Exploratory · single-GPU sweep

Sizing up the Transformers CI on AMD Strix Halo (gfx1151)

An early look at how long the full tests/models suite takes on one Strix Halo APU — run one model per process with slow / integration tests enabled. Memory failures are split into GPU (HIP) and host-RAM (OS OOM-killer), verified against the kernel log.

APU AMD Ryzen AI Max+ 395 w/ Radeon 8060S (gfx1151) Memory 64 GB shared · offered machines: 128 GB torch 2.9.1+rocm7.2 transformers 5.8.0.dev0 RUN_SLOW 1 per-model cap 6h
Download results.csv 473 rows · CSV · refined status + exit codes
Models
473
Passed
151
Failed
305
OOM
17
4 GPU · 13 host RAM
Timeout
0
none hit 6h cap
Wall time
21h 38m
Tests passed
68,014
Tests failed
1,477
Tests skipped
59,130
46% of collected
ModelStatus Time PassFailSkipErr
sorted by time, longest first

How status is decided. Each model ran as its own process (GPU memory reclaimed between models; HF cache wiped after each). FAILED = the suite had ≥1 failing test (usually a few hardware-specific ones), not a whole-suite break. OOM · GPU = an explicit HIP out-of-memory in the log; OOM · host RAM = the process was SIGKILL'd (exit 137) by the Linux OOM-killer — each one matches an Out of memory: Killed process (python) kernel event at the same second. This APU shares VRAM with system RAM, so large models exhaust this test unit's 64 GB of shared memory. The machines AMD is offering have 128 GB, where most of these would likely complete instead of being killed. No run reached the 6-hour cap (exit 124). See exit_code / oom_type in the CSV.