GLM_4_7
Section titled “GLM_4_7”from modal_training_gym.common.models.glm_4_7 import GLM_4_7GLM-4.7 large MoE model from Zhipu AI.
Inherits from: HFModelConfiguration, ModelConfiguration
Fields
Section titled “Fields”| Field | Type | Default | Description |
|---|---|---|---|
model_name | str | "zai-org/GLM-4.7" | HuggingFace repo ID or other model identifier. Default "". |
model_path | `str | None` | None |
architecture | `ModelArchitecture | None` | None |
training | `ModelTrainingConfig | None` | ModelTrainingConfig(gpu_type='H100', n_nodes=4, tensor_model_parallel_size=2, pipeline_model_parallel_size=4, context_parallel_size=1, sequence_parallel=True, expert_model_parallel_size=4, moe_permute_fusion=True, moe_grouped_gemm=True, moe_shared_expert_overlap=True, moe_aux_loss_coeff=0.001, lora_rank=128, lora_alpha=32, target_modules='all-linear', merge_lora=False) |
Methods
Section titled “Methods”download_model(self) -> 'None'
Section titled “download_model(self) -> 'None'”Download or materialize weights into the model volume.