sglang API reference
400 public APIs from sglang (sgl-project/sglang) — 199 classes, 125 functions, 76 methods. Signatures extracted by static analysis of the actual source.
Repository: sgl-project/sglang
| Kind | Count |
|---|---|
| Classes | 199 |
| Functions | 125 |
| Methods | 76 |
API list
class
python.sglang._mps_stub.EventMinimal stand-in for ``torch.cuda.Event``.
class
python.sglang._mps_stub.StreamMinimal stand-in for ``torch.cuda.Stream``.
class
python.sglang._mps_stub.StreamContextMinimal stand-in for ``torch.cuda.StreamContext``.
func
python.sglang._mps_stub.current_device() -> intReturn the index of the current MPS device (always 0).
func
python.sglang._mps_stub.current_stream(device:Any=None) -> StreamReturn the default (and only) MPS stream.
func
python.sglang._mps_stub.device_count() -> intReturn the number of available MPS devices (always 1).
func
python.sglang._mps_stub.get_device_properties(device:Any=0) -> _MPSDevicePropertiesReturn the properties of the MPS device.
func
python.sglang._mps_stub.install() -> NonePatch ``torch.mps`` with the stubs above.
func
python.sglang._mps_stub.set_device(device:Any) -> NoneSet the current device.
func
python.sglang._mps_stub.stream(s:Any) -> StreamReturn a context manager that is a no-op on MPS.
func
python.sglang._triton_stub.install() -> NoneRegister a mock ``triton`` package in *sys.modules*.
class
python.sglang.check_env.BaseEnvBase class for environment check
method
python.sglang.check_env.BaseEnv.get_info() -> dictGet CUDA-related information if available.
method
python.sglang.check_env.BaseEnv.get_package_versions() -> dictGet versions of specified packages.
class
python.sglang.check_env.GPUEnvEnvironment checker for Nvidia GPU
method
python.sglang.check_env.GPUEnv.get_topology()Get GPU topology information.
class
python.sglang.check_env.HIPEnvEnvironment checker for ROCm/HIP
class
python.sglang.check_env.MPSEnvEnvironment checker for Apple Silicon MPS
class
python.sglang.check_env.MUSAEnvEnvironment checker for MThreads GPU
method
python.sglang.check_env.MUSAEnv.get_topology()Get GPU topology information.
class
python.sglang.check_env.NPUEnvEnvironment checker for Ascend NPU
func
python.sglang.cli.utils.get_is_diffusion_model(model_path:str) -> boolDetect whether model_path points to a diffusion model.
class
python.sglang.global_config.GlobalConfigStore some global constants.
func
python.sglang.kernel_api_logging.debug_torch_op(op_func:Callable, op_name:str, *namespace:str='sglang') -> CallableNOTE: For internal use.
func
python.sglang.kernels.aot.python.sgl_kernel.elementwise.rmsnorm(input:torch.Tensor, weight:torch.Tensor, eps:float=1e-06, out:Optional[torch.Tensor]=None, enable_pdl:Optional[bool]=None) -> torch.TensorRoot mean square normalization.
func
python.sglang.kernels.aot.python.sgl_kernel.spatial.create_greenctx_stream_by_value(SM_a:int, SM_b:int, device_id:int=None) -> tuple[ExternalStream, ExternalStream]Create two streams for greenctx.
func
python.sglang.kernels.aot.python.sgl_kernel.spatial.get_sm_available(device_id:int=None) -> intGet the SMs available on the device.
func
python.sglang.kernels.jit.utils.arch.get_jit_cuda_arch() -> ArchInfoGet the current CUDA architecture info.
class
python.sglang.kernels.ops.activation.ReLU2Op``out = relu(input) ** 2`` (single-input, not gated).
func
python.sglang.kernels.ops.activation.gelu_and_mul(input:torch.Tensor, out:Optional[torch.Tensor]=None) -> torch.Tensor``out = gelu(input[..., :d]) * input[..., d:]``.
func
python.sglang.kernels.ops.activation.gelu_quick(input:torch.Tensor, out:Optional[torch.Tensor]=None) -> torch.Tensor``out = input * sigmoid(1.702 * input)``.
func
python.sglang.kernels.ops.activation.relu2(input:torch.Tensor, out:Optional[torch.Tensor]=None) -> torch.Tensor``out = relu(input) ** 2``.
func
python.sglang.kernels.ops.attention.dsa.tilelang_kernel.fp8_index(q:torch.Tensor, q_s:torch.Tensor, k:torch.Tensor, k_s:torch.Tensor) -> torch.TensorPerform index score using FP8 precision.
func
python.sglang.kernels.ops.attention.dsv4.elementwise.fused_q_indexer_rope_first_quant(q_input:torch.Tensor, weight:torch.Tensor, weight_scale:float, cos_sin_cache:torch.Tensor, positions:torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]DeepSeek-V3.2 only.
func
python.sglang.kernels.ops.attention.fla.kda.cdiv(a:int, b:int) -> intCeiling division.
func
python.sglang.kernels.ops.attention.fla.kda.next_power_of_2(n:int) -> intThe next power of 2 (inclusive)
class
python.sglang.kernels.ops.attention.flash_attn.cute.cache_utils.JITCacheIn-memory cache for compiled functions.
method
python.sglang.kernels.ops.attention.flash_attn.cute.cache_utils.JITCache.clear() -> NoneClear in-memory cache of compiled functions
func
python.sglang.kernels.ops.attention.flash_attn.cute.cache_utils.get_jit_cache(name:str | None=None) -> JITCacheJIT cache factory.
func
python.sglang.kernels.ops.attention.flash_attn.cute.cute_dsl_ptxas.patch()Install system ptxas hook.
func
python.sglang.kernels.ops.attention.flash_attn.cute.fa_logging.set_fa_log_level(level:int | str) -> NoneSet the FA log level programmatically.
class
python.sglang.kernels.ops.attention.flash_attn.cute.mask.Sm100MaskEnumEnumeration of mask types for FMHA operations.
func
python.sglang.kernels.ops.attention.flash_attn.cute.pipeline.make_pipeline_state(type:PipelineUserType, stages:int)Creates a pipeline state.
func
python.sglang.kernels.ops.attention.flash_attn.cute.softmax.floor_if_packed(q_idx, qhead_per_kvhead:cutlass.Constexpr[int]) -> cute.TensorConvert q_idx to packed format for Pack-GQA.
func
python.sglang.kernels.ops.attention.position.compute_position_triton(extend_prefix_lens:torch.Tensor, extend_seq_lens:torch.Tensor, extend_seq_lens_sum)Compute positions.
func
python.sglang.kernels.ops.diffusion.residual_gate_add(residual:torch.Tensor, update:torch.Tensor, gate:torch.Tensor) -> torch.TensorFused ``residual + gate * update``.
func
python.sglang.kernels.ops.diffusion.triton.mps_fallback.triton_one_pass_rms_norm_native(x:torch.Tensor, w:torch.Tensor, eps:float=1e-06) -> torch.TensorMLX-accelerated triton_one_pass_rms_norm.
func
python.sglang.kernels.ops.diffusion.triton.sana_wm_gdn.fused_qk_inv_rms(qkv:torch.Tensor, eps:float=1e-05) -> tuple[torch.Tensor, torch.Tensor]Single-pass Triton fused Q+K inverse-RMS.
func
python.sglang.kernels.ops.diffusion.usp_relayout.usp_merge_heads(x:torch.Tensor) -> torch.TensorMerge Ulysses output heads with an exact eager fallback.
