nvalchemi-dynamics-implementation

작성자: nvidia

How to implement a dynamics integrator by subclassing BaseDynamics and overriding pre_update() and post_update() methods. Use when creating a custom…

npx skills add https://github.com/nvidia/nvalchemi-toolkit --skill nvalchemi-dynamics-implementation

nvalchemi Dynamics Implementation

Overview

To implement a dynamics class (integrator) in nvalchemi, subclass BaseDynamics and override two methods: pre_update() and post_update(). The base class handles the model forward pass, hook dispatch, convergence checking, and the step/run loop.

from nvalchemi.dynamics.base import BaseDynamics, ConvergenceHook
from nvalchemi.data import Batch

Step execution flow

Each call to step(batch) executes:

1. BEFORE_STEP hooks
2. BEFORE_PRE_UPDATE hooks  →  pre_update(batch)  →  AFTER_PRE_UPDATE hooks
3. BEFORE_COMPUTE hooks     →  compute(batch)      →  AFTER_COMPUTE hooks
4. BEFORE_POST_UPDATE hooks →  post_update(batch)  →  AFTER_POST_UPDATE hooks
5. AFTER_STEP hooks
6. Check convergence → ON_CONVERGE hooks if converged
7. Increment step_count
  • pre_update() and post_update() run inside torch.no_grad()
  • compute() calls the model forward pass and writes forces/energy to the batch in-place
  • You implement pre_update() and post_update(); everything else is inherited

Implementation guide

1. Define the class

Set __needs_keys__ (model outputs your integrator requires) and __provides_keys__ (state your integrator produces).

class MyDynamics(BaseDynamics):
    __needs_keys__: set[str] = {"forces"}
    __provides_keys__: set[str] = {"velocities", "positions"}

2. Implement __init__

Store integrator parameters. Always call super().__init__() and forward **kwargs (needed for cooperative multiple inheritance with the communication mixin).

def __init__(
    self,
    model: BaseModelMixin,
    n_steps: int,
    dt: float = 1.0,
    hooks: list[Hook] | None = None,
    convergence_hook: ConvergenceHook | dict | None = None,
    **kwargs: Any,
) -> None:
    super().__init__(
        model=model,
        hooks=hooks,
        convergence_hook=convergence_hook,
        n_steps=n_steps,
        **kwargs,
    )
    self.dt = dt

BaseDynamics constructor parameters:

ParameterTypeDescription
modelBaseModelMixinThe neural network potential
hookslist[Hook] | NoneHooks to register (organized by stage)
convergence_hookConvergenceHook | dict | NoneConvergence detection
n_stepsint | NoneDefault step count for run()
exit_statusintStatus value for graduated samples (default: 1)
**kwargsAnyForwarded to communication mixin

3. Implement pre_update(batch)

Update positions based on current velocities and forces. Modify the batch in-place.

def pre_update(self, batch: Batch) -> None:
    positions = batch.positions       # [V, 3]
    velocities = batch.velocities     # [V, 3]
    forces = batch.forces             # [V, 3] or None
    masses = batch.atomic_masses.unsqueeze(-1)  # [V] -> [V, 1]

    with torch.no_grad():
        if forces is not None and not torch.all(forces == 0):
            accelerations = forces / masses
            # x(t+dt) = x(t) + v(t)*dt + 0.5*a(t)*dt^2
            positions.add_(velocities * self.dt + 0.5 * accelerations * self.dt**2)
        else:
            # First step fallback (no forces yet)
            positions.add_(velocities * self.dt)

4. Implement post_update(batch)

Update velocities based on new forces (computed between pre_update and post_update by the inherited compute() method). Modify the batch in-place.

def post_update(self, batch: Batch) -> None:
    velocities = batch.velocities     # [V, 3]
    forces = batch.forces             # [V, 3]
    masses = batch.atomic_masses.unsqueeze(-1)

    with torch.no_grad():
        new_accelerations = forces / masses
        # v(t+dt) = v(t) + a(t+dt)*dt
        velocities.add_(new_accelerations * self.dt)

Inherited methods (do NOT override)

MethodDescription
compute(batch)Model forward pass → validates outputs → writes forces/energy to batch
step(batch)Full step with hook dispatch (see flow above)
run(batch, n_steps=None)Loop calling step() for n_steps iterations
register_hook(hook)Register a hook at its declared stage
_check_convergence(batch)Check convergence criteria, return converged indices
_validate_model_outputs(outputs)Verify __needs_keys__ are present in model output

Inherited attributes

AttributeTypeDescription
modelBaseModelMixinThe wrapped model
step_countintCurrent step (starts at 0, incremented after each step)
hooksdict[DynamicsStage, list[Hook]]Registered hooks by stage
convergence_hookConvergenceHook | NoneConvergence detector
n_stepsint | NoneDefault step count
exit_statusintStatus threshold for graduated samples
model_is_conservativeboolWhether forces use autograd

Usage

from nvalchemi.models.demo import DemoModelWrapper
from nvalchemi.data import AtomicData, Batch
import torch

# Create model and dynamics
model = DemoModelWrapper()
dynamics = MyDynamics(model=model, n_steps=100, dt=0.5)

# Create batch
data = AtomicData(
    atomic_numbers=torch.tensor([6, 6, 8], dtype=torch.long),
    positions=torch.randn(3, 3),
)
batch = Batch.from_data_list([data])

# Initialize required fields (forces/energy must exist for copy_())
batch.forces = torch.zeros(3, 3)
batch.energy = torch.zeros(1, 1)

# Run
result = dynamics.run(batch)
# Or step-by-step
dynamics.step(batch)

Complete example: Velocity Verlet

This mirrors DemoDynamics, the reference implementation.

from __future__ import annotations
from typing import Any, TYPE_CHECKING
import torch
from nvalchemi.data import Batch
from nvalchemi.dynamics.base import BaseDynamics, ConvergenceHook

if TYPE_CHECKING:
    from nvalchemi.dynamics.base import Hook
    from nvalchemi.models.base import BaseModelMixin


class VelocityVerlet(BaseDynamics):
    """Velocity Verlet integrator."""

