Files
sglang/python/sglang/multimodal_gen/runtime/server_args.py
T

1970 lines
74 KiB
Python

# Copied and adapted from: https://github.com/hao-ai-lab/FastVideo
# SPDX-License-Identifier: Apache-2.0
# Inspired by SGLang: https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/server_args.py
"""The arguments of sglang-diffusion Inference."""
import argparse
import dataclasses
import json
import math
import os
import random
import sys
import tempfile
from dataclasses import field
from enum import Enum
from typing import Any, Optional
import addict
import yaml
from sglang.multimodal_gen import envs
from sglang.multimodal_gen.configs.models.encoders import T5Config
from sglang.multimodal_gen.configs.pipeline_configs.base import PipelineConfig
from sglang.multimodal_gen.configs.pipeline_configs.ltx_2 import (
LTX2PipelineConfig,
is_ltx23_native_variant,
)
from sglang.multimodal_gen.configs.quantization.nunchaku import NunchakuSVDQuantArgs
from sglang.multimodal_gen.runtime.disaggregation.disagg_args import (
DisaggArgsMixin,
add_disagg_cli_args,
convert_disagg_role_string,
)
from sglang.multimodal_gen.runtime.disaggregation.roles import RoleType
from sglang.multimodal_gen.runtime.layers.quantization.configs.nunchaku_config import (
NunchakuConfig,
)
from sglang.multimodal_gen.runtime.loader.utils import BYTES_PER_GB
from sglang.multimodal_gen.runtime.managers.memory_managers.layerwise_offload_components import (
LAYERWISE_OFFLOAD_ALL_COMPONENTS,
LAYERWISE_OFFLOAD_DIT_GROUP,
cpu_offload_flags_for_layerwise_components,
layerwise_component_matches_any_selection,
normalize_layerwise_offload_components,
)
from sglang.multimodal_gen.runtime.platforms import (
AttentionBackendEnum,
current_platform,
)
from sglang.multimodal_gen.runtime.server_args_auto_tune import (
PERFORMANCE_MODES,
ServerArgsAutoTuner,
)
from sglang.multimodal_gen.runtime.utils.common import (
is_port_available,
is_valid_ipv6_address,
)
from sglang.multimodal_gen.runtime.utils.logging_utils import (
CYAN,
GREEN,
RED,
RESET,
_sanitize_for_logging,
configure_logger,
init_logger,
)
from sglang.multimodal_gen.utils import (
FlexibleArgumentParser,
StoreBoolean,
expand_path_fields,
)
logger = init_logger(__name__)
LTX2_TWO_STAGE_DEVICE_MODES = ("original", "snapshot", "resident")
LTX2_TWO_STAGE_PIPELINE_NAMES = ("LTX2TwoStagePipeline", "LTX2TwoStageHQPipeline")
# H200-class GPUs (>=130 GiB total) can usually keep both LTX2 DiTs resident.
LTX2_RESIDENT_AUTO_ENABLE_MEM_GB = 130
LORA_MERGE_MODES = ("auto", "merge", "dynamic")
def _normalize_ltx2_two_stage_device_mode(mode: str | None) -> str | None:
if mode is None:
return None
mode = mode.lower()
return mode
def is_ltx2_two_stage_pipeline_name(pipeline_class_name: str | None) -> bool:
return pipeline_class_name in LTX2_TWO_STAGE_PIPELINE_NAMES
class Backend(str, Enum):
"""
Enumeration for different model backends.
- AUTO: Automatically select backend (prefer sglang native, fallback to diffusers)
- SGLANG: Use sglang's native optimized implementation
- DIFFUSERS: Use vanilla diffusers pipeline (supports all diffusers models)
"""
AUTO = "auto"
SGLANG = "sglang"
DIFFUSERS = "diffusers"
@classmethod
def from_string(cls, value: str) -> "Backend":
"""Convert string to Backend enum."""
try:
return cls(value.lower())
except ValueError:
raise ValueError(
f"Invalid backend: {value}. Must be one of: {', '.join([m.value for m in cls])}"
) from None
@classmethod
def choices(cls) -> list[str]:
"""Get all available choices as strings for argparse."""