func
python.sglang.kernels.ops.elementwise.elementwise.fused_sigmoid_mul(attn_output:torch.Tensor, gate:torch.Tensor, inplace:bool=False) -> torch.TensorFused sigmoid-mul for attention output gating.
class
python.sglang.kernels.ops.gemm.cutedsl_bf16_gemm.WorkTileInfoWhich output tile this CTA processes.
func
python.sglang.kernels.ops.gemm.dsv3_fused_a_gemm(mat_a:torch.Tensor, mat_b:torch.Tensor, output:Optional[torch.Tensor]=None) -> torch.TensorDeepSeek-V3 fused QKV-A GEMM.
func
python.sglang.kernels.ops.gemm.kv_b_lora_absorbed.step_a_q_fwd(q_nope:torch.Tensor, B_buf:torch.Tensor, batch_info:LoRABatchInfo, full_K_per_head:int) -> torch.TensorStep A of the q-side correction.
func
python.sglang.kernels.ops.gemm.kv_b_lora_absorbed.step_a_v_fwd(attn_output:torch.Tensor, A_buf:torch.Tensor, batch_info:LoRABatchInfo) -> torch.TensorStep A of the v-side correction.
func
python.sglang.kernels.ops.gemm.trtllm_lora_temp.kv_b_lora_absorbed.step_a_q_fwd(q_nope:torch.Tensor, B_buf:torch.Tensor, batch_info:LoRABatchInfo, full_K_per_head:int) -> torch.TensorStep A of the q-side correction.
func
python.sglang.kernels.ops.gemm.trtllm_lora_temp.kv_b_lora_absorbed.step_a_v_fwd(attn_output:torch.Tensor, A_buf:torch.Tensor, batch_info:LoRABatchInfo) -> torch.TensorStep A of the v-side correction.
func
python.sglang.kernels.ops.kimi_k3.all_reduce.register_comm(comm:Communicator, *pull_sem_mc_ptr:int=0) -> NoneRegister the CustomAllReduceV2 storage plane.
func
python.sglang.kernels.ops.kimi_k3.attn_res_hip.supports_attn_res_hip(hidden_size:int, nvb:int) -> boolWhether this shape fits the register budget.
func
python.sglang.kernels.ops.kimi_k3.gemm_ar.fits(x:torch.Tensor) -> boolWhether this o_proj input can take the fused GEMM+AR path.
func
python.sglang.kernels.ops.layernorm.gemma4_fused_ops.gemma4_fused_routing(gating_output:torch.Tensor, per_expert_scale:torch.Tensor, topk:int) -> tuple[torch.Tensor, torch.Tensor]One-launch Gemma4 router.
func
python.sglang.kernels.ops.lplb.cuda_solver.solve_ipm(A:torch.Tensor, b:torch.Tensor, c:torch.Tensor, num_iters:int=DEFAULT_NUM_ITERS, result:torch.Tensor | None=None) -> torch.TensorRun the fused single-SM IPM kernel.
func
python.sglang.kernels.ops.lplb.shmem_budget.assert_fits(nc:int, nv:int, gpu:str='h100') -> NoneRaise if the fused kernel will not fit on the target GPU.
func
python.sglang.kernels.ops.lplb.shmem_budget.max_nc_for_nv(nv:int, gpu:str='h100') -> intLargest NC that fits for a given NV.
func
python.sglang.kernels.ops.lplb.shmem_budget.report(nc:int, nv:int, gpu:str='h100') -> strHuman-readable summary — used by kernels on init for logging.
class
python.sglang.kernels.ops.mamba.triton_ops.ssu_dispatch.TritonSSUBackendTriton-based selective-state-update backend.
func
python.sglang.kernels.ops.moe.ep_moe_kernels.get_tma_aligned_size(x:int, element_size:int) -> intGlobal memory address of TMA must be 16-byte aligned.
func
python.sglang.kernels.ops.moe.moe_front.get_front_strategy(num_tokens:int, device) -> strReturn the measured front strategy for this exact workload.
func
python.sglang.kernels.ops.sampling.murmur_hash.fmix32(h:tl.uint32) -> tl.uint32final mix of 32-bit hash value for MurmurHash
func
python.sglang.kernels.ops.sampling.murmur_hash.murmur3_mix(h:tl.uint32, k:tl.uint32) -> tl.uint32Mixes a 32-bit key into the hash state.
func
python.sglang.kernels.ops.sampling.murmur_hash.rotl32(x, r:tl.constexpr) -> tl.uint32rotate left 32-bit integer x by r bits e.g.
func
python.sglang.kernels.ops.sampling.top_k_renorm_probs(probs:torch.Tensor, top_k:Union[torch.Tensor, int]) -> torch.TensorRenormalize ``probs`` by top-k thresholding.
func
python.sglang.kernels.ops.sampling.top_p_renorm_probs(probs:torch.Tensor, top_p:Union[torch.Tensor, float]) -> torch.TensorRenormalize ``probs`` by top-p thresholding.
func
python.sglang.kernels.registry.register_kernel(spec:KernelSpec) -> KernelSpecRegister ``spec`` in the process-wide :data:`registry`.
func
python.sglang.kernels.selector.clear_cache() -> NoneDrop the resolved-callable cache (used by tests).
func
python.sglang.kernels.selector.get_kernel(op:str, backend:Optional[KernelBackend]=None) -> CallableResolve ``op`` to a callable kernel and cache it.
class
python.sglang.kernels.spec.DeviceTypeAccelerator device family a kernel can run on.
class
python.sglang.kernels.spec.FormatSignatureA light description of a kernel's data contract.
class
python.sglang.kernels.spec.KernelSpecA single callable kernel implementation and its metadata.
method
python.sglang.kernels.spec.KernelSpec.is_available(platform:PlatformInfo) -> boolWhether this backend can run on ``platform`` (metadata-only check).
method
python.sglang.kernels.spec.KernelSpec.load() -> CallableImport and return the backing callable.
class
python.sglang.kernels.spec.PlatformInfoA minimal snapshot of the runtime accelerator platform.
func
python.sglang.kernels.spec.capabilities_satisfied(capabilities:Union[FrozenSet[CapabilityRequirement], Tuple[CapabilityRequirement, ...], CapabilityRequirement], platform:PlatformInfo) -> boolOR over ``capabilities`` (empty = unrestricted).
class
python.sglang.lang.backend.crusoe.CrusoeSGLang backend for Crusoe managed inference.
class
python.sglang.lang.backend.runtime_endpoint.RuntimeA wrapper for the HTTP server.
class
python.sglang.lang.interpreter.ProgramStateThe state of an SGL program.
class
python.sglang.multimodal_gen.apps.ComfyUI_SGLDiffusion.executors.flux.FluxExecutorExecutor for Flux models in ComfyUI.
class
python.sglang.multimodal_gen.apps.ComfyUI_SGLDiffusion.executors.zimage.ZImageExecutorExecutor for ZImage models in ComfyUI.
method
python.sglang.multimodal_gen.configs.models.base.ModelConfig.update_model_arch(source_model_dict:dict[str, Any]) -> NoneUpdate arch_config with source_model_dict
class
python.sglang.multimodal_gen.configs.models.dits.cosmos3video.Cosmos3VideoConfigDiT config wrapper for Cosmos3 Video model.
class
python.sglang.multimodal_gen.configs.models.dits.ltx_2.LTX2ConfigConfiguration for LTX-2 Video Transformer.
