{ "mean_relative_l2_raw": 0.006937251414878938, "test_samples": 200, "actual_training_epochs": 500, "optimizer_updates": 125000, "paper_table3": 0.007, "signed_gap": -6.274858512106234e-05, "absolute_gap": 6.274858512106234e-05, "relative_gap_percent": -0.8964083588723137, "selection": "final_epoch_500_no_test_selection", "checkpoint": "/public/home/liuyushuang/zhaomingsi/onescience-cfd-repro-packages/MoNo-light/weight/best_model.pth", "checkpoint_sha256": "d20393f9ae89d38f53b8da9e09799db57bef37154ec8c7fddb9a9dd4a6666122", "min": 0.0023765875482458245, "max": 0.02536537307374934, "median": 0.006544329873232485, "training_environment": { "python": "3.10.16 (main, Dec 11 2024, 16:24:50) [GCC 11.2.0]", "torch": "2.4.1", "hip": "6.1.25065", "cuda": null, "device": "BW", "device_count": 1, "platform": "Linux-5.10.134-17.1.3.sga8.x86_64-x86_64-with-glibc2.28", "hostname": "f11r3n13", "slurm_job_id": "1100920", "trainable_parameters": 1696705, "all_parameters": 1734241, "precision": "float32", "code_hashes": { "model/MoNo.py": "7af09c5106dbadc7bac35100c2c852c9739b1fa86a4ab52cc5fe12955b095e72", "config/config.yaml": "c333c872f00ea12eb9bb8c61b85df5614b18e17e49cfb76d8f19e7c038657517", "scripts/train.py": "a1bee1c36f6361c546b4a9ff166681e99baab2f19842ff480f033449673db958", "scripts/common.py": "f4283c26f3bbb3747a6d4a632ac0f499e066c8c411bd5eb73775f0cc01c6e1a9" } }, "inference_environment": { "python": "3.10.16 (main, Dec 11 2024, 16:24:50) [GCC 11.2.0]", "torch": "2.4.1", "hip": "6.1.25065", "cuda": null, "device": "BW", "device_count": 1, "platform": "Linux-5.10.134-17.1.3.sga8.x86_64-x86_64-with-glibc2.28", "hostname": "f11r3n13", "slurm_job_id": "1100920" }, "software_versions": { "torch": "2.4.1+das.opt2.dtk2504", "triton": "3.0.0+das.opt4.dtk2504", "onescience": "0.1.2rc1", "numpy": "1.26.4", "scipy": "1.15.2", "PyYAML": "6.0.2", "matplotlib": "3.10.1" }, "configured_cluster": { "partition": "hx1hdnormal01", "nodes": 1, "gpus_per_node": 1, "cpus_per_task": 8, "memory": "30G", "time_limit": "24:00:00", "gpu_type": "dcu", "ntasks_per_node": 1 }, "gpu_total_memory_bytes": 68702699520, "limitations": [ "Single DCU/PyTorch 2.4.1 HIP backend versus paper RTX4090/PyTorch2.8 CUDA12.8.", "Configured system RAM allocation is 30G, versus up to 600GB reported by the paper; all 1000/200 samples and model dimensions retained.", "Installed Triton is 3.0.0 with DTK customizations, versus paper Triton3.4.0; author fused kernels are not used.", "Independent implementation, packed QKV/SDPA attention, and RNG consumption differ from explicit author attention; not bitwise reproduction.", "Portable PyTorch log-domain Sinkhorn instead of author fused Triton kernels; same finite-step equations and differentiation.", "Identical Darcy geometry is computed once and broadcast over batch; arithmetic ordering differs.", "Paper does not specify seed uncertainty or a tolerance for the Table3 result.", "Final epoch checkpoint named best_model.pth; no test-set model selection.", "Units unspecified by supplied data; metric uses stored pressure scale." ] }