Skip to content

fix(gemma): Replace hardcoded CUDA calls in GemmaFixedRotaryEmbedding for XPU support #4928

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

Open
cheehook wants to merge 1 commit into
base: main
Choose a base branch
from
+8 −6
Open

fix(gemma): Replace hardcoded CUDA calls in GemmaFixedRotaryEmbedding for XPU support #4928

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
14 changes: 8 additions & 6 deletions unsloth/models/gemma.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -300,16 +300,16 @@ def __init__(
for device in range(DEVICE_COUNT):
self._set_cos_sin_cache(
seq_len = self.current_rope_size,
device = torch.device(device),
device = torch.device(DEVICE_TYPE_TORCH, device),
dtype = torch.get_default_dtype(),
)

# dummy so that patch_utils doesn't fail for now
self.cos_cached = torch.empty(
1, device = torch.cuda.current_device(), dtype = torch.get_default_dtype()
1, device = get_current_device(), dtype = torch.get_default_dtype()
)
self.sin_cached = torch.empty(
1, device = torch.cuda.current_device(), dtype = torch.get_default_dtype()
1, device = get_current_device(), dtype = torch.get_default_dtype()
)

def _set_cos_sin_cache(self, seq_len, device, dtype):
Expand Down Expand Up @@ -350,7 +350,7 @@ def forward(self, x, position_ids = None, seq_len = None):

def get_cached(self, seq_len = None, device_index = None):
if device_index is None:
device_index = torch.cuda.current_device()
device_index = get_current_device()
return self.multi_gpu_cos_cached[device_index], self.multi_gpu_sin_cached[
device_index
]
Expand All @@ -362,7 +362,9 @@ def extend_rope_embedding(self, x, seq_len):
self.current_rope_size = math.ceil(seq_len / 8192) * 8192
for device in range(DEVICE_COUNT):
self._set_cos_sin_cache(
self.current_rope_size, device = torch.device(device), dtype = x.dtype
self.current_rope_size,
device = torch.device(DEVICE_TYPE_TORCH, device),
dtype = x.dtype,
)


Expand Down Expand Up @@ -489,5 +491,5 @@ def post_patch(model, tokenizer, correct_dtype = None):

for _ in range(3):
gc.collect()
torch.cuda.empty_cache()
clean_gpu_cache()
return model, tokenizer