Add H30/H35 loading support (config overlay)

h30_h35_support/README.md

@@ -0,0 +1,19 @@
+# Loading H30 / H35 RACE2 weights
+
+`h30_40k/` and `h35_40k/` are full EMA serving models (step 40000 of `h30_40k_s42` /
+`h35_40k_s42`, action horizon 30 / 35, training seed 42, EMA 0.999) plus the exact
+evaluated slim EMA overlay under `<dir>/overlay/`.
+
+Use the OpenPI fork <https://github.com/Seonghoon-Yu/Libero> at revision
+`43bc120d863a2a67fafd3ee2464029ac37040231` with the `h25_support/openpi_overlay/` files
+applied (CPS asset `assets/gauss_timing/vlabench_cps.npz` and the seeded policy-server
+entrypoint are unchanged), and then replace `src/openpi/training/config.py` with the copy in
+this directory's `openpi_overlay/`. The `h25_support` config only registers the H25 configs;
+this one additionally registers `pi05_vlabench_race2_h30`, `pi05_vlabench_race2_h35`
+(and `pi05_vlabench_plainft_h30/h35`). It is a superset of the H25 overlay config.
+
+Configs `pi05_vlabench_race2_h30` / `pi05_vlabench_race2_h35` share the H25 recipe, teacher,
+dataset (`SeonghoonYu/vlabench_primitive_ft_lerobot_224`) and CPS labels; only `action_horizon`
+differs. Serve with `--policy.config pi05_vlabench_race2_h30 --policy.dir h30_40k` (replan 30)
+or the H35 equivalents (replan 35). Environment: Python 3.11, PyTorch 2.7.1+cu126,
+Transformers 4.53.2 with the fork's replacement modules, JAX 0.5.3 (`JAX_PLATFORMS=cpu`).

h30_h35_support/upload_manifest.json

@@ -0,0 +1,66 @@
+{
+ "h30_40k/README.md": {
+  "bytes": 1251,
+  "sha256": "bef338e7e4277530e4064cf21193b9c8197f6b899759d2b9b7cb9867fb10c804"
+ },
+ "h30_40k/assets/vlabench/vlabench_ft_primitive/norm_stats.json": {
+  "bytes": 4424,
+  "sha256": "bce33fc1831830d953ef9bc57e520c02ee44e02b7ad07f27ce92b931a879384f"
+ },
+ "h30_40k/model.safetensors": {
+  "bytes": 7939730180,
+  "sha256": "cf9da150ac5c621d4bc263a77c5f3e46dae7c46bfc0712a57739b7d30ad6edd0"
+ },
+ "h30_40k/overlay/ema.safetensors": {
+  "bytes": 2092955412,
+  "sha256": "f386ba7b8b7c258c9c565ff187f4c74b78b8ed90d0a5e706cb6d9bced35b5bcb"
+ },
+ "h30_40k/overlay/metadata.pt": {
+  "bytes": 3635,
+  "sha256": "1a1682efc584548de15b9fac5f878000cfa6630692de371957e3702917c69522"
+ },
+ "h30_40k/provenance.json": {
+  "bytes": 1647,
+  "sha256": "04abac8dee90ec8bc54cf38c83d7f186824e62115d32e67e1a90e29c13741be0"
+ },
+ "h30_40k/training_config.json": {
+  "bytes": 2864,
+  "sha256": "915975ed8e58c61c4baa279beb34da83359064370ae4cc94d22262c4b8443d79"
+ },
+ "h35_40k/README.md": {
+  "bytes": 1251,
+  "sha256": "6616a53b036bf8d9d98d0af86a3a702d4e779b1898d57a7155b5aade837c8dce"
+ },
+ "h35_40k/assets/vlabench/vlabench_ft_primitive/norm_stats.json": {
+  "bytes": 4424,
+  "sha256": "bce33fc1831830d953ef9bc57e520c02ee44e02b7ad07f27ce92b931a879384f"
+ },
+ "h35_40k/model.safetensors": {
+  "bytes": 7939730180,
+  "sha256": "6c475bf68a73267f673ac2851a4f9bf9290ba2bab7375135040c5aad8b238236"
+ },
+ "h35_40k/overlay/ema.safetensors": {
+  "bytes": 2092955412,
+  "sha256": "8ce31cec60d3ad344785a7f414bc8ce716d07986925278954f7e6e45ac61c3aa"
+ },
+ "h35_40k/overlay/metadata.pt": {
+  "bytes": 3635,
+  "sha256": "e28b60bbaa344c118b3875007fa101184ba7c5f539072c6cfa8ef04162ca2574"
+ },
+ "h35_40k/provenance.json": {
+  "bytes": 1647,
+  "sha256": "417e781e4dbcf8a76367407f42c523a16d9fca9015b175921da30a2aee25f11c"
+ },
+ "h35_40k/training_config.json": {
+  "bytes": 2865,
+  "sha256": "088dd41edb7fc3b781d615788fdb2d9b62e9ba39f801acef688cd4389c329d0e"
+ },
+ "h30_h35_support/README.md": {
+  "bytes": 1315,
+  "sha256": "86d90133b27c59a6e61317293d94b9b75606f143aa689cee9adb9deeb11e94d7"
+ },
+ "h30_h35_support/openpi_overlay/src/openpi/training/config.py": {
+  "bytes": 220052,
+  "sha256": "a3cc788b0cc2564e0956b9ba32deb17461fb55d59fd0a146b79876d1df236669"
+ }
+}