    __needs_keys__: set[str] = {"forces"}
    __provides_keys__: set[str] = {"velocities", "positions"}

    def __init__(
        self,
        model: BaseModelMixin,
        n_steps: int,
        dt: float = 1.0,
        hooks: list[Hook] | None = None,
        convergence_hook: ConvergenceHook | dict | None = None,
        **kwargs: Any,
    ) -> None:
        super().__init__(
            model=model, hooks=hooks, convergence_hook=convergence_hook,
            n_steps=n_steps, **kwargs,
        )
        self.dt = dt
        self._prev_accelerations: torch.Tensor | None = None

    def pre_update(self, batch: Batch) -> None:
        """x(t+dt) = x(t) + v(t)*dt + 0.5*a(t)*dt^2"""
        positions = batch.positions
        velocities = batch.velocities
        forces = batch.forces
        masses = batch.atomic_masses.unsqueeze(-1)

        with torch.no_grad():
            if forces is not None and not torch.all(forces == 0):
                accelerations = forces / masses
                self._prev_accelerations = accelerations.clone()
                positions.add_(velocities * self.dt + 0.5 * accelerations * self.dt**2)
            else:
                positions.add_(velocities * self.dt)

    def post_update(self, batch: Batch) -> None:
        """v(t+dt) = v(t) + 0.5*(a(t) + a(t+dt))*dt"""
        velocities = batch.velocities
        forces = batch.forces
        masses = batch.atomic_masses.unsqueeze(-1)

        with torch.no_grad():
            new_accelerations = forces / masses
            if self._prev_accelerations is not None:
                velocities.add_(
                    0.5 * (self._prev_accelerations + new_accelerations) * self.dt
                )
            else:
                velocities.add_(new_accelerations * self.dt)

Convergence

Use ConvergenceHook to stop early or migrate samples in a pipeline:

from nvalchemi.dynamics.base import ConvergenceHook

hook = ConvergenceHook(
    criteria=[
        {"key": "fmax", "threshold": 0.05},
        {"key": "energy_change", "threshold": 1e-6},
    ],
    source_status=0,   # check samples with this status
    target_status=1,   # migrate converged samples to this status
    frequency=1,       # check every N steps
)

dynamics = MyDynamics(model=model, n_steps=1000, convergence_hook=hook)

Composition with FusedStage

Chain multiple dynamics stages that share a single model forward pass:

relax = MyDynamics(model, n_steps=100, dt=0.5)
md = MyDynamics(model, n_steps=500, dt=0.1)

# Compose with + operator
fused = relax + md

# Samples start in relax, converge, then move to md
fused.run(batch)

Distributed pipeline

Chain stages across ranks with the | operator:

opt_stage = MyDynamics(model, n_steps=100, dt=0.5)   # rank 0
md_stage = MyDynamics(model, n_steps=500, dt=0.1)    # rank 1

pipeline = opt_stage | md_stage

with pipeline:
    pipeline.run()

nvidia의 다른 스킬

compileiq-debug
nvidia
Use when something is wrong: Search() hangs, all evaluations return INVALID_SCORE, scores aren't improving, every config returns the same number, ptxas errors…
official
create-github-pr
nvidia
gh CLI를 사용하여 GitHub 풀 리퀘스트를 생성합니다. 사용자가 새 PR을 만들거나, 코드 리뷰를 제출하거나, 풀 리퀘스트를 열고자 할 때 사용합니다. 트리거 키워드 -…
official
diagnose-perf
nvidia
First-responder performance triage for Isaac Sim and Isaac Lab. Identifies bottleneck category (GPU-bound, CPU-bound, VRAM, loading) using nvidia-smi and…
official
eagle3-review-logs
nvidia
Review EAGLE3 pipeline experiment logs from the launcher's experiments/ directory. Summarizes pass/fail status for all 4 tasks, diagnoses failures with root…
official
nemoclaw-maintainer-cross-issue-sweep
nvidia
다른 열린 이슈들을 스캔하여 주어진 PR이 함께 수정하거나 실수로 망가뜨릴 수 있는 이슈를 찾습니다. 인접 수정 기회와 모순 위험을 file:line…과 함께 출력합니다.
official
karpathy-guidelines
nvidia
일반적인 LLM 코딩 실수를 줄이기 위한 행동 지침입니다. 코드 작성, 검토 또는 리팩토링 시 과도한 복잡성을 피하고 정밀한 변경을 위해 사용하세요.
official
fhir-basics
nvidia
에이전트에게 FHIR R4 API의 작동 방식, 사용 가능한 리소스, 검색 매개변수를 사용한 쿼리 방법, 모든 응답 형식을 올바르게 파싱하는 방법을 가르칩니다…
official
underdeclared-agent
nvidia
A helpful assistant agent
official