SHARED POLICY · EVERY NEW MODEL

Per-Model Training Gate

A lightweight RTX4090 execution path that respects each model's native memory and runtime behavior.

PRIMARY GPU · RTX4090MODEL-SPECIFIC BATCHBEST REPRODUCIBLE CHECKPOINT
Integration
Focused testsPretrained load, optimizer coverage, forward, loss, and backward smoke.
GPU calibration on RTX4090Calibrate the model-specific physical batch while preserving intended effective batch where reasonable.
Full training
Best reproducible checkpointSelect by validation mAP; inspect training-health trajectories diagnostically.
Standalone evaluation
Evidence freeze
COMPLETED

Evidence accepted

A bounded gate or historical result has completed with evidence.

CURRENT

Active training

DEIM-D-FINE-X production training on RTX4090 is the current main-performance step.

NEXT

Queued execution

Work begins after its exact-artifact preflight and available execution slot.

OPTIONAL

Noncritical branch

Skip when time or data budget threatens the submission path.

BLOCKED / CONDITIONAL

Condition not met

Do not execute until the stated diagnostic, evidence, or eligibility condition is satisfied.