ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix: use partial hipinfo output on crash to avoid CPU fallback (RDNA 4 / gfx1200) (#6292)

Browse source 
* fix: use partial hipinfo output on crash to avoid CPU fallback (#6043)

`hipinfo.exe` on some RDNA 4 hosts (e.g. RX 9060 XT / gfx1200) exits
with STATUS_ACCESS_VIOLATION (0xC0000005) after printing the
gcnArchName line.  The previous guard `$LASTEXITCODE -eq 0` in
studio/setup.ps1 and `if result.returncode == 0` in
install_python_stack.py discarded this partial-but-valid output,
causing the installer to fall through to WMI name inference which sets
HasROCm=false and installs CPU PyTorch instead of the ROCm wheel.

Fix: check for gcnArchName in stdout first; accept the arch regardless
of exit code.  Only fall through to the amd-smi / WMI path when no
gcnArchName is present at all (crash before any output, or a genuine
"no device" error).  A cyan INFO substep is emitted when the arch is
recovered from a crashed hipinfo run so users can see what happened.

Adds a regression test covering the crash-with-valid-output path.

Fixes #6043

* Fix/adjust hipinfo crash fallback for PR #6292

---------

Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
This commit is contained in:
Matt Van Horn 2026-06-15 04:26:04 -07:00 • committed by GitHub
commit 08c3878919
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 70 additions and 16 deletions
Download patch file Download diff file Expand all files Collapse all files

24
studio/install_python_stack.py
View file

@ -354,16 +354,20 @@ def _detect_windows_gfx_arch() -> str | None:
stderr = subprocess.DEVNULL,
timeout = 10,
)
if result.returncode == 0:
text = result.stdout.decode(errors = "replace")
# findall gets every gcnArchName line so multi-GPU hosts are
# enumerable and HIP_VISIBLE_DEVICES selects correctly.
_tokens = [
t.strip().lower() for t in re.findall(r"(?im)^\s*gcnArchName\s*:\s*(\S+)", text)
]
_pick = _dedup_pick(_tokens)
if _pick:
return _pick
# Accept partial output even when hipinfo crashes (e.g. exit code
# 0xC0000005 / STATUS_ACCESS_VIOLATION on some RDNA 4 hosts): if
# gcnArchName is present in stdout the device was enumerated before
# the crash, so the arch is trustworthy. Ignoring it causes a
# silent CPU PyTorch fallback (issue #6043).
text = result.stdout.decode(errors = "replace")
# findall gets every gcnArchName line so multi-GPU hosts are
# enumerable and HIP_VISIBLE_DEVICES selects correctly.
_tokens = [
t.strip().lower() for t in re.findall(r"(?im)^\s*gcnArchName\s*:\s*(\S+)", text)
]
_pick = _dedup_pick(_tokens)
if _pick:
return _pick
except Exception:
pass