Skip to content

[intel] add for intel path for llama.py #3012

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 3 commits into from
Jul 22, 2025
Merged

[intel] add for intel path for llama.py #3012

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
edc3a52
fix for intel path
leizhenyuan Jul 21, 2025
56dccfd
remove unuse code
leizhenyuan Jul 21, 2025
d306fbb
Update unsloth/models/llama.py
danielhanchen Jul 22, 2025
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
27 changes: 16 additions & 11 deletions unsloth/models/llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,6 +92,11 @@ def clean_gpu_cache():
else:
torch.cuda.empty_cache()

if DEVICE_TYPE == "xpu":
get_current_device = torch.xpu.current_device
else:
get_current_device = torch.cuda.current_device

def original_apply_qkv(self, X):
Q = self.q_proj(X)
K = self.k_proj(X)
Expand Down Expand Up @@ -1365,8 +1370,8 @@ def __init__(self, dim = None, max_position_embeddings=2048, base=10000, device=
self._set_cos_sin_cache(seq_len=self.current_rope_size, device=torch.device(device_idx), 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())
self.sin_cached = torch.empty(1, device=torch.cuda.current_device(), dtype=torch.get_default_dtype())
self.cos_cached = torch.empty(1, device=get_current_device(), dtype=torch.get_default_dtype())
self.sin_cached = torch.empty(1, device=get_current_device(), dtype=torch.get_default_dtype())
pass

def _set_cos_sin_cache(self, seq_len, device, dtype):
Expand Down Expand Up @@ -1402,7 +1407,7 @@ 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]
pass

Expand Down Expand Up @@ -1484,8 +1489,8 @@ def __init__(self, dim = None, max_position_embeddings=2048, base=10000, device=
self._set_cos_sin_cache(seq_len=self.current_rope_size, device=torch.device(device_idx), 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())
self.sin_cached = torch.empty(1, device=torch.cuda.current_device(), dtype=torch.get_default_dtype())
self.cos_cached = torch.empty(1, device=get_current_device(), dtype=torch.get_default_dtype())
self.sin_cached = torch.empty(1, device=get_current_device(), dtype=torch.get_default_dtype())
pass

def _set_cos_sin_cache(self, seq_len, device, dtype):
Expand Down Expand Up @@ -1518,7 +1523,7 @@ 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]
pass

Expand Down Expand Up @@ -1631,10 +1636,10 @@ def __init__(self,
self.multi_gpu_short_sin_cached[device_idx] = sin_cached

# dummy so that patch_utils doesn't fail for now
self.short_cos_cached = torch.empty(1, device=torch.cuda.current_device(), dtype=torch.get_default_dtype())
self.short_sin_cached = torch.empty(1, device=torch.cuda.current_device(), dtype=torch.get_default_dtype())
self.long_cos_cached = torch.empty(1, device=torch.cuda.current_device(), dtype=torch.get_default_dtype())
self.long_sin_cached = torch.empty(1, device=torch.cuda.current_device(), dtype=torch.get_default_dtype())
self.short_cos_cached = torch.empty(1, device=get_current_device(), dtype=torch.get_default_dtype())
self.short_sin_cached = torch.empty(1, device=get_current_device(), dtype=torch.get_default_dtype())
self.long_cos_cached = torch.empty(1, device=get_current_device(), dtype=torch.get_default_dtype())
self.long_sin_cached = torch.empty(1, device=get_current_device(), dtype=torch.get_default_dtype())
pass

def _set_cos_sin_cache(self, seq_len, device, dtype):
Expand Down Expand Up @@ -1675,7 +1680,7 @@ 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()
if seq_len is not None and seq_len < self.original_max_position_embeddings:
return self.multi_gpu_short_cos_cached[device_index], self.multi_gpu_short_sin_cached[device_index]
return self.multi_gpu_long_cos_cached[device_index], self.multi_gpu_long_sin_cached[device_index]
Expand Down