fix(amdsmi): process visibility, per-GPU attribution, table alignment

[marifamd]

Motivation

Three defects in amd-smi process (and the process table in default amd-smi
output):

1. It could return "No running processes detected" for a caller that lacks
   permission to inspect the running GPU jobs, even though the jobs were clearly
   using the GPUs.
2. The same compute process was reported under every GPU, rather than only
   the GPU(s) it actually uses.
3. The compute-process table mis-rendered its CU % and SDMA columns.

Technical Details

• rocm_smi/src/rocm_smi_kfd.cc

  • IsKfdPidNamespaced() inspected the first KFD PID via PidHasKfdOpen(),
    which opens /proc/<pid>/fd. When that PID belongs to another user, the
    directory is unreadable and the helper returned false, which was
    misinterpreted as evidence of a separate PID namespace. Enumeration then
    fell back to scanning /proc, where the same unreadable fd set dropped the
    process entirely, so the caller saw an empty list.
  • Add a tri-state PidKfdOpenState() distinguishing "has /dev/kfd open",
    "definitely does not", and "inaccessible" (another user's process, or the
    process already exited). Permission failures are now treated as inconclusive
    and skipped rather than flagged as namespaced, so such processes are listed
    with a redacted (N/A) name instead of disappearing.
  • GetProcessGPUs() ran the per-node KFD context-file fallback
    (GetKfdGpuIdsForPid()) unconditionally. Opening /dev/kfd creates
    vram_/stats_/counters_/sdma_<gpuid> files for the whole GPU topology, so a
    job with queues on a single GPU (or none) was attributed to every GPU.
    Attribute a process to a GPU only when it has an active KFD queue on it, or
    a non-zero vram_<gpuid> allocation on it. The zero-valued
    stats_/counters_/sdma_ per-node files are ignored. This lists a process
    against the GPUs it actually uses, aligning with rocm-smi --showpids. The
    container (PID-namespace) vram_* fallback is unchanged.

• amdsmi_cli/amdsmi_logger.py

  • Rebuild the process-table header so labels sit over their right-justified
    value columns (the SDMA header was one column too far left), drop the
    redundant % suffix from the CU value (the unit is in the CU % header),
    and widen SDMA to fit its unit by reclaiming two characters from the
    process-name column. The row stays within the fixed 80-character box. The
    row rendering is extracted into AMDSMILogger._format_process_row /
    PROCESS_TABLE_HEADER so the widths can be unit-tested.

• tests/python/unit/gpu/

  • test_kfd_process_gpus.py: hardware-free spec pinning the queue + non-zero
    VRAM attribution (including that an unreadable vram_* entry is skipped).
  • test_process_table_format.py: hardware-free coverage for the process-table
    column widths (80-character box, dropped % suffix, header-over-value
    alignment).

JIRA ID

JIRA ID: AILITOOLS-283
JIRA ID: ROCM-26000

Test Plan

• On multi-GPU hosts running a compute workload, ran amd-smi process and the
  default amd-smi process table as a non-root user and as root, with
  the installed library and with the fixed library, cross-checking against the
  ground-truth KFD queue gpuids in /sys/class/kfd/kfd/proc/<pid>/queues/ and
  the vram_<gpuid> allocations.
• Verified column alignment for both the N/A and populated CU %/SDMA cases.
• Ran the hardware-free unit tests.

Test Result

• Non-root now lists the root-owned process (previously non-root saw "No running
  processes detected"), with the process name redacted to N/A.
• Per-GPU attribution: on multi-GPU MI300 hosts, a job with queues + VRAM on a
  single GPU is now listed only on that GPU (confirmed against RVS on a gfx125X
  host: RVS with queues/VRAM on one GPU appeared under all 4 GPUs before the fix
  and only the correct GPU after, matching rocm-smi --showpids GPU(s)=1).
• CU % and SDMA headers align over their values within the 80-character box;
  valid values are no longer truncated.
• tests/python/unit/gpu/test_kfd_process_gpus.py and test_process_table_format.py
  pass.

[therock-pr-bot]

✅ All Policy Checks Passed

Check	Status	Details
🌿 Branch Name	✅ Pass	—
📝 PR Title/Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

🎉 All policy checks passed!

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

[therock-pr-bot]

🚫 Please fix the failed policies before requesting reviews.

The following policy checks failed:

• ❌ PR Title/Description

The Not ready to Review label has been added to this PR.
Once all policies pass, the label will be removed automatically.