class
python.sglang.multimodal_gen.configs.models.encoders.qwen3.Qwen3TextArchConfigArchitecture config for Qwen3 text encoder.
class
python.sglang.multimodal_gen.configs.models.encoders.qwen3.Qwen3TextConfigTop-level config for Qwen3 text encoder.
class
python.sglang.multimodal_gen.configs.models.encoders.qwen3vl.Qwen3VLConfigConfiguration for Qwen3-VL text encoder.
class
python.sglang.multimodal_gen.configs.models.vaes.hunyuan3d.Hunyuan3DVAEArchConfigArchitecture config for Hunyuan3D VAE.
class
python.sglang.multimodal_gen.configs.models.vaes.hunyuan3d.Hunyuan3DVAEConfigVAE configuration for Hunyuan3D.
class
python.sglang.multimodal_gen.configs.models.vaes.stablediffusion3.StableDiffusion3VAEConfigConfiguration for StableDiffusion3 VAE.
class
python.sglang.multimodal_gen.configs.pipeline_configs.base.PipelineConfigThe base configuration class for a generation pipeline.
class
python.sglang.multimodal_gen.configs.pipeline_configs.base.STA_ModeSTA (Sliding Tile Attention) modes.
class
python.sglang.multimodal_gen.configs.pipeline_configs.base.SlidingTileAttnConfigConfiguration for sliding tile attention.
class
python.sglang.multimodal_gen.configs.pipeline_configs.base.SpatialImagePipelineConfigBase config for spatial image pipelines (e.g.
func
python.sglang.multimodal_gen.configs.pipeline_configs.base.parse_int_list(value:str) -> list[int]Parse a comma-separated string of integers into a list.
class
python.sglang.multimodal_gen.configs.pipeline_configs.cosmos3.Cosmos3ConfigCosmos3 unified pipeline config.
class
python.sglang.multimodal_gen.configs.pipeline_configs.flux.FluxPipelineConfigConfiguration for the FLUX pipeline.
class
python.sglang.multimodal_gen.configs.pipeline_configs.glm_image.GlmImagePipelineConfigConfiguration for the GlmImage pipeline.
class
python.sglang.multimodal_gen.configs.pipeline_configs.helios.HeliosT2VConfigConfiguration for the Helios T2V pipeline.
class
python.sglang.multimodal_gen.configs.pipeline_configs.krea2.Krea2PipelineConfigKrea-2 single-stream MMDiT, text-to-image.
class
python.sglang.multimodal_gen.configs.pipeline_configs.ltx_2.LTX23PipelineConfigConfiguration overrides for LTX-2.3.
class
python.sglang.multimodal_gen.configs.pipeline_configs.ltx_2.LTX2PipelineConfigConfiguration for LTX-Video pipeline.
class
python.sglang.multimodal_gen.configs.pipeline_configs.qwen_image.QwenImagePipelineConfigConfiguration for the QwenImage pipeline.
func
python.sglang.multimodal_gen.configs.sample.glmimage.align_glm_image_dimension(value:int) -> intRound a GLM-Image dimension up to a supported multiple.
class
python.sglang.multimodal_gen.configs.sample.joy_echo.JoyEchoSamplingParamsSampling parameters for JoyEcho DMD inference.
class
python.sglang.multimodal_gen.configs.sample.ltx_2.LTX2SamplingParamsSampling parameters for LTX-2.
class
python.sglang.multimodal_gen.configs.sample.sampling_params.SamplingParamsSampling parameters for generation.
method
python.sglang.multimodal_gen.configs.sample.sampling_params.SamplingParams.add_cli_args(parser:Any) -> AnyAdd CLI arguments for SamplingParam fields
class
python.sglang.multimodal_gen.configs.sample.sana.SanaSamplingParamsDefaults for SANA 1.5 1024px variant.
class
python.sglang.multimodal_gen.configs.sample.wan.Wan2_2_Base_SamplingParamsSampling parameters for Wan2.2 TI2V 5B model.
class
python.sglang.multimodal_gen.configs.sample.wan.Wan2_2_TI2V_5B_SamplingParamSampling parameters for Wan2.2 TI2V 5B model.
func
python.sglang.multimodal_gen.registry.get_non_diffusers_pipeline_name(model_path:str) -> Optional[str]Get the pipeline name for a known non-diffusers model.
func
python.sglang.multimodal_gen.registry.get_pipeline_config_classes(pipeline_class_name:str) -> Tuple[Type[PipelineConfig], Type[Any]] | NoneGet the configuration classes for a pipeline.
class
python.sglang.multimodal_gen.runtime.cache.cache_dit_integration.CacheDitConfigConfiguration for cache-dit integration.
class
python.sglang.multimodal_gen.runtime.cache.spectrum.SpectrumForecasterChebyshev + discrete Taylor blend forecaster.
func
python.sglang.multimodal_gen.runtime.disaggregation.roles.get_module_role(module_name:str) -> 'RoleType | None'Classify a module name to its primary role.
func
python.sglang.multimodal_gen.runtime.disaggregation.transport.protocol.encode_transfer_msg(msg:Any) -> list[bytes]Encode as [TRANSFER_MAGIC, json_payload_bytes].
func
python.sglang.multimodal_gen.runtime.distributed.group_coordinator.get_local_torch_device() -> torch.deviceReturn the torch device for the current rank.
func
python.sglang.multimodal_gen.runtime.distributed.parallel_state.destroy_model_parallel() -> NoneSet the groups to none and destroy them.
func
python.sglang.multimodal_gen.runtime.distributed.parallel_state.get_dp_rank() -> intReturn my rank for the data parallel group.
func
python.sglang.multimodal_gen.runtime.distributed.parallel_state.get_world_rank() -> intReturn my rank for the world group.
func
python.sglang.multimodal_gen.runtime.distributed.parallel_state.get_world_size() -> intReturn world size for the world group.
class
python.sglang.multimodal_gen.runtime.entrypoints.cli.cli_types.CLISubcommandBase class for CLI subcommands
method
python.sglang.multimodal_gen.runtime.entrypoints.cli.cli_types.CLISubcommand.validate(args:argparse.Namespace) -> NoneValidate the arguments for this command
func
python.sglang.multimodal_gen.runtime.entrypoints.http_server.model_info_endpoint(request:Request)Get model information.
func
python.sglang.multimodal_gen.runtime.entrypoints.http_server.server_info_endpoint(request:Request)Get server information.
class
python.sglang.multimodal_gen.runtime.entrypoints.openai.common_api.ModelCardModel cards.
func
python.sglang.multimodal_gen.runtime.entrypoints.openai.common_api.available_models()Show available models.
func
python.sglang.multimodal_gen.runtime.entrypoints.openai.common_api.retrieve_model(model:str)Retrieve a model instance.
class
python.sglang.multimodal_gen.runtime.entrypoints.openai.realtime.generate_session.GenerateSessionA realtime generation session
func
python.sglang.multimodal_gen.runtime.entrypoints.openai.utils.build_sampling_params(request_id:str, **kwargs) -> SamplingParamsBuild SamplingParams from request parameters.
func
python.sglang.multimodal_gen.runtime.launch_server.parse_url_string(url_str:str) -> list[str]Parse a semicolon-separated URL string into a list.
class
python.sglang.multimodal_gen.runtime.layers.activation.GeluAndMulAn activation function for GeGLU.
class
python.sglang.multimodal_gen.runtime.layers.activation.SiluAndMulAn activation function for SwiGLU.
method
python.sglang.multimodal_gen.runtime.layers.activation.SiluAndMul.forward_native(x:torch.Tensor) -> torch.TensorPyTorch-native implementation equivalent to forward().
