qihuanye
|
e6f2b2b9d4
|
调高batch_size
|
2026-04-09 13:17:39 +00:00 |
|
qihuanye
|
25e4ddb628
|
继续做了通用性能优化，重点从 jepa.py 热路径转到实际的 stable_worldmodel
solver/policy 边界：去掉 CEM 每轮 cpu().tolist() 和结果过早回 CPU，把
plan/warm-start 保持在 GPU，只在 env.step 前最后一步转成 numpy，同时补
了输入张量的 contiguous 处理；
|
2026-04-09 12:33:50 +00:00 |
|
qihuanye
|
995cd8cfec
|
优化 jepa.py 中通用 rollout 热路径：批量预编码动
作、移除循环内
torch.cat，并为 history_size==1 与环形缓冲区更新
添加更轻量实现； 收益不大
|
2026-04-09 11:57:09 +00:00 |
|
qihuanye
|
cd03a0d5cb
|
补充结果
|
2026-04-09 11:11:07 +00:00 |
|
qihuanye
|
96e17a13af
|
补充结果
|
2026-04-09 11:11:07 +00:00 |
|
qihuanye
|
3a94829eac
|
补充评测结果
|
2026-04-09 11:11:07 +00:00 |
|
qihuanye
|
f2750daace
|
取消视频保存
|
2026-04-09 11:11:07 +00:00 |
|
qihuanye
|
9e2407cdc4
|
Wrap eval inference in torch.inference_mode
|
2026-04-09 11:11:07 +00:00 |
|
qihuanye
|
0f85e39690
|
Reduce evaluation overhead with parallel video saving
|
2026-04-08 13:56:57 +00:00 |
|
qihuanye
|
85795bd91d
|
Vectorize image preprocessing in stable_worldmodel policy
|
2026-04-08 13:48:19 +00:00 |
|
qihuanye
|
7c2e341d93
|
fp16
|
2026-04-08 13:40:33 +00:00 |
|
qihuanye
|
12ba4f4352
|
Optimize CEM input transfers before sample expansion
|
2026-04-08 13:01:24 +00:00 |
|
qihuanye
|
fa1c15c896
|
Optimize JEPA eval outputs and inference hot path
|
2026-04-08 12:41:21 +00:00 |
|