From 2e43ca0d54e142fdb2b820bd627b47fe4a23f2b8 Mon Sep 17 00:00:00 2001 From: Pepijn Date: Wed, 15 Jul 2026 15:07:01 +0200 Subject: [PATCH] docs(pi052): describe merged training optimizations --- docs/source/pi052.mdx | 11 +++++------ 1 file changed, 5 insertions(+), 6 deletions(-) diff --git a/docs/source/pi052.mdx b/docs/source/pi052.mdx index 1d5accc35..9308124f8 100644 --- a/docs/source/pi052.mdx +++ b/docs/source/pi052.mdx @@ -141,8 +141,7 @@ pass. ## Training performance -Pi052 includes the optimized training paths from -[PR #3974](https://github.com/huggingface/lerobot/pull/3974). The default path: +Pi052 uses optimized training paths by default: - batches repeated flow targets and suffix projections instead of replaying small operations in Python; @@ -174,10 +173,10 @@ with Flex/manual attention and compiled vision disabled. ### Reported training benchmarks -PR #3974 measured complete optimizer steps with three real camera inputs, BF16 -transformer/action execution, FP32 vision, fused AdamW, and no video decoding or -network I/O. Results vary with GPU, batch shape, annotation mixture, and -checkpointing: +These benchmarks measure complete optimizer steps with three real camera +inputs, BF16 transformer/action execution, FP32 vision, fused AdamW, and no +video decoding or network I/O. Results vary with GPU, batch shape, annotation +mixture, and checkpointing: | Workload | RTX PRO 6000 Blackwell | A100 80 GB | | -------------------------- | -------------------------: | -------------------------: |