Nvidia Nvalchemi Toolkit Nvalchemi Toolkit

nvalchemi Data Structures

Overview

nvalchemi represents atomic systems as graphs using two core classes:

• AtomicData — a single atomic system (molecule, crystal, etc.)
• Batch — an efficient container of multiple AtomicData objects stored as concatenated tensors

Both are Pydantic BaseModel subclasses with DataMixin for device/dtype operations.

from nvalchemi.data import AtomicData, Batch

AtomicData

Construction

Required fields: positions [n_nodes, 3] and atomic_numbers [n_nodes].

import torch

# Minimal
data = AtomicData(
    positions=torch.randn(4, 3),
    atomic_numbers=torch.tensor([1, 6, 6, 1], dtype=torch.long),
)

# With edges (bonds or neighbor list)
data = AtomicData(
    positions=torch.randn(4, 3),
    atomic_numbers=torch.tensor([1, 6, 6, 1], dtype=torch.long),
    neighbor_list=torch.tensor([[0, 1], [1, 0], [1, 2], [2, 1]], dtype=torch.long),
)

# With system-level fields (energy, cell, pbc)
data = AtomicData(
    positions=torch.randn(4, 3),
    atomic_numbers=torch.tensor([1, 6, 6, 1], dtype=torch.long),
    energy=torch.tensor([[0.5]]),
    cell=torch.eye(3).unsqueeze(0),       # [1, 3, 3]
    pbc=torch.tensor([[True, True, False]]),  # [1, 3]
)

From ASE Atoms:

data = AtomicData.from_atoms(
    atoms,                    # ase.Atoms object
    energy_key="energy",      # key in atoms.info / atoms.calc
    forces_key="forces",
    device="cpu",
    dtype=torch.float32,
)

Field reference

Fields are organized by level. All are optional except positions and atomic_numbers.

Custom data can be stored in the info: dict[str, torch.Tensor] field.

Properties

data.num_nodes          # int — number of atoms
data.num_edges          # int — number of edges (0 if None)
data.device             # torch.device
data.dtype              # torch.dtype (of positions)
data.chemical_hash      # str — blake2s hash of structure/composition
data.node_properties    # dict of set node-level fields
data.edge_properties    # dict of set edge-level fields
data.system_properties  # dict of set system-level fields

Dict-like access

data["positions"]                # get attribute by name
data["positions"] = new_tensor   # set attribute by name

Adding custom properties

data.add_node_property("custom_feat", torch.randn(data.num_nodes, 4))
data.add_edge_property("edge_weights", torch.ones(data.num_edges))
data.add_system_property("temperature", torch.tensor([[300.0]]))

Device, clone, serialization

data.to("cuda")                         # move to device
data.to("cpu", dtype=torch.float64)     # move + cast
data.cpu()
data.cuda()
data.clone()                            # deep copy
data.model_dump(exclude_none=True)      # dict
data.model_dump_json()                  # JSON string

Equality

Two AtomicData objects are equal if they have the same chemical_hash:

data1 == data2  # compares by chemical_hash

Batch

Construction

data_list = [
    AtomicData(positions=torch.randn(2, 3), atomic_numbers=torch.ones(2, dtype=torch.long)),
    AtomicData(positions=torch.randn(3, 3), atomic_numbers=torch.ones(3, dtype=torch.long)),
]
batch = Batch.from_data_list(data_list)

# Exclude specific keys
batch = Batch.from_data_list(data_list, exclude_keys=["velocities"])

# Pre-allocated empty buffer (for high-performance use)
buffer = Batch.empty(
    num_systems=40, num_nodes=80, num_edges=80,
    template=data_list[0],  # defines schema
)

Size properties

batch.num_graphs            # number of graphs
batch.batch_size            # alias for num_graphs
batch.num_nodes             # total nodes across all graphs
batch.num_edges             # total edges across all graphs
batch.batch_idx             # Tensor [num_nodes] — per-node graph index
batch.batch_ptr             # Tensor [num_graphs+1] — cumulative node counts
batch.num_nodes_list        # list[int] — per-graph node counts
batch.num_edges_list        # list[int] — per-graph edge counts
batch.num_nodes_per_graph   # Tensor — per-graph node counts
batch.num_edges_per_graph   # Tensor — per-graph edge counts
batch.max_num_nodes         # int — max nodes in any graph
batch.system_capacity       # int — max graphs for pre-allocated batches

