nvalchemi-data-storage

nvalchemi Data Storage

Overview

nvalchemi provides a composable pipeline for persisting and loading atomic data:

Writer                          Reader
(AtomicData/Batch -> Zarr)      (Zarr -> dict[str, Tensor])
                                    |
                                Dataset
                                (dict -> AtomicData, load_batches, prefetch)
                                    |
                    optional MultiDataset composition
                                    |
                                DataLoader
                                (Batch iteration)

from nvalchemi.data.datapipes import (
    AtomicDataZarrWriter,
    AtomicDataZarrReader,
    Dataset,
    MultiDataset,
    DataLoader,
    MultiDatasetBatchSampler,
)

Writing Data

AtomicDataZarrWriter serializes AtomicData, list[AtomicData], or Batch into a Zarr store.

from nvalchemi.data import AtomicData, Batch
from nvalchemi.data.datapipes import AtomicDataZarrWriter
import torch

writer = AtomicDataZarrWriter("dataset.zarr")

# Write a single system
data = AtomicData(
    positions=torch.randn(10, 3),
    atomic_numbers=torch.ones(10, dtype=torch.long),
    energy=torch.tensor([[0.5]]),
)
writer.write(data)

# Write a list of systems
writer.write([data1, data2, data3])

# Write a Batch
batch = Batch.from_data_list([data1, data2])
writer.write(batch)

write() creates the store and refuses to run twice: a second call raises FileExistsError: Zarr store already exists at <path>. To add samples to an existing store use append(); to rebuild from scratch, write to a fresh path. Keep scripts re-runnable by doing one or the other explicitly.

Appending to an existing store

writer = AtomicDataZarrWriter("dataset.zarr")
writer.append(new_data)          # single AtomicData
writer.append([data1, data2])    # list
writer.append(batch)             # Batch

Adding custom arrays

writer.add_custom("my_feature", torch.randn(total_atoms, 32), level="atom")

Deleting and defragmenting

writer.delete([0, 2])   # soft-delete samples 0 and 2 (sets mask=False)
writer.defragment()      # rebuild store without deleted samples

Zarr store layout

dataset.zarr/
├── meta/
│   ├── atoms_ptr       # int64 [N+1] — cumulative node counts
│   ├── edges_ptr       # int64 [N+1] — cumulative edge counts
│   ├── samples_mask    # bool [N] — False = deleted
│   ├── atoms_mask      # bool [V_total]
│   └── edges_mask      # bool [E_total]
├── core/               # AtomicData fields
│   ├── atomic_numbers
│   ├── positions
│   └── ...
├── custom/             # user-defined arrays
└── .zattrs             # root metadata

Reading Data

Low-level: AtomicDataZarrReader

Returns raw dict[str, torch.Tensor] per sample with metadata.

from nvalchemi.data.datapipes import AtomicDataZarrReader

reader = AtomicDataZarrReader(
    "dataset.zarr",
    pin_memory=False,                  # pin tensors to page-locked memory
    include_index_in_metadata=True,    # add "index" key to metadata
)

# Access a sample
data_dict, metadata = reader[0]       # (dict[str, Tensor], dict)

len(reader)          # number of active (non-deleted) samples
reader.field_names   # list of field names in each sample
reader.close()       # release resources
reader.refresh()     # reload after external modifications

Mid-level: Dataset

Wraps a Reader and constructs AtomicData objects, with device transfer and prefetching.

from nvalchemi.data.datapipes import AtomicDataZarrReader, Dataset

reader = AtomicDataZarrReader("dataset.zarr")
ds = Dataset(
    reader,
    device="cuda",       # target device ("auto" picks CUDA if available)
    num_workers=2,       # thread pool size for prefetching
)

# Get a sample
atomic_data, metadata = ds[0]   # AtomicData on target device

# Lightweight metadata (no full construction)
num_atoms, num_edges = ds.get_metadata(0)

# Explicit batch loading. This is the canonical synchronous batch API.
batches = ds.load_batches([[0, 3, 2], [4, 1, 5]])
batch0 = batches[0]

len(ds)    # number of samples
ds.close()

# Context manager
with Dataset(reader, device="cuda") as ds:
    data, meta = ds[0]

Prefetching with CUDA streams

ds = Dataset(reader, device="cuda")

# Prefetch a single sample
stream = torch.cuda.Stream()
ds.prefetch(0, stream=stream)
atomic_data, meta = ds[0]   # waits for prefetch to complete

# Prefetch multiple samples
streams = [torch.cuda.Stream() for _ in range(4)]
ds.prefetch_batch([0, 1, 2, 3], streams=streams)

# Cancel pending prefetches
ds.cancel_prefetch()       # cancel all
ds.cancel_prefetch(0)      # cancel specific index

In-memory datasets

When advising on dataset choice, suggest InMemoryDataset if the full dataset is small enough to fit comfortably in host memory. A good rule of thumb is "on the order of a few GB after batching." This avoids storage I/O after startup and can speed up training or benchmarking.

