Repository navigation
Releases
remote_handler 0.8.0
[0.8.0] — 2026-10-30
Added
on_stage_completehooks forStagedActiveLearningHandler. Each hook
is called with aStageCompletedevent (stage, iteration,calchome,
recorded outputs,is_last) once a stage's outputs are recorded. Use it
for side work such as rendering a dashboard. Hooks are not recorded in the
progress file, so a resume doesn't re-run them for finished steps. An
exception in a hook is logged and does not stop the loop.- ACE warm start.
AceModelTrainer.update()takesinitial_potentials
(e.g. the previous iteration's committee, wired via aStage's
extra_inputs). Membericontinues fitting from
initial_potentials[i % len(...)](pacemaker's-p) instead of starting
from scratch. Ladder fitting (fit.ladder_step/ladder_type) is
supported for fits from scratch; on a warm start those keys are dropped,
since a ladder would restart from an empty basis. keep_bestfor ensemble trainers. Trainnum_modelsmembers but hand
only thekeep_bestwith the lowest final test loss to the next stage,
best first. Implemented for ACE (pacemaker'stest_metrics.txt); other
trainers reject the option. Ensemble trainers also takenum_parallel
(default unchanged:min(num_models, 5)).- Relative high-energy penalty for NeuralIL.
high_energy_penalty_relative
sets the threshold in eV/atom above the training set's lowest energy per
atom, so it no longer depends on the system, POTCARs or functional. It
cannot be combined with the absolutehigh_energy_penalty. ClusterConfig.remote_fs()opens aRemoteFileSystemto the cluster
with itsconnection_policyalready applied.
Changed
- NeuralIL's default high-energy penalty is now relative:
high_energy_penalty_relative = 5.0replaces the absolute
high_energy_penalty = -5.0. Settinghigh_energy_penaltyexplicitly
restores the old behaviour. - The VASP single-point payload rejects unconverged SCF runs. A run whose
last electronic loop stopped at NELM without reaching EDIFF still produced
energies and forces, which were used as labels. It now writes
scf_unconverged.info, no labelled output and no success marker, and
exits non-zero, so the evaluator'srequired_fractiondecides whether the
loop continues. Convergence is read from VASP 6's OUTCAR message, falling
back to comparing OSZICAR's step count against NELM for VASP 5.
Fixed
- Schedule values now reach the actors. The handlers passed
Schedule.get_kwargs()toupdate(**kwargs), but most actors never read
those kwargs, so a scheduledconfig_file,incar_kwargs,n_configs,
... was silently ignored and every iteration used the initial value. Both
StagedActiveLearningHandlerand the legacyActiveLearningHandlernow
apply each iteration's schedule to the actor beforeupdate()(as
attributes, via the newapply_schedule_kwargs) and still forward the
values as keyword arguments. Actors that bake a parameter in at
construction handle it themselves: the VASP/ORCA/EMT evaluators rebuild
calc_config,NeuralilILTrainerre-merges its hyperparameters (and
accepts single hyperparameter keys),SelectorActorforwards keys to the
wrapped selector, andRemoteSelectorforwards them to the selector
constructor as before. - Misspelt schedule keys are flagged at handler construction. A key that
names no attribute of the actor is still only passed toupdate(), as
before, so actors that read schedule values fromupdate(**kwargs)keep
working unchanged. The handler now also emits aScheduleKeyWarningfor
each such key, once, before anything is submitted. Turn it into an error
withwarnings.simplefilter("error", ScheduleKeyWarning). An actor can
declare the keys itsupdate()reads by defining
unknown_schedule_keys(kwargs), which silences the warning (see
docs/source/active_learning.rst).
