{
  "protocol": {
    "operation": "torch.mm (2048x2048 FP32 GEMM, routed through cuBLAS)",
    "trials_per_condition": 20,
    "pairs_per_condition": 190,
    "comparison": "all-pairs bitwise equality on FP32 outputs viewed as int32, plus max elementwise absolute difference",
    "inputs": "generated once on CPU with torch.Generator(seed=42) and transferred; identical bytes across conditions",
    "isolation": "each condition runs in a fresh subprocess so CUBLAS_WORKSPACE_CONFIG is set (or absent) before CUDA initialization",
    "stress_arm": "four CUDA streams issuing concurrent GEMMs sharing the thread's cuBLAS handle; first stream's product tracked across trials",
    "date": "2026-07-18",
    "runtime": "Google Colab",
    "source": "Summary statistics transcribed from the printed cublas_determinism_receipt.json of the Colab run."
  },
  "environment": {
    "pytorch_version": "2.11.0+cu128",
    "cuda_version": "12.8",
    "gpu": "NVIDIA A100-SXM4-80GB",
    "dtype": "float32"
  },
  "conditions": {
    "probe_no_config": {
      "deterministic_mode": true,
      "cublas_workspace_config": "",
      "result": "completed (no RuntimeError)"
    },
    "probe_with_config": {
      "deterministic_mode": true,
      "cublas_workspace_config": ":4096:8",
      "result": "completed"
    },
    "det_run_with_config": {
      "deterministic_mode": true,
      "cublas_workspace_config": ":4096:8",
      "pairs": 190,
      "exact_pairs": 190,
      "max_div": 0.0
    },
    "stress_no_config": {
      "deterministic_mode": false,
      "cublas_workspace_config": "",
      "pairs": 190,
      "exact_pairs": 190,
      "max_div": 0.0
    },
    "stress_with_config": {
      "deterministic_mode": false,
      "cublas_workspace_config": ":4096:8",
      "pairs": 190,
      "exact_pairs": 190,
      "max_div": 0.0
    }
  },
  "interpretation": {
    "finding": "On this stack, torch.mm was bit-identical across all pairs in every condition, and deterministic mode no longer raises without CUBLAS_WORKSPACE_CONFIG.",
    "explanation": "PyTorch allocates cuBLAS workspaces explicitly per handle/stream combination (pytorch/pytorch PR 85447), which satisfies cuBLAS's documented reproducibility conditions; the RuntimeError check requiring CUBLAS_WORKSPACE_CONFIG was removed in pytorch/pytorch PR 161749.",
    "scope": "Stack-scoped observation for PyTorch 2.11.0+cu128 on one A100; older PyTorch versions raise a RuntimeError from torch.mm/mv/bmm under deterministic mode unless the environment variable is set."
  }
}