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

print

Browse source
This commit is contained in:
Daniel Han 2024-12-26 18:45:16 -08:00
commit af7d6cc871
2 changed files with 2 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

1
unsloth/models/_utils.py
View file

@ -1127,7 +1127,6 @@ def patch_gradient_accumulation_fix(Trainer):
r"(.+?)return loss\.detach\(\) \/ self\.args\.gradient_accumulation_steps",
"else:\n"\
"\1print(self.args.gradient_accumulation_steps)\n"
"\2if num_items_in_batch is None:\n"\
"\3loss = loss / self.args.gradient_accumulation_steps\n"\
"\1self.accelerator.backward(loss, **kwargs)",

2
unsloth/models/llama.py
View file

@ -1009,6 +1009,7 @@ def CausalLM_fast_forward(fast_forward_inference):
if not RETURN_LOGITS and HAS_CUT_CROSS_ENTROPY and labels is not None:
n_items = kwargs.get("num_items_in_batch", None) or kwargs.get("n_items", None)
print(0, n_items)
loss = fused_linear_cross_entropy(
hidden_states = hidden_states,
lm_weight = lm_head,
@ -1055,6 +1056,7 @@ def CausalLM_fast_forward(fast_forward_inference):
# Fixes https://github.com/unslothai/unsloth/issues/10
self.extra_ignored_labels = torch.full((self.max_seq_length, 1), -100, device = "cuda:0")
pass
print(1, kwargs.get("num_items_in_batch", None) or kwargs.get("n_items", None))
shift_labels = torch.hstack((labels[..., 1:], self.extra_ignored_labels[:labels.shape[0]]))
loss = fast_cross_entropy_loss(
logits = shift_logits,