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 2024-02-01 17:44:22 +11:00
commit 57044509ad
1 changed files with 16 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

20
unsloth/models/llama.py
View file

@ -239,9 +239,21 @@ 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)
kv_seq_len = K1.shape[-2] + 1
cos, sin = self.rotary_emb(Vn, seq_len = kv_seq_len)
Qn, Kn = inplace_rope_embedding(Qn, Kn, cos, sin, position_ids)
seq_len = K1.shape[-2]
kv_seq_len = seq_len + 1
# 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 = torch.empty((bsz, n_heads, 1, head_dim), dtype = dtype, device = "cuda")
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);
# New KV cache
Kn = torch.cat([K1, Kn], dim = 2)
@ -657,7 +669,7 @@ def LlamaModel_fast_forward(
if output_attentions:
all_self_attns += (layer_outputs[1],)
pass
if past_key_values is not None:
hidden_states = fast_rms_layernorm_inference(self.norm, hidden_states)
else: