`bridge.recipes.glm_vl.glm_45v`#

Module Contents#

Classes#

GLM45VCommonKwargs

Typed options accepted by GLM-4.5V recipe helper functions.

Functions#

`set_glm_45v_pipeline_model_parallel_layout`	Set the GLM-4.5V pipeline model parallel layout.
`glm_45v_finetune_config`	Return a fine-tuning config for GLM-4.5V (based on GLM-4.5 Air 106B).
`_glm_45v_common`	Create a fine-tuning configuration for GLM-4.5V models using a given HuggingFace path.

API#

bridge.recipes.glm_vl.glm_45v.set_glm_45v_pipeline_model_parallel_layout( model_cfg: megatron.bridge.models.gpt_provider.GPTModelProvider, layout: Optional[Union[str, List[List[str]]]] = None, is_peft: bool = False, ) → None#

Set the GLM-4.5V pipeline model parallel layout.

GLM-4.5V (based on GLM-4.5 Air) has 46 decoder layers and no MTP layers. This function sets up predefined layouts for common PP/VP combinations.

Parameters:

model_cfg – The model provider configuration to modify.
layout – Optional custom layout. If None, uses predefined layouts based on PP/VP sizes.
is_peft – Whether the model is trained with PEFT.

class bridge.recipes.glm_vl.glm_45v.GLM45VCommonKwargs#

Bases: typing_extensions.TypedDict

Typed options accepted by GLM-4.5V recipe helper functions.

Initialization

Initialize self. See help(type(self)) for accurate signature.

hf_path: str#: None

dir: Optional[str]#: None

name: str#: None

train_data_path: Optional[List[str]]#: None

valid_data_path: Optional[List[str]]#: None

test_data_path: Optional[List[str]]#: None

dataset_type: Optional[str]#: None

image_folder: Optional[str]#: None

tokenizer_model: Optional[str]#: None

tensor_model_parallel_size: int#: None

pipeline_model_parallel_size: int#: None

pipeline_dtype: Optional[torch.dtype]#: None

virtual_pipeline_model_parallel_size: Optional[int]#: None

expert_model_parallel_size: int#: None

context_parallel_size: int#: None

sequence_parallel: bool#: None

use_megatron_fsdp: bool#: None

train_iters: int#: None

global_batch_size: int#: None

micro_batch_size: int#: None

seq_length: int#: None

lr: float#: None

min_lr: float#: None

lr_warmup_iters: int#: None

lr_decay_iters: Optional[int]#: None

eval_interval: int#: None

save_interval: int#: None

precision_config: Optional[Union[megatron.bridge.training.mixed_precision.MixedPrecisionConfig, str]]#: None

comm_overlap_config: Optional[megatron.bridge.training.comm_overlap.CommOverlapConfig]#: None

freeze_language_model: bool#: None

freeze_vision_model: bool#: None

freeze_vision_projection: bool#: None

pretrained_checkpoint: Optional[str]#: None

layout: Optional[Union[str, List[List[str]]]]#: None

peft: Optional[Union[str, megatron.bridge.peft.base.PEFT]]#: None

finetune_lr: float#: None

wandb_project: Optional[str]#: None

wandb_entity: Optional[str]#: None

wandb_exp_name: Optional[str]#: None

bridge.recipes.glm_vl.glm_45v.glm_45v_finetune_config(

**user_kwargs: typing_extensions.Unpack[bridge.recipes.glm_vl.glm_45v.GLM45VCommonKwargs],

) → megatron.bridge.training.config.ConfigContainer#

Return a fine-tuning config for GLM-4.5V (based on GLM-4.5 Air 106B).

Default configuration:

LoRA/DoRA: TP=1, PP=8, EP=4 (64 GPUs, 8 nodes), LR=1e-4
Full SFT: TP=1, PP=8, EP=16 (512 GPUs, 64 nodes), LR=5e-6

GLM-4.5V is a Vision-Language model with:

106B total parameters (based on GLM-4.5 Air)
Sparse MoE with shared experts
Multi-modality support for images and videos

See _glm_45v_common for the full list of parameters.

bridge.recipes.glm_vl.glm_45v._glm_45v_common( hf_path: str, dir: Optional[str] = None, name: str = 'glm_45v_finetune', pretrained_checkpoint: Optional[str] = None, train_data_path: Optional[List[str]] = None, valid_data_path: Optional[List[str]] = None, test_data_path: Optional[List[str]] = None, dataset_type: Optional[str] = None, image_folder: Optional[str] = None, tokenizer_model: Optional[str] = None, tensor_model_parallel_size: int = 1, pipeline_model_parallel_size: int = 2, pipeline_dtype: Optional[torch.dtype] = None, virtual_pipeline_model_parallel_size: Optional[int] = None, expert_model_parallel_size: int = 4, context_parallel_size: int = 1, sequence_parallel: bool = False, use_megatron_fsdp: bool = False, train_iters: int = 300000, global_batch_size: int = 32, micro_batch_size: int = 1, seq_length: int = 8192, lr: float = 0.0003, min_lr: float = 3e-05, lr_warmup_iters: int = 500, lr_decay_iters: Optional[int] = None, eval_interval: int = 500, save_interval: int = 500, precision_config: Optional[Union[megatron.bridge.training.mixed_precision.MixedPrecisionConfig, str]] = 'bf16_mixed', comm_overlap_config: Optional[megatron.bridge.training.comm_overlap.CommOverlapConfig] = None, freeze_language_model: bool = False, freeze_vision_model: bool = False, freeze_vision_projection: bool = False, layout: Optional[Union[str, List[List[str]]]] = None, peft: Optional[Union[str, megatron.bridge.peft.base.PEFT]] = None, finetune_lr: Optional[float] = None, wandb_project: Optional[str] = None, wandb_entity: Optional[str] = None, wandb_exp_name: Optional[str] = None, ) → megatron.bridge.training.config.ConfigContainer#

Create a fine-tuning configuration for GLM-4.5V models using a given HuggingFace path.

The dataset pipeline is conversation-based. To train multimodal tokens, ensure your preprocessed data includes placeholders (e.g., ) as needed.

GLM-4.5V is a Vision-Language model based on GLM-4.5 Air (106B parameters) with:

Sparse MoE architecture with shared experts
Multi-modal support for images and videos
MRoPE (Multi-Resolution Rotary Position Embedding)

bridge.recipes.glm_vl.glm_45v#

Module Contents#

Classes#

Functions#

API#

`bridge.recipes.glm_vl.glm_45v`#