{ "schema_version": 1, "verified_at_utc": "2026-09-15T00:00:50.565257+00:00", "analyzer_sha256": "fc34c05578da57c94105165cf8ee5af9abc335d18ad4f27221364c4771c1e6f9", "base_protocol_sha256": "ec2c0c3cb8037f11766604c90c1cd7352e8656bea59d4d3ab3da2c7e1ce75496", "plan_sha256": "a42d71925d706406e6089c6c10047ba7dca99a710bf4d38f15404038b9af5873", "frozen_base_sources_unchanged": true, "status": "complete", "all_available_evidence_verified": true, "all_jobs_passed": true, "pending_jobs": 0, "failed_or_unmeasured_jobs": 0, "qualification": { "complete": true, "draws_bit_equal": 11, "predictions_bit_equal": 110000, "draws": [ { "result_path": "qualify-00.json", "result_sha256": "172f9a0079f3b28ecea67767ddd5de877574ae7a22b071b6fd0c0b31c19cfdc9", "result_uncompressed_sha256": "172f9a0079f3b28ecea67767ddd5de877574ae7a22b071b6fd0c0b31c19cfdc9", "result_uncompressed_bytes": 26643, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":16:8", "scope": "task", "status": "passed", "draw": 0, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "qualify-00-result.predictions.npy", "prediction_sha256": "fce8a9458d2158537c7d6207fb0c8e9cfcbba007e7a748976bf3e7cd002729f3", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" } }, { "result_path": "qualify-01.json", "result_sha256": "c6eb56a4d56c4d97755d11ca2b49a72997f6f0cf91799e881f641121fa6e0924", "result_uncompressed_sha256": "c6eb56a4d56c4d97755d11ca2b49a72997f6f0cf91799e881f641121fa6e0924", "result_uncompressed_bytes": 26641, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":16:8", "scope": "task", "status": "passed", "draw": 1, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "qualify-01-result.predictions.npy", "prediction_sha256": "dd55eb6ae12c65660402dbd4a9de852e79682687a686b4d38c94a41e49ce6b1a", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" } }, { "result_path": "qualify-02.json", "result_sha256": "ce45fd2d498af8dad7d8b0cbaf8677ef1bf13c9f6413759a3c5c42b525a2cb8d", "result_uncompressed_sha256": "ce45fd2d498af8dad7d8b0cbaf8677ef1bf13c9f6413759a3c5c42b525a2cb8d", "result_uncompressed_bytes": 26643, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":16:8", "scope": "task", "status": "passed", "draw": 2, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "qualify-02-result.predictions.npy", "prediction_sha256": "75435ffba563c4ff588983f065521d6d8a2adcceb9b2ed5125125faa03c9d601", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" } }, { "result_path": "qualify-03.json", "result_sha256": "17f73993894804358ea4620a9036fda5dea497e692d0cb65be116e0440e3b2a4", "result_uncompressed_sha256": "17f73993894804358ea4620a9036fda5dea497e692d0cb65be116e0440e3b2a4", "result_uncompressed_bytes": 26643, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":16:8", "scope": "task", "status": "passed", "draw": 3, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "qualify-03-result.predictions.npy", "prediction_sha256": "393294bf08f96c82f6ab6e5c88442d9cb6329ca7274b01eed95ec0bcf1ef1fff", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" } }, { "result_path": "qualify-04.json", "result_sha256": "bcbf4958459ff4a0abcaf092aad2060fe01da05c2410546317f17857e4220c98", "result_uncompressed_sha256": "bcbf4958459ff4a0abcaf092aad2060fe01da05c2410546317f17857e4220c98", "result_uncompressed_bytes": 26642, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":16:8", "scope": "task", "status": "passed", "draw": 4, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "qualify-04-result.predictions.npy", "prediction_sha256": "5cb4bed52d62677ca958300d802cb91abf03a6c63722ebfb1f9bf67d0d65044c", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" } }, { "result_path": "qualify-05.json", "result_sha256": "a7df892969cf1a1b7cbe4770d11a41c1ccbbc3b39ad21f2935c04ddff8510027", "result_uncompressed_sha256": "a7df892969cf1a1b7cbe4770d11a41c1ccbbc3b39ad21f2935c04ddff8510027", "result_uncompressed_bytes": 26642, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":16:8", "scope": "task", "status": "passed", "draw": 5, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "qualify-05-result.predictions.npy", "prediction_sha256": "a668e0ec96783c245aaf27bf6edb2170e0efbaf2346851eb30aa737c890d6453", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" } }, { "result_path": "qualify-06.json", "result_sha256": "ceeaf92e6053fcbfb457d7fd55bd60aa5d02385873ae6aaaab0353add503acb8", "result_uncompressed_sha256": "ceeaf92e6053fcbfb457d7fd55bd60aa5d02385873ae6aaaab0353add503acb8", "result_uncompressed_bytes": 26643, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":16:8", "scope": "task", "status": "passed", "draw": 6, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "qualify-06-result.predictions.npy", "prediction_sha256": "081a7ac2a89ee6181fd8dd40f1d167898b9b6c8f0539ae6b15f4a8edb0f2b8d2", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" } }, { "result_path": "qualify-07.json", "result_sha256": "715ebdf8e07da5b9aa780f55b929f7fa5f028fc807c0686543e57b2d2e5f1b60", "result_uncompressed_sha256": "715ebdf8e07da5b9aa780f55b929f7fa5f028fc807c0686543e57b2d2e5f1b60", "result_uncompressed_bytes": 26643, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":16:8", "scope": "task", "status": "passed", "draw": 7, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "qualify-07-result.predictions.npy", "prediction_sha256": "30b0ad45eb3220be4cbf88281ebb4b77ef87e79228d7039a309fbf37f47a804a", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" } }, { "result_path": "qualify-08.json", "result_sha256": "83f6cadf3eeba1dbaed45b7fd7fd1e0fd1a6b97dfe19496c2c2a9ad4bfc1458f", "result_uncompressed_sha256": "83f6cadf3eeba1dbaed45b7fd7fd1e0fd1a6b97dfe19496c2c2a9ad4bfc1458f", "result_uncompressed_bytes": 26642, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":16:8", "scope": "task", "status": "passed", "draw": 8, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "qualify-08-result.predictions.npy", "prediction_sha256": "e58db6fd9957a160c2a340522bfa101a3cd3e798b7e95e7b26bb8819b9ba9d65", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" } }, { "result_path": "qualify-09.json", "result_sha256": "6fb5cf87ff6f534344d464e16509dffe589b09d0590bea1ab035798e29ea150f", "result_uncompressed_sha256": "6fb5cf87ff6f534344d464e16509dffe589b09d0590bea1ab035798e29ea150f", "result_uncompressed_bytes": 26643, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":16:8", "scope": "task", "status": "passed", "draw": 9, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "qualify-09-result.predictions.npy", "prediction_sha256": "bac8a4cbfd9ae48999f195535e4a5dcc17cd475a4b212d39376b7cf937070fda", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" } }, { "result_path": "qualify-10.json", "result_sha256": "b17ad1ad08074f6f00ca263cf7a820e8f2f0f016c7821d132327b54131f2851f", "result_uncompressed_sha256": "b17ad1ad08074f6f00ca263cf7a820e8f2f0f016c7821d132327b54131f2851f", "result_uncompressed_bytes": 26642, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":16:8", "scope": "task", "status": "passed", "draw": 10, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "qualify-10-result.predictions.npy", "prediction_sha256": "82cf23956ebfa10af72e8faa1ca0043bcc26492c17437708cebdc6b46a5a0717", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" } } ], "inherited_accuracy_valid": true, "draws_with_all_state_hashes_bit_equal": 0, "full_state_bit_equality_established": false, "equivalence_scope": "All class predictions, inputs, initial weights and epoch shuffles match. Different final-state or score hashes are explicitly retained; no numerical error bound is inferred from hashes.", "inherited_accuracy": { "correct": 98208, "total": 110000, "accuracy_percent": 89.28, "sample_sd_pp": 0.35908216329971143, "error_percent": 10.719999999999999, "meets_12_percent_error_target": true }, "base_accuracy_sha256": "eb0abcd7613d987074b664ade1e4140699f298d5d58dfa5698b925e0071c7485" }, "comparisons": { "baseline": { "result_path": "baseline.json", "result_sha256": "d3ad2a6e810459e53e9ecc3cacfdec761dca56ea1f6e8a97ad0d5dd66f056ef5", "result_uncompressed_sha256": "d3ad2a6e810459e53e9ecc3cacfdec761dca56ea1f6e8a97ad0d5dd66f056ef5", "result_uncompressed_bytes": 