eval-suite-builder

Use when validating a new model with smoke tests for tensor shapes, forward logic, gradients, loss computation, initialization sanity, and tiny-batch overfit…

Use when reviewing a proposed model architecture for originality, novelty claims, fallback-pattern violations, or similarity to common decoder-only templates.

Use when reading public inspirations and turning them into constraints, opportunities, and anti-copy design notes for a new model idea.

Use when comparing an idea against the internal baseline, prior attempts, and selected public references without turning those references into copy pressure.

Use when designing smoke, small, and production-like training curricula, token budgets, sequencing, or leakage checks for an idea.

Use when PyTorch ROCm runs fail due to device discovery, dtype issues, memory pressure, HIP kernels, compile/runtime mismatches, or unstable fallbacks.

Use before retries or final reports to verify that lineage, artifacts, budgets, and outcomes are written to the experiment ledger.

Use after smoke, small, or full runs to convert metrics, artifacts, and logs into an explicit continue, pivot, or stop recommendation.

Use when defining save cadence, resume policy, crash recovery, and lineage preservation for multi-step training runs and retries.

Use before expensive runs to estimate parameter, optimizer, memory, checkpoint, and token budgets under the single-node ROCm-first setup.