return [backend.value for backend in cls]
@dataclasses.dataclass
class ServerArgs(DisaggArgsMixin):
# Model and path configuration (for convenience)
model_path: str
# explicit model ID override (e.g. "Qwen-Image")
model_id: str | None = None
# Model backend (sglang native or diffusers)
backend: Backend = Backend.AUTO
# Attention
attention_backend: str = None
attention_backend_config: addict.Dict | None = None
component_attention_backends: dict[str, str] | str | None = field(
default_factory=dict
)
cache_dit_config: str | dict[str, Any] | None = (
None # cache-dit config for diffusers
)
# Distributed executor backend
nccl_port: Optional[int] = None
# HuggingFace specific parameters
trust_remote_code: bool = False
revision: str | None = None
# Parallelism
num_gpus: int = 1
performance_mode: str = "auto"
tp_size: Optional[int] = None
sp_degree: Optional[int] = None
# sequence parallelism
ulysses_degree: Optional[int] = None
ring_degree: Optional[int] = None
# data parallelism
# number of data parallelism groups
dp_size: int = 1
# number of gpu in a dp group
dp_degree: int = 1
# cfg parallel (None = auto-decide based on num_gpus)
enable_cfg_parallel: Optional[bool] = None
# number of GPUs in each CFG parallel group (None = auto, 1 = disabled, N > 1 = enabled)
cfg_parallel_degree: Optional[int] = None
hsdp_replicate_dim: int = 1
hsdp_shard_dim: Optional[int] = None
dist_timeout: int | None = 3600 # 1 hour
pipeline_config: PipelineConfig = field(default_factory=PipelineConfig, repr=False)
# Pipeline override
pipeline_class_name: str | None = (
None # Override pipeline class from model_index.json
)
# LoRA parameters
# (Wenxuan) prefer to keep it here instead of in pipeline config to not make it complicated.
lora_path: str | None = None
lora_nickname: str = "default" # for swapping adapters in the pipeline
lora_scale: float = 1.0 # LoRA scale for merging (e.g., 0.125 for Hyper-SD)
lora_merge_mode: str = "auto"
lora_weight_name: str | None = None
# Component path overrides (key = model_index.json component name, value = path)
component_paths: dict[str, str] = field(default_factory=dict)
# path to pre-quantized transformer weights (single .safetensors or directory).
transformer_weights_path: str | None = None
# Quantization method for online quantization
quantization: str | None = None
# Layer name patterns to skip during online quantization
quantization_ignored_layers: list[str] | None = None
# can restrict layers to adapt, e.g. ["q_proj"]
# Will adapt only q, k, v, o by default.
lora_target_modules: list[str] | None = None
# CPU offload parameters
dit_cpu_offload: bool | None = None
# if true, select the DiT layerwise group
dit_layerwise_offload: bool | None = None
layerwise_offload_components: list[str] | None = None
dit_offload_prefetch_size: float = 0.0
text_encoder_cpu_offload: bool | None = None
image_encoder_cpu_offload: bool | None = None
vae_cpu_offload: bool | None = False
use_fsdp_inference: bool | None = None
pin_cpu_memory: bool = True
ltx2_two_stage_device_mode: str | None = None
_explicit_arg_names: set[str] = field(default_factory=set, repr=False)