47833, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16", "--energy" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":4096:8", "scope": "task", "status": "passed", "draw": 0, "predictions_bit_equal": true, "all_state_hashes_bit_equal": true, "state_fields_differing_from_original": [], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "baseline-result.predictions.npy", "prediction_sha256": "fce8a9458d2158537c7d6207fb0c8e9cfcbba007e7a748976bf3e7cd002729f3", "memory": { "peak_allocated_bytes": 82509824, "live_allocated_bytes": 81084416, "peak_reserved_bytes": 96468992, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 67123552, "peak_allocated_mib": 78.6875, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": false, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" }, "energy": { "passed": true, "trials": 3, "invocations_per_trial": 104, "raw_counter_arithmetic_recomputed": true, "summary": { "idle_adjusted_mj_per_task": { "mean": 1002.9797130200805, "sample_sd": 55.63196816552205, "values": [ 1059.7543262993806, 1000.6192837807083, 948.5655289801525 ] }, "unadjusted_mj_per_task": { "mean": 3305.477564102564, "sample_sd": 32.12786539688782, "values": [ 3299.423076923077, 3340.201923076923, 3276.8076923076924 ] }, "cuda_ms_per_task": { "mean": 37.90967422876602, "sample_sd": 0.1610706374635057, "values": [ 37.77359713040865, 38.087512676532455, 37.86791287935697 ] }, "wall_ms_per_task": { "mean": 37.94228616025641, "sample_sd": 0.1615132449200883, "values": [ 37.803932177884604, 38.1197690144231, 37.90315728846152 ] } }, "hardware": { "gpu": "NVIDIA A100-SXM4-40GB", "uuid": "GPU-059d2560-15ff-0871-a190-9065becce1b2", "total_memory_bytes": 42405855232, "power_limit_w": 400.0, "driver": "580.95.05", "cuda_pci_bus_id": "0000:15:00.0", "nvml_pci_bus_id": "00000000:15:00.0", "nvml_handle_selected_by_cuda_pci_bus_id": true }, "scope": "Three paired-idle NVML trials for fresh complete tasks on draw zero" } }, "small-workspace": { "result_path": "small-workspace.json", "result_sha256": "f935630bb4b48f2563e54834f065a1b7f442113fd2e645bd86c9ea1b7ded1e0b", "result_uncompressed_sha256": "f935630bb4b48f2563e54834f065a1b7f442113fd2e645bd86c9ea1b7ded1e0b", "result_uncompressed_bytes": 47828, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16", "--energy" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":16:8", "scope": "task", "status": "passed", "draw": 0, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "small-workspace-result.predictions.npy", "prediction_sha256": "fce8a9458d2158537c7d6207fb0c8e9cfcbba007e7a748976bf3e7cd002729f3", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" }, "energy": { "passed": true, "trials": 3, "invocations_per_trial": 108, "raw_counter_arithmetic_recomputed": true, "summary": { "idle_adjusted_mj_per_task": { "mean": 1018.4190392485816, "sample_sd": 47.31792745681788, "values": [ 1071.6258801087217, 1002.5748159573074, 981.0564216797158 ] }, "unadjusted_mj_per_task": { "mean": 3418.015432098766, "sample_sd": 30.297940864386373, "values": [ 3444.472222222222, 3424.6111111111113, 3384.962962962963 ] }, "cuda_ms_per_task": { "mean": 38.8594066478588, "sample_sd": 0.15432854365317664, "values": [ 38.87688078703704, 39.00425437644676, 38.697084780092595 ] }, "wall_ms_per_task": { "mean": 38.89153690740737, "sample_sd": 0.15269777054414221, "values": [ 38.90498400462961, 39.037066407407345, 38.73256031018515 ] } }, "hardware": { "gpu": "NVIDIA A100-SXM4-40GB", "uuid": "GPU-87335ec7-4260-ce3c-9bea-69a4ecbf5771", "total_memory_bytes": 42405855232, "power_limit_w": 400.0, "driver": "580.95.05", "cuda_pci_bus_id": "0000:0F:00.0", "nvml_pci_bus_id": "00000000:0F:00.0", "nvml_handle_selected_by_cuda_pci_bus_id": true }, "scope": "Three paired-idle NVML trials for fresh complete tasks on draw zero" } }, "zero-workspace": { "result_path": "zero-workspace.json", "result_sha256": "f2bfb50e4594c151bca7b3014419f9b3d35180b76e02b01096fa57f1d4a3d79d", "result_uncompressed_sha256": "f2bfb50e4594c151bca7b3014419f9b3d35180b76e02b01096fa57f1d4a3d79d", "result_uncompressed_bytes": 47798, "result_compressed_sha256": null, "result_compressed_bytes": null, "result_gzip_contents_verified": false, "command": [ "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "measure", "--scope", "task", "--repeats", "16", "--energy" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n" }, "returncode": 0, "workspace": ":0:0", "scope": "task", "status": "passed", "draw": 0, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "zero-workspace-result.predictions.npy", "prediction_sha256": "fce8a9458d2158537c7d6207fb0c8e9cfcbba007e7a748976bf3e7cd002729f3", "memory": { "peak_allocated_bytes": 15400960, "live_allocated_bytes": 13975552, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 14688, "peak_allocated_mib": 14.6875, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" }, "energy": { "passed": true, "trials": 3, "invocations_per_trial": 124, "raw_counter_arithmetic_recomputed": true, "summary": { "idle_adjusted_mj_per_task": { "mean": 994.918786255717, "sample_sd": 5.529763585844365, "values": [ 989.2644849464956, 995.1769015186421, 1000.3149723020133 ] }, "unadjusted_mj_per_task": { "mean": 3203.755376344086, "sample_sd": 17.07544761864355, "values": [ 3186.282258064516, 3204.5806451612902, 3220.4032258064517 ] }, "cuda_ms_per_task": { "mean": 37.375288768481184, "sample_sd": 0.04146411220358215, "values": [ 37.33301568800403, 37.37695706275202, 37.4158935546875 ] }, "wall_ms_per_task": { "mean": 37.40340780913981, "sample_sd": 0.04254968975731325, "values": [ 37.35751584274191, 37.41115768548389, 37.44154989919362 ] } }, "hardware": { "gpu": "NVIDIA A100-SXM4-40GB", "uuid": "GPU-395167ee-ee3d-1e5b-4df9-d552d25cd56f", "total_memory_bytes": 42405855232, "power_limit_w": 400.0, "driver": "580.95.05", "cuda_pci_bus_id": "0000:DA:00.0", "nvml_pci_bus_id": "00000000:DA:00.0", "nvml_handle_selected_by_cuda_pci_bus_id": true }, "scope": "Three paired-idle NVML trials for fresh complete tasks on draw zero" } } }, "reductions": { "small-workspace": { "peak_allocated_bytes_saved": 66846720, "peak_allocated_reduction_percent": 81.01667990468626, "peak_allocated_ratio_to_baseline": 0.1898332009531374, "idle_adjusted_energy_ratio_to_baseline": 1.0153934581408548 }, "zero-workspace": { "peak_allocated_bytes_saved": 67108864, "peak_allocated_reduction_percent": 81.33439237490072, "peak_allocated_ratio_to_baseline": 0.18665607625099284, "idle_adjusted_energy_ratio_to_baseline": 0.9919630211262288 } }, "profiles": { "profile-baseline-task": { "result_path": "profile-baseline-task.json.gz", "result_sha256": "412620275c5cb7c8c5bc7617639d3bda0ac4fb9eed40a80cdede2e9655a5a00c", "result_uncompressed_sha256": "03d14d11ef00cd7aee383fe06ea5c639276bea5e4cb29e49f813214b6430648a", "result_uncompressed_bytes": 80370913, "result_compressed_sha256": "412620275c5cb7c8c5bc7617639d3bda0ac4fb9eed40a80cdede2e9655a5a00c", "result_compressed_bytes": 1587219, "result_gzip_contents_verified": true, "command": [ "/opt/nvidia/nsight-compute/2025.1.1/ncu", "--replay-mode", "app-range", "--cache-control", "none", "--clock-control", "none", "--metrics", "dram__bytes_read.sum,dram__bytes_write.sum,lts__t_sectors_op_read.sum,lts__t_sectors_op_write.sum,lts__t_sector_hit_rate.pct", "--csv", "--page", "raw", "--force-overwrite", "--export", "/tmp/cache-run/profile", "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "profile", "--scope", "task", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n", "metric_query": { "returncode": 0, "raw_query_sha256": "d985c130f3192a845ada05044ebdcac1df547c7bbaa3b3f5a5c712414433125e", "raw_query_retained_in": "profile-baseline-task.json.gz" } }, "returncode": 0, "workspace": ":4096:8", "scope": "task", "status": "passed", "draw": 0, "predictions_bit_equal": true, "all_state_hashes_bit_equal": true, "state_fields_differing_from_original": [], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "profile-baseline-task-result.predictions.npy", "prediction_sha256": "fce8a9458d2158537c7d6207fb0c8e9cfcbba007e7a748976bf3e7cd002729f3", "memory": { "peak_allocated_bytes": 