DownstreamScheduled*wrapper classes that copy kwargs onto the actor
are no longer needed. They still work, since the values are just set
twice. - Reattached array tasks are no longer collected while still running.
from_manifest()never re-attached the executor to the recorded job, so
wait_if_neededpolled the success markers instead of the queue. With a
fraction-basedfinish_check(e.g.RemoteAtomsHandler, 80% by default),
a reattached array was collected as soon as enough tasks had finished,
with the rest still running. Reattached tasks now wait for the job itself
to finish, then applyfinish_checkonce. Old SLURM manifests without a
job_idfall back to polling the markers, with a warning. - Connection policy applies from a task's first SFTP call. Tasks set the
cluster'sconnection_policyon the filesystem only after their first
exists()/makedirs(), and unconditionally replaced it, so a filesystem
shared between tasks could lose an enabled policy. It is now set before
the first call and only ever upgraded, never switched off. - NeuralIL OneCycle schedule is reset every epoch. It spans one epoch
but was never reset, so after the first epoch the learning rate stayed at
the schedule's final value. - NeuralIL high-energy penalty used the padded atom count. The energy
was divided by the padded number of atoms, which pulls the energy per atom
of structures smaller than the largest in the set towards zero, so they
could be down-weighted as high-energy by mistake. It now counts only real
atoms. - Remote commands can no longer hang a submit forever.
RemoteFileSystem.run_commandwaited for the remote exit status with no
limit, so a command stalled on an overloaded login node (e.g. the
run_train.pylookup a MACE trainer does when no
remote_mace_train_runneris given) blocked the submit indefinitely
without logging anything.run_commandon both filesystems now takes a
timeout(defaultCOMMAND_TIMEOUT = 120s) and raises the new
CommandTimeoutError(aRemoteConnectionError), naming the host and the
command, once it expires. The env probes used to locate training runners
get a longer limit (remote_env.PROBE_TIMEOUT = 600s), since importing
torch from a shared filesystem can take minutes. The trainers log the
lookup before it starts and say how to skip it. - A failed runner lookup is no longer reported as a missing package.
resolve_remote_module_file/resolve_remote_executableturned every
error, including connection failures, into "not found", so the trainer
told the user to install mace/pacemaker in the remote env. Timeouts and
connection errors now propagate; only a probe that runs and finds nothing
returns "not found".remote_homealso raises on a timeout instead of
silently falling back to the local home directory.
Behaviour change: everyrun_commandcall now has a limit by default, so
passtimeout=Noneto keep the old unlimited wait.
remote_handler 0.7.0
[0.7.0] — 2026-09-12
Fixed
- Standalone collection of a
collect_only=Truearray task no longer
crashes withAttributeError.ArrayScriptTask.all_calcpathswas only
ever set as a side effect ofallocate_data(), sofinish_check()broke
for anycollect_only=Truetask built in a fresh process that skipped
allocation — includingRemoteAtomsHandler/SinglePointHandler/
VASPHandler/ORCAHandler.all_calcpathsis now a computed property
derived fromnum_array_tasks(explicit, or inferred from
individual_ressources), so it works identically whether or not
allocation happened. - MACE trainers (
MaceModelTrainer/MaceFinetuningTrainer) gained a
collect_onlymode. Previouslyupdate()always allocated a new array
andget_trained_model()always (re-)submitted, so a MACE training run
detached in one process had no way to be collected from another.
Added
ArrayScriptTask.from_manifest()/SingleScriptTask.from_manifest()
— reconstruct a full,collect_only=Truetask from the manifest written by
detach()/write_manifest(), without re-allocating or re-submitting.
Distinct fromTaskHandle(which only rebuilds an executor+filesystem):
this rebuilds the actual task object, sofinish_check()/
collect_results()work exactly as they did in the original process.ArrayScriptTask(num_array_tasks=...)constructor argument — lets a
collect_only=Truearray task be built withoutindividual_ressources
(whichfrom_manifest()relies on). Passing it alongside
individual_ressourcesthat disagree raisesTaskErrorinstead of
silently picking one.submit_detached()on both task classes — the second half of
detach()(start the job, persist the manifest, return at once) without
the allocation step, for a caller that already allocated separately (e.g.
