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 04:35:39 +11:00
commit 36d29078cd
1 changed files with 13 additions and 28 deletions
Download patch file Download diff file Expand all files Collapse all files

41
unsloth/models/gemma.py
View file

@ -68,19 +68,6 @@ 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(dtype=torch.bfloat16)
self.sin_cached = emb.sin().to(dtype=torch.bfloat16)
print(self.cos_cached)
print(self.sin_cached)
def forward(self, x, position_ids, seq_len=None):
# x: [bs, num_attention_heads, seq_len, head_size]
if self.inv_freq is None:
@ -88,20 +75,18 @@ class FastGemmaRotaryEmbedding(torch.nn.Module):
self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim)
)
length = position_ids.shape[1]
#if self.cos_cached is None:
position_ids = torch.arange(self.max_position_embeddings, device=x.device, 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(dtype=x.dtype)
self.sin_cached = emb.sin().to(dtype=x.dtype)
# pass
print(self.cos_cached)
print(self.sin_cached)
raise
return self.cos_cached[:,:length], self.sin_cached[:,:length]
# length = position_ids.shape[1]
if self.cos_cached is None:
position_ids = torch.arange(self.max_position_embeddings, device=x.device, 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(dtype=x.dtype)
self.sin_cached = emb.sin().to(dtype=x.dtype)
pass
# return self.cos_cached[:,:length], self.sin_cached[:,:length]
return self.cos_cached, self.sin_cached
pass
@ -141,7 +126,7 @@ def GemmaAttention_fast_forward(
K = K.view(bsz, q_len, n_kv_heads, head_dim).transpose(1, 2)
V = V.view(bsz, q_len, n_kv_heads, head_dim).transpose(1, 2)
if True:#position_ids is None:
if True:# position_ids is None:
# cos = self.rotary_emb.cos_cached
# sin = self.rotary_emb.sin_cached
cos, sin = self.rotary_emb(V, position_ids, seq_len = q_len)