chore: drop one-off bench/build/train scripts from the PR

Remove development-only tooling that doesn't belong in the PR:
- examples/benchmark/* (pi052 step/kernel benchmark slurm + harness)
- examples/port_datasets/slurm_build_robocasa_composite_seen.py and
  src/lerobot/scripts/build_robocasa_composite_seen.py (composite_seen
  dataset build scripts)
- scripts/build_episode_filter.py, scripts/build_robocasa_smoke.sh,
  scripts/train_pi052_human300_exclude_unannotated.sh

None are imported by the library, tests, or entry points.

Co-authored-by: Cursor <cursoragent@cursor.com>

scripts/build_episode_filter.py

@@ -1,162 +0,0 @@
-#!/usr/bin/env python
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Emit the ``--dataset.episodes`` include-list for a LeRobot dataset, minus a
-set of excluded episode indices.
-
-``LeRobotDatasetConfig.episodes`` is an *include* list (train only on the listed
-episodes), so "exclude episode X" means "pass every episode except X". This
-helper builds that complement.
-
-For ``pepijn223/robocasa_pretrain_human300_v4`` the default exclusion set is the
-63 episodes that carry NO ``subtask`` annotation (in fact no persistent language
-rows at all) — see the scan in this PR's discussion. Training the steerable
-SmolVLA/pi052 policy on those episodes would feed it frames with empty subtask
-targets, so we drop them.
-
-Usage (prints a compact ``[0,1,2,...]`` list to stdout, logs to stderr):
-
-    python scripts/build_episode_filter.py \
-        --repo-id pepijn223/robocasa_pretrain_human300_v4
-
-    # capture in a shell script
-    EPISODES=$(python scripts/build_episode_filter.py --repo-id <id>)
-    lerobot-train ... --dataset.episodes="$EPISODES"
-
-The helper reads ``meta/info.json`` from the Hub to learn ``total_episodes`` and
-validates that every excluded index is in ``[0, total_episodes)`` before emitting
-the complement. Pass ``--no-validate-hub`` to skip the network round-trip and use
-``--total-episodes`` directly (e.g. for an offline / local dataset).
-"""
-
-from __future__ import annotations
-
-import argparse
-import json
-import sys
-
-# Episodes in pepijn223/robocasa_pretrain_human300_v4 with no `subtask`
-# annotation (no persistent language rows at all). 63 episodes / 179,009 frames.
-DEFAULT_EXCLUDE: tuple[int, ...] = (
-    1065, 2972, 6971, 8129, 9167, 9170, 9171, 9177, 9190, 9196, 9199, 9204,
-    9207, 9208, 9210, 9217, 9232, 9234, 9240, 9243, 9254, 9256, 9258, 9259,
-    9261, 9263, 9264, 15928, 16350, 18729, 20026, 21703, 25314, 25319, 25321,
-    25324, 25333, 25340, 25356, 25366, 25374, 25388, 25392, 25825, 25893,
-    26347, 26357, 26374, 26375, 26388, 26394, 26398, 26400, 26409, 26422,
-    26423, 26426, 26895, 26905, 26915, 26954, 27064, 30812,
-)
-
-
-def _log(msg: str) -> None:
-    print(msg, file=sys.stderr, flush=True)
-
-
-def _total_episodes_from_hub(repo_id: str, revision: str | None) -> int:
-    """Return ``total_episodes`` from the dataset's ``meta/info.json`` on the Hub."""
-    from huggingface_hub import hf_hub_download
-
-    path = hf_hub_download(
-        repo_id=repo_id,
-        filename="meta/info.json",
-        repo_type="dataset",
-        revision=revision,
-    )
-    with open(path) as f:
-        info = json.load(f)
-    total = int(info["total_episodes"])