# ComfyUI integration
comfyui_mode: bool = False
# Compilation
enable_torch_compile: bool = False
# warmup
warmup: bool = False
warmup_resolutions: list[str] = None
warmup_steps: int = 1
disable_autocast: bool | None = None
# Explicit quantization method override (e.g. "mxfp8", "fp8", "modelslim").
# When set, the transformer loader will use this instead of auto-detection.
quantization: str | None = None
# Quantization / Nunchaku SVDQuant configuration
nunchaku_config: NunchakuSVDQuantArgs | NunchakuConfig | None = field(
default_factory=NunchakuSVDQuantArgs, repr=False
)
# Master port for distributed inference
master_port: int = 30005
# http server endpoint config
host: str | None = "127.0.0.1"
port: int | None = 30000
# TODO: webui and their endpoint, check if webui_port is available.
webui: bool = False
webui_port: int | None = 12312
scheduler_port: int = 5555
batching_mode: str = "dynamic"
batching_max_size: int = 1
batching_delay_ms: float = 0.0
batching_config: str | None = None
enable_batching_metrics: bool = False
# Strict port mode: fail if requested port is unavailable instead of auto-selecting
strict_ports: bool = False
output_path: str | None = "outputs/"
input_save_path: str | None = "inputs/uploads"
# Prompt text file for batch processing
prompt_file_path: str | None = None
# model paths for correct deallocation
model_paths: dict[str, str] = field(default_factory=dict)
model_loaded: dict[str, bool] = field(
default_factory=lambda: {
"transformer": True,
"vae": True,
"video_vae": True,
"audio_vae": True,
"video_dit": True,
"audio_dit": True,
"dual_tower_bridge": True,
}
)
# # DMD parameters
# dmd_denoising_steps: List[int] | None = field(default=None)
# MoE parameters used by Wan2.2
boundary_ratio: float | None = None
# Disaggregation — fields defined here, methods in DisaggArgsMixin,
# CLI registration in disagg_args.add_disagg_cli_args().
base_gpu_id: int = 0
disagg_role: RoleType = RoleType.MONOLITHIC
disagg_timeout: int = 600
disagg_dispatch_policy: str = "round_robin"
disagg_mode: bool = False
disagg_server_addr: str | None = None
encoder_urls: str | None = None
denoiser_urls: str | None = None
decoder_urls: str | None = None
encoder_tp: int | None = None
denoiser_tp: int | None = None
denoiser_sp: int | None = None
denoiser_ulysses: int | None = None
denoiser_ring: int | None = None
decoder_tp: int | None = None
disagg_transfer_pool_size: int = 256 * 1024 * 1024
disagg_p2p_hostname: str = "127.0.0.1"
disagg_ib_device: str | None = None
pool_work_endpoint: str | None = None
pool_result_endpoint: str | None = None
# Logging
log_level: str = "info"
uvicorn_access_log_exclude_prefixes: list[str] = field(default_factory=list)
# Tracing
enable_trace: bool = False
otlp_traces_endpoint: str = "localhost:4317"
# get_role_parallelism, derive_pool_*_endpoint — from DisaggArgsMixin
@property
def broker_port(self) -> int:
return self.port + 1
@property
def is_local_mode(self) -> bool:
"""
If no server is running when a generation task begins, 'local_mode' will be enabled: a dedicated server will be launched
"""
return self.host is None or self.port is None
def _adjust_path(self):
expand_path_fields(self)
self._adjust_save_paths()
def _adjust_parameters(self):
"""set defaults and normalize values."""
auto_tuner = ServerArgsAutoTuner(self)
auto_tuner.adjust_based_on_performance_mode()
if auto_tuner.could_override_server_args():
self._adjust_offload()
auto_tuner.maybe_adjust_auto_default_layerwise_offload()
self._adjust_ltx2_two_stage_device_mode()
if auto_tuner.could_override_server_args():
auto_tuner.maybe_adjust_auto_component_residency_after_offload()
auto_tuner.maybe_adjust_auto_fsdp_with_offload_enabled()
auto_tuner.maybe_replace_cpu_offloaded_components_with_layerwise()
self._adjust_path()
self._adjust_quant_config()
self._adjust_warmup()
self._adjust_network_ports()
# adjust parallelism before attention backend
self._adjust_parallelism()
self._adjust_attention_backend()
self._adjust_platform_specific()
self._adjust_layerwise_offload_components()
self._adjust_autocast()
auto_tuner.finalize_auto_flags()
self.adjust_pipeline_config()
def _validate_parameters(self):
"""check consistency and raise errors for invalid configs"""
self._validate_pipeline()
self._validate_offload()
if not current_platform.is_cpu():
self._validate_parallelism()
self._validate_cfg_parallel()
self._validate_batching()
def _adjust_save_paths(self):
"""Normalize empty-string save paths to None (disabled)."""