func
python.sglang.multimodal_gen.runtime.layers.activation.get_act_fn(act_fn_name:str) -> nn.ModuleGet an activation function by name.
class
python.sglang.multimodal_gen.runtime.layers.attention.backends.attention_backend.AttentionBackendAbstract class for attention backends.
class
python.sglang.multimodal_gen.runtime.layers.attention.layer.LocalAttentionAttention layer.
class
python.sglang.multimodal_gen.runtime.layers.attention.turbo_layer.DistributedAttentionInitialization.
class
python.sglang.multimodal_gen.runtime.layers.custom_op.CustomOpBase class for custom ops.
class
python.sglang.multimodal_gen.runtime.layers.layernorm.RMSNormRoot mean square normalization.
class
python.sglang.multimodal_gen.runtime.layers.linear.ColumnParallelLinearLinear layer with column parallelism.
class
python.sglang.multimodal_gen.runtime.layers.linear.LinearBaseBase linear layer.
class
python.sglang.multimodal_gen.runtime.layers.linear.LinearMethodBaseBase class for different (maybe quantized) linear methods.
method
python.sglang.multimodal_gen.runtime.layers.linear.LinearMethodBase.apply(layer:torch.nn.Module, x:torch.Tensor, bias:torch.Tensor | None=None) -> torch.TensorApply the weights in layer to the input tensor.
class
python.sglang.multimodal_gen.runtime.layers.linear.MergedColumnParallelLinearPacked linear layers with column parallelism.
class
python.sglang.multimodal_gen.runtime.layers.linear.ReplicatedLinearReplicated linear layer.
class
python.sglang.multimodal_gen.runtime.layers.linear.RowParallelLinearLinear layer with row parallelism.
class
python.sglang.multimodal_gen.runtime.layers.linear.UnquantizedLinearMethodLinear method without quantization.
class
python.sglang.multimodal_gen.runtime.layers.mlp.FeedForwardA feed-forward layer.
class
python.sglang.multimodal_gen.runtime.layers.mlp.MLPMLP for DiT blocks, NO gated linear units
func
python.sglang.multimodal_gen.runtime.layers.parallel_conv.gather_height_sizes(x:torch.Tensor) -> list[int]gather heights of sharded feature_maps from peers
class
python.sglang.multimodal_gen.runtime.layers.quantization.configs.base_config.QuantizationConfigBase class for quantization configs.
class
python.sglang.multimodal_gen.runtime.layers.quantization.configs.base_config.QuantizeMethodBaseBase class for different quantized methods.
method
python.sglang.multimodal_gen.runtime.layers.quantization.configs.base_config.QuantizeMethodBase.process_weights_after_loading(layer:nn.Module) -> NoneProcess the weight after loading.
class
python.sglang.multimodal_gen.runtime.layers.quantization.fp8.Fp8ConfigConfig class for FP8.
class
python.sglang.multimodal_gen.runtime.layers.quantization.fp8.Fp8LinearMethodLinear method for FP8.
class
python.sglang.multimodal_gen.runtime.layers.quantization.modelopt_quant.ModelOptFp4ConfigConfig class for NVFP4.
class
python.sglang.multimodal_gen.runtime.layers.rotary_embedding.base.RotaryEmbeddingOriginal rotary positional embedding.
class
python.sglang.multimodal_gen.runtime.layers.rotary_embedding.mrope.NDRotaryEmbeddingN-dimensional rotary positional embedding.
class
python.sglang.multimodal_gen.runtime.layers.rotary_embedding.mrope.OneDRotaryEmbedding1D rotary positional embedding with caching.
class
python.sglang.multimodal_gen.runtime.layers.visual_embedding.ModulateProjectionModulation layer for DiT blocks.
class
python.sglang.multimodal_gen.runtime.layers.vocab_parallel_embedding.UnquantizedEmbeddingMethodUnquantized method for embeddings.
func
python.sglang.multimodal_gen.runtime.layers.vocab_parallel_embedding.pad_vocab_size(vocab_size:int, pad_to:int=DEFAULT_VOCAB_PADDING_SIZE) -> intPad the vocab size to the given value.
class
python.sglang.multimodal_gen.runtime.loader.component_loaders.component_loader.AutoProcessorLoaderLoader for auto processor.
class
python.sglang.multimodal_gen.runtime.loader.component_loaders.component_loader.ImageProcessorLoaderLoader for image processor.
class
python.sglang.multimodal_gen.runtime.loader.component_loaders.component_loader.TokenizerLoaderLoader for tokenizers.
class
python.sglang.multimodal_gen.runtime.loader.component_loaders.scheduler_loader.SchedulerLoaderLoader for scheduler.
class
python.sglang.multimodal_gen.runtime.loader.component_loaders.text_encoder_loader.TextEncoderLoaderLoader for text encoders.
class
python.sglang.multimodal_gen.runtime.loader.component_loaders.text_encoder_loader.TextEncoderLoader.SourceA source for weights.
func
python.sglang.multimodal_gen.runtime.loader.weight_utils.default_weight_loader(param:torch.Tensor, loaded_weight:torch.Tensor) -> NoneDefault weight loader.
func
python.sglang.multimodal_gen.runtime.loader.weight_utils.maybe_remap_kv_scale_name(name:str, params_dict:dict) -> str | NoneRemap the name of FP8 k/v_scale parameters.
func
python.sglang.multimodal_gen.runtime.managers.forward_context.get_forward_context() -> 'ForwardContext'Get the current forward context.
class
python.sglang.multimodal_gen.runtime.managers.gpu_worker.GPUWorkerA worker that executes the model on a single GPU.
method
python.sglang.multimodal_gen.runtime.managers.gpu_worker.GPUWorker.execute_forward(batch:List[Req], return_req:bool=False) -> OutputBatch | ReqExecute a forward pass.
method
python.sglang.multimodal_gen.runtime.managers.gpu_worker.GPUWorker.merge_lora_weights(target:str='all', strength:float=1.0) -> OutputBatchMerge LoRA weights.
method
python.sglang.multimodal_gen.runtime.managers.gpu_worker.GPUWorker.release_realtime_session(session_id:str) -> OutputBatchrelease the session of a realtime connection
method
python.sglang.multimodal_gen.runtime.managers.gpu_worker.GPUWorker.unmerge_lora_weights(target:str='all') -> OutputBatchUnmerge LoRA weights.
class
python.sglang.multimodal_gen.runtime.managers.scheduler.SchedulerRuns the main event loop for the rank 0 worker.
class
python.sglang.multimodal_gen.runtime.models.bridges.mova_dual_tower.PerFrameAttentionPoolingPer-frame multi-head attention pooling.
method
python.sglang.multimodal_gen.runtime.models.bridges.mova_dual_tower.PerFrameAttentionPooling.forward(x:torch.Tensor, grid_size:Tuple[int, int, int]) -> torch.TensorForward pass.
func
python.sglang.multimodal_gen.runtime.models.dits.common.modulate(x:torch.Tensor, shift:torch.Tensor | None=None, scale:torch.Tensor | None=None) -> torch.TensorModulate by shift and scale.
class
python.sglang.multimodal_gen.runtime.models.dits.cosmos3video.Cosmos3GatedMLPGated MLP (SwiGLU-style) for Cosmos3.
class
python.sglang.multimodal_gen.runtime.models.dits.cosmos3video.Cosmos3OmniTransformerCosmos3 Omni transformer.
