423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-08-23 22:32:54 +08:00
Code Issues Packages Projects Releases Wiki Activity

add some

Browse Source 
Former-commit-id: 785cc70ff205f5962c3ca67f453589e4a471ba8c
This commit is contained in:
fzc8578 2025-01-06 19:32:39 +08:00
parent a0188a430f
commit b9eeaa9706
4 changed files with 15 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
src/llamafactory/data/mm_plugin.py
View File

@ -366,6 +366,7 @@ class CpmOPlugin(BasePlugin):
position_ids_ = list(range(input_ids_.size(0))) position_ids_ = list(range(input_ids_.size(0)))
# print(input_ids_.shape, len(position_ids_) # print(input_ids_.shape, len(position_ids_)
position_ids.append(position_ids_) position_ids.append(position_ids_)
#TODO add pad
position_ids = torch.tensor(position_ids, dtype=torch.int64) position_ids = torch.tensor(position_ids, dtype=torch.int64)
mm_inputs.update({ mm_inputs.update({
"image_bound": image_bounds_list, "image_bound": image_bounds_list,

6
src/llamafactory/model/model_utils/visual.py
View File

@ -142,6 +142,10 @@ def get_forbidden_modules(config: "PretrainedConfig", finetuning_args: "Finetuni
forbidden_modules.update({"visual.patch_embed", "visual.blocks", "model", "lm_head"}) forbidden_modules.update({"visual.patch_embed", "visual.blocks", "model", "lm_head"})
elif finetuning_args.freeze_vision_tower: elif finetuning_args.freeze_vision_tower:
forbidden_modules.add("visual") forbidden_modules.add("visual")
elif model_type == "minicpmv":
if finetuning_args.freeze_vision_tower:
forbidden_modules.add("vpm")
return forbidden_modules return forbidden_modules
@ -196,6 +200,8 @@ def patch_target_modules(
return "^(?!.*vision_model).*(?:{}).*".format("|".join(target_modules)) return "^(?!.*vision_model).*(?:{}).*".format("|".join(target_modules))
elif model_type == "qwen2_vl": elif model_type == "qwen2_vl":
return "^(?!.*visual).*(?:{}).*".format("|".join(target_modules)) return "^(?!.*visual).*(?:{}).*".format("|".join(target_modules))
elif model_type == "minicpmv":
return "^(?!.*vpm).*(?:{}).*".format("|".join(target_modules))
else: else:
return target_modules return target_modules
else: else:

13
src/llamafactory/model/patcher.py
View File

@ -138,12 +138,13 @@ def patch_model(
add_valuehead: bool, add_valuehead: bool,
) -> None: ) -> None:
gen_config = model.generation_config # check and fix generation config gen_config = model.generation_config # check and fix generation config
if not gen_config.do_sample and ( if gen_config is not None:
(gen_config.temperature is not None and gen_config.temperature != 1.0) if not gen_config.do_sample and (
or (gen_config.top_p is not None and gen_config.top_p != 1.0) (gen_config.temperature is not None and gen_config.temperature != 1.0)
or (gen_config.typical_p is not None and gen_config.typical_p != 1.0) or (gen_config.top_p is not None and gen_config.top_p != 1.0)
): or (gen_config.typical_p is not None and gen_config.typical_p != 1.0)
gen_config.do_sample = True ):
gen_config.do_sample = True
if "GenerationMixin" not in str(model.generate.__func__): if "GenerationMixin" not in str(model.generate.__func__):
model.generate = MethodType(PreTrainedModel.generate, model) model.generate = MethodType(PreTrainedModel.generate, model)

2
src/llamafactory/train/sft/trainer.py
View File

@ -24,6 +24,7 @@ import numpy as np
import torch import torch
from transformers import Seq2SeqTrainer from transformers import Seq2SeqTrainer
from typing_extensions import override from typing_extensions import override
import copy
from ...extras import logging from ...extras import logging
from ...extras.constants import IGNORE_INDEX from ...extras.constants import IGNORE_INDEX
@ -122,7 +123,6 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
labels = inputs.pop("labels", None) labels = inputs.pop("labels", None)
else: else:
labels = inputs.get("labels") labels = inputs.get("labels")
loss, generated_tokens, _ = super().prediction_step( loss, generated_tokens, _ = super().prediction_step(
model, inputs, prediction_loss_only=prediction_loss_only, ignore_keys=ignore_keys, **gen_kwargs model, inputs, prediction_loss_only=prediction_loss_only, ignore_keys=ignore_keys, **gen_kwargs
) )