SDK reference

Types and functions in the modal-dojo Python SDK.

Models

Name	Description
Beam_1_0_501B_A23B	Reflection Beam 1.0: 501B total parameters, 23B active.
Beam_1_0_501B_A23B_LoRA	Select the rank-16 Bridge LoRA preset.
DeepSeek_V4_1_Flash	DeepSeek-V4.1-Flash sparse-attention MoE model, 40 layers and 384 routed experts.
DeepSeek_V4_1_Flash_LoRA	Pinned 1017 release used for the rank-32 native LoRA recipe.
Gemma4_26B_A4B	Google Gemma-4-26B-A4B-it multimodal MoE model with 26B total and 4B active parameters.
GLM_4_7	Zhipu AI GLM-4.7 MoE model with 355B total and 32B active parameters.
GLM_5_3_Flash	GLM-5.3-Flash hybrid MoE; the Miles preset trains its text backbone.
GLM_5_3_Flash_LoRA	GLM-5.3-Flash with the Miles native LoRA preset.
HFModelConfiguration	Downloads Hugging Face model weights with snapshot_download.
Inkling_Small	Thinking Machines Lab Inkling-Small MoE model with 276B total and 12B active parameters.
Inkling_Small_LoRA	Selects Inkling_Small_LoRA_Recipe.
Kimi_K3	Moonshot Kimi-K3 hybrid KDA/MLA MoE model, 93 layers and 896 routed experts.
ModelArchitecture	Megatron transformer architecture parameters.
ModelConfig	Defines model identity, weight download, and response parsing.
Moonlight_16B_A3B_Instruct	Moonshot AI Moonlight model with 16B total and 3B active parameters.
Muse_Spark_Open	816B total / 42B active MoE; architecture comes from Miles’ provider.
ParsedResponse	Structured result of parsing raw model output.
Qwen3_0_6B	Alibaba Qwen3-0.6B model.
Qwen3_1_7B	Alibaba Qwen3-1.7B model.
Qwen3_30B	Alibaba Qwen3-30B-A3B MoE model with 30B total and 3B active parameters.
Qwen3_4B	Alibaba Qwen3-4B model.
Qwen3_5_0_8B	Alibaba Qwen3.5-0.8B model.
Qwen3_5_2B	Alibaba Qwen3.5-2B model.
Qwen3_5_4B	Alibaba Qwen3.5-4B model.
Qwen3_5_9B	Alibaba Qwen3.5-9B model.
Qwen3_6_27B	Qwen3.6-27B dense hybrid Gated DeltaNet/attention model.
Qwen3_6_35B	Alibaba Qwen3.6-35B-A3B model.
Qwen3_8_27B	Alibaba Qwen3.8-27B model.
Qwen3_8B	Alibaba Qwen3-8B model.
Qwen3_ASR_1_7B	Alibaba Qwen3-ASR-1.7B speech recognition model.
Qwen3_VL_8B	Alibaba Qwen3-VL-8B-Instruct model.
ToolCall	Tool invocation parsed from model output.

Datasets

Name	Description
DatasetConfig	Dataset fields and materialization behavior shared across training frameworks.
HarborDataset	A dataset loaded from Harbor tasks.
HuggingFaceDataset	A dataset loaded from a Hugging Face datasets repository.
MultimodalDataset	Dataset of text prompts paired with image, audio, or video data.
OnlineRollout	Placeholder rows that size a live generate batch.