a MACE trainer'supdate()).docs/source/examples/example_3_standalone.py— submit a DFT array and
a MACE training array in one run, collect both from a later, separate run.AceModelTrainer/GraceModelTrainer— ensemble-array trainers
fitting ACE (pacemaker) / GRACE (gracemaker) potentials, the same
shape asMaceModelTrainer. Newace/graceinstall extras (they
cannot coexist in one environment — both resolve to a top-level
tensorpotentialpackage, but from two unrelated projects; see the
comments inpyproject.toml), plusace_tensorpot_gpufor the
GPU-acceleratedpacemakerbackend.- Crash-resume now reattaches to an already-submitted job instead of
resubmitting a duplicate, closing the gap where a driver restart
mid-iteration (not just between iterations) orphaned a still-running job
on the cluster.Sampler/Selector/GroundTruthEvaluator/
ModelTrainergained apending_manifest()hook (default: reattach
unsupported, so every existing actor keeps today's behaviour);
ActiveLearningHandlerandStagedActiveLearningHandlerboth gained a
_reattach_if_pending()helper that flips a pending actor to
collect_onlyfor one step when it finds one (duck-typed for the
latter, so a stage actor need not subclass the actor ABCs at all — see
StagedActiveLearningHandler's module docstring).DetachableTaskMixin
gainedsubmit_if_needed()/wait_if_needed()so an actor'supdate()
waits correctly either way — a single immediatefinish_check()when
freshly submitted, or polling until done (or an optionalmax_wait) when
reattached — and everyDetachableTaskMixinactor (the MACE/ACE/GRACE/
NeuralilIL trainers, the VASP/ORCA/EMT evaluators) gained a matching
reattach_max_waitconstructor kwarg. docs/source/active_learning.rst— a dedicated page for the
active-learning loop: the four actor interfaces,StagedActiveLearningHandler's
stage/ring/DAG model and resume, and every shipped evaluator/trainer/
selector with an unfoldable, signature-checked usage example (a real
minimal config file for the trainers that need one).docs/source/examples.rst
gained matching Examples 4–7 (ORCA, NeuralilIL, ACE, GRACE).
Changed
MaceModelTrainer/MaceFinetuningTrainer/AceModelTrainer/
GraceModelTrainernow share one_EnsembleArrayTrainerbase (was a
MACE-only_MaceArrayTrainer), so the three ensemble-array trainers
duplicate only their tool-specific hooks (train-runner resolution,
artifact naming, stub scaffolding) instead of the whole submit/poll/
collect machinery. Public signatures unaffected.docs/source/concepts.rst's active-learning section is now a short
teaser pointing at the newactive_learning.rst, instead of the only
place the loop was documented.
Deprecated
ActiveLearningHandler— raises aDeprecationWarningon
construction and will be removed in a future release. It hardcodes the
fixed sampler/evaluator/trainer triple with no selector stage; use
StagedActiveLearningHandlerinstead (from_actors(...)is a drop-in
replacement for this constructor).
remote_handler 0.6.0
[0.6.0] — 2026-08-26
Sampling moves out of remote_handler. Every concrete Sampler and the
remote payload behind it now lives in its own downstream project repository;
the package keeps the Sampler interface and the orchestration around it.
This is a breaking change — see Removed / Breaking for the import
paths and extras that are gone. In exchange the host install and the test
matrix shrink substantially, and the Steinhardt processor moves in.
Added
SteinhardtProcessor/SteinhardtArgs— per-atom Steinhardt
bond-orientational order parameters (q<l>/w<l>, plus the per-frame
meansq<l>_mean/w<l>_meaninatoms.info) as a built-inprocessing
step, re-exported fromremote_handler.processing. The compute is a JAX
port of QUIP'ssteinhardt_nelson_qw_module::calc_qwtogether with the
angular machinery it relies on (SphericalYCartesian/SolidRCartesian,
wigner3j), validated against QUIP'sget_qwbinary to machine precision
with the reference data checked in undertest/data/steinhardt/. The
jit-compiled core runs in float64 / complex128 (the spherical-harmonic
sums lose all signal in float32), opted in for the scope of the computation
rather than by flipping the globaljax_enable_x64flag, so a surrounding
float32 pipeline is unaffected. Previously this lived downstream and was
cited as the example of an out-of-package processor.[steinhardt]install extra (jax>=0.9.0).jaxis imported lazily
inside the cluster-siderun, so the module stays importable — and cheap —
on a jax-free host, like the other built-in processors.