if self.output_path is not None and self.output_path.strip() == "":
self.output_path = None
if self.input_save_path is not None and self.input_save_path.strip() == "":
self.input_save_path = None
def _adjust_quant_config(self):
"""
resolve, validate and adjust quantization config
handles only nunchaku for now
"""
ncfg = self.nunchaku_config
if ncfg is None or isinstance(ncfg, NunchakuConfig):
return
resolution = ncfg.resolve_runtime_config()
if resolution.transformer_weights_path:
self.transformer_weights_path = resolution.transformer_weights_path
self.nunchaku_config = resolution.nunchaku_config
def adjust_pipeline_config(self):
# enable parallel folding when SP is enabled
if self.tp_size != 1 or self.sp_degree <= 1:
return
enabled = False
for text_encoder_config in self.pipeline_config.text_encoder_configs:
if isinstance(text_encoder_config, T5Config):
text_encoder_config.parallel_folding = True
enabled = True
text_encoder_config.parallel_folding_mode = "sp"
if enabled:
logger.info(
"Enabled T5 text encoder parallel folding (mode=sp) for %s (tp_size=%s, sp_degree=%s).",
self.__class__.__name__,
self.tp_size,
self.sp_degree,
)
def _adjust_offload(self):
if current_platform.is_cpu():
# CPU platform does not need offload
return
# TODO: to be handled by each platform
if current_platform.get_device_total_memory() / BYTES_PER_GB < 30:
logger.info(
"Enabling large component offloading for GPU with low device memory"
)
if self.dit_cpu_offload is None:
self.dit_cpu_offload = True
if self.text_encoder_cpu_offload is None:
self.text_encoder_cpu_offload = True
if self.image_encoder_cpu_offload is None:
self.image_encoder_cpu_offload = True
elif self.pipeline_config.task_type.is_image_gen():
logger.info(
"Disabling some offloading (except dit, text_encoder) for image generation model"
)
if self.dit_cpu_offload is None:
self.dit_cpu_offload = True
if self.text_encoder_cpu_offload is None:
self.text_encoder_cpu_offload = True
if self.image_encoder_cpu_offload is None:
self.image_encoder_cpu_offload = False
else:
if self.dit_cpu_offload is None:
self.dit_cpu_offload = True
if self.text_encoder_cpu_offload is None:
self.text_encoder_cpu_offload = True
if self.image_encoder_cpu_offload is None:
self.image_encoder_cpu_offload = True
def _adjust_ltx2_two_stage_device_mode(self):
if not self._is_ltx23_two_stage_pipeline():
return
mode = self.ltx2_two_stage_device_mode
if mode is None:
env_mode = os.getenv("SGLANG_LTX2_TWO_STAGE_DEVICE_MODE")
mode = (
_normalize_ltx2_two_stage_device_mode(env_mode)
if env_mode
else self._resolve_default_ltx2_two_stage_device_mode()
)
else:
mode = _normalize_ltx2_two_stage_device_mode(mode)
if mode not in LTX2_TWO_STAGE_DEVICE_MODES:
raise ValueError(
f"Invalid ltx2_two_stage_device_mode={mode!r}. "
f"Expected one of {LTX2_TWO_STAGE_DEVICE_MODES}."
