SDK reference

"# SDK reference\n\nClasses and methods in the `modal-training-gym` Python SDK.\n\n## Models\n\n| Name | Description |\n|------|-------------|\n| [`Gemma-4-26B-A4B-it`](/reference/gemma4_26b_a4b) | Google Gemma-4-26B-A4B-it multimodal MoE model with 25.2B total and 3.8B active parameters. |\n| [`GLM-4.7`](/reference/glm_4_7) | Zhipu AI GLM-4.7 MoE model with 355B total and 32B active parameters. |\n| [`HFModelConfiguration`](/reference/hfmodelconfiguration) | Downloads Hugging Face model weights with `snapshot_download`. |\n| [`Inkling-Small`](/reference/inkling_small) | Thinking Machines Lab Inkling-Small MoE model with 276B total and 12B active parameters. |\n| [`ModelArchitecture`](/reference/modelarchitecture) | Megatron transformer architecture parameters. |\n| [`ModelConfig`](/reference/modelconfig) | Defines model identity, weight download, and response parsing. |\n| [`Moonlight-16B-A3B-Instruct`](/reference/moonlight_16b_a3b_instruct) | Moonshot AI Moonlight model with 16B total and 3B active parameters. |\n| [`Qwen3-0.6B`](/reference/qwen3_0_6b) | Alibaba Qwen3-0.6B model. |\n| [`Qwen3-1.7B`](/reference/qwen3_1_7b) | Alibaba Qwen3-1.7B model. |\n| [`Qwen3-30B-A3B`](/reference/qwen3_30b) | Alibaba Qwen3-30B-A3B MoE model with 30B total and 3B active parameters. |\n| [`Qwen3-4B`](/reference/qwen3_4b) | Alibaba Qwen3-4B model. |\n| [`Qwen3-8B`](/reference/qwen3_8b) | Alibaba Qwen3-8B model. |\n| [`Qwen3-ASR-1.7B`](/reference/qwen3_asr_1_7b) | Alibaba Qwen3-ASR-1.7B speech recognition model. |\n| [`Qwen3-VL-8B`](/reference/qwen3_vl_8b) | Alibaba Qwen3-VL-8B-Instruct model. |\n| [`Qwen3.5-0.8B`](/reference/qwen3_5_0_8b) | Alibaba Qwen3.5-0.8B model. |\n| [`Qwen3.5-2B`](/reference/qwen3_5_2b) | Alibaba Qwen3.5-2B model. |\n| [`Qwen3.5-4B`](/reference/qwen3_5_4b) | Alibaba Qwen3.5-4B model. |\n| [`Qwen3.5-9B`](/reference/qwen3_5_9b) | Alibaba Qwen3.5-9B model. |\n| [`Qwen3.6-27B`](/reference/qwen3_6_27b) | Qwen3.6-27B dense hybrid Gated DeltaNet/attention model. |\n| [`Qwen3.6-35B-A3B`](/reference/qwen3_6_35b) | Alibaba Qwen3.6-35B-A3B model. |\n| [`Qwen3.8-27B`](/reference/qwen3_8_27b) | Alibaba Qwen3.8-27B model. |\n| [`ToolCall`](/reference/toolcall) | Tool invocation parsed from model output. |\n\n## Datasets\n\n| Name | Description |\n|------|-------------|\n| [`DatasetConfig`](/reference/datasetconfig) | Dataset fields and materialization behavior shared across training frameworks. |\n| [`HarborDataset`](/reference/harbordataset) | A dataset loaded from Harbor tasks. |\n| [`HuggingFaceDataset`](/reference/huggingfacedataset) | A dataset loaded from a Hugging Face `datasets` repository. |\n\n## Recipes\n\n| Name | Description |\n|------|-------------|\n| [`Gemma4_26B_A4B_Recipe`](/reference/gemma4_26b_a4b_recipe) | Gemma-4-26B-A4B GRPO recipe on 1 node with 8 H200 GPUs. |\n| [`GLM_4_7_Recipe`](/reference/glm_4_7_recipe) | GLM-4.7 MoE GSPO recipe for 8 trainer nodes and 8 rollout nodes, each with 8 H200 GPUs. |\n| [`Inkling_Small_LoRA_Recipe`](/reference/inkling_small_lora_recipe) | Inkling-Small rank-32 LoRA GRPO recipe for 4 nodes with 8 H200 GPUs each. |\n| [`Inkling_Small_Recipe`](/reference/inkling_small_recipe) | Inkling-Small full-parameter GRPO recipe for 4 nodes with 8 H200 GPUs each. |\n| [`Moonlight_16B_A3B_Recipe`](/reference/moonlight_16b_a3b_recipe) | Moonlight-16B-A3B DAPO recipe for 1 node with 8 H100 GPUs. |\n| [`Qwen3.5-0.8B_Recipe`](/reference/qwen3_5_0_8b_recipe) | Qwen3.5-0.8B GRPO recipe for 1 node with 8 H100 GPUs. |\n| [`Qwen3.5-2B_Recipe`](/reference/qwen3_5_2b_recipe) | Qwen3.5-2B GRPO recipe for 1 node with 8 H100 GPUs. |\n| [`Qwen3.5-4B_Recipe`](/reference/qwen3_5_4b_recipe) | Qwen3.5-4B GRPO recipe for 1 node with 8 H100 GPUs. |\n| [`Qwen3.5-9B_Recipe`](/reference/qwen3_5_9b_recipe) | Qwen3.5-9B GRPO recipe for 1 node with 8 H100 GPUs. |\n| [`Qwen3_0_6B_Recipe`](/reference/qwen3_0_6b_recipe) | Qwen3-0.6B GRPO recipe for 1 node with 8 H100 GPUs. |\n| [`Qwen3_1_7B_Recipe`](/reference/qwen3_1_7b_recipe) | Qwen3-1.7B GRPO recipe for 1 node with 8 H100 GPUs. |\n| [`Qwen3_4B_Recipe`](/reference/qwen3_4b_recipe) | Qwen3-4B GRPO recipe for 1 node with 8 H100 GPUs. |\n| [`Qwen3_5_4B_Miles_Recipe`](/reference/qwen3_5_4b_miles_recipe) | Qwen3.5-4B GRPO recipe for 1 node with 8 H100 GPUs. |\n| [`Qwen3_6_27B_Recipe`](/reference/qwen3_6_27b_recipe) | Qwen3.6-27B GRPO recipe for 4 nodes with 8 H100 GPUs each. |\n| [`Qwen3_6_35B_Recipe`](/reference/qwen3_6_35b_recipe) | Qwen3.6-35B-A3B MoE recipe for 1 node with 8 H100 GPUs. |\n| [`Qwen3_8_27B_Recipe`](/reference/qwen3_8_27b_recipe) | Qwen3.8-27B GRPO recipe for 4 nodes with 8 H100 GPUs each. |\n| [`Qwen3_8B_Recipe`](/reference/qwen3_8b_recipe) | Qwen3-8B GRPO recipe for 1 node with 8 H100 GPUs. |\n| [`Qwen3_ASR_1_7B_Recipe`](/reference/qwen3_asr_1_7b_recipe) | Qwen3-ASR-1.7B audio GRPO recipe for 1 node with 2 H100 GPUs. |\n| [`Qwen3_VL_8B_Recipe`](/reference/qwen3_vl_8b_recipe) | Qwen3-VL-8B GRPO recipe for 1 node with 8 H100 GPUs. |\n\n## Training\n\n| Name | Description |\n|------|-------------|\n| [`MetricConfig`](/reference/metricconfig) | Defines metric tracker metadata, environment variables, and links. |\n| [`MilesRecipe`](/reference/milesrecipe) | Miles training and Modal resource settings. |\n| [`ModalRayCluster`](/reference/modalraycluster) | Starts and coordinates a Ray cluster inside a Modal clustered function. |\n| [`SlimeRecipe`](/reference/slimerecipe) | Slime training and Modal