Changed
convert_to_jax/jax_output_formatmoved to the shared MACE trainer
base, soMaceModelTrainer(train from scratch) supports the torch →
mace-jax conversion too, not justMaceFinetuningTrainer. Conversion is a
property of training a MACE model, not of finetuning specifically. The
cluster-side step is now the sharedmace_utils.maybe_convert_to_jax, called
by both themace_training.pyandmace_finetuning.pypayloads, and
MaceTrainingInputgained the two matching fields. Existing
MaceFinetuningTrainerusage is unchanged.- The
[all]extra is recomposed to
dev,mace_freeze,quests,neuralil,mace_jax,steinhardt:torch_simand
clinamenare gone with the samplers,steinhardtis new. CI still
installs[all], so every optional tier keeps un-skipping. - Docs and README follow the split. The installation table drops the
removed extras and the "torch_simneeds a mace source" warning, the
examples page covers DFT labelling and MACE training, and the README gains
an Example projects using remote_handler section linking the downstream
sampler repositories.
Fixed
FixAtomsconstraints no longer corrupt forces on the aggregation
paths either.scripts/mace/mace_utils.write_with_keysand
AseCalculatorEvaluator._collect_resultsnow call
get_forces(apply_constraint=False). Both re-derive forces from anAtoms
deserialized from extxyz after the singlepoint already stored the correct
values, so aFixAtomsconstraint carried in the file silently re-zeroed
the frozen atoms' forces on the way intoREF_forces— the training labels
— even though the payload that wrote them was already fixed. The 0.5.0 entry
covered onlydft_singlepoint.py/orca_singlepoint.py/
ase_singlepoint.py; this completes it. Regression tests added for both
call sites.
Removed
- Every concrete sampler and its remote payload.
workflows/samplers.py
is gone in full —LammpsSampler,TorchSimSampler,CMAESSampler,
BatchCMAESSampler,SSPDSampler,BatchSSPDSampler— along with the
payloads they shipped (scripts/lammps/,scripts/torch_sim/including
mace_ensemble.py,parallel_tempering.py,create_interface.pyand
simulation_pipelines.py,scripts/clinamen2/,scripts/sspd/) and their
input schemas (LammpsSamplingInput,TorchSimSamplingInput,
CmaesSamplingInput,CmaesOptimizeInput).remote_handlernow provides
only theSamplerinterface, inworkflows/active_learning_base.py. - The
[torch_sim]and[clinamen]install extras. - The samplers' examples —
examples/mace_md_AL/,
examples/torch_sim_AL/,examples/cmaes_evolution_AL/,
examples/mace_sspd_AL/, the staleexamples/cluster_handler.py, and the
torch-sim / active-learning docs example pages.examples/local_emt_AL/
(no cluster, no licensed software) remains the runnable reference. - The samplers' tests —
test_cmaes_sampling_integration.py,
test_torch_sim_integration.py,test_mace_ensemble.pyand
test_active_learning_workflows.py. Theintegrationtier now means
"needs mace", not "needs mace/torch_sim/clinamen".
Breaking
from remote_handler.workflows import LammpsSampler, TorchSimSampler(both
were inworkflows.__all__) and any import from
remote_handler.workflows.samplersnow raiseImportError. Depend on the
relevant downstream project instead — the actors keep their constructor
signatures there, so only the import path moves.pip install ".[torch_sim]"/".[clinamen]"no longer resolve; pip warns
that the extra is not provided and installs nothing for it.".[all]"still
resolves but no longer brings torch-sim or Clinamen2 into the environment.
remote_handler 0.5.0
[0.5.0] — 2026-07-26
Added
- Durable task handles (crash resume + detach-on-submit). A submitted
SingleScriptTask/ArrayScriptTasknow persists everything needed to
await and collect its job to a small JSONTaskManifest(core/manifest.py),
written both locally (<local_savepath>.manifest.json) and into the remote
job directory. This happens by default on everysubmit/run— right
after submission and before the blocking wait — so if the driver process
dies mid-wait the running job is no longer orphaned (the original motivation).
Opt out per task withwrite_manifest_on_submit=False. The write is
best-effort (the local copy is the contract; a remote-copy failure only
warns) so persisting the safety net can never take down the job it protects.
task.detach()additionally returns without waiting, so submitting a
remote job feels like a local backgrounded command.
