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

inference

Browse source
This commit is contained in:
Daniel Han-Chen 2024-02-01 23:17:53 +11:00
commit e4b5e38800
2 changed files with 15 additions and 15 deletions
Download patch file Download diff file Expand all files Collapse all files

24
unsloth/models/llama.py
View file

@ -244,19 +244,19 @@ def LlamaAttention_fast_forward_inference(
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
# 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 = Xn.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_Q = torch.empty((bsz, n_heads, 1, head_dim), dtype = Xn.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);
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)
@ -741,7 +741,7 @@ def LlamaForCausalLM_fast_forward(
hidden_states = outputs[0]
bsz, q_len, hd = hidden_states.shape
if False:#bsz == 1 and q_len == 1:
if bsz == 1 and q_len == 1:
logits = torch.mv(self.lm_head.weight, hidden_states.ravel())
logits = logits.unsqueeze(0).unsqueeze(0)
else:

2
unsloth/models/mistral.py
View file

@ -214,7 +214,7 @@ def MistralForCausalLM_fast_forward(
hidden_states = outputs[0]
bsz, q_len, hd = hidden_states.shape
if False:#bsz == 1 and q_len == 1:
if bsz == 1 and q_len == 1:
logits = torch.mv(self.lm_head.weight, hidden_states.ravel())
logits = logits.unsqueeze(0).unsqueeze(0)
else: