ContextQMD
Back to Eliza

Nebius H200 AlphaChip Runbook

packages/chip/scripts/alphachip/nebius_h200_runbook.md

2.0.11.6 KB
Original Source

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

sh
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:

sh
ALPHACHIP_GPU_IMAGE=1 scripts/alphachip/build_container.sh

First cloud smoke

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

E1 training shape

Once NETLIST_FILE and INIT_PLACEMENT point to converted E1 protobuf/PLC artifacts:

sh
export ROOT_DIR=/shared/alphachip/e1/run_001
export REVERB_SERVER=<reverb-host-ip>:8008
export NETLIST_FILE=/shared/alphachip/e1/netlist.pb.txt
export INIT_PLACEMENT=/shared/alphachip/e1/initial.plc

Run ppo_reverb_server on the Reverb host, many ppo_collect jobs on CPU hosts, and train_ppo --use_gpu on the H200 host. Match the upstream docs/ARIANE.md job split and tune sequence_length to the number of movable E1 macros or soft macros.