resource settings. |\n| [`TrackioConfig`](/reference/trackioconfig) | Trackio logging configuration shared across all frameworks. |\n| [`TrainConfig`](/reference/trainconfig) | A dataset, model, and recipe for one training run. |\n| [`TrainingGroup`](/reference/traininggroup) | A parameter sweep over a base `TrainConfig`. |\n| [`TrainingRun`](/reference/trainingrun) | A launched training run that can be inspected, awaited, or loaded by ID. |\n| [`TrainResult`](/reference/trainresult) | Result metadata and checkpoints for a completed training run. |\n| [`WandbConfig`](/reference/wandbconfig) | Weights & Biases run metadata and credentials. |\n\n## Deployment\n\n| Name | Description |\n|------|-------------|\n| [`CustomDeployment`](/reference/customdeployment) | A model deployed with an SGLang or vLLM recipe. |\n| [`DeepSeek_V4_Flash_SglangRecipe`](/reference/deepseek_v4_flash_sglangrecipe) | DeepSeek-V4-Flash 284B MoE SGLang recipe with 13B active parameters for 4×B200. |\n| [`Endpoint`](/reference/endpoint) | Controls a [Modal Endpoint](https://modal.com/docs/guide/endpoints) that persists until stopped. |\n| [`GLM_4_7_SglangRecipe`](/reference/glm_4_7_sglangrecipe) | GLM-4.7 355B-A32B MoE SGLang recipe for 8×H200. |\n| [`ParsedResponse`](/reference/parsedresponse) | Structured result of parsing raw model output. |\n| [`Qwen3_0_6B_SglangRecipe`](/reference/qwen3_0_6b_sglangrecipe) | Qwen3-0.6B SGLang recipe for 1×H100. |\n| [`Qwen3_0_6B_VllmRecipe`](/reference/qwen3_0_6b_vllmrecipe) | Qwen3-0.6B vLLM recipe for 1×H100. |\n| [`Qwen3_1_7B_SglangRecipe`](/reference/qwen3_1_7b_sglangrecipe) | Qwen3-1.7B SGLang recipe for 1×H100. |\n| [`Qwen3_1_7B_VllmRecipe`](/reference/qwen3_1_7b_vllmrecipe) | Qwen3-1.7B vLLM recipe for 1×H100. |\n| [`Qwen3_30B_SglangRecipe`](/reference/qwen3_30b_sglangrecipe) | Qwen3-30B-A3B (MoE) SGLang recipe for 4×H100. |\n| [`Qwen3_30B_VllmRecipe`](/reference/qwen3_30b_vllmrecipe) | Qwen3-30B-A3B (MoE) vLLM recipe for 4×H100. |\n| [`Qwen3_4B_SglangRecipe`](/reference/qwen3_4b_sglangrecipe) | Qwen3-4B SGLang recipe for 1×H100. |\n| [`Qwen3_4B_VllmRecipe`](/reference/qwen3_4b_vllmrecipe) | Qwen3-4B vLLM recipe for 1×H100. |\n| [`Qwen3_6_35B_SglangRecipe`](/reference/qwen3_6_35b_sglangrecipe) | Qwen3.6-35B-A3B (MoE) SGLang recipe for 4×H100. |\n| [`Qwen3_6_35B_VllmRecipe`](/reference/qwen3_6_35b_vllmrecipe) | Qwen3.6-35B-A3B (MoE) vLLM recipe. |\n| [`Qwen3_8B_SglangRecipe`](/reference/qwen3_8b_sglangrecipe) | Qwen3-8B SGLang recipe for 1×H100. |\n| [`Qwen3_8B_VllmRecipe`](/reference/qwen3_8b_vllmrecipe) | Qwen3-8B vLLM recipe for 1×H100. |\n| [`SglangRecipe`](/reference/sglangrecipe) | SGLang server settings. |\n| [`VllmRecipe`](/reference/vllmrecipe) | vLLM server settings. |\n"

