challenge-baseline-model

name	challenge-baseline-model
description	Provision and launch the Simulation Challenge baseline inference model end to end: clone the inference code from a given git repo/branch, download the checkpoints from ModelScope into the repo's local checkpoints path, install deps, and start the inference agent. Trigger: When the user asks to "拉取 baseline 模型", "下载推理代码/权重", "搭一个 baseline", "clone the inference repo", "download ckpts", "set up the baseline model", "跑起来 baseline", "provision the baseline", or hands over a repo URL to stand up an inference server. Load even on light phrasings: "部署一下 demo", "部署 demo", "跑个 demo", "deploy a/the demo", "下载一下 baseline", "下载 baseline", "拉一下 baseline".
metadata	{"author":"zy","version":"1.0"}

challenge-baseline-model — Provision & run the baseline inference model

End-to-end setup of a contestant inference model: clone code → download checkpoints → install deps → run agent. Once running, the agent serves a Simulation Challenge job — hand off to challenge-run-agent for the launch contract and challenge-inference-protocol for the obs/action wire format.

⚠️ The CONFIG block below holds DEBUG/example values (provided for bring-up). Replace the repo URL, branch, and checkpoint specs with the official ones when ready — everything downstream reads from these variables, so it's a single-place edit.

---

CONFIG — edit these (current = debug example)

# --- code ---
REPO_URL="https://github.com/Anonymous-694/ACoT-VLA.git"   # clone URL (strip GitHub /tree/<branch>)
REPO_BRANCH="agibot_world_challenge"
# Local clone target (= inference repo root). If the user doesn't specify a path, default to the
# current directory — i.e. assume you're already inside (or want the repo placed in) the cwd.
REPO_DIR="${REPO_DIR:-$(pwd)}"

# --- checkpoints ---
# Downloaded from ModelScope (agibot_world/GenieSim3.0-Dataset, checkpoints/<name>) via the bundled
# scripts/download_checkpoint.sh — NO rsync / SSH needed, only `pip install modelscope`.
# The downloader ships with this skill at scripts/download_checkpoint.sh and is also in the
# genie-sim repo at ./scripts/download_checkpoint.sh.
DL_CKPT="$(dirname "$0")/scripts/download_checkpoint.sh"   # or hard-code this skill's scripts/ path

# Which checkpoints to fetch. Each name maps to a board (see table below). Pull only what the job
# needs — list one or more of: instruction_and_robust_pi05 / manipulation_pi05 / spatial_pi05
CKPT_NAMES=(
  "instruction_and_robust_pi05"
  # "manipulation_pi05"
  # "spatial_pi05"
)

Checkpoint ↔ board map (config.board → ModelScope checkpoint name):

Job board	Checkpoint name (checkpoints/<name>)
instruction	instruction_and_robust_pi05
robust	instruction_and_robust_pi05
manip	manipulation_pi05
spatial	spatial_pi05

---

Step 1 — Clone the inference code

if [ -d "$REPO_DIR/.git" ]; then
  echo "repo exists at $REPO_DIR — fetching $REPO_BRANCH"
  git -C "$REPO_DIR" fetch origin "$REPO_BRANCH" && git -C "$REPO_DIR" checkout "$REPO_BRANCH" && git -C "$REPO_DIR" pull --ff-only
else
  git clone -b "$REPO_BRANCH" "$REPO_URL" "$REPO_DIR"
fi

(/tree/<branch> in a GitHub web URL just means that branch — clone the .git URL with -b.)

Step 2 — Download checkpoints (ModelScope)

Fetch each needed checkpoint into $REPO_DIR/checkpoints/<name> with the bundled download_checkpoint.sh. It pulls from ModelScope — no SSH/rsync, no password — only the modelscope CLI (pip install modelscope if missing).

for name in "${CKPT_NAMES[@]}"; do
  "$DL_CKPT" "$name" "$REPO_DIR/checkpoints"
done

Each lands at $REPO_DIR/checkpoints/<name>/ (e.g. checkpoints/instruction_and_robust_pi05/). Single-board example:

"$DL_CKPT" instruction_and_robust_pi05 "$REPO_DIR/checkpoints"

Script signature: download_checkpoint.sh [CHECKPOINT_NAME] [LOCAL_DIR] — valid names are instruction_and_robust_pi05 / manipulation_pi05 / spatial_pi05; output goes to <LOCAL_DIR>/<name>/. download_checkpoint.sh -h lists them.

Layout note: there's no step subdir — params/ (and assets/, _CHECKPOINT_METADATA) sit directly under checkpoints/<name>_pi05/. tunnel.sh selects the ckpt by board via PI05_BOARD and points at the _pi05-suffixed dirs directly (instruction/robust → checkpoints/instruction_and_robust_pi05, spatial → checkpoints/spatial_pi05, manip → checkpoints/manipulation_pi05), so just download into checkpoints/ and no symlink is needed. Override the resolved path with PI05_CKPT_DIR if your layout differs. See challenge-inference-protocol §3 for the per-board ckpt layout.

