PublishedRunScenario
Training scenarios · Implementation: mlsysim/engine/scenario_definitions/training.py.
Construct this scenario with keyword inputs. The fields below show the types and defaults declared by this class.
A published training run: its total compute spread over its accelerators and days.
Optional (added 2026-10-20): a run reported as accelerator_hours (a model card’s GPU hours) at device_power per accelerator. Its compute is the workload’s training_ops or, when unset, 6PD from its training tokens.
Notes
Equation: Compose mlsysim.physics.transformer.calc_transformer_training_flops.
Assumptions: Published workload fields and externally supplied accelerator-time describe the same training run; the fallback work estimate uses parameter-token accounting.
Excludes: Reconstructed energy is not meter data, and missing reference run fields are not inferred.
References
Derivation: this scenario binds its input fields to the referenced implementation contracts: mlsysim.physics.transformer.calc_transformer_training_flops. The methods below identify any additional arithmetic.
Examples
>>> from mlsysim import Scenarios
>>> case = Scenarios.Gpt3PublishedRun
>>> str(case.training_operations)
'3.14×10²³ flop'Attributes
| Name | Description |
|---|---|
| accelerator | |
| accelerator_hours | |
| achieved_rate | Training work amortized over accelerator-time. Equation: work divided by supplied |
| device_power | |
| energy | Energy implied by supplied accelerator-time and constant device power. Equation: |
| peak_fraction | Achieved per-accelerator arithmetic rate relative to catalog peak. Equation: |
| training_operations | Recorded training work, falling back to parameter-token accounting when absent. |
| workload |
workload
Type: Workload.
Declared default: required.
accelerator
Type: HardwareNode.
Declared default: required.
accelerator_hours
Type: Optional[Quantity].
Declared default: None.
device_power
Type: Optional[Quantity].
Declared default: None.
training_operations
engine.scenario_types.PublishedRunScenario.training_operations
Recorded training work, falling back to parameter-token accounting when absent. Equation: stored training_ops or mlsysim.physics.transformer.calc_transformer_training_flops(parameters, training_tokens). Units: FLOPs; the fallback is modeled, not measured.
achieved_rate
engine.scenario_types.PublishedRunScenario.achieved_rate
Training work amortized over accelerator-time. Equation: work divided by supplied accelerator-hours, otherwise recorded work divided by reference accelerator count and run duration. Units: TFLOP/s per accelerator; the fallback requires recorded training work.
energy
engine.scenario_types.PublishedRunScenario.energy
Energy implied by supplied accelerator-time and constant device power. Equation: accelerator_hours * device_power. Units: GWh; facility overhead and varying power are excluded.
peak_fraction
engine.scenario_types.PublishedRunScenario.peak_fraction
Achieved per-accelerator arithmetic rate relative to catalog peak. Equation: achieved_rate / accelerator.compute.peak_flops. Units: dimensionless; the peak precision must match the interpretation of training work.