82509824, "live_allocated_bytes": 81084416, "peak_reserved_bytes": 96468992, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 67123552, "peak_allocated_mib": 78.6875, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": false, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" }, "dram_traffic_measured": true, "profile": { "scope": "task", "warmup_complete_tasks": 2, "requested_repeats": 16, "fresh_tasks_in_range": 16, "fresh_training_sequences_in_range": 16, "epochs_per_sequence": 2, "training_scope_ignores_repeats": false, "reset_inside_range": true, "prediction_inside_range": true, "epoch_order_copies_inside_range": true, "output_serialization_inside_range": false, "cpu_validation_inside_range": false, "synchronized_before_stop": true }, "ncu_report_path": "profile-baseline-task-profile.ncu-rep.gz", "ncu_report_sha256": "d4d72f72149a32ac9285e48294e7c2b8b42b5adfb3ff3fd82e1a9ec495e0aeb9", "ncu_report_uncompressed_sha256": "d63ce9073e7b4564cf1de5b514d1823bdb573fa9cfede2dd1051bc0a3fcd2f66", "ncu_report_uncompressed_bytes": 88576540, "ncu_report_compressed_sha256": "d4d72f72149a32ac9285e48294e7c2b8b42b5adfb3ff3fd82e1a9ec495e0aeb9", "ncu_report_compressed_bytes": 11959402, "ncu_report_gzip_contents_verified": true, "metric_rows": { "dram__bytes_read.sum": { "value": 49300864.0, "raw_value": "49300864", "raw_unit": "byte", "range_id": "0" }, "dram__bytes_write.sum": { "value": 29329408.0, "raw_value": "29329408", "raw_unit": "byte", "range_id": "0" }, "lts__t_sectors_op_read.sum": { "value": 474067036.0, "raw_value": "474067036", "raw_unit": "sector", "range_id": "0" }, "lts__t_sectors_op_write.sum": { "value": 237557538.0, "raw_value": "237557538", "raw_unit": "sector", "range_id": "0" }, "lts__t_sector_hit_rate.pct": { "value": 95.49, "raw_value": "95.49", "raw_unit": "%", "range_id": "0" } }, "pass_counts": [ 2 ], "profile_device_attributes": { "device__attribute_display_name": "NVIDIA A100-SXM4-80GB", "device__attribute_l2_cache_size": "41943040", "device__attribute_total_memory": "85094825984", "device__attribute_memory_clock_rate": "1593000", "device__attribute_multiprocessor_count": "108", "profiler__replayer_passes": "2", "profiler__replayer_passes_type_warmup": "0", "gpu__time_duration.sum": "2584336832" }, "pass_count_retained": true, "range_training_sequences": 16, "dram_read_bytes_per_sequence": 3081304.0, "dram_write_bytes_per_sequence": 1833088.0, "dram_total_bytes_per_sequence": 4914392.0, "l2_read_request_bytes_per_sequence": 948134072.0, "l2_write_request_bytes_per_sequence": 475115076.0, "l2_sector_hit_rate_percent": 95.49, "counter_unit": "per complete fresh reset/train/predict task" }, "profile-baseline-training": { "result_path": "profile-baseline-training.json.gz", "result_sha256": "426de66111db1a5fd1e9b4b5d4ea7e2ce56ceb3570aa23c882977edfaa1e8431", "result_uncompressed_sha256": "5aebbd9227c77209ff8cbc6693ed96be7d4e574b2fed9aff15a88ab58f37c2df", "result_uncompressed_bytes": 80370909, "result_compressed_sha256": "426de66111db1a5fd1e9b4b5d4ea7e2ce56ceb3570aa23c882977edfaa1e8431", "result_compressed_bytes": 1587229, "result_gzip_contents_verified": true, "command": [ "/opt/nvidia/nsight-compute/2025.1.1/ncu", "--replay-mode", "app-range", "--cache-control", "none", "--clock-control", "none", "--metrics", "dram__bytes_read.sum,dram__bytes_write.sum,lts__t_sectors_op_read.sum,lts__t_sectors_op_write.sum,lts__t_sector_hit_rate.pct", "--csv", "--page", "raw", "--force-overwrite", "--export", "/tmp/cache-run/profile", "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "profile", "--scope", "training", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n", "metric_query": { "returncode": 0, "raw_query_sha256": "adaab0cf189570bd1f0e332014f1dd49b40842687eedc0929b6867e8723c54c5", "raw_query_retained_in": "profile-baseline-training.json.gz" } }, "returncode": 0, "workspace": ":4096:8", "scope": "training", "status": "passed", "draw": 