docs: JEPA roadmap round-2 done items; queue GPU-resume re-upload and TP/PP

Co-Authored-By: Claude Fable 5 <[email protected]>
2026-07-10 03:51:50 -07:00
parent ac3f2af06b
commit 68481ef314
1 changed files with 3 additions and 1 deletions
@@ -150,7 +150,9 @@ All in `crates/training/rtx-transformers/src/ssl/`. 163 tests passing.
 **Done (2026-07-10):** multi-node TCP AllReduce (`JepaGradSync::tcp`, wired into the runner for world_size > 1, two-rank loopback training tested); `rtx-jepa` training CLI (`crates/tooling/rtx-jepa-cli`: train/bench/plan/validate); gzip WebDataset shards; full-depth GPU and CPU ViT (all bring-up block caps removed, parity verified); `ViTSizeStr::Micro` test size.
-**Next:** NCCL/RNCCL GPU-direct AllReduce (TCP backend is CPU-buffer parameter-server), tensor/pipeline parallel execution (config exists, execution is DP-only), linear-probe eval wired into the training loop.
+**Done (2026-07-10, round 2):** NCCL GPU-direct AllReduce (`nccl` feature; id bootstrap over the TCP rendezvous; verified single-rank on RTX 5060 Ti); k-NN + linear-probe eval wired into the training loop (results in summary/CSV/CLI); GPU-run checkpointing + eval via `context_encoder_cpu_weights()`.
 **Next:** GPU weight re-upload on checkpoint resume (GPU resume currently restores only the step counter, with a warning); tensor/pipeline parallel execution (config exists, execution is DP-only — needs a multi-GPU node to build honestly); multi-rank NCCL verification on real multi-GPU hardware.
 ## Current State