Classes and methods in the modal-training-gym Python SDK.

Models

Name	Description
`Gemma-4-26B-A4B-it`	Google Gemma-4-26B-A4B-it multimodal MoE model with 25.2B total and 3.8B active parameters.
`GLM-4.7`	Zhipu AI GLM-4.7 MoE model with 355B total and 32B active parameters.
`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.
`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.
`Qwen3-0.6B`	Alibaba Qwen3-0.6B model.
`Qwen3-1.7B`	Alibaba Qwen3-1.7B model.
`Qwen3-30B-A3B`	Alibaba Qwen3-30B-A3B MoE model with 30B total and 3B active parameters.
`Qwen3-4B`	Alibaba Qwen3-4B 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.
`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-A3B`	Alibaba Qwen3.6-35B-A3B model.
`Qwen3.8-27B`	Alibaba Qwen3.8-27B 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.

Recipes

Name	Description
`Gemma4_26B_A4B_Recipe`	Gemma-4-26B-A4B GRPO recipe on 1 node with 8 H200 GPUs.
`GLM_4_7_Recipe`	GLM-4.7 MoE GSPO recipe for 8 trainer nodes and 8 rollout nodes, each with 8 H200 GPUs.
`Inkling_Small_LoRA_Recipe`	Inkling-Small rank-32 LoRA GRPO recipe for 4 nodes with 8 H200 GPUs each.
`Inkling_Small_Recipe`	Inkling-Small full-parameter GRPO recipe for 4 nodes with 8 H200 GPUs each.
`Moonlight_16B_A3B_Recipe`	Moonlight-16B-A3B DAPO recipe for 1 node with 8 H100 GPUs.
`Qwen3.5-0.8B_Recipe`	Qwen3.5-0.8B GRPO recipe for 1 node with 8 H100 GPUs.
`Qwen3.5-2B_Recipe`	Qwen3.5-2B GRPO recipe for 1 node with 8 H100 GPUs.
`Qwen3.5-4B_Recipe`	Qwen3.5-4B GRPO recipe for 1 node with 8 H100 GPUs.
`Qwen3.5-9B_Recipe`	Qwen3.5-9B GRPO recipe for 1 node with 8 H100 GPUs.
`Qwen3_0_6B_Recipe`	Qwen3-0.6B GRPO recipe for 1 node with 8 H100 GPUs.
`Qwen3_1_7B_Recipe`	Qwen3-1.7B GRPO recipe for 1 node with 8 H100 GPUs.
`Qwen3_4B_Recipe`	Qwen3-4B GRPO recipe for 1 node with 8 H100 GPUs.
`Qwen3_5_4B_Miles_Recipe`	Qwen3.5-4B GRPO recipe for 1 node with 8 H100 GPUs.
`Qwen3_6_27B_Recipe`	Qwen3.6-27B GRPO recipe for 4 nodes with 8 H100 GPUs each.
`Qwen3_6_35B_Recipe`	Qwen3.6-35B-A3B MoE recipe for 1 node with 8 H100 GPUs.
`Qwen3_8_27B_Recipe`	Qwen3.8-27B GRPO recipe for 4 nodes with 8 H100 GPUs each.
`Qwen3_8B_Recipe`	Qwen3-8B GRPO recipe for 1 node with 8 H100 GPUs.
`Qwen3_ASR_1_7B_Recipe`	Qwen3-ASR-1.7B audio GRPO recipe for 1 node with 2 H100 GPUs.
`Qwen3_VL_8B_Recipe`	Qwen3-VL-8B GRPO recipe for 1 node with 8 H100 GPUs.

Training

Name	Description
`MetricConfig`	Defines metric tracker metadata, environment variables, and links.
`MilesRecipe`	Miles training and Modal resource settings.
`ModalRayCluster`	Starts and coordinates a Ray cluster inside a Modal clustered function.
`SlimeRecipe`	Slime training and Modal resource settings.
`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.
`TrainResult`	Result metadata and checkpoints for a completed training run.
`WandbConfig`	Weights & Biases run metadata and credentials.

Deployment

Name	Description
`CustomDeployment`	A model deployed with an SGLang or vLLM recipe.
`DeepSeek_V4_Flash_SglangRecipe`	DeepSeek-V4-Flash 284B MoE SGLang recipe with 13B active parameters for 4×B200.
`Endpoint`	Controls a Modal Endpoint that persists until stopped.
`GLM_4_7_SglangRecipe`	GLM-4.7 355B-A32B MoE SGLang recipe for 8×H200.
`ParsedResponse`	Structured result of parsing raw model output.
`Qwen3_0_6B_SglangRecipe`	Qwen3-0.6B SGLang recipe for 1×H100.
`Qwen3_0_6B_VllmRecipe`	Qwen3-0.6B vLLM recipe for 1×H100.
`Qwen3_1_7B_SglangRecipe`	Qwen3-1.7B SGLang recipe for 1×H100.
`Qwen3_1_7B_VllmRecipe`	Qwen3-1.7B vLLM recipe for 1×H100.
`Qwen3_30B_SglangRecipe`	Qwen3-30B-A3B (MoE) SGLang recipe for 4×H100.
`Qwen3_30B_VllmRecipe`	Qwen3-30B-A3B (MoE) vLLM recipe for 4×H100.
`Qwen3_4B_SglangRecipe`	Qwen3-4B SGLang recipe for 1×H100.
`Qwen3_4B_VllmRecipe`	Qwen3-4B vLLM recipe for 1×H100.
`Qwen3_6_35B_SglangRecipe`	Qwen3.6-35B-A3B (MoE) SGLang recipe for 4×H100.
`Qwen3_6_35B_VllmRecipe`	Qwen3.6-35B-A3B (MoE) vLLM recipe.
`Qwen3_8B_SglangRecipe`	Qwen3-8B SGLang recipe for 1×H100.
`Qwen3_8B_VllmRecipe`	Qwen3-8B vLLM recipe for 1×H100.
`SglangRecipe`	SGLang server settings.
`VllmRecipe`	vLLM server settings.