{ "device": "NVIDIA GB10", "torch": "2.9.1+cu130", "attention": "sage2", "resolution": [ 1344, 768 ], "frames": 124, "steps": 1, "seed": 440420, "text_tokens": 100, "warmup_runs": 1, "cuda_profiler_capture": true, "elapsed_seconds": 23.707971603027545, "stage_trace": [], "checksums": [ -197605.3125, 528.5980224609375 ], "peak_allocated_bytes": 15570401280, "peak_reserved_bytes": 18538823680, "measurement_policy": "uninstrumented production sampling wall time inside a whole-run NVTX range" }