Skip to content

remove prev thread fix, replaced by main changes #1916

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Sep 26, 2025
Merged

remove prev thread fix, replaced by main changes #1916

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
47 changes: 1 addition & 46 deletions gptqmodel/quantization/gptq.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,45 +27,6 @@

log = setup_logger()

# TODO: move this to a locking class
# --------------------------------------------------------------------------------------
# Per-device lock registry to guard device-specific critical sections (like tensor moves)
# --------------------------------------------------------------------------------------
_device_locks = {} # {(device_type, index): threading.Lock()}
_device_locks_guard = threading.Lock() # guards the registry itself


def _device_key(dev) -> tuple:
"""
Normalize a device into a hashable (type, index) key.
Examples:
torch.device('cuda', 0) -> ('cuda', 0)
torch.device('xpu') -> ('xpu', -1)
'cuda:1' -> ('cuda', 1)
'cpu' -> ('cpu', -1)
"""
if isinstance(dev, torch.device):
return (dev.type, dev.index if dev.index is not None else -1)
if isinstance(dev, str):
try:
d = torch.device(dev)
return _device_key(d)
except Exception:
return ("str", dev) # last-resort string key
# Unknown type — stringify
return ("unknown", str(dev))


def _get_device_lock(dev) -> threading.Lock:
key = _device_key(dev)
with _device_locks_guard:
lk = _device_locks.get(key)
if lk is None:
lk = threading.Lock()
_device_locks[key] = lk
return lk
# --------------------------------------------------------------------------------------

lock = threading.Lock()
torch.backends.cuda.matmul.allow_tf32 = False
torch.backends.cudnn.allow_tf32 = False
Expand Down Expand Up @@ -616,13 +577,7 @@ def quantize(
scale = torch.cat(scale, dim=1)
zero = torch.cat(zero, dim=1)

target_device = self.module.weight.data.device

# limit one sync tensor move action per device due to cuda limits
if Q.device != target_device:
dev_lock = _get_device_lock(target_device)
with dev_lock:
Q = Q.to(device=target_device, non_blocking=False)
Q = Q.to(device=self.module.weight.data.device, non_blocking=False)

duration = time.time() - start

Expand Down