class
python.sglang.multimodal_gen.runtime.models.dits.flux.FluxTransformer2DModelThe Transformer model introduced in Flux.
class
python.sglang.multimodal_gen.runtime.models.dits.flux_2.Flux2Transformer2DModelThe Transformer model introduced in Flux 2.
class
python.sglang.multimodal_gen.runtime.models.dits.glm_image.GlmImageKVCacheContainer for all layers' KV caches.
class
python.sglang.multimodal_gen.runtime.models.dits.glm_image.GlmImageLayerKVCacheKV cache for GlmImage model.
class
python.sglang.multimodal_gen.runtime.models.dits.joy_image.ModulateWanModulation layer for WanX.
func
python.sglang.multimodal_gen.runtime.models.dits.joy_image.fused_add_gate(residual:torch.Tensor, x:torch.Tensor, gate:torch.Tensor) -> torch.TensorFused residual addition with gate.
class
python.sglang.multimodal_gen.runtime.models.dits.mova_audio_dit.Conv1dLocalIslandInherits from Conv1d and overrides forward.
class
python.sglang.multimodal_gen.runtime.models.dits.mova_video_dit.CrossAttentionCross-Attention module for MOVA DiT.
class
python.sglang.multimodal_gen.runtime.models.dits.qwen_image.QwenImageTransformer2DModelThe Transformer model introduced in Qwen.
class
python.sglang.multimodal_gen.runtime.models.dits.sana.GLUMBConvGated Linear Unit with Multi-Branch Convolution.
class
python.sglang.multimodal_gen.runtime.models.dits.sana_wm.SanaWMBlockOne transformer block of SANA-WM.
class
python.sglang.multimodal_gen.runtime.models.dits.sana_wm.SanaWMTransformer3DModelSANA-WM 2.6B TI2V world model.
class
python.sglang.multimodal_gen.runtime.models.dits.sana_wm_components.TimestepEmbedderUpstream ``TimestepEmbedder``.
class
python.sglang.multimodal_gen.runtime.models.dits.sana_wm_refiner_transformer.SanaWMRefinerBlockVideo-only LTX-2 transformer block.
class
python.sglang.multimodal_gen.runtime.models.encoders.gemma2.Gemma2ModelGemma2 text encoder model for SANA pipeline.
class
python.sglang.multimodal_gen.runtime.models.encoders.gemma_3.Gemma3MultiModalProjectorProjector for Gemma3 multimodal.
class
python.sglang.multimodal_gen.runtime.models.encoders.qwen3.Qwen3DecoderLayerQwen3 transformer decoder layer.
class
python.sglang.multimodal_gen.runtime.models.encoders.t5.AttentionTypeAttention type.
class
python.sglang.multimodal_gen.runtime.models.parameter.BasevLLMParameterBase parameter for vLLM linear layers.
class
python.sglang.multimodal_gen.runtime.models.parameter.ModelWeightParameterParameter class for linear layer weights.
class
python.sglang.multimodal_gen.runtime.models.schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteSchedulerEuler scheduler.
class
python.sglang.multimodal_gen.runtime.models.vaes.autoencoder_kl_qwenimage.QwenImageDecoder3dA 3D decoder module.
class
python.sglang.multimodal_gen.runtime.models.vaes.autoencoder_kl_qwenimage.QwenImageEncoder3dA 3D encoder module.
class
python.sglang.multimodal_gen.runtime.models.vaes.autoencoder_kl_qwenimage.QwenImageRMS_normA custom RMS normalization layer.
class
python.sglang.multimodal_gen.runtime.models.vaes.autoencoder_kl_qwenimage.QwenImageResidualBlockA custom residual block module.
class
python.sglang.multimodal_gen.runtime.models.vaes.hunyuan3d_vae.Latent2MeshOutputContainer for mesh output from VAE decoder.
class
python.sglang.multimodal_gen.runtime.models.vaes.hunyuan3d_vae.MCSurfaceExtractorMarching Cubes surface extractor.
class
python.sglang.multimodal_gen.runtime.models.vaes.hunyuan3d_vae.SurfaceExtractorBase class for surface extraction algorithms.
class
python.sglang.multimodal_gen.runtime.models.vaes.hunyuan3d_vae.VectsetVAEBase VAE class for vector set encoding.
method
python.sglang.multimodal_gen.runtime.models.vaes.hunyuan3d_vae.VectsetVAE.latents2mesh(latents:torch.FloatTensor, **kwargs)Convert latents to mesh.
class
python.sglang.multimodal_gen.runtime.models.vaes.ltx_2_audio.LTX2AudioAudioPatchifierPatchifier for spectrogram/audio latents.
class
python.sglang.multimodal_gen.runtime.models.vaes.ltx_2_vae.AutoencoderKLCausalLTX2VideoStreaming causal LTX-2 VAE.
class
python.sglang.multimodal_gen.runtime.models.vaes.ltx_2_vae.LTX2VideoDownBlock3DDown block used in the LTXVideo model.
class
python.sglang.multimodal_gen.runtime.models.vaes.ltx_2_vae.LTX2VideoMidBlock3dA middle block used in the LTXVideo model.
class
python.sglang.multimodal_gen.runtime.models.vaes.ltx_2_vae.LTX2VideoUpBlock3dUp block used in the LTXVideo model.
func
python.sglang.multimodal_gen.runtime.models.vaes.minimax_h3_audio_vae.alias_free.sinc(x:torch.Tensor)Implementation of sinc, i.e.
class
python.sglang.multimodal_gen.runtime.models.vaes.wanvae.WanAttentionBlockCausal self-attention with a single head.
class
python.sglang.multimodal_gen.runtime.models.vaes.wanvae.WanDecoder3dA 3D decoder module.
class
python.sglang.multimodal_gen.runtime.models.vaes.wanvae.WanEncoder3dA 3D encoder module.
class
python.sglang.multimodal_gen.runtime.models.vaes.wanvae.WanMidBlockMiddle block for WanVAE encoder and decoder.
class
python.sglang.multimodal_gen.runtime.models.vaes.wanvae.WanRMS_normA custom RMS normalization layer.
class
python.sglang.multimodal_gen.runtime.models.vaes.wanvae.WanResampleA custom resampling module for 2D and 3D data.
class
python.sglang.multimodal_gen.runtime.models.vaes.wanvae.WanResidualBlockA custom residual block module.
class
python.sglang.multimodal_gen.runtime.pipelines.hunyuan3d_pipeline.Hunyuan3D2PipelineHunyuan3D 2.0 image-to-mesh pipeline.
class
python.sglang.multimodal_gen.runtime.pipelines.ltx_2_pipeline.LTX2TwoStageResidencyControllerLTX-2.3 two-stage residency controller.
class
python.sglang.multimodal_gen.runtime.pipelines.mova_pipeline.MOVAPipelineMOVA pipeline with SGLang stage orchestration.
method
python.sglang.multimodal_gen.runtime.pipelines.mova_pipeline.MOVAPipeline.initialize_pipeline(server_args:ServerArgs) -> NoneInitialize the pipeline.
class
python.sglang.multimodal_gen.runtime.pipelines.sana_wm_pipeline.SanaWMPipelineSANA-WM TI2V pipeline (single-stage).
class
python.sglang.multimodal_gen.runtime.pipelines.sana_wm_realtime_pipeline.SanaWMRealtimePipelineSANA-WM realtime interactive pipeline.
class
python.sglang.multimodal_gen.runtime.pipelines.wan_pipeline.WanPipelineWan video diffusion pipeline with LoRA support.