TaskHandle.from_file(<manifest>)rebuilds the filesystem + execution backend
in a different process (re-binding to the running job via the new
Executor.attach) to poll (is_finished),wait, andcollectthe results. ClusterConfig.to_dict()— JSON-round-trippable snapshot, embedded in the
manifest so a resume does not depend on the original cluster JSON.Executor.is_finished()— non-blocking completion poll (SLURMsqueue;
directsentinel file);ArrayScriptTask.submit(only_start=…)for parity
withSingleScriptTask.- CLI
status/collect.remote_handler status [PATH]reports whether
a detached task has finished (exit0finished /2running, for shell
loops);remote_handler collect [PATH] [--wait] [--force]fetches its
results.PATHis the manifest file or a directory containing one (defaults
to the cwd) — so the workflow is: launch, walk away, and the next daycd
back and runremote_handler collect --wait. [all]install extra —pip install ".[all]"pulls the full stack in
one shot (dev,mace_freeze,torch_sim,clinamen,quests,neuralil,mace_jax).
It deliberately uses themace_freezefork as the solemace-torchsource
(not[mace]), so it stays internally consistent. CI's install step now
uses it, so every optional test tier un-skips.- Runnable local MD→EMT→Morse AL example (
examples/local_emt_AL/) — a
self-contained active-learning loop (MD sampling → EMT ground truth → Morse
MLFF) that runs locally with no cluster or licensed software, plus the
accompanying docs. memoryjob parameter onSlurmHandlerBase, threaded through the tasks
layer and wired to the DFT engines (dft_runner) and the SLURM templates
(datalab/dirona/musica), so a per-job memory request can be set
centrally instead of hardcoded in a header.
Changed
shard_sizeis now optional.AtomsShardManager,
RemoteTrajectoryProcessor, andRemoteOptimizationHandleraccept
shard_size=None(or a negative value such as-1), meaning "no sharding":
the entire configuration set is processed as a single shard, resolved once
the total count is known. Previously a shard size was mandatory.
Fixed
- Conda env no longer shadowed by module-loaded interpreters. The generated
activation now appends[ -n "$CONDA_PREFIX" ] && export PATH="$CONDA_PREFIX/bin:$PATH"afterconda activate. On clusters where a
job'senvironment/slurm_header_addrunsmodule load(e.g. EESSI/Lmod,
which prepend a compat-layerpythontoPATH), conda would re-insert the
env'sbinbehind that entry, so a barepythonran the module's
interpreter and the env's packages (ase, …) were invisible —
ModuleNotFoundErrordespite a correctconda_env. Forcing the env'sbin
to the front fixes it centrally, so users never have to hardcode the env path
in their header (it is already declared once, in the cluster config). Applies
to both the SLURM anddirectbackends; guarded so a blank activation is a
no-op. FixAtomsconstraints no longer corrupt written forces. The singlepoint
payloads (dft_singlepoint.py,orca_singlepoint.py,ase_singlepoint.py)
readatoms.calc.results["forces"]instead ofatoms.get_forces(), which
appliesFixAtomsconstraints deserialized from the extxyz and silently
zeroes the forces on frozen atoms. The calculator's stored results are
unaffected, so the real forces are now recorded.
remote_handler 0.4.1
[0.4.1] — 2026-07-09
A patch/minor increment on top of 0.4.0: two optional per-cluster config
knobs, a bug fix in trajectory sharding, and the SLURM-availability probe moved
behind the connection-resilience wrapper. No breaking changes.
Added
conda_env_dir— an optionalClusterConfigfield naming the directory
on the remote that holds the named conda environments, so activation no
longer assumes a default install location. Prompted for by
remote_handler configure clusterand threaded throughremote_env.slurm_header_add— reintroduced as an optionalClusterConfigfield for
extra lines injected verbatim into the generated SLURM job header (e.g.
account/partition/QoS directives). Exposed via the CLI and rendered by the
SLURM handler.
Changed
validate_slurm_availablemoved fromcluster_handlertoslurm_utilsand
now runs through the same resilient wrapper as other remote commands, so a
transient credential/connection failure while probingsbatchpauses and
retries instead of crashing.