-    if total <= 0:
-        raise ValueError(f"info.json reports non-positive total_episodes={total!r}")
-    return total
-
-
-def build_include_list(total_episodes: int, exclude: set[int]) -> list[int]:
-    """Return ``[0, total_episodes)`` with ``exclude`` removed, ascending."""
-    out_of_range = sorted(e for e in exclude if e < 0 or e >= total_episodes)
-    if out_of_range:
-        raise ValueError(
-            f"{len(out_of_range)} excluded index(es) outside [0, {total_episodes}): "
-            f"{out_of_range[:10]}{'...' if len(out_of_range) > 10 else ''}. "
-            "The dataset may have changed — re-run the subtask scan before training."
-        )
-    return [e for e in range(total_episodes) if e not in exclude]
-
-
-def main() -> int:
-    p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
-    p.add_argument("--repo-id", default="pepijn223/robocasa_pretrain_human300_v4")
-    p.add_argument("--revision", default=None, help="Dataset revision/branch (default: main).")
-    p.add_argument(
-        "--exclude-file",
-        default=None,
-        help="Optional JSON file with a list of episode indices to exclude. "
-        "Overrides the built-in default set.",
-    )
-    p.add_argument(
-        "--total-episodes",
-        type=int,
-        default=None,
-        help="Total episode count. If omitted, read from meta/info.json on the Hub.",
-    )
-    p.add_argument(
-        "--no-validate-hub",
-        action="store_true",
-        help="Do not fetch info.json from the Hub; requires --total-episodes.",
-    )
-    p.add_argument(
-        "--out",
-        default=None,
-        help="Write the list to this file instead of stdout.",
-    )
-    args = p.parse_args()
-
-    if args.exclude_file:
-        with open(args.exclude_file) as f:
-            data = json.load(f)
-        # Accept either a bare list or the {"missing_episode_indices": [...]} report shape.
-        exclude = set(data["missing_episode_indices"] if isinstance(data, dict) else data)
-    else:
-        exclude = set(DEFAULT_EXCLUDE)
-
-    if args.total_episodes is not None:
-        total = args.total_episodes
-        if not args.no_validate_hub:
-            hub_total = _total_episodes_from_hub(args.repo_id, args.revision)
-            if hub_total != total:
-                raise ValueError(
-                    f"--total-episodes={total} disagrees with Hub info.json total_episodes={hub_total}."
-                )
-    else:
-        if args.no_validate_hub:
-            raise SystemExit("--no-validate-hub requires --total-episodes.")
-        total = _total_episodes_from_hub(args.repo_id, args.revision)
-
-    include = build_include_list(total, exclude)
-    _log(
-        f"[build_episode_filter] repo={args.repo_id} total={total} "
-        f"excluded={len(exclude)} kept={len(include)}"
-    )
-
-    # Compact JSON (no spaces) so the resulting CLI arg stays as short as possible.
-    payload = "[" + ",".join(map(str, include)) + "]"
-    if args.out:
-        with open(args.out, "w") as f:
-            f.write(payload)
-        _log(f"[build_episode_filter] wrote {len(payload)} bytes to {args.out}")
-    else:
-        sys.stdout.write(payload)
-    return 0
-
-
-if __name__ == "__main__":
-    raise SystemExit(main())

