{
  "date": "2026-09-11",
  "experiment": "Equal8GiB GPU allocation, device visibility and process count",
  "hardware": "2x Intel Arc Pro B70",
  "host": {
    "os": "Ubuntu26.04.1",
    "kernel": "7.0.0-31-generic",
    "compute_runtime": "26.27.39122.11-0",
    "level_zero_loader": "1.28.2-2",
    "reported_memtotal_kib": 62850252
  },
  "container_image": "ghcr.io/rmacy/qwen38-fp8-dspark@sha256:e5b9b8b494d94f344f04c795c202a83be3ddb3ba63bf9d852335d4d20fb26d24",
  "packages": {
    "vllm": "0.21.1.dev0+gad7125a43.d20260810.xpu",
    "torch": "2.11.0+xpu",
    "transformers": "5.8.0",
    "triton-xpu": "3.7.0"
  },
  "method": {
    "allocation": "4GiB uint8 tensor per card, filled with1 and synchronized; held8seconds",
    "isolation": "ZE_AFFINITY_MASK=0 and1 set separately before each child imports torch; both use local xpu:0",
    "control": "Two subprocesses without a visibility mask, using xpu:0 and xpu:1 respectively",
    "accounting": "Sum post-allocation drm-resident-gtt across unique (physical device, DRM client ID); retain original units during conversion. Snapshots per child are asynchronous. Do not interpret this as unique whole-system RAM usage or add it to RSS.",
    "repetitions": "Three fresh container runs for each layout, nine total",
    "network": "disabled; no model weights mounted"
  },
  "limitations": [
    "Allocation probe only: no vLLM serving, model quality, oneCCL collectives or DFlash validation.",
    "Tested only on this Linux B70 host and cached runtime.",
    "Driver counters can include runtime overhead and shared mappings.",
    "Visibility selection must occur before runtime initialization; distributed worker mapping requires separate integration."
  ],
  "results": [
    {
      "run": "2026-09-11-memory-probe",
      "layout": "one-process-both",
      "started_utc": "2026-09-11T08:10:55.711859+00:00",
      "finished_utc": "2026-09-11T08:11:06.599591+00:00",
      "exit_code": 0,
      "explicit_gpu_allocation_bytes": 8589934592,
      "summed_drm_resident_gtt_bytes": 8735072256,
      "drm_client_count": 4,
      "raw_log_sha256": "7a51da189f6c7bec82a8be271c451e559662cc72dfde3d114beba618f5a554b5"
    },
    {
      "run": "2026-09-11-memory-probe",
      "layout": "two-process-isolated",
      "started_utc": "2026-09-11T08:11:16.339307+00:00",
      "finished_utc": "2026-09-11T08:11:25.680779+00:00",
      "exit_code": 0,
      "explicit_gpu_allocation_bytes": 8589934592,
      "summed_drm_resident_gtt_bytes": 149209088,
      "drm_client_count": 4,
      "raw_log_sha256": "2f5493514cc464e4a2f204edcc65b5e689acc17f5cbf7332035976201b14daef"
    },
    {
      "run": "2026-09-11-memory-probe",
      "layout": "two-process-visible",
      "started_utc": "2026-09-11T08:11:06.599741+00:00",
      "finished_utc": "2026-09-11T08:11:16.339177+00:00",
      "exit_code": 0,
      "explicit_gpu_allocation_bytes": 8589934592,
      "summed_drm_resident_gtt_bytes": 8806547456,
      "drm_client_count": 8,
      "raw_log_sha256": "2a28036a315e0e3873bd85127a44454b654ca1dbc3adecc70bee1b61fc183f08"
    },
    {
      "run": "2026-09-11-memory-probe-2",
      "layout": "one-process-both",
      "started_utc": "2026-09-11T08:12:14.340009+00:00",
      "finished_utc": "2026-09-11T08:12:24.198355+00:00",
      "exit_code": 0,
      "explicit_gpu_allocation_bytes": 8589934592,
      "summed_drm_resident_gtt_bytes": 8735072256,
      "drm_client_count": 4,
      "raw_log_sha256": "589e81bb1f2a3ef5ea5969fcadb40fa477916db19c2c76777ce481ab59a2daa7"
    },
    {
      "run": "2026-09-11-memory-probe-2",
      "layout": "two-process-isolated",
      "started_utc": "2026-09-11T08:12:33.985214+00:00",
      "finished_utc": "2026-09-11T08:12:43.319142+00:00",
      "exit_code": 0,
      "explicit_gpu_allocation_bytes": 8589934592,
      "summed_drm_resident_gtt_bytes": 149209088,
      "drm_client_count": 4,
      "raw_log_sha256": "489754c4e6a9bbabfbab66aa8ede8bb8c5c22b97b32cb277144990e614338b38"
    },
    {
      "run": "2026-09-11-memory-probe-2",
      "layout": "two-process-visible",
      "started_utc": "2026-09-11T08:12:24.208364+00:00",
      "finished_utc": "2026-09-11T08:12:33.975962+00:00",
      "exit_code": 0,
      "explicit_gpu_allocation_bytes": 8589934592,
      "summed_drm_resident_gtt_bytes": 8806547456,
      "drm_client_count": 8,
      "raw_log_sha256": "629d93d08697b00893ea482811074ab7970e83c23abb36c027feb403504a1f68"
    },
    {
      "run": "2026-09-11-memory-probe-3",
      "layout": "one-process-both",
      "started_utc": "2026-09-11T08:12:43.346472+00:00",
      "finished_utc": "2026-09-11T08:12:53.157410+00:00",
      "exit_code": 0,
      "explicit_gpu_allocation_bytes": 8589934592,
      "summed_drm_resident_gtt_bytes": 8735072256,
      "drm_client_count": 4,
      "raw_log_sha256": "ecea0366d34b8016e65f30d589d325c8f2d152550474e823d99e2ccb7be2b7e3"
    },
    {
      "run": "2026-09-11-memory-probe-3",
      "layout": "two-process-isolated",
      "started_utc": "2026-09-11T08:13:03.045100+00:00",
      "finished_utc": "2026-09-11T08:13:12.396728+00:00",
      "exit_code": 0,
      "explicit_gpu_allocation_bytes": 8589934592,
      "summed_drm_resident_gtt_bytes": 149209088,
      "drm_client_count": 4,
      "raw_log_sha256": "1862e14f6c01ca630ea32d48f6d4758a114f6e88e0caaad7ca3fbe6ed2e79cc1"
    },
    {
      "run": "2026-09-11-memory-probe-3",
      "layout": "two-process-visible",
      "started_utc": "2026-09-11T08:12:53.167386+00:00",
      "finished_utc": "2026-09-11T08:13:03.036511+00:00",
      "exit_code": 0,
      "explicit_gpu_allocation_bytes": 8589934592,
      "summed_drm_resident_gtt_bytes": 8806547456,
      "drm_client_count": 8,
      "raw_log_sha256": "bee5d03a1e586e23632ae1259b1b77d5dc111c84fb1674f0e3fe269f990ee9d0"
    }
  ]
}