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

Update cross_entropy_loss.py

Browse source
This commit is contained in:
Daniel Han 2024-10-28 15:01:04 -07:00
commit 8e30e2e646
1 changed files with 3 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

6
unsloth/kernels/cross_entropy_loss.py
View file

@ -269,7 +269,7 @@ class Fast_CrossEntropyLoss(torch.autograd.Function):
div, mod = divmod(vocab_size, MAX_FUSED_SIZE)
n_chunks = div + (mod != 0)
losses = torch.empty(n_rows, dtype = torch.float32, device = "cuda:0")
losses = torch.empty(n_rows, dtype = torch.float32, device = logits.device)
DO_SOFTCAPPING = (logit_softcapping != 0)
DO_LOGIT_SCALING = (logit_scaling != 0)
@ -277,7 +277,7 @@ class Fast_CrossEntropyLoss(torch.autograd.Function):
if n_chunks == 1:
# For small vocabs <= 65336 like Llama, Mistral
BLOCK_SIZE, num_warps = calculate_settings(vocab_size)
logsumexp = torch.empty(n_rows, dtype = torch.float32, device = "cuda:0")
logsumexp = torch.empty(n_rows, dtype = torch.float32, device = logits.device)
_cross_entropy_forward[(n_rows,)](
logits, logits.stride(0),
@ -294,7 +294,7 @@ class Fast_CrossEntropyLoss(torch.autograd.Function):
)
else:
# For large vocabs > 65336 like Gemma 256K
logsumexp = torch.empty((n_rows, n_chunks,), dtype = torch.float32, device = "cuda:0")
logsumexp = torch.empty((n_rows, n_chunks,), dtype = torch.float32, device = logits.device)
_chunked_cross_entropy_forward[(n_rows, n_chunks,)](
logits, logits.stride(0),