gladit-cylinderflow-550k-ema09999

GLaDiT CylinderFlow: update 550000, EMA 0.9999

Selected GLaDiT inference weights and the matching frozen UVP VGAE for initial-frame-conditioned generation of 64 future CylinderFlow frames. The source archive includes same-machine latency, Validation100 and plotting tools.

学长运行入口 / Complete handoff

完整运行说明：五套源码 → 同机测速和质量评测 → 绘图

一键下载全部源码和说明 · 在线查看五套源码 · 固定版本权重下载器

只需分享本仓库链接。DiT 权重及配套 VGAE 已公开；baseline 沿用既有已选权重和 prepared 数据，详见交接说明。

Model and selection

• DiT: width512, depth24, 8 heads, one-hop graph attention, latent width4, condition width512, 1000 training diffusion steps; DDIM eta0 sampling.
• Checkpoint: fd1160b9-4d66-4bd3-b515-4b152d802ea7:550000, EMA0.9999.
• Selection: lowest recorded Validation24 mean UV relative RMSE across evaluated updates and raw/EMA candidates, score0.08477181532805404 under the recorded 20-step, three independently scored sample protocol.
• Parent training was paused at update811685 of the planned1000000 updates. This is the selected550k checkpoint, not the pause recovery checkpoint.
• The accompanying VGAE was selected by its declared Validation UVP reconstruction objective. The representation, normalizers and DiT latent statistics are bound in inference_manifest.json and artifacts/artifact.json.
• Dataset: phase-zero stride8, stored75 frames, input frame0 and future frames1–64, dt0.08, physical UVP. Inlet/wall UV uses the original frame0 writeback rule. Normalization is fitted on Train. Test remains sealed.

Files

• dit_ema.pt: selected tensors and minimal configuration/identity, without optimizer or other EMAs.
• artifacts/autoencoder.pt: matching VGAE weights and necessary metadata.
• artifacts/artifact.json: representation and Train latent statistics; no training cache required.
• artifacts/dataset_manifest.json: sanitized dataset/normalization metadata.
• source.zip: compatible source including graph_dit.pareto_run and graph_dit.pareto_plot.
• PARETO.md: full same-machine instructions for all five methods.

Use a fixed Hub commit revision from the handoff receipt when downloading. The source base commit plus local changes are recorded in the inference manifest; the archive is the delivered source snapshot. Reuse an existing matching CUDA environment with this source's dependencies. No training or latent-cache build is needed.

Run

Extract source.zip into a source directory and run from that directory:

export CUDA_VISIBLE_DEVICES=0
export OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2
python -m graph_dit.pareto_run \
  --checkpoint /shared/gladit/dit_ema.pt --weights ema_0.9999 \
  --artifacts /shared/gladit/artifacts --data-dir /shared/cylinderflow_stride8 \
  --sampling-steps 6 20 --ensemble-sizes 1 2 4 8 16 \
  --campaign-id cylinderflow_same_machine_20260920 --device cuda:0 --threads 2 \
  --output-dir /shared/cylinderflow_same_machine_20260920/gladit

Each point measures a complete serialK-member physical-field mean, and separately evaluates that mean on Validation100. Timing uses fixedValidation24, two warmups and three measured repeats, FP32, noTF32/autocast, synchronized CPU-to-CPU wall time. Model loading, disk I/O and quality scoring are outside the latency interval.

Run each baseline's pareto_run.py in its own repository on the same machine, then use python -m graph_dit.pareto_plot as documented in PARETO.md. The plot includes two GLaDiT configuration curves and four baseline points; it does not imply that connecting segments form a strict Pareto frontier.

Validation status and intended use

The handoff was checked through static source inspection, trusted checkpoint content inspection, and rendering existing archived data for layout. No target GPU execution was performed during packaging. Fresh same-machine quality and latency results must be generated before using the plot as comparison evidence. The recorded selection score above is historical Validation evidence, not a same-machine measurement of the newS/K grid. Research use on the stated dataset; other geometries and operating conditions have not been established here.