"```python\nfrom modal_training_gym.deploy_recipes.sglang_recipe import SglangRecipe\n```\n\nSGLang serving configuration.\n\n**Inherits from:** `BaseDeployRecipe`\n\n## Fields\n\n| Field | Type | Default | Description |\n|-------|------|---------|-------------|\n| `recipe_type` | `DeployRecipeType` | `sglang` | |\n| `gpu` | `Literal['H100', 'H200', 'B200', 'B300']` | `\"H100\"` | GPU type for the serving container. Default `\"H100\"`. |\n| `tp` | `int \\| None` | `None` | Tensor parallelism degree. Default `None` (SGLang infers from GPU count). |\n| `dp` | `int \\| None` | `None` | Data parallelism degree. Emitted as `--dp-size` and enables `--enable-dp-attention` when set. Default `None`. |\n| `context_length` | `int \\| None` | `None` | Maximum context length. Default `None` (model default). |\n| `mem_fraction_static` | `float \\| None` | `None` | Fraction of GPU memory for KV cache. Default `None` (SGLang default). |\n| `chunked_prefill_size` | `int \\| None` | `None` | Chunked prefill token budget. Default `None`. |\n| `max_running_requests` | `int \\| None` | `None` | Max concurrent requests per worker. Default `None`. |\n| `sglang_image` | `str` | `\"lmsysorg/sglang:v0.5.12\"` | Docker image tag for the SGLang container. Default is a recent nightly. |\n| `extra_server_args` | `dict[str, str] \\| None` | `None` | Additional `--flag value` pairs passed to `sglang.launch_server`. Use an empty string value for boolean flags (e.g. `{\"--trust-remote-code\": \"\"}`). Default `None`. |\n| `env_vars` | `dict[str, str]` | `{}` | Extra environment variables baked into the serving image (e.g. DeepGEMM MegaMoE knobs). Merged on top of the base HF cache env. Default empty. |\n| `install_transformers_from_git` | `bool` | `True` | If `True` (default), the serve image `pip install`s transformers from GitHub so brand-new architectures (historically DeepSeek-V4 on older SGLang tags) are recognized by `AutoConfig`. Set `False` when the chosen `sglang_image` already ships a compatible transformers — otherwise the git install double-registers configs (e.g. `qwen3_asr`) and the server crashloops on import. |\n| `environment_name` | `str \\| None` | `None` | Modal environment to deploy into. Default `None`. |\n| `deploy_strategy` | `str` | `\"rolling\"` | Modal deployment strategy. Default `\"rolling\"`. |\n| `startup_timeout` | `int` | `1200` | Seconds the server container is allowed to spend in startup before Modal kills it — gates both Modal's container `startup_timeout` and the SGLang health-check poll. Bump this for very large models whose weight load exceeds the default (e.g. GLM-4.7 at 355B, Kimi-K2.5 at ~1T). Default `1200` (20 minutes). |\n\n## Methods\n\n### `server_args(self, *, served_model_name: 'str') -> 'dict[str, str]'`\n\nBuild the `--flag value` dict for the SGLang launch command.\n\n**Source:** [`modal_training_gym/deploy_recipes/sglang_recipe.py`](https://github.com/modal-projects/training-gym/blob/main/modal_training_gym/deploy_recipes/sglang_recipe.py)\n"

SglangRecipe

SGLang serving configuration.

from modal_training_gym.deploy_recipes.sglang_recipe import SglangRecipe

SGLang serving configuration.

Inherits from: BaseDeployRecipe

Fields

Field	Type	Default	Description
`recipe_type`	`DeployRecipeType`	`sglang`
`gpu`	`Literal['H100', 'H200', 'B200', 'B300']`	`"H100"`	GPU type for the serving container. Default `"H100"`.
`tp`	`int \| None`	`None`	Tensor parallelism degree. Default `None` (SGLang infers from GPU count).
`dp`	`int \| None`	`None`	Data parallelism degree. Emitted as `--dp-size` and enables `--enable-dp-attention` when set. Default `None`.
`context_length`	`int \| None`	`None`	Maximum context length. Default `None` (model default).
`mem_fraction_static`	`float \| None`	`None`	Fraction of GPU memory for KV cache. Default `None` (SGLang default).
`chunked_prefill_size`	`int \| None`	`None`	Chunked prefill token budget. Default `None`.
`max_running_requests`	`int \| None`	`None`	Max concurrent requests per worker. Default `None`.
`sglang_image`	`str`	`"lmsysorg/sglang:v0.5.12"`	Docker image tag for the SGLang container. Default is a recent nightly.
`extra_server_args`	`dict[str, str] \| None`	`None`	Additional `--flag value` pairs passed to `sglang.launch_server`. Use an empty string value for boolean flags (e.g. `{"--trust-remote-code": ""}`). Default `None`.
`env_vars`	`dict[str, str]`	`{}`	Extra environment variables baked into the serving image (e.g. DeepGEMM MegaMoE knobs). Merged on top of the base HF cache env. Default empty.
`install_transformers_from_git`	`bool`	`True`	If `True` (default), the serve image `pip install`s transformers from GitHub so brand-new architectures (historically DeepSeek-V4 on older SGLang tags) are recognized by `AutoConfig`. Set `False` when the chosen `sglang_image` already ships a compatible transformers — otherwise the git install double-registers configs (e.g. `qwen3_asr`) and the server crashloops on import.
`environment_name`	`str \| None`	`None`	Modal environment to deploy into. Default `None`.
`deploy_strategy`	`str`	`"rolling"`	Modal deployment strategy. Default `"rolling"`.
`startup_timeout`	`int`	`1200`	Seconds the server container is allowed to spend in startup before Modal kills it — gates both Modal’s container `startup_timeout` and the SGLang health-check poll. Bump this for very large models whose weight load exceeds the default (e.g. GLM-4.7 at 355B, Kimi-K2.5 at ~1T). Default `1200` (20 minutes).

Methods

`server_args(self, *, served_model_name: 'str') -> 'dict[str, str]'`

Build the --flag value dict for the SGLang launch command.

Source: modal_training_gym/deploy_recipes/sglang_recipe.py