0, "predictions_bit_equal": true, "all_state_hashes_bit_equal": true, "state_fields_differing_from_original": [], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "profile-baseline-training-result.predictions.npy", "prediction_sha256": "fce8a9458d2158537c7d6207fb0c8e9cfcbba007e7a748976bf3e7cd002729f3", "memory": { "peak_allocated_bytes": 82509824, "live_allocated_bytes": 81084416, "peak_reserved_bytes": 96468992, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 67123552, "peak_allocated_mib": 78.6875, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": false, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" }, "dram_traffic_measured": true, "profile": { "scope": "training", "warmup_complete_tasks": 2, "requested_repeats": 16, "fresh_tasks_in_range": 0, "fresh_training_sequences_in_range": 1, "epochs_per_sequence": 2, "training_scope_ignores_repeats": true, "reset_inside_range": false, "prediction_inside_range": false, "epoch_order_copies_inside_range": true, "output_serialization_inside_range": false, "cpu_validation_inside_range": false, "synchronized_before_stop": true }, "ncu_report_path": "profile-baseline-training-profile.ncu-rep.gz", "ncu_report_sha256": "8942e31975b4da1c2930d3f8c2e6e60f51bccf7da1a5ae614356355b7da9679b", "ncu_report_uncompressed_sha256": "edf26ff2704e497147a22047d01225e3953b7b4a913e49c63a4f3c633ef6976c", "ncu_report_uncompressed_bytes": 6742899, "ncu_report_compressed_sha256": "8942e31975b4da1c2930d3f8c2e6e60f51bccf7da1a5ae614356355b7da9679b", "ncu_report_compressed_bytes": 870848, "ncu_report_gzip_contents_verified": true, "metric_rows": { "dram__bytes_read.sum": { "value": 2265984.0, "raw_value": "2265984", "raw_unit": "byte", "range_id": "0" }, "dram__bytes_write.sum": { "value": 1896832.0, "raw_value": "1896832", "raw_unit": "byte", "range_id": "0" }, "lts__t_sectors_op_read.sum": { "value": 26466194.0, "raw_value": "26466194", "raw_unit": "sector", "range_id": "0" }, "lts__t_sectors_op_write.sum": { "value": 12477848.0, "raw_value": "12477848", "raw_unit": "sector", "range_id": "0" }, "lts__t_sector_hit_rate.pct": { "value": 95.29, "raw_value": "95.29", "raw_unit": "%", "range_id": "0" } }, "pass_counts": [ 2 ], "profile_device_attributes": { "device__attribute_display_name": "NVIDIA A100-SXM4-40GB", "device__attribute_l2_cache_size": "41943040", "device__attribute_total_memory": "42405855232", "device__attribute_memory_clock_rate": "1215000", "device__attribute_multiprocessor_count": "108", "profiler__replayer_passes": "2", "profiler__replayer_passes_type_warmup": "0", "gpu__time_duration.sum": "144168640" }, "pass_count_retained": true, "range_training_sequences": 1, "dram_read_bytes_per_sequence": 2265984.0, "dram_write_bytes_per_sequence": 1896832.0, "dram_total_bytes_per_sequence": 4162816.0, "l2_read_request_bytes_per_sequence": 846918208.0, "l2_write_request_bytes_per_sequence": 399291136.0, "l2_sector_hit_rate_percent": 95.29, "counter_unit": "per original two-epoch training sequence, reset and prediction outside range" }, "profile-small-task": { "result_path": "profile-small-task.json.gz", "result_sha256": "e7c527c7b1ada773d6b176f30e84c97e36886f3521558de8f80ab2040c56c23c", "result_uncompressed_sha256": "1a4cd304ffb3d28ee33b841da45b404bd0d47c00c35366cb134f21439c6d0c51", "result_uncompressed_bytes": 80370903, "result_compressed_sha256": "e7c527c7b1ada773d6b176f30e84c97e36886f3521558de8f80ab2040c56c23c", "result_compressed_bytes": 1587223, "result_gzip_contents_verified": true, "command": [ "/opt/nvidia/nsight-compute/2025.1.1/ncu", "--replay-mode", "app-range", "--cache-control", "none", "--clock-control", "none", "--metrics", "dram__bytes_read.sum,dram__bytes_write.sum,lts__t_sectors_op_read.sum,lts__t_sectors_op_write.sum,lts__t_sector_hit_rate.pct", "--csv", "--page", "raw", "--force-overwrite", "--export", "/tmp/cache-run/profile", "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "profile", "--scope", "task", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n", "metric_query": { "returncode": 