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 04:08:04 +11:00
commit 6238f16625
1 changed files with 4 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

8
unsloth/models/gemma.py
View file

@ -83,8 +83,8 @@ class FastGemmaRotaryEmbedding(torch.nn.Module):
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()
print(position_ids_expanded.to(torch.float64))
print(inv_freq_expanded.to(torch.float64))
print(position_ids_expanded.shape, position_ids_expanded.to(torch.float64))
print(inv_freq_expanded.shape, inv_freq_expanded.to(torch.float64))
freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
print(freqs.to(torch.int64))
emb = torch.cat((freqs, freqs), dim=-1)
@ -111,8 +111,8 @@ class FastGemmaRotaryEmbedding(torch.nn.Module):
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(1, -1, 1).to("cuda")
position_ids_expanded = t[:, None, :].float()
print(position_ids_expanded.to(torch.float64))
print(inv_freq_expanded.to(torch.float64))
print(position_ids_expanded.shape, position_ids_expanded.to(torch.float64))
print(inv_freq_expanded.shape, inv_freq_expanded.to(torch.float64))
freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
print(freqs.to(torch.int64))
emb = torch.cat((freqs, freqs), dim=-1)