ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Update llama.py

Browse source
This commit is contained in:
Daniel Han-Chen 2024-02-04 03:57:29 +11:00
commit 7ab3426608
1 changed files with 14 additions and 14 deletions
Download patch file Download diff file Expand all files Collapse all files

24
unsloth/models/llama.py
View file

@ -149,19 +149,19 @@ def LlamaAttention_fast_forward_inference(
Kn = Kn.view(bsz, 1, n_kv_heads, head_dim).transpose(1, 2)
Vn = Vn.view(bsz, 1, n_kv_heads, head_dim).transpose(1, 2)
cos, sin = self.rotary_emb(Vn, seq_len = kv_seq_len)
Qn, Kn = inplace_rope_embedding(Qn, Kn, cos, sin, position_ids)
# cos = self.rotary_emb.cos_cached[seq_len]
# sin = self.rotary_emb.sin_cached[seq_len]
# h = head_dim // 2
# cos, sin = self.rotary_emb(Vn, seq_len = kv_seq_len)
# Qn, Kn = inplace_rope_embedding(Qn, Kn, cos, sin, position_ids)
cos = self.rotary_emb.cos_cached[seq_len]
sin = self.rotary_emb.sin_cached[seq_len]
h = head_dim // 2
# RH_Q = self.RH_Q
# RH_Q[:,:,:,:h] = Qn[:,:,:,h:]; RH_Q[:,:,:,h:] = Qn[:,:,:,:h]; torch.neg(RH_Q[:,:,:,:h], out = RH_Q[:,:,:,:h]);
# Qn *= cos; Qn.addcmul_(RH_Q, sin);
RH_Q = self.RH_Q
RH_Q[:,:,:,:h] = Qn[:,:,:,h:]; RH_Q[:,:,:,h:] = Qn[:,:,:,:h]; torch.neg(RH_Q[:,:,:,:h], out = RH_Q[:,:,:,:h]);
Qn *= cos; Qn.addcmul_(RH_Q, sin);
# RH_K = RH_Q[:,:n_kv_heads,:,:] # torch.empty((n_kv_heads, 1, head_dim), dtype = dtype, device = "cuda")
# RH_K[:,:,:,:h] = Kn[:,:,:,h:]; RH_K[:,:,:,h:] = Kn[:,:,:,:h]; torch.neg(RH_K[:,:,:,:h], out = RH_K[:,:,:,:h]);
# Kn *= cos; Kn.addcmul_(RH_K, sin);
RH_K = RH_Q[:,:n_kv_heads,:,:] # torch.empty((n_kv_heads, 1, head_dim), dtype = dtype, device = "cuda")
RH_K[:,:,:,:h] = Kn[:,:,:,h:]; RH_K[:,:,:,h:] = Kn[:,:,:,:h]; torch.neg(RH_K[:,:,:,:h], out = RH_K[:,:,:,:h]);
Kn *= cos; Kn.addcmul_(RH_K, sin);
# New KV cache
# Kn = torch.cat([K1, Kn], dim = 2)
@ -471,7 +471,7 @@ def LlamaModel_fast_forward(
pass
# We already handle KV cache position_ids ourselves.
if (past_key_values_length != 0):
if False:#(past_key_values_length != 0):
position_ids = torch.arange(
past_key_values_length, seq_length + past_key_values_length,
dtype = torch.int32,