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 2024-02-03 01:22:53 +11:00
commit 705bbba5c5
1 changed files with 6 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

9
unsloth/models/llama.py
View file

@ -123,7 +123,7 @@ def LlamaAttention_fast_forward_inference(
# Prefill phase
# if not hasattr(self, "paged_attention"):
if do_prefill:
self.paged_attention = torch.empty((self.config.max_position_embeddings+1, 2, bsz, n_kv_heads, head_dim), dtype = dtype, device = "cuda")
self.paged_attention = torch.empty((128+1, 2, bsz, n_kv_heads, head_dim), dtype = dtype, device = "cuda")
self.paged_attention_K = self.paged_attention[:,0]
self.paged_attention_V = self.paged_attention[:,1]
self.paged_attention_K[:seq_len] = K1.permute(2, 0, 1, 3)
@ -131,10 +131,13 @@ def LlamaAttention_fast_forward_inference(
self.temp_QA = torch.empty((2, bsz, 1, hd), dtype = dtype, device = "cuda")
self.temp_KV = torch.empty((2, bsz, 1, n_kv_heads*head_dim), dtype = dtype, device = "cuda")
self.RH_Q = torch.empty((bsz, n_heads, 1, head_dim), dtype = dtype, device = "cuda")
self.attention = torch.empty((bsz, n_heads, 1, self.config.max_position_embeddings), dtype = dtype, device = "cuda")
self.attention = torch.empty((bsz, n_heads, 1, 128), dtype = dtype, device = "cuda")
self.scalar = 1.0 / math_sqrt(self.head_dim)
elif kv_seq_len >= self.paged_attention.shape[0]:
self.paged_attention.resize_((self.paged_attention.shape[0]+128, 2, bsz, n_kv_heads, head_dim))
self.paged_attention_K = self.paged_attention[:,0]
self.paged_attention_V = self.paged_attention[:,1]
pass
# pass
Qn = fast_linear_forward(self.q_proj, Xn, out = self.temp_QA[0])
Kn = fast_linear_forward(self.k_proj, Xn, out = self.temp_KV[0])