Functions — github.com/2U1/Qwen-VL-Series-Finetune

hub / github.com/2U1/Qwen-VL-Series-Finetune / functions

Functions211 in github.com/2U1/Qwen-VL-Series-Finetune

⨍ Functions 211◇ Types & classes 28

↓ 15 callersFunction_flatten_vision_features

(vision_outputs)

src/train/monkey_patch_forward.py:19

↓ 14 callersMethodget_image_features

(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None)

src/model/modeling_cls.py:108

↓ 10 callersMethodget_input_embeddings

(self)

src/model/modeling_cls.py:91

↓ 9 callersFunctionget_mm_token_type_ids

(inputs, input_ids)

src/dataset/data_utils.py:77

↓ 9 callersFunctionget_qwen_vl_generation_backbone

(model)

src/model/load_model.py:54

↓ 9 callersMethodget_video_features

( self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = None

src/model/modeling_cls.py:103

↓ 9 callersFunctionpad_sequence

Pad a list of sequences to the same length. sequences: list of tensors in [seq_len, *] shape

src/dataset/data_utils.py:57

↓ 8 callersFunctionget_peft_state_non_lora_maybe_zero_3

(named_params, require_grad_only=True)

src/train/train_utils.py:50

↓ 6 callersFunction_get_deepstack_features

(vision_outputs)

src/train/monkey_patch_forward.py:28

↓ 6 callersFunctionload_qwen_vl_generation_model

( model_name_or_path: str, *, config: PretrainedConfig | None = None, **kwargs: Any )

src/model/load_model.py:73

↓ 5 callersMethodlanguage_model

(self)

src/model/modeling_cls.py:112

↓ 5 callersFunctionset_requires_grad

(parameters, requires_grad)

src/train/train_sft.py:40

↓ 5 callersFunctionset_requires_grad

(parameters, requires_grad)

src/train/train_grpo.py:42

↓ 5 callersFunctionset_requires_grad

(parameters, requires_grad)

src/train/train_dpo.py:40

↓ 4 callersMethod__init__

(self, config)

src/model/modeling_cls.py:63

↓ 4 callersFunction_ensure_mm_token_type_ids_generate_compat

(model)

src/trainer/grpo_trainer.py:97

↓ 4 callersMethod_get_per_token_logps_and_entropies

Compute log-probs and (optionally) entropies for each token.

src/trainer/grpo_trainer.py:640

↓ 4 callersFunction_get_text_hidden_size

(config)

src/model/modeling_cls.py:45

↓ 4 callersFunctionformat_assistant_response

( content, reasoning=None, *, enable_reasoning=False, use_reasoning_prefill=False, use

src/dataset/data_utils.py:116

↓ 4 callersMethodget_decoder

(self)

src/model/modeling_cls.py:100

↓ 4 callersFunctionget_peft_state_maybe_zero_3

(named_params, bias)

src/train/train_utils.py:25

↓ 4 callersFunctionget_qwen_multimodal_settings

(model_id_or_path)

src/dataset/data_utils.py:85

↓ 4 callersFunctionsafe_save_model_for_hf_trainer

Collects the state dict and dump to disk.

src/train/train_utils.py:57

↓ 4 callersMethodset_decoder

(self, decoder)

src/model/modeling_cls.py:97

↓ 4 callersMethodset_input_embeddings

(self, value)

src/model/modeling_cls.py:94

↓ 4 callersFunctionset_requires_grad

(parameters, requires_grad)

src/train/train_cls.py:53

↓ 4 callersFunctionuse_default_system_message

(model_type)

src/dataset/data_utils.py:92

↓ 3 callersFunction_expand_video_grid_to_frames

(video_grid_thw)

src/train/monkey_patch_forward.py:50

↓ 3 callersFunction_make_dummy_qwen3_visual_inputs

(visual)

src/train/monkey_patch_forward.py:36

↓ 3 callersFunctionchat_template_uses_reasoning_prefill

(processor, model_type=None)

src/dataset/data_utils.py:96

↓ 3 callersFunctionget_image_info

(image_path, min_pixel, max_pixel, width, height, image_patch_size)

src/dataset/data_utils.py:138

↓ 3 callersFunctionget_video_info

(video_path, min_pixels, max_pixels, width, height, fps, image_patch_size, return_video_metadata=False)

src/dataset/data_utils.py:163

↓ 3 callersFunctionmodel_supports_optional_reasoning

(model_type)

src/dataset/data_utils.py:112

↓ 3 callersFunctionrank0_print

(*args)

src/train/train_cls.py:32

↓ 2 callersMethod_gather_predictions

Gather string predictions across all processes.

