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 03:49:55 +11:00
commit 890d73e6d0
1 changed files with 3 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

3
unsloth/models/gemma.py
View file

@ -79,6 +79,7 @@ class FastGemmaRotaryEmbedding(torch.nn.Module):
inv_freq = 1.0 / (
self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device="cpu").float() / self.dim)
)
print(inv_freq)
t = torch.arange(self.max_position_embeddings, device="cpu", dtype=torch.int64).float().to("cuda").unsqueeze(0)
inv_freq_expanded = inv_freq[None, :, None].float().expand(1, -1, 1).to("cuda")
position_ids_expanded = t[:, None, :].float()
@ -103,10 +104,12 @@ 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()
freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
print(freqs.to(torch.int64))
emb = torch.cat((freqs, freqs), dim=-1)
seq_len = position_ids.shape[1]