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

Merge branch 'main' into pixtral-patch

Browse Source 
Former-commit-id: 26f45829b4
This commit is contained in:
Zhangchi Feng
2024-09-30 12:37:03 +08:00
committed by GitHub
parent a2452d0b1c 8d20d6c95c
commit 69e801d456
12 changed files with 690 additions and 219 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
src/llamafactory/model/loader.py
View File

@@ -25,7 +25,7 @@ from .model_utils.misc import register_autoclass
from .model_utils.mod import convert_pretrained_model_to_mod, load_mod_pretrained_model
from .model_utils.unsloth import load_unsloth_pretrained_model
from .model_utils.valuehead import load_valuehead_params
from .model_utils.visual import get_image_seqlen
from .model_utils.visual import get_image_seqlen, get_patch_size, get_vision_feature_select_strategy
from .patcher import patch_config, patch_model, patch_tokenizer, patch_valuehead_model
@@ -100,9 +100,11 @@ def load_tokenizer(model_args: "ModelArguments") -> "TokenizerModule":
setattr(processor, "tokenizer", tokenizer)
setattr(processor, "image_seqlen", get_image_seqlen(config))
setattr(processor, "image_resolution", model_args.image_resolution)
setattr(processor, "patch_size", get_patch_size(config))
setattr(processor, "video_resolution", model_args.video_resolution)
setattr(processor, "video_fps", model_args.video_fps)
setattr(processor, "video_maxlen", model_args.video_maxlen)
setattr(processor, "vision_feature_select_strategy", get_vision_feature_select_strategy(config))
except Exception:
processor = None
@@ -156,7 +158,6 @@ def load_model(
load_class = AutoModelForVision2Seq
else:
load_class = AutoModelForCausalLM
if model_args.train_from_scratch:
model = load_class.from_config(config)
else:

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

@@ -37,10 +37,11 @@ def configure_attn_implementation(
if is_flash_attn_2_available():
require_version("transformers>=4.42.4", "To fix: pip install transformers>=4.42.4")
require_version("flash_attn>=2.6.3", "To fix: pip install flash_attn>=2.6.3")
logger.warning("Gemma-2 should use flash attention 2, change `flash_attn` to fa2.")
model_args.flash_attn = "fa2"
if model_args.flash_attn != "fa2":
logger.warning("Gemma-2 should use flash attention 2, change `flash_attn` to fa2.")
model_args.flash_attn = "fa2"
else:
logger.warning("Gemma-2 should use eager attention, change `flash_attn` to disabled.")
logger.warning("FlashAttention-2 is not installed, use eager attention.")
model_args.flash_attn = "disabled"
elif model_args.flash_attn == "sdpa":
logger.warning("Gemma-2 should use soft-capping attention, while the SDPA attention does not support it.")

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

@@ -34,7 +34,7 @@ def find_all_linear_modules(model: "PreTrainedModel", freeze_vision_tower: bool)
forbidden_modules.add("output_layer")
elif model_type == "internlm2":
forbidden_modules.add("output")
elif model_type in ["llava", "paligemma"]:
elif model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
forbidden_modules.add("multi_modal_projector")
elif model_type == "qwen2_vl":
forbidden_modules.add("merger")

32
src/llamafactory/model/model_utils/visual.py
View File

@@ -92,7 +92,7 @@ def autocast_projector_dtype(model: "PreTrainedModel", model_args: "ModelArgumen
if getattr(model, "quantization_method", None):
model_type = getattr(model.config, "model_type", None)
if model_type in ["llava", "paligemma"]:
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
mm_projector: "torch.nn.Module" = getattr(model, "multi_modal_projector")
elif model_type == "qwen2_vl":
mm_projector: "torch.nn.Module" = getattr(getattr(model, "visual"), "merger")
@@ -110,7 +110,13 @@ def configure_visual_model(config: "PretrainedConfig") -> None:
Patches VLMs before loading them.
"""
model_type = getattr(config, "model_type", None)
if model_type == "llava": # required for ds zero3 and valuehead models
if model_type in [
"llava",
"llava_next",
"llava_next_video",
"paligemma",
"video_llava",
]: # required for ds zero3 and valuehead models
setattr(config, "hidden_size", getattr(config.text_config, "hidden_size", None))
if getattr(config, "is_yi_vl_derived_model", None):
@@ -124,7 +130,7 @@ def get_forbidden_modules(config: "PretrainedConfig", finetuning_args: "Finetuni
"""
model_type = getattr(config, "model_type", None)
forbidden_modules = set()
if model_type in ["llava", "paligemma", "pixtral"]:
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "pixtral", "video_llava"]:
if finetuning_args.freeze_vision_tower:
forbidden_modules.add("vision_tower")
forbidden_modules.add("vision_encoder")
@@ -153,12 +159,28 @@ def get_image_seqlen(config: "PretrainedConfig") -> int:
image_seqlen += 1
elif model_type == "paligemma":
image_seqlen = config.vision_config.num_image_tokens
elif model_type in ["qwen2_vl", "pixtral"]: # variable length
else:
image_seqlen = -1
return image_seqlen
def get_patch_size(config: "PretrainedConfig") -> int:
r"""
Computes the patch size of the vit.
"""
patch_size = getattr(config.vision_config, "patch_size", -1)
return patch_size
def get_vision_feature_select_strategy(config: "PretrainedConfig") -> int:
r"""
Get the vision_feature_select_strategy.
"""
vision_feature_select_strategy = getattr(config, "vision_feature_select_strategy", "default")
return vision_feature_select_strategy
def patch_target_modules(
config: "PretrainedConfig", finetuning_args: "FinetuningArguments", target_modules: Sequence[str]
) -> Union[str, List[str]]:
@@ -167,7 +189,7 @@ def patch_target_modules(
"""
model_type = getattr(config, "model_type", None)
if finetuning_args.freeze_vision_tower:
if model_type in ["llava", "paligemma"]:
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
return "^(?!.*vision_tower).*(?:{}).*".format("|".join(target_modules))
elif model_type == "qwen2_vl":
return "^(?!.*visual).*(?:{}).*".format("|".join(target_modules))