src/trainer/sft_trainer.py:445

↓ 2 callersFunction_qwen3_5_mixed_modality_forward_impl

( self, *, output_cls, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.

src/train/monkey_patch_forward.py:86

↓ 2 callersFunctionapply_qwen_vl_monkey_patches

(model_type: str)

src/model/load_model.py:60

↓ 2 callersFunctionget_model_name_from_path

(model_path)

src/utils.py:109

↓ 2 callersFunctionis_lora_model

Check if a model directory contains LoRA adapter files. Args: model_path: Path to the model directory Returns:

src/utils.py:96

↓ 2 callersFunctionis_video_file

(filename)

src/serve/app.py:13

↓ 2 callersFunctionllava_to_openai

(conversations, is_video=False)

src/dataset/data_utils.py:29

↓ 2 callersFunctionload_pretrained_model

(model_path, model_base, model_name, load_8bit=False, load_4bit=False, device_map="

src/utils.py:25

↓ 2 callersFunctionmaybe_zero_3

(param, ignore_status=False, name=None, device=torch.device('cpu'))

src/train/train_utils.py:6

↓ 2 callersFunctionrank0_print

(*args)

src/train/train_sft.py:19

↓ 2 callersFunctionrank0_print

(*args)

src/train/train_dpo.py:19

↓ 2 callersFunctionreplace_image_tokens

(input_string, is_video=False)

src/dataset/data_utils.py:19

↓ 1 callersMethod__init__

(self, samples_per_cls, beta=0.9995, gamma=1.5, reduction="mean")

src/loss/class_balance_loss.py:35

↓ 1 callersMethod_apply_liger_grpo_loss_type_override

If the user passed --liger_grpo_loss_type, swap the loss variant on the LigerFusedLinearGRPOLoss instance that TRL's GRPOTrainer set up.

src/trainer/grpo_trainer.py:138

↓ 1 callersMethod_extract_prompt_and_reference

Extract prompt (question only) and reference (answer) from input_ids and labels. In SFT dataset, labels == IGNORE_INDEX for prompt t

src/trainer/sft_trainer.py:193

↓ 1 callersMethod_get_dataloader

Create a [`~torch.utils.data.DataLoader`] from the given dataset.

src/trainer/cls_trainer.py:78

↓ 1 callersMethod_get_last_hidden_state

( self, unwrapped_model, input_ids, attention_mask, logits_to_keep,

src/trainer/grpo_trainer.py:727

↓ 1 callersFunction_iter_generate_models

(model)

src/trainer/grpo_trainer.py:58

↓ 1 callersFunction_liger_exposes_mm_token_type_ids

liger-kernel >= 0.8.0 restored `mm_token_type_ids` on the patched VL forward (see linkedin/Liger-Kernel#1120, #1140). Older releases drop it from

src/trainer/grpo_trainer.py:83

↓ 1 callersFunction_normalize_trl_optional_flags

()

src/trainer/dpo_trainer.py:23

↓ 1 callersFunction_normalize_trl_optional_flags

()

src/trainer/grpo_trainer.py:27

↓ 1 callersMethod_prepare_generation_inputs

Prepare inputs for generation by padding prompts and including vision inputs.

src/trainer/sft_trainer.py:226

↓ 1 callersMethodconcatenated_inputs

( batch: dict[str, Union[list, torch.LongTensor]], padding_value: int )

src/trainer/dpo_trainer.py:67

↓ 1 callersFunctionconfigure_llm

(model, training_args)

src/train/train_sft.py:60

↓ 1 callersFunctionconfigure_llm

(model, training_args)

src/train/train_grpo.py:62

↓ 1 callersFunctionconfigure_llm

(model, training_args)

src/train/train_dpo.py:60

↓ 1 callersFunctionconfigure_llm

(model, training_args)

src/train/train_cls.py:69

↓ 1 callersFunctionconfigure_vision_tower

(model, training_args, compute_dtype, device)

src/train/train_sft.py:44

↓ 1 callersFunctionconfigure_vision_tower

(model, training_args, compute_dtype, device)

src/train/train_grpo.py:46

↓ 1 callersFunctionconfigure_vision_tower

(model, training_args, compute_dtype, device)

src/train/train_dpo.py:44

↓ 1 callersFunctionconfigure_vision_tower

(model, training_args, compute_dtype, device)

src/train/train_cls.py:57

↓ 1 callersFunctiondisable_torch_init

Disable the redundant torch default initialization to accelerate model creation.

src/utils.py:17

↓ 1 callersFunctionfind_target_linear_names

(model, num_lora_modules=-1, lora_namespan_exclude=[], verbose=True)

src/train/train_sft.py:23

↓ 1 callersFunctionfind_target_linear_names

(model, num_lora_modules=-1, lora_namespan_exclude=[], verbose=True)

src/train/train_grpo.py:25

↓ 1 callersFunctionfind_target_linear_names

(model, num_lora_modules=-1, lora_namespan_exclude=[], verbose=True)

src/train/train_dpo.py:23

↓ 1 callersFunctionfind_target_linear_names

(model, num_lora_modules=-1, lora_namespan_exclude=[], verbose=True)

src/train/train_cls.py:36

↓ 1 callersFunctionget_image_content

(image_path, min_pixel, max_pixel, width, height)

src/dataset/cls_dataset.py:30

↓ 1 callersFunctionget_loss_function

(training_args, samples_per_class=None)

src/loss/loss_factory.py:5

↓ 1 callersFunctionget_qwen_vl_sequence_classification_model_cls

(model_type: str)

src/model/load_model.py:92

↓ 1 callersFunctionget_video_content

(video_path, min_pixels, max_pixels, width, height, fps, nframes)

src/dataset/cls_dataset.py:44

↓ 1 callersMethodget_window_index

(self, grid_thw)

src/train/monkey_patch_vision.py:80

↓ 1 callersFunctionload_qwen_vl_sequence_classification_model

( model_name_or_path: str, *, config: PretrainedConfig | None = None, **kwargs: Any )

src/model/load_model.py:102

↓ 1 callersFunctionload_reward_funcs

( module_path: str = "train.reward_funcs", *, name_pred = lambda n: n.endswith("_reward"), obj

src/utils.py:117

↓ 1 callersFunctionmain

(args)

src/serve/app.py:82

↓ 1 callersFunctionmake_classification_data_module

(model_id, processor, data_args)

src/dataset/cls_dataset.py:254

↓ 1 callersFunctionmake_dpo_data_module

Make dataset and collator for DPO fine-tuning.

src/dataset/dpo_dataset.py:378

↓ 1 callersFunctionmake_grpo_data_module

Make dataset and collator for supervised fine-tuning.

src/dataset/grpo_dataset.py:179

↓ 1 callersFunctionmake_supervised_data_module

Make dataset and collator for supervised fine-tuning.

src/dataset/sft_dataset.py:368

↓ 1 callersFunctionmerge_lora

(args)

src/merge_lora_weights.py:4

↓ 1 callersFunctionrank0_print

(*args)

src/train/train_grpo.py:21

↓ 1 callersFunctionreplace_qwen2_5_vision

()

src/train/monkey_patch_vision.py:16

↓ 1 callersFunctionreplace_qwen2_5_with_mixed_modality_forward

()

src/train/monkey_patch_forward.py:70

↓ 1 callersFunctionreplace_qwen3_5_moe_with_mixed_modality_forward

()

src/train/monkey_patch_forward.py:79

↓ 1 callersFunctionreplace_qwen3_5_with_mixed_modality_forward

()

src/train/monkey_patch_forward.py:76

↓ 1 callersFunctionreplace_qwen3_with_mixed_modality_forward

()

src/train/monkey_patch_forward.py:73

↓ 1 callersFunctionreplace_qwen_2_with_mixed_modality_forward

()

src/train/monkey_patch_forward.py:67

↓ 1 callersMethodrot_pos_emb

(self, grid_thw)

src/train/monkey_patch_vision.py:51

↓ 1 callersFunctionsamples_per_class_from_ids

(label_ids, num_classes)

src/dataset/data_utils.py:194

↓ 1 callersFunctiontrain

()

src/train/train_sft.py:82

↓ 1 callersFunctiontrain

()

src/train/train_grpo.py:85

↓ 1 callersFunctiontrain

()

src/train/train_dpo.py:81

↓ 1 callersFunctiontrain

()

src/train/train_cls.py:84

↓ 1 callersFunctionunfreeze_topk_layers

(model, k_llm: int = 0, k_vis: int = 0)

src/train/train_sft.py:68

↓ 1 callersFunctionunfreeze_topk_layers

(model, k_llm: int = 0, k_vis: int = 0)

src/train/train_grpo.py:70

↓ 1 callersFunctionunfreeze_topk_layers

(model, k_llm: int = 0, k_vis: int = 0)

src/train/train_dpo.py:68

↓ 1 callersFunctionunfreeze_topk_layers

(model, k_llm: int = 0, k_vis: int = 0)

src/train/train_cls.py:73

next →1–100 of 211, ranked by callers