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

Update gemma.py

Browse source
This commit is contained in:
Daniel Han-Chen 2024-02-25 02:37:48 +11:00
commit 49a124243e
1 changed files with 3 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

6
unsloth/models/gemma.py
View file

@ -70,10 +70,10 @@ class FastGemmaRotaryEmbedding(torch.nn.Module):
# x: [bs, num_attention_heads, seq_len, head_size]
if self.inv_freq is None:
self.inv_freq = 1.0 / (
self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device="cpu").float() / self.dim)
self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device="cuda").float() / self.dim)
)
t = torch.arange(position_ids.shape[1], device="cpu", dtype=torch.int64).float()
t = torch.arange(position_ids.shape[1], device="cuda", dtype=torch.int64).float()
freqs = torch.outer(t, self.inv_freq)
emb_old = torch.cat((freqs, freqs), dim=-1).to("cuda")
@ -82,7 +82,7 @@ class FastGemmaRotaryEmbedding(torch.nn.Module):
freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
emb_new = torch.cat((freqs, freqs), dim=-1)
print(torch.dist(emb_old, emb_new))
logger.warning_once(str(torch.dist(emb_old, emb_new)))
return emb_new.cos().to(dtype=x.dtype), emb_new.sin().to(dtype=x.dtype)
pass