Indexing

# Single graph -> AtomicData
batch[0]
batch[-1]
batch.get_data(0)

# Sub-batch -> Batch
batch[1:3]                          # slice
batch[torch.tensor([0, 2])]        # int tensor
batch[[0, 2]]                       # list
batch[torch.tensor([True, False, True])]  # bool mask

# Attribute -> Tensor
batch["positions"]                  # concatenated positions from all graphs

# Reconstruct all graphs
all_graphs = batch.to_data_list()   # list[AtomicData]

Containment, length, iteration

"positions" in batch       # True
len(batch)                 # num_graphs
for key, tensor in batch:  # iterate (key, value) pairs
    ...

Mutation

# Add a new key (one value per graph)
batch.add_key("node_feat", [torch.randn(2, 4), torch.randn(3, 4)], level="node")
batch.add_key("temperature", [torch.tensor([[300.0]]), torch.tensor([[350.0]])], level="system")
batch.add_key("edge_attr", [torch.randn(1, 4), torch.randn(2, 4)], level="edge")

# Overwrite an existing key
batch.add_key("node_feat", new_values, level="node", overwrite=True)

# Concatenate batches (in-place)
batch.append(other_batch)
batch.append_data([more_atomic_data])

Pre-allocated buffer operations

For high-throughput workflows (e.g. streaming dynamics), use pre-allocated buffers:

# Create buffer
buffer = Batch.empty(num_systems=40, num_nodes=80, num_edges=80, template=data)

# Copy selected graphs into buffer
mask = torch.tensor([True, False])           # which src graphs to copy
copied_mask = torch.zeros(2, dtype=torch.bool)  # updated in-place: which actually fit
dest_mask = torch.zeros(buffer.system_capacity, dtype=torch.bool)
buffer.put(src_batch, mask, copied_mask=copied_mask, dest_mask=dest_mask)

# Remove copied graphs from source (compact in-place)
src_batch.defrag(copied_mask=copied_mask)

# Reset buffer for reuse
buffer.zero()

Device, clone, memory

batch.to("cuda")
batch.cpu()
batch.cuda()
batch.clone()
batch.contiguous()     # make all tensors contiguous
batch.pin_memory()     # pin for async host-to-device transfer

Serialization

batch.model_dump()                    # flat dict of all tensors + metadata
batch.model_dump(exclude_none=True)   # drop None-valued keys
batch.model_dump_json()               # JSON string

Distributed communication

Batch supports point-to-point distributed communication via torch.distributed. Data is sent in three phases: a metadata header (num_graphs, num_nodes, num_edges), per-group segment lengths, and bulk tensor data.

Blocking send/recv:

import torch.distributed as dist

# Sender (rank 0)
batch.send(dst=1, tag=0, group=None)

# Receiver (rank 1) — template provides schema (keys, dtypes, group structure)
received = Batch.recv(src=0, device="cuda", template=template_batch, tag=0)

Non-blocking send/recv:

# Sender — returns _BatchSendHandle
handle = batch.isend(dst=1, tag=0, group=None)
# ... do other work ...
handle.wait()  # block until all sends complete

# Receiver — returns _BatchRecvHandle
handle = Batch.irecv(src=0, device="cuda", template=template_batch, tag=0)
# ... do other work ...
received = handle.wait()  # block until data arrives, returns Batch

Key details:

• template is required on the receiver to know the attribute keys, dtypes, and group structure (atoms/edges/system). Cache it across calls.
• A 0-graph sentinel batch can be sent or received. Only the metadata header is transmitted.
• tag is a base tag incremented internally per group. Use distinct base tags for concurrent send/recv pairs.
• empty_like(batch) creates a 0-graph batch with the same schema, which is useful for sentinel signals.

sentinel = Batch.empty_like(batch, device="cuda")  # 0-graph, same schema
sentinel.send(dst=1)  # signal "no more data"

Round-trip

reconstructed = batch.to_data_list()
batch_again = Batch.from_data_list(reconstructed)

Source: NVIDIA/nvalchemi-toolkit — distributed by TomeVault.