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 2024-02-26 04:20:19 +11:00
commit b0b38f770b
1 changed files with 22 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

22
unsloth/models/gemma.py
View file

@ -68,6 +68,18 @@ class FastGemmaRotaryEmbedding(torch.nn.Module):
self.register_buffer("cos_cached", None, persistent=False)
self.register_buffer("sin_cached", None, persistent=False)
self.inv_freq = 1.0 / (
self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device="cuda").float() / self.dim)
)
position_ids = torch.arange(self.max_position_embeddings, device="cuda", dtype=torch.int64).unsqueeze(0)
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(1, -1, 1)
position_ids_expanded = position_ids[:, None, :].float()
freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
emb = torch.cat((freqs, freqs), dim=-1)
self.cos_cached = emb.cos().to(torch.bfloat16)
self.sin_cached = emb.sin().to(torch.bfloat16)
def forward(self, x, position_ids, seq_len=None):
# x: [bs, num_attention_heads, seq_len, head_size]
if self.inv_freq is None:
@ -577,6 +589,16 @@ class FastGemmaModel(FastLlamaModel):
lm_head.out_features = lm_head.weight.shape[0]
model.lm_head = lm_head
# Gemma has tied weights! This means lm_head == embed_tokens
if model.model.embed_tokens.weight.data_ptr() != model.lm_head.weight.data_ptr():
lm_head = torch.nn.Linear(1, 1, bias = None)
del lm_head.weight
lm_head.weight = model.model.embed_tokens.weight
lm_head.in_features = lm_head.weight.shape[1]
lm_head.out_features = lm_head.weight.shape[0]
model.lm_head = lm_head
pass
# Also patch all dtypes - BnB seems to not allocate the correct type?
# BnB default dtype seems to be float16!
correct_dtype = lm_head.weight.dtype