DeepSeek_V4_1_Flash_LoRA_Recipe
"# DeepSeek_V4_1_Flash_LoRA_Recipe\n\n```python\nfrom modal_dojo import DeepSeek_V4_1_Flash_LoRA_Recipe\n```\n\nRank-32 attention/shared-expert LoRA on 32 H200 GPUs with a 32K window.\n\nCovers all five attention projections and the shared expert's gate/up/down\nprojections supported by Miles #4205 and SGLang #45231. Routed experts,\ncompressor, indexer and Engram remain frozen. Workload and reward are\nsupplied by the training config; generation uses ordinary stopping.\n\n## `cli_args`\n\n```python\ncli_args(dataset: DatasetConfig | None = None, eval_dataset: DatasetConfig | None = None, dataset_path: str | None = None, eval_dataset_path: str | None = None, model: ModelConfig | None = None) -> list[str]\n```\n\n## `download_model`\n\n```python\ndownload_model(revision: str | None = None) -> None\n```\n\n## `effective_num_epoch`\n\n```python\neffective_num_epoch() -> int | None\n```\n\n## `get_base_recipe`\n\n```python\nget_base_recipe(model_config: ModelConfig) -> MilesRecipe | None\n```\n\nReturn the model preset for `model_config`.\n\nCall this explicitly. `TrainConfig` uses the recipe it receives\nwithout applying a preset.\n\n**Returns**\n\nThe model preset, or `None` when no preset is registered.\n\n**Raises**\n\n- `DojoConfigError`: The recipe accepts only registered models and `model_config` is not registered.\n\n## `gpu_allocation`\n\n```python\ngpu_allocation: GpuAllocation\n```\n\n## `post_process_data`\n\n```python\npost_process_data() -> None\n```\n\n## `post_process_model`\n\n```python\npost_process_model() -> None\n```\n\n## `total_nodes`\n\n```python\ntotal_nodes: int\n```\n\n## `train_async`\n\n```python\ntrain_async: bool\n```\n\n## `validate_model_parallelism`\n\n```python\nvalidate_model_parallelism(model: ModelConfig) -> None\n```\n\nValidate the model's parallelism settings.\n"
from modal_dojo import DeepSeek_V4_1_Flash_LoRA_RecipeRank-32 attention/shared-expert LoRA on 32 H200 GPUs with a 32K window.
Covers all five attention projections and the shared expert’s gate/up/down projections supported by Miles #4205 and SGLang #45231. Routed experts, compressor, indexer and Engram remain frozen. Workload and reward are supplied by the training config; generation uses ordinary stopping.
cli_args
Section titled “cli_args”cli_args(dataset: DatasetConfig | None = None, eval_dataset: DatasetConfig | None = None, dataset_path: str | None = None, eval_dataset_path: str | None = None, model: ModelConfig | None = None) -> list[str]download_model
Section titled “download_model”download_model(revision: str | None = None) -> Noneeffective_num_epoch
Section titled “effective_num_epoch”effective_num_epoch() -> int | Noneget_base_recipe
Section titled “get_base_recipe”get_base_recipe(model_config: ModelConfig) -> MilesRecipe | NoneReturn the model preset for model_config.
Call this explicitly. TrainConfig uses the recipe it receives
without applying a preset.
Returns
The model preset, or None when no preset is registered.
Raises
DojoConfigError: The recipe accepts only registered models andmodel_configis not registered.
gpu_allocation
Section titled “gpu_allocation”gpu_allocation: GpuAllocationpost_process_data
Section titled “post_process_data”post_process_data() -> Nonepost_process_model
Section titled “post_process_model”post_process_model() -> Nonetotal_nodes
Section titled “total_nodes”total_nodes: inttrain_async
Section titled “train_async”train_async: boolvalidate_model_parallelism
Section titled “validate_model_parallelism”validate_model_parallelism(model: ModelConfig) -> NoneValidate the model’s parallelism settings.