)
self.ltx2_two_stage_device_mode = mode
def _resolve_default_ltx2_two_stage_device_mode(self) -> str:
if not current_platform.is_cuda():
logger.info(
"Automatically set ltx2_two_stage_device_mode=snapshot on non-CUDA platform"
)
return "snapshot"
device_name = str(current_platform.get_device_name(0)).upper()
device_total_memory_gb = (
current_platform.get_device_total_memory() / BYTES_PER_GB
)
if (
"H200" in device_name
or device_total_memory_gb >= LTX2_RESIDENT_AUTO_ENABLE_MEM_GB
):
logger.info(
"Automatically set ltx2_two_stage_device_mode=resident for high-memory CUDA GPU (%s, %.2f GiB total)",
device_name,
device_total_memory_gb,
)
return "resident"
logger.info(
"Automatically set ltx2_two_stage_device_mode=snapshot for CUDA GPU (%s, %.2f GiB total)",
device_name,
device_total_memory_gb,
)
return "snapshot"
def _is_ltx23_two_stage_pipeline(self) -> bool:
return is_ltx2_two_stage_pipeline_name(self.pipeline_class_name) and (
self._is_ltx23_model_path(self.model_path)
or is_ltx23_native_variant(self.pipeline_config.vae_config.arch_config)
)
def _uses_ltx23_snapshot_two_stage_residency(self) -> bool:
return (
self.ltx2_two_stage_device_mode == "snapshot"
and self._is_ltx23_two_stage_pipeline()
)
def _adjust_attention_backend(self):
if self.attention_backend in ["fa3", "fa4"]:
self.attention_backend = "fa"
self.component_attention_backends = (
self._normalize_component_attention_backends(
self.component_attention_backends
)
)
# attention_backend_config
if self.attention_backend_config is None:
self.attention_backend_config = addict.Dict()
elif isinstance(self.attention_backend_config, str):
self.attention_backend_config = addict.Dict(
self._parse_attention_backend_config(self.attention_backend_config)
)
if self.backend != Backend.DIFFUSERS and isinstance(
self.pipeline_config, LTX2PipelineConfig
):
text_backend = self.component_attention_backends.get("text_encoder")
if text_backend != "torch_sdpa":
if text_backend is None:
logger.info(
"Automatically set torch_sdpa backend for component text_encoder to preserve LTX2 official attention semantics"
)
else:
logger.warning(
"Overriding %s backend with torch_sdpa for component text_encoder to preserve LTX2 official attention semantics",
text_backend,
)
self.component_attention_backends["text_encoder"] = "torch_sdpa"
if self.ring_degree > 1:
if self.attention_backend is not None and self.attention_backend not in (
"fa",
"sage_attn",
):
raise ValueError(
"Ring Attention is only supported for flash attention or sage attention backend for now"
)
if self.attention_backend is None:
self.attention_backend = "fa"
logger.info(
"Ring Attention is currently only supported for flash attention or sage attention; "
"attention_backend has been automatically set to flash attention"
)
if self.attention_backend is None and self.backend != Backend.DIFFUSERS:
if (
current_platform.is_cuda()
and self.pipeline_class_name is None
and self.num_gpus == 1
and self.tp_size == 1
and self.sp_degree == 1
and self.ulysses_degree == 1
and self.ring_degree == 1
and self._is_ltx23_model_path(self.model_path)
):
self.attention_backend = "fa"
logger.info(
"Automatically set attention_backend=fa for LTX-2.3 one-stage on 1 GPU to preserve precision"
)
return
self._set_default_attention_backend()
@staticmethod
def _normalize_attention_backend_name(backend: str) -> str:
if not isinstance(backend, str):
raise ValueError("Attention backend name must be a string")
normalized = backend.strip().lower()
if normalized in ("fa3", "fa4"):
normalized = "fa"
try:
return AttentionBackendEnum[normalized.upper()].name.lower()
except KeyError:
raise ValueError(
f"Invalid attention backend '{backend}'. "
f"Available options are: {[e.name.lower() for e in AttentionBackendEnum]}"
) from None
@staticmethod
def _parse_component_attention_backend_map(
value: dict[str, str] | str | None,
) -> dict[str, str]:
if value is None or value == "":
return {}
if isinstance(value, dict):
return dict(value)
if not isinstance(value, str):
raise ValueError(
"component_attention_backends must be a dict or a comma-separated component=backend string"
)
try:
parsed = json.loads(value)
if not isinstance(parsed, dict):
raise ValueError
return parsed
except (json.JSONDecodeError, ValueError):
pass
result: dict[str, str] = {}
for pair in value.split(","):
pair = pair.strip()
if not pair:
continue
if "=" not in pair:
raise ValueError(
"component_attention_backends must use component=backend entries"
)
component, backend = pair.split("=", 1)
result[component.strip()] = backend.strip()
return result
@classmethod
def _normalize_component_attention_backends(
cls, value: dict[str, str] | str | None
) -> dict[str, str]:
raw = cls._parse_component_attention_backend_map(value)
normalized: dict[str, str] = {}
for component, backend in raw.items():
if not isinstance(component, str):
raise ValueError("Component attention backend key must be a string")
component_name = component.strip().replace("-", "_")
if not component_name:
raise ValueError("Component attention backend key must not be empty")
normalized[component_name] = cls._normalize_attention_backend_name(backend)
return normalized
def resolve_component_attention_backend(
self, *component_names: str | None
) -> tuple[AttentionBackendEnum | None, str | None]:
for component_name in component_names:
if component_name is None:
continue
key = component_name.replace("-", "_")
fallback_keys = [key]
if key.endswith("_2"):
