423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-09-12 16:12:48 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix #4160

Browse Source 
The split heads should be concatenated in dim=2


Former-commit-id: a793e8456b664ea0b48f0ba162999f18d06b4c2f
This commit is contained in:
hiyouga 2024-06-11 00:37:17 +08:00
parent d984776d35
commit 2f164c2c41
1 changed files with 7 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
src/llamafactory/model/model_utils/longlora.py
View File

@ -96,7 +96,8 @@ def llama_attention_forward(
(
attn_output[:, :, : self.num_heads // 2],
attn_output[:, :, self.num_heads // 2 :].roll(groupsz // 2, dims=1),
)
),
dim=2,
)
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
@ -194,7 +195,8 @@ def llama_flash_attention_2_forward(
(
attn_output[:, :, : self.num_heads // 2],
attn_output[:, :, self.num_heads // 2 :].roll(groupsz // 2, dims=1),
)
),
dim=2,
)
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
@ -293,7 +295,8 @@ def llama_sdpa_attention_forward(
(
attn_output[:, :, : self.num_heads // 2],
attn_output[:, :, self.num_heads // 2 :].roll(groupsz // 2, dims=1),
)
),
dim=2,
)
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
@ -303,7 +306,7 @@ def llama_sdpa_attention_forward(
def _apply_llama_patch() -> None:
require_version("transformers==4.40.2", "To fix: pip install transformers==4.40.2")
require_version("transformers==4.41.2", "To fix: pip install transformers==4.41.2")
LlamaAttention.forward = llama_attention_forward
LlamaFlashAttention2.forward = llama_flash_attention_2_forward
LlamaSdpaAttention.forward = llama_sdpa_attention_forward