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 03:07:20 +11:00
commit 39a58f7d2e
1 changed files with 4 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

13
unsloth/models/gemma.py
View file

@ -73,19 +73,14 @@ class FastGemmaRotaryEmbedding(torch.nn.Module):
self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device="cuda").float() / self.dim)
)
t = torch.arange(position_ids.shape[1], device="cuda", dtype=torch.int64).float().unsqueeze(0)
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)
emb_old = torch.cat((freqs, freqs), dim=-1)
emb = torch.cat((freqs, freqs), dim=-1)
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to("cuda")
position_ids_expanded = position_ids[:, None, :].float()
freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
emb_new = torch.cat((freqs, freqs), dim=-1)
logger.warning_once(str(torch.dist(emb_old, emb_new)))
return emb_old.cos().to(dtype=x.dtype), emb_old.sin().to(dtype=x.dtype)
seq_len = position_ids.shape[1]
return emb.cos().to(dtype=x.dtype)[:seq_len], emb.sin().to(dtype=x.dtype)[:seq_len]
pass