Recipes

Name	Description
Beam_1_0_501B_A23B_LoRA_256k_Recipe	256k Beam LoRA on 16 training and 16 rollout H200s.
DeepSeek_V4_1_Flash_LoRA_Recipe	Rank-32 attention/shared-expert LoRA on 32 H200 GPUs with a 32K window.
DeepSeek_V4_1_Flash_Recipe	DeepSeek-V4.1-Flash GRPO recipe for 8 nodes with 8 H200 GPUs each.
Gemma4_26B_A4B_Recipe	Gemma-4-26B-A4B recipe.
GLM_4_7_Recipe	GLM-4.7 recipe.
GLM_5_3_Flash_LoRA_256k_Recipe	GLM-5.3-Flash rank-16 bridge LoRA for 256k context on 4 H200:8 nodes.
GLM_5_3_Flash_LoRA_Recipe	GLM-5.3-Flash rank-16 native LoRA, 32k context, on 16 H200s.
Inkling_Small_LoRA_Recipe	Inkling-Small rank-32 LoRA recipe.
Inkling_Small_Recipe	Inkling-Small full-parameter recipe.
Kimi_K3_LoRA_64k_Recipe	Kimi-K3 rank-32 LoRA recipe for 64k context on 8 B300:8 nodes.
Kimi_K3_LoRA_Recipe	Kimi-K3 rank-32 LoRA recipe for 8 nodes with 8 B300 GPUs each.
MilesRecipe	Miles training and Modal resource settings.
Moonlight_16B_A3B_Recipe	Moonlight-16B-A3B recipe.
Muse_Spark_Open_64k_Recipe	Full-parameter GRPO, 65,536-token context, 4 nodes × 8 B300 GPUs.
Qwen3_0_6B_Recipe	Qwen3-0.6B recipe.
Qwen3_1_7B_Recipe	Qwen3-1.7B recipe.
Qwen3_4B_Recipe	Qwen3-4B recipe.
Qwen3_5_0_8B_Recipe	Qwen3.5-0.8B recipe.
Qwen3_5_2B_Recipe	Qwen3.5-2B recipe.
Qwen3_5_4B_Miles_Recipe	Qwen3.5-4B recipe.
Qwen3_5_4B_Recipe	Qwen3.5-4B recipe.
Qwen3_5_9B_Recipe	Qwen3.5-9B recipe.
Qwen3_6_27B_Recipe	Qwen3.6-27B recipe.
Qwen3_6_35B_Recipe	Qwen3.6-35B-A3B recipe.
Qwen3_8_27B_Recipe	Qwen3.8-27B recipe.
Qwen3_8B_Recipe	Qwen3-8B recipe.
Qwen3_ASR_1_7B_Recipe	Qwen3-ASR-1.7B recipe.
Qwen3_VL_8B_Recipe	Qwen3-VL-8B recipe.
SlimeRecipe	Slime training and Modal resource settings.

Training

Name	Description
Checkpoint	A complete training checkpoint discovered on a Modal Volume.
CheckpointType	Whether a checkpoint is Hugging Face or Megatron weights.
DashboardComponent	Supported run-scoped dashboard component slots.
DashboardMetricConfig	Log framework metrics to the Modal Dojo dashboard only.
DojoConfigError	Raised when a training or deploy config is invalid.
DojoError	Base error for Modal Dojo.
GpuAllocationError	Raised when a recipe’s cluster or parallelism settings are invalid.
MetricConfig	Defines metric tracker metadata, environment variables, and links.
ModalCaptureError	Raised when a cloudpickled user callback captures live Modal state.
Sample	A prompt, response, parsed structure, score, and metadata from one model call.
TrackioConfig	Trackio logging configuration shared across all frameworks.
TrainConfig	A dataset, model, and recipe for one training run.
TrainingGroup	A parameter sweep over a base TrainConfig.
TrainingRun	A launched training run that can be inspected, awaited, or loaded by ID.
WandbConfig	Weights & Biases run metadata and credentials.
convert_megatron_checkpoint_to_hf	Convert a Megatron checkpoint to Hugging Face format.
convert_native_lora_checkpoint_to_peft	Convert a Miles native-LoRA iter_*/adapter directory to a PEFT adapter at <adapter_path>_hf.
extract_code	Extract Python code from an LLM response.
score_in_sandbox	Run code against test cases in a Modal sandbox.

Deployment

Name	Description
CustomDeployment	A model deployed with an SGLang or vLLM recipe.
Endpoint	Controls a Modal Endpoint that persists until stopped.
Sandbox	Controls a Modal Sandbox in the app_name app.
SandboxResult	Outcome of a single Sandbox.run invocation.
SglangRecipe	SGLang server settings.
VllmRecipe	vLLM server settings.