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 03:50:59 +11:00
commit 33a72ba122
1 changed files with 1 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

2
unsloth/models/gemma.py
View file

@ -84,6 +84,7 @@ class FastGemmaRotaryEmbedding(torch.nn.Module):
inv_freq_expanded = inv_freq[None, :, None].float().expand(1, -1, 1).to("cuda")
position_ids_expanded = t[:, None, :].float()
freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
print(freqs.to(torch.int64))
emb = torch.cat((freqs, freqs), dim=-1)
self.cos_cached = emb.cos().to(dtype=torch.bfloat16)
@ -104,7 +105,6 @@ class FastGemmaRotaryEmbedding(torch.nn.Module):
self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device="cuda").float() / self.dim)
)
print(self.inv_freq)
t = torch.arange(self.max_position_embeddings, device="cpu", dtype=torch.int64).float().to("cuda").unsqueeze(0)
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to("cuda")
position_ids_expanded = t[:, None, :].float()