# Self-directed benchmark: ssm_mamba strategy stress test

- Paper ID: self.20260509080324.001
- Hypothesis: Benchmark a simplified recovered baseline against an ablated variant.
- Plan: Run the deterministic recovery benchmark and capture VRAM plus throughput telemetry.
- Expected Signal: Recovered benchmark should execute and emit standard telemetry fields.
- Concept Combo: recovery_baseline
- Note: Generated fallback because the architect returned no content.

--- ATTEMPT: preflight (code=2) ---
--- STDOUT ---
--- RUNTIME PROFILE ---
Device policy: gpu_required
Torch: 2.11.0+rocm7.1
Accelerator backend: rocm
Torch CUDA build: None
Torch HIP build: 7.1.52802
CUDA available: False
CUDA device count: 0
Recommended autocast dtype: fp32
Recommended DataLoader pin_memory: False
Recommended DataLoader num_workers: 12
Recommended starting batch size: 16
Recommended CPU threads: 24
/dev/kfd present: True

--- STDERR ---
GPU_REQUIRED policy blocked benchmark execution.
Reason: torch.cuda did not expose a usable accelerator in this Python environment.
Install a CUDA- or ROCm-enabled PyTorch build (not +cpu), then retry.
Tip: In your current env, torch reports a CPU-only build.

--- HUMAN SUMMARY (LAYMAN) ---
Result: The test did not finish successfully.
Main failure observed: GPU_REQUIRED policy blocked benchmark execution.
Plain-English interpretation: The script hit a runtime problem before it could report final metrics.