Commit Graph

Select branches

Hide Pull Requests

main

qhy

Mono Color

• a639fdefca 没啥用 qhy qihuanye 2026-05-18 02:09:19 +08:00
• 28f2fba0e8 加入一个提前停止的机制 还有减少环境步中间步骤传递至cpu qihuanye 2026-05-18 00:48:59 +08:00
• 113e591899 多机调整 qihuanye 2026-05-17 20:49:33 +08:00
• 0164e21f48 多机 qihuanye 2026-05-17 19:23:31 +08:00
• 02080e2564 在正式测试前添加warm up qihuanye 2026-05-16 14:53:58 +00:00
• d86aeb2df0 修改默认求解器 qihuanye 2026-05-14 08:53:10 +00:00
• 5e55727901 增加脚本 qihuanye 2026-05-14 04:27:10 +00:00
• 02c3cea3f9 amd构建说明 qihuanye 2026-05-14 03:52:50 +00:00
• f08f2b82f4 Parameter Tuning qihuanye 2026-05-04 08:20:23 +00:00
• e84074d6d6 调整ignore的文件 qihuanye 2026-05-04 08:05:47 +00:00
• cf43af0729 更改求解器 step=150时 成功率更高 step=125时 速度更快 成功率持平 qihuanye 2026-05-04 07:55:13 +00:00
• 4c3fdbcce6 调参 qihuanye 2026-05-04 07:01:33 +00:00
• 75a5d86966 调整配置 启动视频写入开关 qihuanye 2026-04-10 03:44:09 +00:00
• 46cb2177bc pusht数据集配置，优化后测试结果 qihuanye 2026-04-10 03:40:20 +00:00
• 8ba5bc8b0b 多卡 qihuanye 2026-04-10 03:13:54 +00:00
• e6f2b2b9d4 调高batch_size qihuanye 2026-04-09 13:17:39 +00:00
• 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； qihuanye 2026-04-09 12:33:50 +00:00
• 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 qihuanye 2026-04-09 11:57:09 +00:00
• cd03a0d5cb 补充结果 qihuanye 2026-04-09 10:19:06 +00:00
• 20ffb3492b Disable Gym passive checker by default in stable_worldmodel env creation qihuanye 2026-04-09 10:18:54 +00:00
• 96e17a13af 补充结果 qihuanye 2026-04-09 10:15:08 +00:00
• 006102d00c 减少循环里的张量形状重排和临时对象 qihuanye 2026-04-09 10:14:58 +00:00
• 3a94829eac 补充评测结果 qihuanye 2026-04-09 10:00:25 +00:00
• 38be7d3bef Optimize inference path: add predictor-only torch.compile with reduce-overhead qihuanye 2026-04-09 10:00:13 +00:00
• f2750daace 取消视频保存 qihuanye 2026-04-09 09:52:40 +00:00
• 9e2407cdc4 Wrap eval inference in torch.inference_mode qihuanye 2026-04-09 09:18:35 +00:00
• 0f85e39690 Reduce evaluation overhead with parallel video saving qihuanye 2026-04-08 13:56:57 +00:00
• 85795bd91d Vectorize image preprocessing in stable_worldmodel policy qihuanye 2026-04-08 13:48:19 +00:00
• 7c2e341d93 fp16 qihuanye 2026-04-08 13:40:33 +00:00
• 12ba4f4352 Optimize CEM input transfers before sample expansion qihuanye 2026-04-08 13:01:24 +00:00
• fa1c15c896 Optimize JEPA eval outputs and inference hot path qihuanye 2026-04-08 12:41:21 +00:00
• 8b84251eb9 add profile frame and bf15/fp16 switch qihuanye 2026-03-31 11:09:02 +00:00
• ca231f9f9d updating readme with hugging face datasets main quentinll 2026-03-26 22:53:48 -04:00
• 19399be69f Merge pull request #1 from Hoiyeuhng/fix/device-agnostic-and-readme Lucas Maes 2026-03-24 14:40:26 +01:00
• d6475e6133 fix: use proj.device instead of hardcoded cuda, fix README typos Haiyang Luo 2026-03-23 23:30:53 -07:00
• 83f97d72ad Initial commit Lucas Maes 2026-03-12 22:56:21 -04:00