# Secondary two-stage components inherit the base component
# backend unless explicitly overridden.
fallback_keys.append(key[:-2])
for backend_key in fallback_keys:
backend = self.component_attention_backends.get(backend_key)
if backend is not None:
return AttentionBackendEnum[backend.upper()], backend_key
return None, None
def _adjust_warmup(self):
if self.warmup_resolutions is not None:
self.warmup = True
if self.warmup:
logger.info(
"Warmup enabled, the launch time is expected to be longer than usual"
)
@staticmethod
def _require_port(port: int, name: str) -> None:
"""Raise if *port* is occupied (used under ``--strict-ports``)."""
if not is_port_available(port):
raise RuntimeError(
f"{name} port {port} is unavailable and --strict-ports is enabled. "
f"Either use a different port or disable --strict-ports."
)
def _adjust_network_ports(self):
# Disagg role instances (encoder/denoiser/decoder) don't serve HTTP,
# so skip settling the HTTP port to avoid unnecessary port collisions.
needs_http = self.disagg_role in (
RoleType.MONOLITHIC,
RoleType.SERVER,
)
if self.strict_ports:
requested_ports = []
if needs_http:
requested_ports.append((self.port, "HTTP"))
requested_ports.append((self.scheduler_port, "Scheduler"))
if self.master_port is not None:
requested_ports.append((self.master_port, "Master"))
seen_ports: dict[int, str] = {}
for port, name in requested_ports:
if port in seen_ports:
raise RuntimeError(
f"{name} port {port} duplicates {seen_ports[port]} port and "
"--strict-ports is enabled."
)
seen_ports[port] = name
self._require_port(port, name)
else:
settled_ports: set[int] = set()
if needs_http:
self.port = self.settle_port(self.port)
settled_ports.add(self.port)
initial_scheduler_port = self.scheduler_port + (
random.randint(0, 100) if self.scheduler_port == 5555 else 0
)
self.scheduler_port = self.settle_port(
initial_scheduler_port, avoid=settled_ports
)
settled_ports.add(self.scheduler_port)
if self.master_port is not None:
self.master_port = self.settle_port(
self.master_port, 37, avoid=settled_ports
)
def _adjust_parallelism(self):
sp_unspecified = self.sp_degree is None
ulysses_unspecified = self.ulysses_degree is None
ring_unspecified = self.ring_degree is None
cfg_unspecified = self.enable_cfg_parallel is None
if current_platform.is_cpu() and (self.tp_size or 1) > 1:
# CPU platform reuse num_gpus to represent num cpu numa nodes as devices
self.num_gpus = self.tp_size
if self.hsdp_shard_dim is None:
self.hsdp_shard_dim = self.num_gpus
if self.tp_size is None:
self.tp_size = 1
# --cfg-parallel-size takes precedence over --enable-cfg-parallel bool.
if self.cfg_parallel_degree is not None:
if self.cfg_parallel_degree == 1:
self.enable_cfg_parallel = False
elif self.cfg_parallel_degree > 1:
self.enable_cfg_parallel = True
cfg_unspecified = False
# Auto-enable CFG parallel when user hasn't set any parallelism flags
# and there are enough GPUs. Only auto-enable for models whose default
# SamplingParams use classifier-free guidance (negative_prompt is not None),
# because non-CFG models (e.g. FLUX) crash when CFG parallel splits ranks.
if cfg_unspecified:
cfg_group_size = self.dp_size * self.tp_size * 2
if (
self.performance_mode != "manual"
and self.num_gpus >= 2
and self.num_gpus % cfg_group_size == 0
and sp_unspecified
and ulysses_unspecified
and ring_unspecified
and self._model_default_uses_cfg()
):
self.enable_cfg_parallel = True
logger.info(
"Automatically enabled CFG parallel for %d GPUs. "
"Use --sp-degree / --ulysses-degree to use sequence "
"parallelism instead.",
self.num_gpus,
)
else:
self.enable_cfg_parallel = False
# Resolve cfg_parallel_degree to a concrete int now that enable_cfg_parallel is settled.
if self.cfg_parallel_degree is None:
self.cfg_parallel_degree = 2 if self.enable_cfg_parallel else 1
# adjust sp_degree: allocate all remaining GPUs after TP and DP
if self.sp_degree is None:
num_gpus_per_group = self.dp_size * self.tp_size
if self.enable_cfg_parallel:
num_gpus_per_group *= self.cfg_parallel_degree
if self.num_gpus % num_gpus_per_group == 0:
self.sp_degree = self.num_gpus // num_gpus_per_group
else:
# Will be validated later
self.sp_degree = 1
if (
self.ulysses_degree is None
and self.ring_degree is None
and self.sp_degree != 1
):
self.ulysses_degree = self.sp_degree
logger.info(
f"Automatically set ulysses_degree=sp_degree={self.ulysses_degree} for best performance"
)
if self.ulysses_degree is None:
self.ulysses_degree = 1
logger.debug(
f"Ulysses degree not set, using default value {self.ulysses_degree}"
)
if self.ring_degree is None:
self.ring_degree = 1
logger.debug(f"Ring degree not set, using default value {self.ring_degree}")
def _model_default_uses_cfg(self) -> bool:
"""
Check whether the model uses classifier-free guidance by default.
CFG is active when *both* ``negative_prompt is not None`` and ``guidance_scale > 1``.
"""
from sglang.multimodal_gen.registry import get_model_info
model_info = get_model_info(self.model_path, self.backend, self.model_id)
if model_info is None:
return False
default_params = model_info.sampling_param_cls()
return (
getattr(default_params, "negative_prompt", None) is not None
and getattr(default_params, "guidance_scale", 0) > 1.0
)
@staticmethod
def _is_ltx23_model_path(model_path: str | None) -> bool:
if not model_path:
return False
normalized = model_path.lower()
return any(
token in normalized
for token in (
"lightricks/ltx-2.3",
"models--lightricks--ltx-2.3",
"lightricks__ltx-2.3",
)
)
def _adjust_platform_specific(self):
if current_platform.is_mps():
self.use_fsdp_inference = False
self.dit_layerwise_offload = False
self.layerwise_offload_components = None
def is_arg_explicitly_set(self, arg_name: str) -> bool:
return arg_name in self._explicit_arg_names
def should_configure_layerwise_offload_for_lazy_component(
self, component_name: str
) -> bool:
"""Return whether a lazy-loaded component should try layerwise offload.
Lazy components are loaded after the normal pipeline-wide configuration
pass, so they should only attempt layerwise configuration when their
component name is covered by the selected layerwise scope.