0, "raw_query_sha256": "adaab0cf189570bd1f0e332014f1dd49b40842687eedc0929b6867e8723c54c5", "raw_query_retained_in": "profile-small-task.json.gz" } }, "returncode": 0, "workspace": ":16:8", "scope": "task", "status": "passed", "draw": 0, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "profile-small-task-result.predictions.npy", "prediction_sha256": "fce8a9458d2158537c7d6207fb0c8e9cfcbba007e7a748976bf3e7cd002729f3", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" }, "dram_traffic_measured": true, "profile": { "scope": "task", "warmup_complete_tasks": 2, "requested_repeats": 16, "fresh_tasks_in_range": 16, "fresh_training_sequences_in_range": 16, "epochs_per_sequence": 2, "training_scope_ignores_repeats": false, "reset_inside_range": true, "prediction_inside_range": true, "epoch_order_copies_inside_range": true, "output_serialization_inside_range": false, "cpu_validation_inside_range": false, "synchronized_before_stop": true }, "ncu_report_path": "profile-small-task-profile.ncu-rep.gz", "ncu_report_sha256": "b23501d03bba541bf48a02cbbf09481762144553f69b09654e99f93d33aee4f8", "ncu_report_uncompressed_sha256": "bd4bb547c58a0625203d5992fe5bf0830cb322022f627ba272899660cf849afa", "ncu_report_uncompressed_bytes": 85003041, "ncu_report_compressed_sha256": "b23501d03bba541bf48a02cbbf09481762144553f69b09654e99f93d33aee4f8", "ncu_report_compressed_bytes": 11416744, "ncu_report_gzip_contents_verified": true, "metric_rows": { "dram__bytes_read.sum": { "value": 52088576.0, "raw_value": "52088576", "raw_unit": "byte", "range_id": "0" }, "dram__bytes_write.sum": { "value": 29900800.0, "raw_value": "29900800", "raw_unit": "byte", "range_id": "0" }, "lts__t_sectors_op_read.sum": { "value": 429809710.0, "raw_value": "429809710", "raw_unit": "sector", "range_id": "0" }, "lts__t_sectors_op_write.sum": { "value": 226110828.0, "raw_value": "226110828", "raw_unit": "sector", "range_id": "0" }, "lts__t_sector_hit_rate.pct": { "value": 95.45, "raw_value": "95.45", "raw_unit": "%", "range_id": "0" } }, "pass_counts": [ 2 ], "profile_device_attributes": { "device__attribute_display_name": "NVIDIA A100-SXM4-40GB", "device__attribute_l2_cache_size": "41943040", "device__attribute_total_memory": "42405855232", "device__attribute_memory_clock_rate": "1215000", "device__attribute_multiprocessor_count": "108", "profiler__replayer_passes": "2", "profiler__replayer_passes_type_warmup": "0", "gpu__time_duration.sum": "2451604224" }, "pass_count_retained": true, "range_training_sequences": 16, "dram_read_bytes_per_sequence": 3255536.0, "dram_write_bytes_per_sequence": 1868800.0, "dram_total_bytes_per_sequence": 5124336.0, "l2_read_request_bytes_per_sequence": 859619420.0, "l2_write_request_bytes_per_sequence": 452221656.0, "l2_sector_hit_rate_percent": 95.45, "counter_unit": "per complete fresh reset/train/predict task" }, "profile-small-training": { "result_path": "profile-small-training.json.gz", "result_sha256": "ebb2dfa1fd5162034234f0fb064b6a51ee58965a0682cab88776a7f8a0b47f5c", "result_uncompressed_sha256": "c2a2d33a2a0f4a823a7a24fb5643d0ae23d01ead2920dd1832ec0dbc66bace8e", "result_uncompressed_bytes": 80370901, "result_compressed_sha256": "ebb2dfa1fd5162034234f0fb064b6a51ee58965a0682cab88776a7f8a0b47f5c", "result_compressed_bytes": 1587197, "result_gzip_contents_verified": true, "command": [ "/opt/nvidia/nsight-compute/2025.1.1/ncu", "--replay-mode", "app-range", "--cache-control", "none", "--clock-control", "none", "--metrics", "dram__bytes_read.sum,dram__bytes_write.sum,lts__t_sectors_op_read.sum,lts__t_sectors_op_write.sum,lts__t_sector_hit_rate.pct", "--csv", "--page", "raw", "--force-overwrite", "--export", "/tmp/cache-run/profile", "/usr/local/bin/python", "/workspace/cache/fixture.py", "--input", "/tmp/cache-run/input.npz", "--config", "/tmp/cache-run/config.json", "--output", "/tmp/cache-run/result.json", "--mode", "profile", "--scope", "training", "--repeats", "16" ], "capabilities": { "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n", "metric_query": { "returncode": 