libero: lint launch headers (drop GPU counts), drop sweep mention

fwd4 · fwd4 · commit 4d351ddcbb1a · 2026-06-26T22:09:40.000+08:00
diff --git a/docs/action_policy_libero_sft.md b/docs/action_policy_libero_sft.md
@@ -46,8 +46,7 @@ bash examples/launch_sft_action_policy_libero.sh   # HSDP 2x8; set NNODES/NODE_R
 
 Recipe knobs live in `action_policy_libero_nano`; the TOML sets run-level scalars
 (lr 5e-5, warmup 500, cycle 16000, `save_iter=500`, HSDP 2x8). Global batch is
-2048 = `max_samples_per_batch` 128 × 16 ranks × grad_accum 1. Sweep the saved
-checkpoints to pick the best iteration.
+2048 = `max_samples_per_batch` 128 × 16 ranks × grad_accum 1.
 
 ## 3. Closed-loop eval
 
diff --git a/examples/launch_sft_action_policy_libero.sh b/examples/launch_sft_action_policy_libero.sh
@@ -23,7 +23,7 @@
 #   hf download nvidia/LIBERO_LeRobot_v3 --repo-type dataset --include 'libero_10/**' --local-dir <dir>
 #   export LIBERO_ROOT=<dir>/libero_10
 #
-# Usage (8-GPU allocation, inside the training container, from the repo root):
+# Usage (HSDP 2x8; set NNODES/NODE_RANK/MASTER_ADDR per node):
 #   LIBERO_ROOT=<dir>/libero_10 bash examples/launch_sft_action_policy_libero.sh
 
 TOML_FILE="examples/toml/sft_config/action_policy_libero_repro.toml"

Original file line number	Diff line number	Diff line change
`@@ -23,7 +23,7 @@`
`23`	`23`	`# hf download nvidia/LIBERO_LeRobot_v3 --repo-type dataset --include 'libero_10/**' --local-dir <dir>`
`24`	`24`	`# export LIBERO_ROOT=<dir>/libero_10`
`25`	`25`	`#`
`26`		`-# Usage (8-GPU allocation, inside the training container, from the repo root):`
	`26`	`+# Usage (HSDP 2x8; set NNODES/NODE_RANK/MASTER_ADDR per node):`
`27`	`27`	`# LIBERO_ROOT=<dir>/libero_10 bash examples/launch_sft_action_policy_libero.sh`
`28`	`28`
`29`	`29`	`TOML_FILE="examples/toml/sft_config/action_policy_libero_repro.toml"`