{
  "schema_version": 1,
  "what": "MPS active-thread cap enforcement with the WG3 GPU services MANAGED",
  "harness": "scripts/cotenancy/mps-cap-sweep.py",
  "harness_note": "Runs against the STACK's MPS control daemon, not a throwaway private one, and records the daemon's own client list around every cell. Re-runnable: re-execute this file.",
  "host": "gb10-ref",
  "recorded_at": "2026-08-04T12:29:57.922893Z",
  "pipe_directory": "/tmp/nvidia-mps",
  "cap_semantics": "CUDA_MPS_ACTIVE_THREAD_PERCENTAGE PROVISIONS but does not RESERVE. NVIDIA documents that kernels from different clients may still execute on the same SM. A cap shaping throughput is NOT evidence of isolation, and MPS gives no memory isolation at all.",
  "protocol": {
    "caps_percent": "100,75,50,25,10",
    "iters": 80,
    "matrix_size": 4096,
    "dtype": "fp16",
    "op": "matmul",
    "client_runtime": "container amini/wg3-sr-worker:latest as uid 0"
  },
  "managed_clients_before": {
    "value": [
      1746902,
      1809847
    ],
    "state": "measured",
    "source": "nvidia-cuda-mps-control get_client_list",
    "measured_at": "2026-08-04T12:29:57.922893Z",
    "evidence": {
      "basis": "the daemon's own client list"
    },
    "unknown_reason": null
  },
  "managed_clients_after": {
    "value": [
      1746902,
      1809847
    ],
    "state": "measured",
    "source": "nvidia-cuda-mps-control get_client_list",
    "measured_at": "2026-08-04T12:29:57.922893Z",
    "evidence": {
      "basis": "the daemon's own client list"
    },
    "unknown_reason": null
  },
  "unmanaged_tenants": {
    "value": [
      {
        "pid": 4355,
        "used_memory_mb": 655,
        "process_name": "~/engines/llama.cpp/build/bin/llama-server"
      },
      {
        "pid": 4349,
        "used_memory_mb": 5723,
        "process_name": "~/.venvs/voxcpm-trial/bin/python"
      },
      {
        "pid": 24093,
        "used_memory_mb": 1106,
        "process_name": "/app/.venv/bin/python3"
      },
      {
        "pid": 32440,
        "used_memory_mb": 209,
        "process_name": "~/Ulap/ULAP-ONE/radio-planner/.venv/bin/python"
      },
      {
        "pid": 2459827,
        "used_memory_mb": 489,
        "process_name": "python"
      }
    ],
    "state": "measured",
    "source": "nvidia-smi --query-compute-apps minus get_client_list",
    "measured_at": "2026-08-04T12:29:57.922893Z",
    "evidence": {
      "note": "foreign tenants belonging to other projects; NOT stopped by this script, which is why this is a co-tenancy number and not a clean-box number"
    },
    "unknown_reason": null
  },
  "cells": [
    {
      "cap": "100",
      "iters": 80,
      "size": 4096,
      "seconds": 0.22185,
      "tflops": 49.56,
      "client_reported_pid": 1,
      "client_pid_is_namespaced": true,
      "mps_clients_joined_during_cell": [
        2025497
      ],
      "self_was_managed": true,
      "managed_clients_after_cell": [
        1746902,
        1809847
      ]
    },
    {
      "cap": "75",
      "iters": 80,
      "size": 4096,
      "seconds": 0.30176,
      "tflops": 36.44,
      "client_reported_pid": 1,
      "client_pid_is_namespaced": true,
      "mps_clients_joined_during_cell": [
        2027478
      ],
      "self_was_managed": true,
      "managed_clients_after_cell": [
        1746902,
        1809847
      ]
    },
    {
      "cap": "50",
      "iters": 80,
      "size": 4096,
      "seconds": 0.48426,
      "tflops": 22.71,
      "client_reported_pid": 1,
      "client_pid_is_namespaced": true,
      "mps_clients_joined_during_cell": [
        2029584
      ],
      "self_was_managed": true,
      "managed_clients_after_cell": [
        1746902,
        1809847
      ]
    },
    {
      "cap": "25",
      "iters": 80,
      "size": 4096,
      "seconds": 0.77205,
      "tflops": 14.24,
      "client_reported_pid": 1,
      "client_pid_is_namespaced": true,
      "mps_clients_joined_during_cell": [
        2031922
      ],
      "self_was_managed": true,
      "managed_clients_after_cell": [
        1746902,
        1809847
      ]
    },
    {
      "cap": "10",
      "iters": 80,
      "size": 4096,
      "seconds": 1.2846,
      "tflops": 8.56,
      "client_reported_pid": 1,
      "client_pid_is_namespaced": true,
      "mps_clients_joined_during_cell": [
        2033827
      ],
      "self_was_managed": true,
      "managed_clients_after_cell": [
        1746902,
        1809847
      ]
    }
  ],
  "tflops_by_cap": {
    "value": {
      "100": 49.56,
      "75": 36.44,
      "50": 22.71,
      "25": 14.24,
      "10": 8.56
    },
    "state": "measured",
    "source": "fp16 matmul inside a managed MPS client",
    "measured_at": "2026-08-04T12:29:57.922893Z",
    "evidence": {
      "iters": 80,
      "size": 4096,
      "every_cell_self_managed": true
    },
    "unknown_reason": null
  },
  "complete_sweep": true,
  "partial_run_note": null
}