0, "raw_query_sha256": "adaab0cf189570bd1f0e332014f1dd49b40842687eedc0929b6867e8723c54c5", "raw_query_retained_in": "profile-small-training.json.gz" } }, "returncode": 0, "workspace": ":16:8", "scope": "training", "status": "passed", "draw": 0, "predictions_bit_equal": true, "all_state_hashes_bit_equal": false, "state_fields_differing_from_original": [ "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "scores_sha256" ], "state_fields_checked": [ "input_hashes", "initial_parameter_sha256", "epoch_permutation_sha256", "final_parameter_sha256", "final_velocity_sha256", "state_vector_sha256", "predictions_sha256", "scores_sha256" ], "prediction_path": "profile-small-training-result.predictions.npy", "prediction_sha256": "fce8a9458d2158537c7d6207fb0c8e9cfcbba007e7a748976bf3e7cd002729f3", "memory": { "peak_allocated_bytes": 15663104, "live_allocated_bytes": 14237696, "peak_reserved_bytes": 29360128, "named_storage_bytes": 13960864, "unnamed_live_allocated_bytes": 276832, "peak_allocated_mib": 14.9375, "nominal_l2_reference_bytes": 41943040, "peak_allocated_below_nominal_l2_capacity": true, "cache_residency_established": false, "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup." }, "software": { "python": "3.11.15", "torch": "2.5.1+cu124", "cuda": "12.4", "numpy": "2.2.6" }, "dram_traffic_measured": true, "profile": { "scope": "training", "warmup_complete_tasks": 2, "requested_repeats": 16, "fresh_tasks_in_range": 0, "fresh_training_sequences_in_range": 1, "epochs_per_sequence": 2, "training_scope_ignores_repeats": true, "reset_inside_range": false, "prediction_inside_range": false, "epoch_order_copies_inside_range": true, "output_serialization_inside_range": false, "cpu_validation_inside_range": false, "synchronized_before_stop": true }, "ncu_report_path": "profile-small-training-profile.ncu-rep.gz", "ncu_report_sha256": "e7171b7daf0cfa33b3dc594db3539f64b0bc198f57cdec570e9b5ac9dc7257fc", "ncu_report_uncompressed_sha256": "eb058c1179676d6c709258997877650be716f1f682a0acd1d3c3895b37c38b3c", "ncu_report_uncompressed_bytes": 6534891, "ncu_report_compressed_sha256": "e7171b7daf0cfa33b3dc594db3539f64b0bc198f57cdec570e9b5ac9dc7257fc", "ncu_report_compressed_bytes": 836089, "ncu_report_gzip_contents_verified": true, "metric_rows": { "dram__bytes_read.sum": { "value": 2036480.0, "raw_value": "2036480", "raw_unit": "byte", "range_id": "0" }, "dram__bytes_write.sum": { "value": 1595648.0, "raw_value": "1595648", "raw_unit": "byte", "range_id": "0" }, "lts__t_sectors_op_read.sum": { "value": 23587372.0, "raw_value": "23587372", "raw_unit": "sector", "range_id": "0" }, "lts__t_sectors_op_write.sum": { "value": 11740492.0, "raw_value": "11740492", "raw_unit": "sector", "range_id": "0" }, "lts__t_sector_hit_rate.pct": { "value": 97.58, "raw_value": "97.58", "raw_unit": "%", "range_id": "0" } }, "pass_counts": [ 2 ], "profile_device_attributes": { "device__attribute_display_name": "NVIDIA A100-SXM4-40GB", "device__attribute_l2_cache_size": "41943040", "device__attribute_total_memory": "42405855232", "device__attribute_memory_clock_rate": "1215000", "device__attribute_multiprocessor_count": "108", "profiler__replayer_passes": "2", "profiler__replayer_passes_type_warmup": "0", "gpu__time_duration.sum": "149245664" }, "pass_count_retained": true, "range_training_sequences": 1, "dram_read_bytes_per_sequence": 2036480.0, "dram_write_bytes_per_sequence": 1595648.0, "dram_total_bytes_per_sequence": 3632128.0, "l2_read_request_bytes_per_sequence": 754795904.0, "l2_write_request_bytes_per_sequence": 375695744.0, "l2_sector_hit_rate_percent": 97.58, "counter_unit": "per original two-epoch training sequence, reset and prediction outside range" } }, "scope": "Independent CPU comparison against frozen predictions and state, storage accounting, and retained NVML/Ncu counters. No test labels or GPU retraining used.", "cache_claim": "Capacity and measured traffic are separate. No claim that every access hits L2 or that profiling proves universal cache residency." }