1 lightning.finite_horizon.diagnostics.commissioning

lightning.finite_horizon.diagnostics.commissioning

Observed optimizer milestones for opt-in engineering commissioning.

These gates inspect successful transactions, never proxy batch/global-step counts. They do not certify learning quality or replace the authenticated final fit receipt. Cloned states are CPU artifacts; live tensors remain untouched.

1.1 Classes

Name Description
QhCommissioningMilestoneReceipt Typed observed gate bound to CPU state bytes and declared context.
QhCommissioningMilestonesCallback Observe live finite FP32 state and strict CPU scorer reconstruction.
QhCommissioningMilestonesConfig Declare observed gates and source/input context before fitting.
QhMpsMemoryBudgetCallback Guard observed live MPS tensor allocation after each training batch.
QhMpsMemoryObservation Current allocator values and sampled maxima, never an unsampled peak.

1.1.1 QhCommissioningMilestoneReceipt

lightning.finite_horizon.diagnostics.commissioning.QhCommissioningMilestoneReceipt(
)

Typed observed gate bound to CPU state bytes and declared context.

1.1.1.1 Attributes

Name Description
schema_version Fixed serialized milestone schema; version one binds observed gates rather than batch/global-step
successful_optimizer_updates Positive exact completed-update gate observed after a training batch; identifies this milestone and
expected_device Declared required device of live scorer parameters, gradients and moments at the gate; CPU is a
context_ids Composition-provided nonempty source/config/input identities copied from the commissioning
state_filename Milestone-local filename of the saved CPU online/target/optimizer/gradient state, published once
state_sha256 SHA-256 of the exact saved state-file bytes; authenticates the referenced CPU artifact independently
strict_reload Always true after strict scorer reconstruction and exact state equality succeed under a forked CPU
optimizer_evidence Detached successful-transaction identity, tensor counts, tensor residency/dtypes and finite-support
memory_observation Latest sampled MPS allocation evidence when the live-tensor guard is enabled; None for CPU gates or

1.1.2 QhCommissioningMilestonesCallback

lightning.finite_horizon.diagnostics.commissioning.QhCommissioningMilestonesCallback(
    config,
    output_dir,
)

Observe live finite FP32 state and strict CPU scorer reconstruction.

A gate failure interrupts engineering fitting. Construction and reload run under a forked CPU RNG, preserving the training RNG. No live parameter, gradient, moment, forward result or optimizer state is modified.

1.1.2.1 Methods

Name Description
on_fit_end Require the complete engineering budget; never infer absent gates.
1.1.2.1.1 on_fit_end
lightning.finite_horizon.diagnostics.commissioning.QhCommissioningMilestonesCallback.on_fit_end(
    trainer,
    pl_module,
)

Require the complete engineering budget; never infer absent gates.

1.1.3 QhCommissioningMilestonesConfig

lightning.finite_horizon.diagnostics.commissioning.QhCommissioningMilestonesConfig(
)

Declare observed gates and source/input context before fitting.

1.1.3.1 Attributes

Name Description
milestones Strictly increasing successful-update gates; final count is engineering budget.
expected_device Required live tensor residency; CPU is for focused verification only.
max_allocated_bytes Opt-in live MPS tensor-allocation limit; driver/cache bytes are separate evidence.
context_ids Composition-provided source/config/input identities sealed in each receipt.

1.1.4 QhMpsMemoryBudgetCallback

lightning.finite_horizon.diagnostics.commissioning.QhMpsMemoryBudgetCallback(
    max_allocated_bytes,
    output_dir,
)

Guard observed live MPS tensor allocation after each training batch.

Torch driver allocation includes its caching pool and other driver resources; it is recorded independently rather than compared with the live tensor limit. Sampling does not establish a true operation peak. No synchronization, cache eviction, model change or optimizer change is introduced.

1.1.5 QhMpsMemoryObservation

lightning.finite_horizon.diagnostics.commissioning.QhMpsMemoryObservation()

Current allocator values and sampled maxima, never an unsampled peak.

1.1.5.1 Attributes

Name Description
schema_version Fixed serialized observation schema used by the authenticated memory reader; defaults to version
sampling_scope Maxima and sample_count cover this callback instance lifetime, including its train-batch-end
sample_count Positive number of completed callback memory samples represented by this observation, incremented
current_allocated_bytes Nonnegative live MPS tensor allocation in bytes at the latest callback sample; excludes allocator
driver_allocated_bytes Nonnegative MPS driver allocation in bytes at the same latest sample, including caching and other
max_sampled_current_allocated_bytes Largest live-tensor allocation in bytes among this callback instance samples; a sampled maximum,
max_sampled_driver_allocated_bytes Largest driver allocation in bytes among this callback instance samples; not compared against the
max_allocated_bytes Positive caller-declared live MPS tensor budget in bytes, bound into the record and checked
within_live_tensor_budget Whether the latest live-tensor sample is at most max_allocated_bytes. The reader also checks the

1.2 Functions

Name Description
read_commissioning_milestones Authenticate every declared observed gate; final fit admission is separate.
read_mps_memory_observation Authenticate sampled memory evidence against one declared tensor budget.

1.2.1 read_commissioning_milestones

lightning.finite_horizon.diagnostics.commissioning.read_commissioning_milestones(
    config,
    output_dir,
)

Authenticate every declared observed gate; final fit admission is separate.

1.2.2 read_mps_memory_observation

lightning.finite_horizon.diagnostics.commissioning.read_mps_memory_observation(
    output_dir,
    *,
    max_allocated_bytes,
)

Authenticate sampled memory evidence against one declared tensor budget.

Back to top