,
"quantized_modules": 400,
"paired_checks": {
"development": {
"n": 608,
"agreement_percent": 95.88815789473684,
"bf16_accuracy_percent": 80.92105263157895,
"fp4_accuracy_percent": 81.57894736842105,
"mean_KL_bf16_to_fp4": 0.015593464629358985
},
"confirmation": {
"n": 544,
"agreement_percent": 97.24264705882354,
"bf16_accuracy_percent": 77.57352941176471,
"fp4_accuracy_percent": 76.83823529411765,
"mean_KL_bf16_to_fp4": 0.011581205562651279
}
},
"hardware": "AMD Instinct MI355X",
"serving_endpoints_changed": false,
"notes": [
"Full Decision Index edition0.2.1:150317successful requests across44benchmarks; all native scores independently recomputed.",
"Requires included FP4DecisionModel adapter. BF16 activations and GEMM; no native FP4 Tensor Core acceleration."
]
}
|