scripts/build_robocasa_smoke.sh

@@ -1,47 +0,0 @@
-#!/bin/bash
-# Build a tiny RoboCasa smoke dataset (2 short atomic tasks, all episodes) for
-# fast end-to-end training validation before the real run.
-#
-# Defaults: target/human, OpenStandMixerHead + NavigateKitchen (~1k episodes,
-# ~131k frames, ~109 min @ 20 fps), 2 SLURM workers on hopper-cpu.
-#
-# Override via env: TASKS, REPO_ID, WORK_DIR, WORKERS, CPUS, PARTITION, LOCAL=1.
-
-set -euo pipefail
-
-cd "${LEROBOT_ROOT:-$HOME/lerobot}"
-source ~/miniconda3/etc/profile.d/conda.sh
-conda activate lerobot
-
-REPO_ID="${REPO_ID:-${HF_USER:?HF_USER is unset}/robocasa_smoke_2atomic_v3}"
-WORK_DIR="${WORK_DIR:-/fsx/${USER}/robocasa/datasets/v1.0}"
-ROBOCASA_ROOT="${ROBOCASA_ROOT:-/fsx/${USER}/robocasa}"
-LOGS_DIR="${LOGS_DIR:-/fsx/${USER}/logs/robocasa}"
-TASKS="${TASKS:-OpenStandMixerHead NavigateKitchen}"
-WORKERS="${WORKERS:-2}"
-CPUS="${CPUS:-8}"
-PARTITION="${PARTITION:-hopper-cpu}"
-LOCAL="${LOCAL:-0}"
-
-ARGS=(
-    examples/port_datasets/slurm_build_robocasa_composite_seen.py
-    --repo-id="$REPO_ID"
-    --work-dir="$WORK_DIR"
-    --robocasa-root="$ROBOCASA_ROOT"
-    --split=target --source=human
-    --tasks $TASKS
-    --workers="$WORKERS"
-    --cpus-per-task="$CPUS"
-    --partition="$PARTITION"
-    --mem-per-cpu=4G
-    --time=04:00:00
-    --logs-dir="$LOGS_DIR"
-    --job-name=port_robocasa_smoke
-)
-if [[ "$LOCAL" == "1" ]]; then
-    ARGS+=(--slurm=0)
-fi
-
-echo "Smoke dataset: $REPO_ID"
-echo "Tasks: $TASKS"
-python "${ARGS[@]}"

