Skip to navigation

Training Report Schema

View as Markdown

training_report.json is the unified training output for NeMoRun, MegatronRun, and Megatron-Bridge.

Versioning

The top-level schema_version field uses MAJOR.MINOR versioning:

  • Increment MAJOR when removing, renaming, or incompatibly changing a field.

  • Increment MINOR when adding an optional field or otherwise making a backward-compatible schema change.

  • Do not change the schema version for implementation fixes that leave the JSON contract unchanged.

Consumers should reject unsupported major versions and tolerate unknown fields within a supported major version.

Version History

1.0 — 2026-07-14

Added to root:

  • schema_version: str

Added to root.config:

  • test_id: str

  • test_name: str

  • description: str

  • test_scenario_name: str

  • system_path: str

  • tests_dir_path: str

  • test_scenario_path: str

  • container_image: str

  • cloudai_execution_node: str

  • env_vars: dict[str,`` ``Any]

  • gpus_per_node: Optional[int]

  • nodes: list[str]

  • clique_size: Optional[int]

  • fp8: Optional[str]

  • fp8_recipe: Optional[str]

  • expert_tensor_parallel_size: int

Training Report Models

class cloudai.report_generator.training.models.TrainingResults(

*,

schema_version: str = ‘1.0’,

config: TrainingConfig,

steps: List[TrainingStep],

aggregation: StepAggregation | None = None,

)

Container for parsed training output.

class cloudai.report_generator.training.models.TrainingConfig(

*,

test_id: str,

test_name: str,

description: str,

test_scenario_name: str,

test_template_name: str,

system_path: str,

tests_dir_path: str,

test_scenario_path: str,

container_image: str = ”,

cloudai_execution_node: str,

env_vars: dict[str, Any],

world_size: int | None = None,

num_nodes: int,

gpus_per_node: int | None = None,

nodes: list[str],

clique_size: int | None = None,

fp8: str | None = None,

fp8_recipe: str | None = None,

micro_batch_size: int,

global_batch_size: int,

seq_length: int,

tensor_parallel_size: int,

pipeline_parallel_size: int,

context_parallel_size: int | None,

virtual_pipeline_parallel_size: int | None,

sequence_parallel: bool,

expert_parallel_size: int,

expert_tensor_parallel_size: int,

data_parallel_size: int | None = None,

num_layers: int,

hidden_size: int,

num_attention_heads: int,

num_query_groups: int,

ffn_hidden_size: int,

kv_channels: int,

normalization: str,

position_embedding_type: str,

model_name: str = ”,

num_experts: int | None,

moe_router_topk: int | None,

moe_ffn_hidden_size: int | None,

moe_grouped_gemm: bool | None,

profiling_enabled: bool = False,

profiling_start_step: int | None = None,

profiling_stop_step: int | None = None,

exclude_start_steps: int = 5,

exclude_post_profiling_steps: int = 2,

)

Resolved training configuration from the framework artifact + CloudAI.

CloudAI-computed fields are supplied by the parser during construction.

class cloudai.report_generator.training.models.TrainingStep(

*,

iteration: int,

step_time_sec: float,

loss: float,

memory_reserved_bytes: float,

memory_allocated_bytes: float,

tflops_per_gpu: float | None = None,

)

Results for a single training iteration.

class cloudai.report_generator.training.models.StepAggregation(

*,

step_time_sec: MetricStats,

loss: MetricStats,

memory_reserved_bytes: MetricStats,

memory_allocated_bytes: MetricStats,

tflops_per_gpu: MetricStats | None = None,

)

Per-metric aggregated statistics over the filtered steps.

class cloudai.report_generator.training.models.MetricStats(

mean: float,

min: float,

max: float,

std: float,

t99: float,

t95: float,

)

Aggregated statistics for one metric over the filtered steps.