423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-12-15 03:10:35 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix incorrect loss value for vlms

Browse Source 
Former-commit-id: 30567a1487
This commit is contained in:
hiyouga
2024-10-30 08:56:46 +00:00
parent 1b02915d19
commit 584ce3a105
12 changed files with 48 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/llamafactory/model/model_utils/longlora.py
View File

@@ -353,7 +353,7 @@ def llama_sdpa_attention_forward(
def _apply_llama_patch() -> None:
require_version("transformers>=4.41.2,<=4.46.0", "To fix: pip install transformers>=4.41.2,<=4.46.0")
require_version("transformers>=4.41.2,<=4.46.1", "To fix: pip install transformers>=4.41.2,<=4.46.1")
LlamaAttention.forward = llama_attention_forward
LlamaFlashAttention2.forward = llama_flash_attention_2_forward
LlamaSdpaAttention.forward = llama_sdpa_attention_forward

2
src/llamafactory/model/model_utils/packing.py
View File

@@ -114,7 +114,7 @@ def get_unpad_data(attention_mask: "torch.Tensor") -> Tuple["torch.Tensor", "tor
def _patch_for_block_diag_attn(model_type: str) -> None:
require_version("transformers>=4.41.2,<=4.46.0", "To fix: pip install transformers>=4.41.2,<=4.46.0")
require_version("transformers>=4.41.2,<=4.46.1", "To fix: pip install transformers>=4.41.2,<=4.46.1")
if is_transformers_version_greater_than_4_43():
import transformers.modeling_flash_attention_utils