scripts/train_pi052_human300_exclude_unannotated.sh

@@ -1,115 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=pi052-hirobot-robocasa-human300
-#SBATCH --partition=hopper-prod
-#SBATCH --qos=high
-#SBATCH --time=48:00:00
-#SBATCH --ntasks=1
-#SBATCH --gpus-per-task=8
-
-set -euo pipefail
-
-cd "${LEROBOT_ROOT:-$HOME/lerobot}"
-
-export LEROBOT_DEBUG_PREDS_EVERY=1000
-export PATH="$HOME/miniconda3/bin:$HOME/.local/bin:$PATH"
-export LD_LIBRARY_PATH="$HOME/miniconda3/lib:${LD_LIBRARY_PATH:-}"
-export NCCL_TIMEOUT="${NCCL_TIMEOUT:-1800}"
-export HF_HUB_DOWNLOAD_TIMEOUT="${HF_HUB_DOWNLOAD_TIMEOUT:-120}"
-export WANDB_INIT_TIMEOUT="${WANDB_INIT_TIMEOUT:-300}"
-export PYTORCH_CUDA_ALLOC_CONF="${PYTORCH_CUDA_ALLOC_CONF:-expandable_segments:True}"
-# Compile path: pin triton + inductor caches node-local. The shared
-# /fsx cache mixes kernels built against different glibc versions and
-# trips ``GLIBC_2.34 not found`` on hopper nodes (bench v3 confirmed).
-export TRITON_CACHE_DIR="/tmp/triton_${SLURM_JOB_ID}"
-export TORCHINDUCTOR_CACHE_DIR="/tmp/torchinductor_${SLURM_JOB_ID}"
-mkdir -p "$TRITON_CACHE_DIR" "$TORCHINDUCTOR_CACHE_DIR"
-
-# Non-fatal so an unstaged local hotfix doesn't kill the job. CI / clean
-# checkouts still fast-forward as before; dirty trees just keep their
-# in-flight changes (the working tree is what runs).
-git pull --ff-only || echo "[warn] git pull skipped — keeping working tree."
-python -m pip install -q --upgrade -e .
-python -m pip install -q --upgrade -e '.[pi]'
-python -m pip install -q --upgrade 'liger-kernel'
-
-# FlashAttention-2 is NOT installed. The pi052 dual-expert layer compute
-# uses SDPA (the block-bidirectional mask is unsupported by FA2 anyway),
-# and the only other consumer would be liger-kernel — which gracefully
-# degrades when flash_attn is absent. The previously-installed wheel was
-# built against a newer GLIBC than some hopper compute nodes provide
-# (job 22162586 on ip-26-0-162-14 hit ``GLIBC_2.32 not found``), so the
-# safest configuration is "not installed". To re-enable for the
-# downstream HF Gemma ``generate`` path, install a wheel matching the
-# node's libc — but verify on every assigned node first.
-
-DATASET="pepijn223/robocasa_pretrain_human300_v4"
-DATASET_REVISION="${DATASET_REVISION:-main}"
-POLICY_REPO_ID="pepijn223/pi052_robocasa_human300"
-JOB_NAME="pi052-hirobot-robocasa-human300"
-NUM_PROCESSES=8
-# BS=36 — fits ~72 GB / 80 GB, BS=36 × 8 GPUs = 288 effective.
-BATCH_SIZE=${BATCH_SIZE:-36}
-STEPS=${STEPS:-5000}
-RUN_ID="${SLURM_JOB_ID:-$(date +%Y%m%d_%H%M%S)}"
-OUTPUT_DIR="/fsx/pepijn/outputs/train/pi052_robocasa_human300_${RUN_ID}"
-
-# --- Exclude un-annotated episodes -----------------------------------------
-# 63 episodes in this dataset carry NO `subtask` annotation (no persistent
-# language rows at all). `--dataset.episodes` is an INCLUDE list, so we pass
-# the complement: every episode index except those 63. The helper reads
-# meta/info.json from the Hub to confirm total_episodes (32043) and validates
-# the excluded indices are in range before emitting the list. If the dataset
-# version changes such that the indices fall out of range, the helper aborts
-# the job rather than silently training on the wrong episodes.
-echo "Building episode include-list (excluding un-annotated episodes)..."
-EPISODES=$(python scripts/build_episode_filter.py \
-    --repo-id "$DATASET" \
-    --revision "$DATASET_REVISION")
-
-echo "Training pi052 on $DATASET with ${NUM_PROCESSES} GPUs, batch size ${BATCH_SIZE}/GPU, ${STEPS} steps"
-echo "Output directory: $OUTPUT_DIR"
-export LEROBOT_DUMP_RECIPE_SAMPLES=8
-
-accelerate launch --multi_gpu --num_processes="$NUM_PROCESSES" \
-    -m lerobot.scripts.lerobot_train \
-    --policy.type=pi052 \
-    --policy.pretrained_path=lerobot/pi05_base \
-    --policy.recipe_path=recipes/subtask_mem_vqa_robocasa.yaml \
-    --dataset.repo_id="$DATASET" \
-    --dataset.revision="$DATASET_REVISION" \
-    --dataset.episodes="$EPISODES" \
-    --dataset.video_backend=pyav \
-    --output_dir="$OUTPUT_DIR" \
-    --job_name="$JOB_NAME" \
-    --policy.repo_id="$POLICY_REPO_ID" \
-    --policy.compile_model=true \
-    --policy.compile_mode=default \
-    --policy.gradient_checkpointing=true \
-    --policy.device=cuda \
-    --policy.tokenizer_max_length=256 \
-    --policy.action_tokenizer_name=lerobot/fast-action-tokenizer \
-    --policy.chunk_size=30 \
-    --policy.n_action_steps=30 \
-    --policy.max_action_tokens=256 \
-    --steps="$STEPS" \
-    --policy.scheduler_decay_steps="$STEPS" \
-    --batch_size="$BATCH_SIZE" \
-    --wandb.enable=true \
-    --policy.dtype=bfloat16 \
-    --policy.optimizer_lr=5e-5 \
-    --policy.optimizer_grad_clip_norm=1.0 \
-    --policy.scheduler_decay_lr=5e-6 \
-    --policy.lm_head_lr_scale=5.0 \
-    --ema.enable=true \
-    --wandb.disable_artifact=true \
-    --wandb.project=hirobot \
-    --log_freq=100 \
-    --save_freq=5000 \
-    --num_workers=4 \
-    --prefetch_factor=4 \
-    --persistent_workers=true \
-    --dataset.image_transforms.enable=true \
-    --dataset.image_transforms.max_num_transforms=3 \
-    --dataset.image_transforms.random_order=true \
-    --policy.auto_fit_fast_tokenizer=true \
-    --policy.knowledge_insulation=true