func
python.sglang.multimodal_gen.runtime.pipelines_core.build_pipeline(server_args:ServerArgs) -> PipelineWithLoRAOnly works with valid hf diffusers configs.
class
python.sglang.multimodal_gen.runtime.pipelines_core.schedule_batch.OutputBatchFinal output (after pipeline completion)
class
python.sglang.multimodal_gen.runtime.pipelines_core.stages.base.PipelineStageAbstract base class for all pipeline stages.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.base.PipelineStage.device() -> torch.deviceGet the device for this stage.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.base.PipelineStage.forward(batch:Req, server_args:ServerArgs) -> ReqForward pass of the stage's processing.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.base.PipelineStage.verify_input(batch:Req, server_args:ServerArgs) -> VerificationResultVerify the input for the stage.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.base.PipelineStage.verify_output(batch:Req, server_args:ServerArgs) -> VerificationResultVerify the output for the stage.
class
python.sglang.multimodal_gen.runtime.pipelines_core.stages.causal_denoising.CausalDMDDenoisingStageDenoising stage for causal diffusion.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.causal_denoising.CausalDMDDenoisingStage.verify_input(batch:Req, server_args:ServerArgs) -> VerificationResultVerify denoising stage inputs.
class
python.sglang.multimodal_gen.runtime.pipelines_core.stages.denoising_dmd.DmdDenoisingStageDenoising stage for DMD.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.denoising_dmd.DmdDenoisingStage.forward(batch:Req, server_args:ServerArgs) -> ReqRun the denoising loop.
class
python.sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.cosmos3.Cosmos3DecodingStageVAE decoding stage for Cosmos3.
class
python.sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.cosmos3.Cosmos3TokenizationStageTokenization stage for Cosmos3.
class
python.sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.hunyuan3d.paint.DDIMSolverDDIM solver for fast sampling.
class
python.sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.minimax_h3.task_profiles.MiniMaxH3TaskProfileOne row of the task table.
class
python.sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.mova.MOVADenoisingStageRun MOVA dual-tower denoising loop.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.mova.MOVADenoisingStage.verify_input(batch:Req, server_args:ServerArgs) -> VerificationResultVerify denoising stage inputs.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.mova.MOVADenoisingStage.verify_output(batch:Req, server_args:ServerArgs) -> VerificationResultVerify denoising stage outputs.
func
python.sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.sana_wm.parity_probe.dump_obj(dirpath:Path | str | None, name:str, obj) -> NoneSave a picklable object (e.g.
func
python.sglang.multimodal_gen.runtime.pipelines_core.stages.progressive_resolution.spectral_ops.dct_1d(x:torch.Tensor, norm:str='ortho') -> torch.Tensor1-D DCT-II via torch.fft.
class
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidatorsCommon validators for pipeline stages.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.bool_value(value:Any) -> boolCheck if value is a boolean.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.divisible_by(value:Any, divisor:int) -> boolCheck if value is divisible by divisor.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.is_list(value:Any) -> boolCheck if value is a list (can be empty).
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.is_tuple(value:Any) -> boolCheck if value is a tuple.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.list_length(value:Any, length:int) -> boolCheck if list has specific length.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.list_not_empty(value:Any) -> boolCheck if value is a non-empty list.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.non_negative_float(value:Any) -> boolCheck if value is a non-negative float.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.non_negative_int(value:Any) -> boolCheck if value is a non-negative float.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.none_or_list(value:Any) -> boolCheck if value is None or a list.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.none_or_positive_int(value:Any) -> boolCheck if value is None or a positive integer.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.not_none(value:Any) -> boolCheck if value is not None.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.positive_float(value:Any) -> boolCheck if value is a positive float.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.positive_int(value:Any) -> boolCheck if value is a positive integer.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.string_not_empty(value:Any) -> boolCheck if value is a non-empty string.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.StageValidators.string_not_none(value:Any) -> boolCheck if value is a non-empty string.
class
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.ValidationFailureDetails about a specific validation failure.
class
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.VerificationResultWrapper class for stage verification results.
method
python.sglang.multimodal_gen.runtime.pipelines_core.stages.validators.VerificationResult.is_valid() -> boolCheck if all validations passed.
func
python.sglang.multimodal_gen.runtime.platforms.cpu_platform_plugin() -> str | NoneDetect if CPU platform should be used.
method
python.sglang.multimodal_gen.runtime.platforms.interface.Platform.get_current_memory_usage(device:torch.types.Device | None=None) -> floatReturn the memory usage in bytes.
method
python.sglang.multimodal_gen.runtime.platforms.interface.Platform.get_device_name(device_id:int=0) -> strGet the name of a device.
method
python.sglang.multimodal_gen.runtime.platforms.interface.Platform.get_device_total_memory(device_id:int=0) -> intGet the total memory of a device in bytes.
method
python.sglang.multimodal_gen.runtime.platforms.interface.Platform.get_device_uuid(device_id:int=0) -> strGet the uuid of a device, e.g.
method
python.sglang.multimodal_gen.runtime.platforms.interface.Platform.seed_everything(seed:int | None=None) -> NoneSet the seed of each random module.
func
python.sglang.multimodal_gen.runtime.platforms.xpu_platform_plugin() -> str | NoneDetect if Intel XPU platform is available.
class
python.sglang.multimodal_gen.runtime.postprocess.realesrgan_upscaler.ImageUpscalerLazy-loaded Real-ESRGAN upscaler.
method
python.sglang.multimodal_gen.runtime.postprocess.realesrgan_upscaler.ImageUpscaler.upscale(frames:list[np.ndarray]) -> list[np.ndarray]Upscale a list of HWC uint8 frames.
class
python.sglang.multimodal_gen.runtime.postprocess.realesrgan_upscaler.RRDBResidual in Residual Dense Block.
method
python.sglang.multimodal_gen.runtime.postprocess.rife_interpolator.Model.load_model(path:str, strip_module_prefix:bool=True) -> NoneLoad weights from {path}/flownet.pkl.
func
python.sglang.multimodal_gen.runtime.postprocess.rife_interpolator.warp(tenInput:torch.Tensor, tenFlow:torch.Tensor) -> torch.TensorWarp tenInput by tenFlow using grid_sample.
class
python.sglang.multimodal_gen.runtime.realtime.session.BaseRealtimeStateper-session state owned by pipeline stages
class
python.sglang.multimodal_gen.runtime.server_args.server_args.BackendEnumeration for different model backends.
method
python.sglang.multimodal_gen.runtime.server_args.server_args.Backend.from_string(value:str) -> 'Backend'Convert string to Backend enum.
func
python.sglang.multimodal_gen.runtime.server_warmup.get_first_generation_req(req_or_group:Any) -> Req | NoneExtract the first req
func
python.sglang.multimodal_gen.runtime.server_warmup.is_realtime_serving(server_args:ServerArgs) -> boolSynthetic warmup has no realtime session state.
func
python.sglang.multimodal_gen.runtime.server_warmup.is_warmup_req(req_or_group:Any) -> booleither server-based or req-based
func
python.sglang.multimodal_gen.runtime.utils.common.add_prefix(name:str, prefix:str) -> strAdd a weight path prefix to a module name.
func
python.sglang.multimodal_gen.runtime.utils.hf_diffusers_utils.check_gguf_file(model:str | os.PathLike) -> boolCheck if the file is a GGUF model.
