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

device

Browse source
This commit is contained in:
Daniel Han 2025-03-03 00:04:08 -08:00
commit 54121bb4b3
2 changed files with 3 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

7
unsloth/kernels/utils.py
View file

@ -452,7 +452,9 @@ pass
def matmul_lora(X, W, W_quant, A, B, s, out = None):
dtype = X.dtype
W = fast_dequantize(W.t(), W_quant, use_global_buffer = False)
if X.device != W.device:
print(X.device, W.device, torch.cuda.current_device())
if X.dim() == 3:
batch, seq_len, d = X.shape
X = X.view(-1, X.shape[-1])
@ -460,9 +462,6 @@ def matmul_lora(X, W, W_quant, A, B, s, out = None):
else:
reshape = False
pass
if X.device != W.device:
print(X.device, W.device, torch.cuda.current_device())
out = torch_matmul(X, W, out = out)
if W_quant is not None: del W

1
unsloth/models/llama.py
View file

@ -385,7 +385,6 @@ def LlamaAttention_fast_forward(
head_dim = self.head_dim
assert(n_kv_heads * n_groups == n_heads)
print(hidden_states.device, torch.cuda.current_device())
Q, K, V = self.apply_qkv(self, hidden_states)
Q = Q.view(bsz, q_len, n_heads, head_dim).transpose(1, 2)
K = K.view(bsz, q_len, n_kv_heads, head_dim).transpose(1, 2)