If the dataset is larger than host memory, or if keeping an extra resident copy would pressure the training job, recommend regular reader-backed Dataset instead so samples are loaded from storage on demand.

from nvalchemi.data.datapipes import AtomicDataZarrReader, InMemoryDataset

reader = AtomicDataZarrReader("dataset.zarr")
ds = InMemoryDataset(
    reader=reader,
    chunk_size=32768,
    device="cuda",          # emitted batch target; resident cache stays on CPU
    skip_validation=True,   # only for trusted toolkit-written stores
)

High-level: DataLoader

Iterates over a batch-loadable dataset in batches, producing Batch objects.

from nvalchemi.data.datapipes import AtomicDataZarrReader, Dataset, DataLoader

reader = AtomicDataZarrReader("dataset.zarr", pin_memory=True)
ds = Dataset(reader, device="cuda", num_workers=1)

loader = DataLoader(
    ds,
    batch_size=32,
    shuffle=True,
    drop_last=False,
    sampler=None,              # optional torch Sampler
    prefetch_factor=16,        # fuse 16 batches per read_many call
    num_streams=2,             # CUDA streams for prefetching
    use_streams=True,          # enable stream prefetching
)

# For throughput tuning (skip_validation, prefetch_factor, chunk/shard
# sizing), see the nvalchemi-zarr-perf skill.

for batch in loader:
    # batch is a Batch with concatenated tensors on target device
    print(batch.num_graphs, batch.num_nodes)

len(loader)                    # number of batches
loader.set_epoch(epoch)        # for distributed sampler

Use prefetch_factor=0 to disable async fused prefetch while still reading each emitted batch through Dataset.load_batches([indices]). For explicit/manual batch reads, use load_batches(...).

Composing multiple datasets

Use MultiDataset to concatenate multiple batch-loadable datasets, including reader-backed Dataset and InMemoryDataset, behind one global index space while keeping the same load_batches(...) fast path:

from nvalchemi.data.datapipes import (
    AtomicDataZarrReader,
    DataLoader,
    Dataset,
    MultiDataset,
    MultiDatasetBatchSampler,
)

ds_a = Dataset(AtomicDataZarrReader("dataset_a.zarr"), device="cuda")
ds_b = Dataset(AtomicDataZarrReader("dataset_b.zarr"), device="cuda")
dataset = MultiDataset(ds_a, ds_b, output_strict=True)

batch_sampler = MultiDatasetBatchSampler.balanced(
    dataset,
    batch_size=64,
    epoch_policy="max_size",  # oversample smaller datasets when replacement=True
    replacement=True,
)

loader = DataLoader(dataset, batch_sampler=batch_sampler, prefetch_factor=16)

Sampler notes:

• samples_per_dataset accepts integer counts or float ratios.
• epoch_policy="min_size" stops at the smallest contributing dataset.
• epoch_policy="max_size" covers the largest dataset and oversamples smaller datasets when replacement=True.

Custom Readers

Subclass Reader to support additional storage formats.

from nvalchemi.data.datapipes.backends.base import Reader

class MyReader(Reader):
    def __init__(self, path, **kwargs):
        super().__init__(**kwargs)
        self.path = path

    def _load_sample(self, index: int) -> dict[str, torch.Tensor]:
        """Load raw tensor dict for a single sample."""
        ...

    def _load_many_samples(self, indices) -> list[dict[str, torch.Tensor]]:
        """Optional fast path for coalesced batch reads."""
        ...

    def __len__(self) -> int:
        """Total number of samples."""
        ...

    # Optional overrides:
    def _get_sample_metadata(self, index: int) -> dict[str, Any]:
        """Per-sample metadata (default: empty dict)."""
        ...

    def _get_field_names(self) -> list[str]:
        """List of field names in each sample."""
        ...

    def close(self):
        """Release resources."""
        ...

Custom readers plug directly into Dataset and DataLoader:

reader = MyReader("data/", pin_memory=True)
ds = Dataset(reader, device="cuda")
loader = DataLoader(ds, batch_size=16)

Full Workflow Example

import torch
from nvalchemi.data import AtomicData, Batch
from nvalchemi.data.datapipes import (
    AtomicDataZarrWriter,
    AtomicDataZarrReader,
    Dataset,
    DataLoader,
)

# --- Write ---
data_list = [
    AtomicData(
        positions=torch.randn(n, 3),
        atomic_numbers=torch.ones(n, dtype=torch.long),
        energy=torch.tensor([[float(i)]]),
    )
    for i, n in enumerate([5, 8, 3, 12])
]

writer = AtomicDataZarrWriter("train.zarr")
writer.write(data_list)

# Append more later
writer.append(AtomicData(
    positions=torch.randn(6, 3),
    atomic_numbers=torch.ones(6, dtype=torch.long),
))

# --- Read & Train ---
reader = AtomicDataZarrReader("train.zarr")
ds = Dataset(reader, device="cuda", num_workers=4)
loader = DataLoader(ds, batch_size=2, shuffle=True, prefetch_factor=2)

for epoch in range(10):
    loader.set_epoch(epoch)
    for batch in loader:
        energy = batch["energy"]          # [batch_size, 1]
        positions = batch["positions"]    # [total_nodes, 3]
        # ... model forward pass ...