method
python.sglang.multimodal_gen.runtime.utils.logging_utils.ColoredFormatter.format(record:logging.LogRecord) -> strAdds color to the log
class
python.sglang.multimodal_gen.runtime.utils.mesh3d_utils.ImageProcessorV2Image processor for Hunyuan3D single-view input.
class
python.sglang.multimodal_gen.runtime.utils.mesh3d_utils.MVImageProcessorV2Multi-view image processor for Hunyuan3D.
func
python.sglang.multimodal_gen.runtime.utils.mesh3d_utils.export_to_trimesh(mesh_output:Any) -> AnyConvert mesh output to trimesh format.
func
python.sglang.multimodal_gen.runtime.utils.mesh3d_utils.get_perspective_projection_matrix(fovy:float, aspect_wh:float, near:float, far:float) -> np.ndarrayCompute perspective projection matrix.
func
python.sglang.multimodal_gen.runtime.utils.mesh3d_utils.mesh_uv_wrap(mesh:Any) -> AnyApply UV unwrapping to mesh.
func
python.sglang.multimodal_gen.runtime.utils.mesh3d_utils.stride_from_shape(shape:Tuple[int, ...]) -> List[int]Compute stride from shape for scatter operations.
func
python.sglang.multimodal_gen.runtime.utils.mesh3d_utils.transform_pos(mtx:Union[np.ndarray, torch.Tensor], pos:torch.Tensor, keepdim:bool=False) -> torch.TensorTransform positions by a matrix.
class
python.sglang.multimodal_gen.runtime.utils.trace_wrapper.DiffStageNamed trace stages for the diffusion pipeline.
func
python.sglang.multimodal_gen.runtime.utils.vision.load_image(image:str | bytes | PIL.Image.Image, convert_method:Callable[[PIL.Image.Image], PIL.Image.Image] | None=None) -> PIL.Image.ImageLoads `image` to a PIL Image.
func
python.sglang.multimodal_gen.runtime.utils.vision.normalize(images:np.ndarray | torch.Tensor) -> np.ndarray | torch.TensorNormalize an image array to [-1,1].
func
python.sglang.multimodal_gen.runtime.utils.vision.numpy_to_pt(images:np.ndarray) -> torch.TensorConvert a NumPy image to a PyTorch tensor.
func
python.sglang.multimodal_gen.runtime.utils.vision.resize(image:PIL.Image.Image | np.ndarray | torch.Tensor, height:int, width:int, resize_mode:str='default', resample:str='lanczos') -> PIL.Image.Image | np.ndarray | torch.TensorResize image.
func
python.sglang.multimodal_gen.runtime.vla.observation.tensor_fingerprint(tensor:torch.Tensor) -> strHash tensor content with SRT's CPU/CUDA implementation.
func
python.sglang.multimodal_gen.utils.get_mixed_precision_state() -> MixedPrecisionStateGet the current mixed precision state.
func
python.sglang.multimodal_gen.utils.resolve_obj_by_qualname(qualname:str) -> AnyResolve an object by its fully qualified name.
func
python.sglang.multimodal_gen.utils.warn_for_unimplemented_methods(cls:type[T]) -> type[T]A replacement for `abc.ABC`.
func
python.sglang.srt.arg_groups.deepseek_v4_hook.validate_deepseek_v4_cp(server_args:ServerArgs) -> NoneValidate DeepSeek V4 context-parallel configuration.
func
python.sglang.srt.arg_groups.overrides.run_post_process_pass(server_args:Any, fn:Callable[..., dict]) -> NoneInvoke one pass at its legacy handler slot.
func
python.sglang.srt.batch_invariant_ops.batch_invariant_ops.log_softmax(input:torch.Tensor, dim:int=-1) -> torch.TensorCompute log_softmax using Triton kernel.
func
python.sglang.srt.batch_invariant_ops.batch_invariant_ops.rms_norm(input:torch.Tensor, weight:torch.Tensor, eps:float=1e-06) -> torch.TensorCompute RMS normalization using Triton kernel.
func
python.sglang.srt.compilation.backend.set_model_tag(tag:str)Context manager to set the model tag.
func
python.sglang.srt.compilation.inductor_pass.get_pass_context() -> PassContextGet the current pass context.
class
python.sglang.srt.compilation.pass_manager.PostGradPassManagerThe pass manager for post-grad passes.
class
python.sglang.srt.configs.dbrx.DbrxAttentionConfigConfiguration class for Dbrx Attention.
class
python.sglang.srt.configs.dbrx.DbrxConfigConfiguration class for Dbrx.
class
python.sglang.srt.configs.dbrx.DbrxFFNConfigConfiguration class for Dbrx FFN.
class
python.sglang.srt.configs.embedding_model_spec.EmbeddingExecutionThe model path that produces an embedding.
class
python.sglang.srt.configs.internvl.InternLM2TokenizerConstruct a InternLM2 tokenizer.
method
python.sglang.srt.configs.internvl.InternLM2Tokenizer.get_vocab()Returns vocab as a dict
method
python.sglang.srt.configs.internvl.InternLM2Tokenizer.vocab_size()Returns vocab size
class
python.sglang.srt.configs.kimi_k25.KimiK25ConfigK2-VL model configuration.
method
python.sglang.srt.configs.kimi_k25.KimiK25Config.vocab_size() -> intGet vocab size from text config for compatibility.
class
python.sglang.srt.configs.lfm2.Lfm2ConfigSGLang configuration for LFM2 models.
method
python.sglang.srt.configs.lfm2.Lfm2Config.mamba_chunk_size() -> intReturn chunk size for Mamba2 backend.
class
python.sglang.srt.configs.lfm2_vl.Lfm2VlConfigSGLang configuration for LFM2-VL models.
method
python.sglang.srt.configs.lfm2_vl.Lfm2VlConfig.mamba_chunk_size() -> intReturn chunk size for Mamba2 backend.
method
python.sglang.srt.configs.model_config.ModelConfig.get_num_kv_heads(tensor_parallel_size) -> intReturns the number of KV heads per GPU.
method
python.sglang.srt.configs.model_config.ModelConfig.get_total_num_kv_heads() -> intReturns the total number of KV heads.
class
python.sglang.srt.connector.azure.AzureBlobConnectorFile connector for Azure Blob Storage.
method
python.sglang.srt.connector.serde.serde.Deserializer.from_bytes(bs:bytes) -> torch.TensorDeserialize a pytorch tensor from bytes.
method
python.sglang.srt.connector.serde.serde.Serializer.to_bytes(t:torch.Tensor) -> bytesSerialize a pytorch tensor to bytes.
func
python.sglang.srt.connector.utils.parse_model_name(url:str) -> strParse the model name from the url.
class
python.sglang.srt.debug_utils.comparator.aligner.token_aligner.smart.types.TokenAlignerPlanToken alignment plan.
func
python.sglang.srt.debug_utils.comparator.dims_spec.tensor_naming.find_dim_index(dim_specs:list[DimSpec], name:str) -> Optional[int]Find index by name.
func
python.sglang.srt.debug_utils.comparator.dims_spec.tensor_naming.get_dim_names(tensor:torch.Tensor) -> tuple[Optional[str], ...]Get dimension names attached to a tensor.
method
python.sglang.srt.debug_utils.comparator.dims_spec.types.DimSpec.sub_dims() -> list[str]Sub-dim names.
func
python.sglang.srt.debug_utils.comparator.entrypoint.parse_args(argv:list[str]) -> argparse.NamespaceParse CLI arguments from an argv list.
class
python.sglang.srt.debug_utils.comparator.report_sink.ReportSinkUnified entry point for all record output.
