Filling a block from another machine
One process fills blocks in a pipeline: it reads the recipe from this repo, claims a block, walks its voxels on
the local GPU, packs the shard in a subprocess on the CPU and uploads it in a thread -- while the next block
already walks. Machines coordinate through claim files in the repo, nothing else; a claim carries a heartbeat
(the walk's progress, ETA and memory, every ten minutes). Every walk is spooled batch by batch into its run
record, so a killed worker resumes a block from its finished batches, and --drain uploads what an earlier
worker on the host left finished.
# 1. dmipy-sim (public, github.com/dmrai-lab/dmipy-sim) at the manifest's commit, with a CUDA jaxlib
git clone https://github.com/dmrai-lab/dmipy-sim && cd dmipy-sim && git checkout 64bb9e7 # manifest code.commit
python3.11 -m venv ~/dmipy-venv && . ~/dmipy-venv/bin/activate
pip install -e ".[cuda12]" huggingface_hub
# the CUDA libraries jaxlib loads are in the venv's nvidia/*/lib directories
export LD_LIBRARY_PATH=$(ls -d ~/dmipy-venv/lib/python3.11/site-packages/nvidia/*/lib | tr '\n' ':')$LD_LIBRARY_PATH
export XLA_PYTHON_CLIENT_PREALLOCATE=false
python -c "import jax; print(jax.devices())" # must list the GPU
# 2. log in with a token that can write to SubstrateCommons/disco-replay
hf auth login
# 3. prove the machine on a scaled-down block (a few minutes; a proving run, not part of the pack)
python worker/disco_fill.py --block 0 --budget 30000 --smoke
# (once per variant, done for ground-truth) the certifying walk: block 0's voxels at 2e5 walkers, the full battery
python worker/disco_fill.py --block 0 --budget 200000 --certify
# 4. fill blocks until none is open, pipelined (a session with a time limit: --hours H claims nothing new after H hours);
# a plan in passes (manifest plan.passes) fills pass 1 of every block before pass 2; --pass K takes only pass K
python worker/disco_fill.py --next --loop
# after a crash or a kill: upload what was finished here, release the rest, then fill again
python worker/disco_fill.py --drain && python worker/disco_fill.py --next --loop
--variant selects a recipe from manifest.json (default default_variant); a pass of a block is claimed, walked
(from its own seed stream, at its scale of the block's counts) and uploaded like a block, as block-NNNN.pK.rpk;
the shards are staging: a finished pass is appended to the lossless columnar layout disco/columns
(python -m dmipy_sim.fill.consolidate --append, the union's weights) and its shards removed; --batch lowers the walker batch
on a small card (a GH200 runs 100k); --workdir holds the shard while it is built (5 GB free per block).
A worker claims --claim-batch blocks (default 3) in one commit and fills them in turn -- the hub allows 128
commits per repository per hour, and a block costs its claim and its upload -- and releases what is left of the
batch when it stops. Its claim carries a heartbeat every 15 min; a claim not refreshed for 45 min is a dead
worker's, and the next worker that claims releases it (--drain on the dead worker's own host does the same at
once, and uploads what it had finished). python -m dmipy_sim.fill.status --repo SubstrateCommons/disco-replay reads the hub -- every shard's summary, every claim's
heartbeat, the plan -- and prints the fill's state per contributor (shards, rate, the block in flight, the
heartbeat's age), the time to the end of each pass at the last hour's rate and the commit budget; with
--publish it commits that as STATUS.md. Every shard carries its own per-voxel
certificate, a .json summary with the timings and the sha256 of the shard, and a .run/ record of the walk
(python -m dmipy_sim.run <dir> reads it: the timeline, memory per phase, how it ended);
certificate/disco-block-0000.run/ is one.
A Kaggle GPU session as a worker
worker/kaggle/ holds a script kernel; every kaggle kernels push -p worker/kaggle starts a session that fills
blocks for about 11 hours in rounds of 60k walkers (a 31 GB host, 20 GB of disk, two T4s) and stops. A pushed version cannot
reach Kaggle's secrets service (measured), so the hub token lives in a PRIVATE dataset of the account, hf-token,
holding one file hf_token.txt, mounted into the session by dataset_sources; only the account sees it. Once:
pip install kaggle && kaggle auth login # the account's API login
mkdir hf-token && cp <a hub token with write access to this dataset> hf-token/hf_token.txt
echo '{"title": "hf-token", "id": "<user>/hf-token", "licenses": [{"name": "other"}]}' > hf-token/dataset-metadata.json
kaggle datasets create -p hf-token # private by default
Then, per session (from any machine with the login):
kaggle kernels push -p worker/kaggle # rhjfick/disco-fill; your own: edit the id and the dataset
kaggle kernels status rhjfick/disco-fill
kaggle kernels output rhjfick/disco-fill -p out/ # the session's log