Fixed
AtomsShardManagercrash on a default slice. Whenevery/start/end
were left at theirNonedefaults (asVASPHandlerand the other DFT
handlers do),combine_trajscomputediteration_idx * indices.step—
int * None— and raisedTypeErrorbefore any job was submitted. The
slice fields are now normalized (step→1,start→0) anditeration_index
is computed asstart + iteration_idx * step, which also fixes a latent
offset bug where a non-Nonestartwas ignored. Regression tests added.
remote_handler 0.4.0
[0.4.0] — 2026-07-02
A feature release on top of 0.3.0. The headline is a rewrite of trajectory
post-processing into a self-contained plugin framework
(remote_handler.processing), plus a mace-jax ASE calculator, a staged
active-learning handler, and connection resilience for the subprocess-ssh path.
Breaking: the old optimizer entry points are gone, replaced by
remote_handler.processing (no compatibility shims).
Added
remote_handler.processing— a self-contained, extensible trajectory-
processing framework, organized by feature rather than by layer. A
RemoteTrajectoryProcessorshards a trajectory and runs an ordered list of
Processorsteps over each shard (viaprocessing/driver.pyon the cluster);
RemoteOptimizationHandleris a thin factory for the optimize step.
Structural analysis is just the core with the processors you want, e.g.
processors=[RDFProcessor(), SteinhardtProcessor()].- Processor plugin contract — each processor keeps its whole trio in one
module: anArgsschema (single source of truth), a cluster-siderun, and
a host-sideProcessorspec, wired viaregister(name, run, schema=...).
Downstream packages self-register the same way; the driver imports each
processor'sregistration_modulebefore resolving steps, so a downstream
processor works with no custom driver. Built-ins:optimize,rdf. - Per-processor argument validation — the driver validates every step's
args against its schema at job start (schema.from_dict(args).validate()),
so a malformed composed payload fails fast instead of deep inside an array
task. The hostbuild_argsfills the same schema, so emit and check can't
drift. - mace-jax ASE calculator (
scripts/mace/mace_jax_ase.py,
MaceJaxEnsembleCalculator) — single model or stacked-params ensemble, with
amax_neighborsbucket ladder to bound JIT recompiles and per-structure
energy/force uncertainty from the ensemble spread. StagedActiveLearningHandler(+Stage,InputSource,SelectorActor)
— a staged active-learning loop with explicit per-stage inputs and selection.- Selector as an actor —
SelectorActorplus novelty/uncertainty-based
selection wired through the loop; selectors take an explicitreference
argument rather than mutating shared state. - Connection resilience — an opt-in
ConnectionPolicy(per-cluster
resilient_connection) that pauses-and-retries SSH operations when a
credential expires instead of crashing, resuming once it is refreshed. configure_logging— a utility for timestamped logging.remote_handler show extrasnow also reports the installedremote_handler
version and conda env, per configured machine.
Changed
UniformSelector→StrideSelector. The old name remains as a subclass
alias that emits aDeprecationWarning; it will be removed in a future
release.- Every subprocess
sshinvocation is now non-interactive by construction
(stdinclosed,-o BatchMode=yes -o ConnectTimeout=…, detached session)
and bounded byConnectionPolicy.command_timeout(default 120s, applies even
when resilience is off), so a stalled connection surfaces as an error rather
than an indefinite hang.
Fixed
- Subprocess
sshno longer hangs on an interactive QR/MFA prompt when a
stepcertificate expires: the prompt can't grab the terminal, and a bounded
timeout turns the stall into aTimeoutExpiredthat is classified transient
and retried under an enabled policy. squeuestatus polling treats a purged (completed) job —squeue -j
exiting with "Invalid job id specified" — as finished, instead of crashing
the driver (the failure seen when a job completes while the connection is
briefly down).- Collection of finetuned MACE models.
lock.shnow includestorch-sim; the uv lock was refreshed.
Removed
- The old optimizer entry points, replaced by
remote_handler.processingwith
no compatibility shims:workflows/optimizer.py,
workflows/processors.py,workflows/run_optimizer.py, and the deprecated
scripts/jaxnest/optimize_traj.py. Import handlers and processors from
remote_handler.processinginstead.
remote_handler 0.3.0
[0.3.0] — 2026-06-22
A feature release on top of 0.2.0: optional-dependency ergonomics, YAML
config inputs for the DFT and NeuralIL paths, a choice of mace-jax artifact
format, and a batch of SLURM/template fixes.