func
python.sglang.srt.debug_utils.comparator.utils.calc_per_token_rel_diff(x:torch.Tensor, y:torch.Tensor, *seq_dim:int) -> torch.TensorCosine-distance-like metric per token position.
class
python.sglang.srt.debug_utils.source_patcher.types.PatchApplicationErrormatch text not found or not unique in source.
class
python.sglang.srt.disaggregation.base.conn.BaseKVManagerBase class for managing transfer states
class
python.sglang.srt.disaggregation.common.staging_handler.DecodeStagingContextStaging-specific context for decode mode.
method
python.sglang.srt.disaggregation.common.staging_handler.DecodeStagingHandler.create(kv_manager, scheduler, tp_rank:int) -> DecodeStagingHandlerFactory: create handler.
method
python.sglang.srt.disaggregation.common.staging_handler.DecodeStagingHandler.is_staging_room(room:int) -> boolCheck if a room is registered for staging scatter.
method
python.sglang.srt.disaggregation.common.staging_handler.DecodeStagingHandler.submit_last_scatter_async(room:int) -> boolRecord all-ranks Success.
class
python.sglang.srt.disaggregation.common.staging_handler.PrefillStagingContextStaging-specific context for prefill mode.
class
python.sglang.srt.disaggregation.decode.DecodePreallocQueueStore the requests that are preallocating.
method
python.sglang.srt.disaggregation.decode.DecodePreallocQueue.add(req:Req, is_retracted:bool=False, is_rebootstrap:bool=False) -> NoneAdd a request to the pending queue.
method
python.sglang.srt.disaggregation.decode.DecodePreallocQueue.extend(reqs:List[Req], is_retracted:bool=False) -> NoneAdd a request to the pending queue.
class
python.sglang.srt.disaggregation.decode.DecodeTransferQueueStore the requests that is polling kv
class
python.sglang.srt.disaggregation.encode_grpc_server.SGLangEncoderServergRPC service implementation for SGLang encoder.
func
python.sglang.srt.disaggregation.encode_receiver.extract_original_req_id(part_req_id:str) -> strExtract the original request ID from a part request ID.
func
python.sglang.srt.disaggregation.encode_receiver.video_meta_attrs_for(model_type:Optional[str]) -> tupleVideo-meta attrs for model_type.
class
python.sglang.srt.disaggregation.kv_events.KVCacheEventBase class for all KV cache-related events
class
python.sglang.srt.disaggregation.kv_events.KVEventsConfigConfiguration for KV event publishing.
method
python.sglang.srt.disaggregation.kv_events.KVEventsConfig.from_cli(cli_value:str) -> 'KVEventsConfig'Parse the CLI value for the event publisher config.
class
python.sglang.srt.disaggregation.kv_events.NullEventPublisherNo-op implementation (default when disabled).
class
python.sglang.srt.disaggregation.kv_events.StorageMediumStorage tier for KV cache events.
class
python.sglang.srt.disaggregation.nixl.conn.TransferInfoContains indices for a transfer, sent by KVReceiver.
class
python.sglang.srt.disaggregation.nixl.conn.TransferStatusUsed by KV Receiver to know when a transfer is done.
class
python.sglang.srt.disaggregation.prefill.PrefillBootstrapQueueStore the requests in bootstrapping
method
python.sglang.srt.disaggregation.prefill.PrefillBootstrapQueue.finalize_bootstrap(req:Req) -> boolInitialize the sender after bootstrap completes.
func
python.sglang.srt.disaggregation.prefill.should_force_retry(req:Req) -> boolTest hook to force a request into optimistic prefill retry.
func
python.sglang.srt.distributed.device_communicators.custom_all_reduce_utils.gpu_p2p_access_check(src:int, tgt:int) -> boolCheck if GPU src can access GPU tgt.
func
python.sglang.srt.distributed.utils.get_global_tcp_store() -> Optional[TCPStore]Get the existing global TCPStore.
func
python.sglang.srt.distributed.utils.get_pp_indices(num_hidden_layers:int, pp_rank:int, pp_size:int) -> Tuple[int, int]Try to evenly distribute layers across partitions.
class
python.sglang.srt.dllm.mixin.scheduler.DllmManagerManager for Diffusion LLM request scheduling.
method
python.sglang.srt.dllm.mixin.scheduler.DllmManager.add_waiting_reqs(reqs:Union[Req, List[Req]]) -> NoneAdd requests to waiting queue with redundancy check.
method
python.sglang.srt.dllm.mixin.scheduler.DllmManager.any_staging_reqs() -> boolCheck if there are requests in staging queue.
method
python.sglang.srt.dllm.mixin.scheduler.DllmManager.filter_finished_reqs() -> NoneRemove finished requests from both queues.
method
python.sglang.srt.dllm.mixin.scheduler.DllmManager.get_decode_requests() -> List[Req]Get all decode requests from waiting queue.
method
python.sglang.srt.dllm.mixin.scheduler.DllmManager.get_prefill_requests() -> List[Req]Get all prefill requests from waiting queue.
func
python.sglang.srt.elastic_ep.elastic_ep.try_admit_scale_ranks(global_ranks:List[int]) -> boolAdmit append-only ranks into the expandable WORLD group.
class
python.sglang.srt.entrypoints.anthropic.protocol.AnthropicBashToolAnthropic ``bash_*`` server tool family.
class
python.sglang.srt.entrypoints.anthropic.protocol.AnthropicComputerToolAnthropic ``computer_*`` server tool family.
class
python.sglang.srt.entrypoints.anthropic.protocol.AnthropicCountTokensRequestAnthropic count_tokens API request.
class
python.sglang.srt.entrypoints.anthropic.protocol.AnthropicCountTokensResponseAnthropic count_tokens API response.
class
python.sglang.srt.entrypoints.anthropic.protocol.AnthropicErrorError structure for Anthropic API.
class
python.sglang.srt.entrypoints.anthropic.protocol.AnthropicErrorResponseError response structure for Anthropic API.
class
python.sglang.srt.entrypoints.anthropic.protocol.AnthropicMessageEndDeltaDelta carried on ``message_delta`` events.
class
python.sglang.srt.entrypoints.anthropic.protocol.AnthropicMessagesRequestAnthropic Messages API request.
class
python.sglang.srt.entrypoints.anthropic.protocol.AnthropicMessagesResponseAnthropic Messages API response.
class
python.sglang.srt.entrypoints.anthropic.protocol.AnthropicOutputConfigClaude 4.7 ``output_config`` block.
class
python.sglang.srt.entrypoints.anthropic.protocol.AnthropicToolChoiceTool choice strategy.
class
python.sglang.srt.entrypoints.anthropic.protocol.AnthropicUsageToken usage information.
class
python.sglang.srt.entrypoints.anthropic.protocol.AnthropicWebSearchToolAnthropic ``web_search_*`` server tool family.
class
python.sglang.srt.entrypoints.anthropic.serving.AnthropicServingHandler for Anthropic Messages API requests.
func
python.sglang.srt.entrypoints.elastic_ep.scale_elastic_ep(raw_request:Request)Request an asynchronous EP scale-up.
class
python.sglang.srt.entrypoints.engine.EngineThe entry point to the inference engine.
method
python.sglang.srt.entrypoints.engine.Engine.close_session(session_id:str) -> NoneClose a session and release its resources.
About this data
These signatures were extracted from the public source of sgl-project/sglang
using Python's ast module. Argument names, default values,
type annotations and return types are taken verbatim from the code.
Implementation bodies are never stored. See
how it works for details.