Downloads · 30 days
0
Riasok/gr00t-n15-libero-spatial-ppo-fp32head-rankhook
gr00t-n15-libero-spatial-ppo-fp32head-rankhook is a machine learning model from Riasok. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Trained with RLinf (FSDP1, 4xH100, 64 envs x 8 rollout epochs, lr 5e-6 / value 1e-4, AdamW wd 0.01, clip 1.0, 240-step horizon, flow-SDE noise 0.5, updateepoch 4). Start: RLinf/RLinf-Gr00t-SFT-Spatial (N1.5). Actor pr…
Downloads · 30 days
0
Access
Public
Updated Sep 23, 2026
Repo size
25.2 GB
Likes
0
Public
Click a slice to open those files.
.distcp16.2 GB · 64%
From the Hugging Face model README
Trained with RLinf (FSDP1, 4xH100, 64 envs x 8 rollout epochs, lr 5e-6 / value 1e-4, AdamW wd 0.01, clip 1.0, 240-step horizon,
flow-SDE noise 0.5, update_epoch 4). Start: RLinf/RLinf-Gr00t-SFT-Spatial (N1.5). Actor precision: action head kept in fp32
(masters + optimizer state), backbone bf16, bf16 compute via FSDP mixed precision (param_dtype bf16, reduce fp32).
Pure fp32 was infeasible: the Eagle/SigLIP backbone hard-requires flash-attention-2, which rejects fp32.
global_step_N/actor/model_state_dict/full_weights.pt : full state dict (action_head.* fp32, everything else bf16)global_step_N/actor/dcp_checkpoint/ : FSDP distributed checkpoint (model + optimizer) for resumerank_hook/captures/optimizer_step_{001,005,020}.pt : per-target weight / gradient / Adam-moment slices (fp32)rank_hook/analysis/ : rank_metrics.csv, optimizer_geometry.csv, rank95_by_optimizer_stage.{png,pdf}, README.mdrank_hook/code/ : sitecustomize.py (FSDP1-aware hook), summarize_ranks.py, unit test| global step | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|
| success | 0.471 | 0.508 | 0.531 | 0.557 | 0.537 |
| stage | W2 timestep (emb 31) | AdaNorm shift (L0/5/10/15) |
|---|---|---|
| PPO gradient | 2 / 3 / 2 | 1-2 |
| Adam direction | 14 / 4 / 3 | 2-6 |
| fp32 cumulative update | 14 / 8 / 5 | 3-6 |
| bf16 endpoint difference | 899 / 657 / 614 | 880-908 / 628-651 / 576-589 |
The true update is rank <= 14; the "high rank" of a bf16 checkpoint diff is bf16 quantisation noise. AdamW-formula prediction matches the actual fp32 delta with cosine > 0.9999997.