Added
MaceFinetuningTrainer(jax_output_format=...)— choose the converted
mace-jax artifact: the opaque.pklbundle ("pkl", default — unchanged
behaviour) or a directory bundle ("bundle": a human-readableconfig.json
plusparams.msgpack) that mace-jax's ownload_model_bundleconsumes.- mace-jax conversion utilities:
convert_torch_model_to_jax_bundle()(torch
.model→ directory bundle) andgenerate_jax_foundation_model()(download
a MACE foundation model and emit a mace-jax bundle). remote_handler show extras— probes (by actually importing) which optional
dependency groups are installed.- INCAR via YAML:
incar_config_fileonVASPHandler/VASPEvaluator,
layered under any explicitincar_kwargs. NeuralilILTrainer(config_file=...)— training hyperparameters from a YAML
file, layered defaults <config_file<training_kwargs.
Changed
remote_mace_train_runneris now optional on the MACE trainers: when unset
it is auto-discovered from the remote conda env (mace.cli.run_train).- The
torch_simextra no longer depends on mace — torch_sim runs without it.
The redundanttorch_sim_freezeextra was removed (a runtime guard reports
the missing piece instead). ~is accepted again inconda_sourceand remote paths, expanding to the
remote$HOME(good defaults without hardcoding a home directory).
Fixed
map_indicesis now honoured on the SLURM array path (the rendered
body_array.sh.j2MAP=(...)lookup was dropped).MaceFinetuningTrainerno longer requests 2 GPUs by default —
n_processorsis threaded through to the SLURM request.NeuralilILTrainerno longer mutates its class-level default training kwargs
across instances (defaults are deep-copied).- Corrected the
musicaSLURM template.
remote_handler 0.2.0
[0.2.0] — 2026-06-18
First deliberately versioned release. Prior 0.1.0 was never tracked; this
draws the line.
Added
- Pluggable execution backends: a scheduler-free
directexecutor alongside
SLURM, selected viaClusterConfig.executor;run_commandon the
filesystem handlers. remote_handlercommand-line tool (configure/check/show) and a
persisted global config —REMOTE_HANDLER_CONFIG_DIRis now optional.ClusterConfig.from_name()/ClusterConfig.available_names().EMTEvaluatorand a genericase_singlepoint.pypayload (shared
AseCalculatorEvaluatorbase) — pure-ASE ground truth that runs with no
cluster and no licensed software.- Opt-in connection resilience (
resilient_connection), useful for MFA authentication via certificates which are only valid for a certain amount of time. NeuralilILTrainer;MaceFinetuningTrainer(convert_to_jax=True).- A custom SLURM header template may be given as an explicit path, not only a
registered name. - Docs: Core Concepts, Contributing, expanded Installation/CLI, API-checked
examples, mermaid diagrams; a weekly scheduled CI run.
Changed
- Unified the DFT layer:
SinglePointHandlerbase withVASPHandler/
ORCAHandlerpresets;AseCalculatorEvaluatorbase withVASPEvaluator/
ORCAEvaluatorpresets. Public signatures preserved. - SLURM templates now render conda activation from the configured
conda_source(micromamba works on SLURM too) instead of a hardcoded
~/miniconda3. - Core dependencies relaxed from exact pins to floors.
- The package version is now derived from git tags (setuptools-scm).
ClusterConfignow lives incore.cluster_config(re-exported from
core.cluster_handler, so the old import path still works — non-breaking).- Test harness: the suite runs without any configuration (config state is
isolated per-test); CI exercises a real SSH/SFTP + scheduler-free round-trip
via thesshlivetier.
Removed
- The
heavypytest marker. - The driver-side
input_modelsvalidation layer — each script now validates
its own input viavalidate(calcpath).
Fixed
conda_sourcenow actually drives SLURM job activation (was ignored).- Auto-patch a paramiko 3.5.1 crash (
AttributeError: public_blob) with
agent-provided SSH certificates, applied at import — no action needed. ORCAEvaluatorgained stub mode; cleaned bare-excepts, typos, and a batch of
docstring/lint issues.
Breaking
- Several modules moved out of the package root:
atoms_handler→
core.atoms_handler/core.atoms_shard;dft_runner/optimizer/
run_optimizer→workflows.*. Update imports accordingly.