[bong-water-water-bong]

A fix onto #2023's branch (1bit/i669-2018-tests-py) for its red build. Refs #2018. Merging it into that branch is what turns #2023 green; it is not meant for main by itself.

Why #2023's build is red

It is not flakiness and not a benchmark timing. #2023 registers tests/hyperloom_benchmark_test.py as a ctest row, and that test asserts something only true on a box with the GPU:

49/200 Test #49: hyperloom_benchmark ...***Failed    0.06 sec
  File "tests/hyperloom_benchmark_test.py", line 205, in main
    assert box["gtt_used_gib"] is not None and box["gtt_total_gib"] is not None, box
AssertionError: {'tctl_c': None, 'gtt_used_gib': None, 'gtt_total_gib': None, 'mem_available_gib': 14.57, ...}

CI's container has no /sys/class/drm/card0/device/mem_info_gtt_*. The test had never run in CI before, which is the gap #2023 exists to close, so this is the first thing the new row found.

The fix (one file, the test)

• Where the device exists, the box's own reading is asserted as before. Where it does not, the test asserts the other half of the rule its own comment states: absent is null.
• The numbers are now checked on a fake device directory (3 GiB of 120 GiB in bytes reads back as 3.0 / 120.0; an unreadable counter is null), so the reader is covered on every box.

Proof

$ python3 tests/hyperloom_benchmark_test.py                         # this box, device present
hyperloom_benchmark_test: OK
$ # GPU_SYSFS pointed at a path that does not exist (what CI has)
hyperloom_benchmark_test: OK
$ # same, with the test as it is at 5ba2afe
AssertionError: {'tctl_c': 70.25, 'gtt_used_gib': None, 'gtt_total_gib': None, ...}     <- CI's failure
$ # reader broken on purpose (divides by 2 more)
AssertionError: {... 'gtt_used_gib': 1.5, 'gtt_total_gib': 60.0 ...}                    <- the fake-device arm catches it

CI on this PR is the real check: it runs the row in the container that failed.

🤖 Generated with Claude Code

[bong-water-water-bong]

VERDICT: READY at 7a4fe1c — "absent is a null" is the right rule, and moving the numeric checks to a fake device is the part that makes this a fix rather than a skip.

Why it is the right shape: the old assertion required gtt_used_gib/gtt_total_gib to be present and ordered, which is a statement about this box rather than about the rule under test. A container has no amdgpu, so the honest expectation there is None, and assert box["gtt_used_gib"] is None is a stronger check than skipping: it fails if the collector invents a zero. And because the numbers are now exercised against a fake device, the 0 <= used <= total invariant is still tested on every box instead of only on the one with a GPU.

Two asks:

1. Prove the invariant can fail on the fake device. The 0 <= used <= total assertion is now only reachable through the fake path — so a fake case where used > total (and one where a value is negative) must exist, or the assertion has never fired anywhere and the suite would pass with it deleted. That is the same vacuous-oracle problem this PR is fixing, one layer in.
2. State the coverage matrix in the body — which assertions run on a GPU box, which in a container, which on the fake device. The failure this PR fixes appeared because nobody had written that down; a line of it prevents the next one.

This also answers a question the release gate cares about: a test that fails only on the box with the hardware is a test the whole fleet has been treating as green.

[bong-water-water-bong]

VERDICT at 7a4fe1c: READY

This is the fix for the thing blocking #2023, and it does the right two things at once:

• hyperloom_benchmark_test now passes without this box's device — a benchmark row in CI is a flaky gate by construction unless it degrades honestly when the hardware is absent, and this is that treatment;
• it registers three Python tests that had no runner (check_pins, gguf_to_q4nx, hrx_long_prompt) via add_python_test. That is #1800's defect in a third file kind, fixed while passing.

A benchmark that skips without its device and a test that runs are opposite outcomes, and getting both in one PR is what unblocks the checker work behind it.

[bong-water-water-bong]

ready for review at 7a4fe1c

CI build is green at this head, in the container where #2023's row failed: 49/200 Test #49: hyperloom_benchmark ... Passed, 200 of 200 (the eight rows #2023 registers included). Base is #2023's branch, not main.

[bong-water-water-bong]

Closing: #2023's author pushed their own fix for the same red row at e0ccae3 (the test skips with 77 on a box with no GPU device), so this is superseded. One thing from here that their route does not carry, offered and not pushed: with a skip, the GTT byte-to-GiB reader is exercised only on a box that has the device; the fake-device arm in this PR checks it everywhere. #2023 now conflicts with main and needs a merge of main.