Step 3 — Install dependencies

cd "$REPO_DIR"
uv sync            # or: pip install -e packages/openpi-client && pip install -r requirements

Step 4 — Run the inference agent

Needs a live job (challenge-submit-job → JOB_UUID, TUNNEL_ENDPOINT) and CHALLENGE_TOKEN. Pick the board to match the job's config.board:

[ -f ~/.simubotix-challenge.env ] && . ~/.simubotix-challenge.env
cd "$REPO_DIR"
PI05_BOARD=instruction ./scripts/tunnel.sh 0 "$JOB_UUID" "$TUNNEL_ENDPOINT"

Verify it reaches state -> RUNNING, then cases stream in. For concurrency, GPU pinning, and the drain/exit lifecycle, follow challenge-run-agent. For decoding what flows over the wire, see challenge-inference-protocol.

VRAM footprint & packing multiple instances on one GPU

A single pi05 inference service uses < 8 GB of VRAM at runtime, so a 24 GB card (e.g. RTX 4090) fits at most 3 services concurrently (3 is the practical max for pi05 — more will OOM under load). Pass the same GPU index as the first arg to each tunnel.sh and launch them as separate processes (different JOB_UUIDs):

PI05_BOARD=instruction ./scripts/tunnel.sh 0 "$JOB_UUID_A" "$TUNNEL_ENDPOINT" &
PI05_BOARD=instruction ./scripts/tunnel.sh 0 "$JOB_UUID_B" "$TUNNEL_ENDPOINT" &
PI05_BOARD=instruction ./scripts/tunnel.sh 0 "$JOB_UUID_C" "$TUNNEL_ENDPOINT" &

The reference tunnel.sh already sets XLA_PYTHON_CLIENT_PREALLOCATE=false and XLA_PYTHON_CLIENT_ALLOCATOR=platform (on-demand allocation, good for sharing), but it hardcodes XLA_PYTHON_CLIENT_MEM_FRACTION=0.48 (≈11.5 GB per process) — too high to pack three.

Derive the per-process cap from the concurrency the user asks for, instead of a fixed number. The user tells you how many services they want on one GPU (call it N, max 3 for pi05 on 24 GB); split the GPU evenly, keeping ~10% headroom:

mem_fraction = 0.9 / N      # clamped to [0.10, 0.90], then leave PREALLOCATE=false

User-requested N (per GPU)	XLA_PYTHON_CLIENT_MEM_FRACTION	per-process cap on 24 GB
1	0.90	~21.6 GB (effectively uncapped)
2	0.45	~10.8 GB
3 (max)	0.30	~7.2 GB

N > 3 on a single 24 GB card is not supported for pi05 — runtime usage (~8 GB each) overruns the card even though the fraction math would compute a smaller ceiling.

Since one service only touches < 8 GB at runtime and PREALLOCATE=false, the fraction is just a ceiling — sizing it to 0.9/N reserves an even slice per neighbor without wasting memory.

Make tunnel.sh compute this from a concurrency env var (honor an explicit override if set):

# in scripts/tunnel.sh, replacing the hardcoded `export XLA_PYTHON_CLIENT_MEM_FRACTION=0.48`
N=${PI05_PARALLELISM:-1}                       # services sharing this GPU (from the user's request)
frac=$(awk -v n="$N" 'BEGIN{f=0.9/n; if(f>0.9)f=0.9; if(f<0.1)f=0.1; printf "%.2f", f}')
export XLA_PYTHON_CLIENT_MEM_FRACTION=${XLA_PYTHON_CLIENT_MEM_FRACTION:-$frac}

Then launch N services on the same GPU index, telling each how many share the card:

N=3
for i in $(seq 1 "$N"); do
  PI05_PARALLELISM=$N PI05_BOARD=instruction \
    ./scripts/tunnel.sh 0 "${JOB_UUIDS[$i]}" "$TUNNEL_ENDPOINT" &
done

(PI05_PARALLELISM is the divisor; an explicit XLA_PYTHON_CLIENT_MEM_FRACTION=... still wins if you want to override the computed value for a specific launch.)

---

Notes

• Repo path default: if the user doesn't give an inference-code path, REPO_DIR defaults to the current directory ($(pwd)) — assume you're already in (or want the repo in) the cwd. Override by exporting REPO_DIR or editing the CONFIG line.
• Swapping to the official model: edit only the CONFIG block (repo URL/branch + CKPT_NAMES). Add manipulation_pi05 / spatial_pi05 to CKPT_NAMES to fetch those boards.
• Idempotent: Step 1 re-fetches if the repo already exists; Step 2's modelscope download caches/resumes, so re-running continues rather than restarting.
• No SSH needed: checkpoints come from ModelScope over HTTP — the only prerequisite is pip install modelscope. Downloads can be large; run Step 2 with run_in_background if driving from the assistant.
• One service < 8 GB VRAM: a 24 GB GPU (RTX 4090) hosts at most 3 pi05 services at once. Don't hardcode the mem fraction — derive it from the concurrency the user asks for: XLA_PYTHON_CLIENT_MEM_FRACTION = 0.9 / N (N = services per GPU, max 3; 2→0.45, 3→0.30), keep PREALLOCATE=false. tunnel.sh reads this from PI05_PARALLELISM. See Step 4's VRAM footprint note.