Nebius H200 AlphaChip Runbook

Use this when local 16 GB VRAM is too small or too slow. AlphaChip's published Ariane-scale recipe used 8x V100 for training plus many CPU collect workers; a single H200 should be enough for first E1 experiments, but walltime will still depend on how many CPU collect jobs feed Reverb.

Machine shape

1 GPU training host: H200, Docker, NVIDIA container runtime, 200 GB disk.
1 CPU/Reverb host: 32+ vCPU, 100 GB disk.
Optional CPU collect pool: start with 32-96 vCPU total; scale collect jobs until the learner is no longer waiting for replay data.
Shared storage: mounted filesystem or object storage for ROOT_DIR.

Setup

git clone <this-repo> e1-chip
cd e1-chip/packages/chip
git clone https://github.com/google-research/circuit_training.git external/circuit_training
git -C external/circuit_training checkout r0.0.4
scripts/alphachip/build_container.sh

For a GPU image:

ALPHACHIP_GPU_IMAGE=1 scripts/alphachip/build_container.sh

The lighter current path is to build the normal AlphaChip image and add the CUDA 12.2 user-space wheels expected by TensorFlow 2.15:

scripts/alphachip/build_container.sh
ALPHACHIP_IMAGE=circuit_training:e1-r0.0.4-cuda-pip \
  scripts/alphachip/build_cuda_runtime_image.sh

If Docker's NVIDIA runtime is unavailable but the GPU devices are visible, the local wrappers can use direct device mounts:

ALPHACHIP_GPU_MODE=manual USE_GPU=True scripts/alphachip/run_toy_training.sh

First cloud smoke

USE_GPU=True NUM_COLLECT_JOBS=8 scripts/alphachip/run_toy_training.sh

E1 training shape

Prepare the E1 soft-macro benchmark from an OpenLane DEF:

ALPHACHIP_BENCH_DIR=/shared/alphachip/e1_softmacro_full \
  scripts/alphachip/prepare_e1_softmacro_benchmark.sh \
    --def pd/openlane/runs/<run>/46-openroad-detailedrouting/e1_chip_top.def \
    --cols 16 \
    --rows 16

Measure the OpenROAD-derived baseline:

ALPHACHIP_COMPARE_DIR=/shared/alphachip/e1_softmacro_full/compare \
  scripts/alphachip/compare_proxy_costs.sh /shared/alphachip/e1_softmacro_full

Run a first integrated GPU experiment:

ALPHACHIP_BENCH_DIR=/shared/alphachip/e1_softmacro_full \
ALPHACHIP_RUN_DIR=/shared/alphachip/e1_softmacro_full_train \
USE_GPU=True \
NUM_COLLECT_JOBS=8 \
SEQUENCE_LENGTH=257 \
OBS_MAX_NUM_NODES=512 \
OBS_MAX_NUM_EDGES=8192 \
OBS_MAX_GRID_SIZE=16 \
TRAIN_ITERATIONS=5 \
EPISODES_PER_ITERATION=16 \
PER_REPLICA_BATCH_SIZE=16 \
  scripts/alphachip/run_e1_softmacro_training.sh

The integrated runner starts Reverb, CPU collectors, the learner, and the upstream evaluator. The evaluator writes:

text

<ALPHACHIP_RUN_DIR>/run_00/eval_output/rl_opt_placement.plc

Compare the exported placement:

ALPHACHIP_PLC=/shared/alphachip/e1_softmacro_full_train/run_00/eval_output/rl_opt_placement.plc \
ALPHACHIP_COMPARE_DIR=/shared/alphachip/e1_softmacro_full/compare \
  scripts/alphachip/compare_proxy_costs.sh /shared/alphachip/e1_softmacro_full

Payload from this workstation

The current full E1 benchmark can be packaged for upload:

scripts/alphachip/package_nebius_payload.sh /tmp/e1-alphachip/e1_softmacro_full

This writes:

text

build/alphachip/nebius/e1_alphachip_payload.tar.gz

After extracting that archive on an H200 host, the remote one-command runner is:

NUM_COLLECT_JOBS=8 \
TRAIN_ITERATIONS=5 \
EPISODES_PER_ITERATION=16 \
PER_REPLICA_BATCH_SIZE=16 \
  scripts/alphachip/run_h200_payload.sh

run_h200_payload.sh builds circuit_training:e1-r0.0.4 first, then derives circuit_training:e1-r0.0.4-cuda-pip, runs the OpenROAD proxy baseline, trains, and re-runs the comparison against the evaluator-exported AlphaChip PLC.

The Nebius CLI is configured in this workspace, but federation auth may need to be refreshed before VM commands work:

nebius profile list
nebius compute instance list --parent-id project-e00kfz6cpr00q21z892vec

For larger runs, either keep the integrated runner or split jobs manually: ppo_reverb_server on the Reverb host, many ppo_collect jobs on CPU hosts, train_ppo --use_gpu on the H200 host, and learning.eval pointed at the same variable-container server. Match the upstream docs/ARIANE.md job split and tune sequence_length to the number of movable E1 soft macros.