EMEP Configuration Specification

YAML configuration schema for EMEP platform, registry, merge, evolution, evaluation, storage, GPU, deployment, security, and observability settings.

EMEP uses a single YAML configuration file to control platform behavior. The configuration is loaded at startup and validated against a JSON Schema. Invalid values prevent startup. This page defines all sections and provides a full example.

Schema Sections

platform

Key	Type	Default	Description
name	string	”emep”	Platform instance name
version	string	required	EMEP version string
log_level	string	”info”	One of: debug, info, warn, error
max_concurrent_jobs	integer	4	Global job concurrency limit

registry

Key	Type	Default	Description
backend	string	”postgres”	ModelRegistry backend: postgres, sqlite, memory
connection_string	string	required	Database connection string
cache_ttl	integer	300	Registry metadata cache TTL in seconds

merge

Key	Type	Default	Description
default_strategy	string	”slerp”	Default MergeStrategy
max_models_per_merge	integer	8	Maximum source models in one merge
validation_enabled	boolean	true	Run MergeValidation after every merge
artifact_retention_days	integer	30	Days to retain intermediate merge artifacts

evolution

Key	Type	Default	Description
algorithm	string	”nsga2”	Evolution algorithm: nsga2 (Deb et al. 2002), nsga3 (Deb & Jain 2014), cmaes (Hansen 2001)
population_size	integer	50	Initial population size
max_generations	integer	100	Hard stop after N generations
crossover_rate	float	0.8	Probability of crossover per pair
mutation_rate	float	0.1	Probability of mutation per genome

evaluation

Key	Type	Default	Description
benchmark_backend	string	”vllm”	Inference backend for benchmarks
deterministic_seeds	boolean	true	Fix random seeds for reproducibility
timeout_seconds	integer	3600	Per-benchmark timeout

storage

Key	Type	Default	Description
artifact_store	string	”s3”	ArtifactStore backend: s3, gcs, azure, local
bucket	string	required	Storage bucket or root path
encryption	string	”aes256”	At-rest encryption algorithm

gpu

Key	Type	Default	Description
orchestrator	string	”kubernetes”	GPU orchestrator: kubernetes, slurm, manual
default_dtype	string	”bf16”	Default compute dtype: fp32, bf16, fp16, fp8
memory_fraction	float	0.9	Fraction of GPU memory to allocate per job

deployment

Key	Type	Default	Description
inference_backend	string	”vllm”	Default InferenceBackend
quantization_enabled	boolean	false	Enable QuantizationEngine for deployment
max_batch_size	integer	16	Maximum inference batch size

security

Key	Type	Default	Description
artifact_signing	boolean	true	Enable ArtifactStore signing
require_signed_models	boolean	true	Reject unsigned models in ModelRegistry
license_scanning	boolean	true	Enable license compliance checks

observability

Key	Type	Default	Description
metrics_backend	string	”prometheus”	Metrics sink
tracing_backend	string	”jaeger”	Distributed tracing backend
alert_webhook	string	""	URL for critical alert webhooks

Full Example

platform:
  name: "emep-production"
  version: "0.1.0"
  log_level: "info"
  max_concurrent_jobs: 8

registry:
  backend: "postgres"
  connection_string: "postgresql://user:pass@db:5432/emep"
  cache_ttl: 300

merge:
  default_strategy: "ties"
  max_models_per_merge: 4
  validation_enabled: true
  artifact_retention_days: 30

evolution:
  algorithm: "nsga2"
  population_size: 50
  max_generations: 100
  crossover_rate: 0.8
  mutation_rate: 0.1

evaluation:
  benchmark_backend: "vllm"
  deterministic_seeds: true
  timeout_seconds: 3600

storage:
  artifact_store: "s3"
  bucket: "emep-artifacts-prod"
  encryption: "aes256"

gpu:
  orchestrator: "kubernetes"
  default_dtype: "bf16"
  memory_fraction: 0.9

deployment:
  inference_backend: "vllm"
  quantization_enabled: true
  max_batch_size: 16

security:
  artifact_signing: true
  require_signed_models: true
  license_scanning: true

observability:
  metrics_backend: "prometheus"
  tracing_backend: "jaeger"
  alert_webhook: "https://alerts.emep.dev/webhook"

Validation

The configuration is validated on load. Unknown keys are rejected. Missing required keys prevent startup. Type mismatches produce a detailed error message with the offending key and expected type.

Integration Points

• API Specification: the REST API reads the same config file for defaults.
• CLI Specification: emep commands use config values when flags are omitted.
• GPU Orchestration: the gpu section is consumed by the GPU orchestrator layer.
• Security: the security section drives ArtifactStore signing and license compliance.