ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

refactor mm training

Browse Source 
Former-commit-id: 179c0558699e287cbf38a2d73bff47e86d589c5a
This commit is contained in:
hiyouga
2024-08-30 02:14:31 +08:00
parent 77c2c7076b
commit c62a6ca59d
29 changed files with 499 additions and 312 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -27,7 +27,7 @@ logger = get_logger(__name__)
def configure_liger_kernel(config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool) -> None:
if not is_trainable or not model_args.use_liger_kernel:
if not is_trainable or not model_args.enable_liger_kernel:
return
if getattr(config, "model_type", None) == "gemma":

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,<=4.43.4", "To fix: pip install transformers>=4.41.2,<=4.43.4")
require_version("transformers>=4.41.2,<=4.44.3", "To fix: pip install transformers>=4.41.2,<=4.44.3")
LlamaAttention.forward = llama_attention_forward
LlamaFlashAttention2.forward = llama_flash_attention_2_forward
LlamaSdpaAttention.forward = llama_sdpa_attention_forward

7
src/llamafactory/model/model_utils/misc.py
View File

@@ -36,11 +36,14 @@ def find_all_linear_modules(model: "PreTrainedModel", freeze_vision_tower: bool)
forbidden_modules.add("output")
elif model.config.model_type in ["llava", "paligemma"]:
forbidden_modules.add("multi_modal_projector")
elif model.config.model_type in ["qwen2_vl"]:
elif model.config.model_type == "qwen2_vl":
forbidden_modules.add("merger")
if freeze_vision_tower:
forbidden_modules.add("vision_tower")
if model.config.model_type == "qwen2_vl":
forbidden_modules.add("visual")
else:
forbidden_modules.add("vision_tower")
module_names = set()
for name, module in model.named_modules():

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,<=4.43.4", "To fix: pip install transformers>=4.41.2,<=4.43.4")
require_version("transformers>=4.41.2,<=4.44.3", "To fix: pip install transformers>=4.41.2,<=4.44.3")
if is_transformers_version_greater_than_4_43():
import transformers.modeling_flash_attention_utils