"""
component_names = normalize_layerwise_offload_components(
self.layerwise_offload_components
)
if not component_names:
return False
if LAYERWISE_OFFLOAD_ALL_COMPONENTS in component_names:
return True
return layerwise_component_matches_any_selection(
component_name, component_names
)
@property
def is_dit_layerwise_offload_selected(self) -> bool:
"""returns if dit is selected to be layerwise-offload"""
component_names = self.layerwise_offload_components
return bool(
component_names
and "dit_cpu_offload"
in cpu_offload_flags_for_layerwise_components(component_names)
)
def _adjust_layerwise_offload_components(self):
explicitly_set_component_names = normalize_layerwise_offload_components(
self.layerwise_offload_components
)
if self.dit_layerwise_offload:
if explicitly_set_component_names is None:
explicitly_set_component_names = [LAYERWISE_OFFLOAD_DIT_GROUP]
elif LAYERWISE_OFFLOAD_DIT_GROUP not in explicitly_set_component_names:
explicitly_set_component_names = [
LAYERWISE_OFFLOAD_DIT_GROUP,
*explicitly_set_component_names,
]
if explicitly_set_component_names is not None:
self.layerwise_offload_components = explicitly_set_component_names
self._disable_cpu_offload_for_layerwise_components(
explicitly_set_component_names
)
return
def _disable_cpu_offload_for_layerwise_components(
self, component_names: list[str]
) -> None:
# Layerwise offload owns H2D/D2H for selected component weights.
flag_names = cpu_offload_flags_for_layerwise_components(component_names)
disabled_flag_names: list[str] = []
if "dit_cpu_offload" in flag_names and self.dit_cpu_offload is not False:
self.dit_cpu_offload = False
disabled_flag_names.append("dit_cpu_offload")
if (
"text_encoder_cpu_offload" in flag_names
and self.text_encoder_cpu_offload is not False
):
self.text_encoder_cpu_offload = False
disabled_flag_names.append("text_encoder_cpu_offload")
if (
"image_encoder_cpu_offload" in flag_names
and self.image_encoder_cpu_offload is not False
):
self.image_encoder_cpu_offload = False
disabled_flag_names.append("image_encoder_cpu_offload")
if "vae_cpu_offload" in flag_names and self.vae_cpu_offload is not False:
self.vae_cpu_offload = False
disabled_flag_names.append("vae_cpu_offload")
if disabled_flag_names:
logger.info(
"Disabling %s because the selected layerwise offload components "
"manage the same weights.",
", ".join(disabled_flag_names),
)
def _adjust_autocast(self):
if self.disable_autocast is None:
self.disable_autocast = not self.pipeline_config.enable_autocast
def _parse_attention_backend_config(self, config_str: str) -> dict[str, Any]:
"""parse attention backend config from string."""
if not config_str:
return {}
# 1. treat as file path
if os.path.exists(config_str):
if config_str.endswith((".yaml", ".yml")):
with open(config_str, "r") as f:
return yaml.safe_load(f)
elif config_str.endswith(".json"):
with open(config_str, "r") as f:
return json.load(f)
# 2. treat as JSON string
try:
return json.loads(config_str)
except json.JSONDecodeError:
pass
# 3. treat as k=v pairs (simple implementation). e.g., "sparsity=0.5,enable_x=true"
try:
config = {}
pairs = config_str.split(",")
for pair in pairs:
k, v = pair.split("=", 1)
k = k.strip()
v = v.strip()
if v.lower() == "true":
v = True
elif v.lower() == "false":
v = False
elif v.replace(".", "", 1).isdigit():
v = float(v) if "." in v else int(v)
config[k] = v
return config
except Exception:
raise ValueError(f"Could not parse attention backend config: {config_str}")
def __post_init__(self):
# configure logger before use
configure_logger(server_args=self)
# Convert string disagg_role to enum (from CLI/config)
convert_disagg_role_string(self.__dict__)
# 1. adjust parameters
self._adjust_parameters()
# 2. Validate parameters
self._validate_parameters()
# log clean server_args
try:
safe_args = _sanitize_for_logging(self, key_hint="server_args")
logger.info("server_args: %s", json.dumps(safe_args, ensure_ascii=False))
except Exception:
# Fallback to default repr if sanitization fails
logger.info(f"server_args: {self}")
@staticmethod
def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
# Model and path configuration
parser.add_argument(
"--model-path",
type=str,
help="The path of the model weights. This can be a local folder or a Hugging Face repo ID.",
)
parser.add_argument(
"--model-id",
type=str,
default=ServerArgs.model_id,
help=(
"Override the model ID used for config resolution. "
"Useful when --model-path is a local directory whose name does not match "
"any registered HF repo name. Should be the repo name portion of the HF ID "
"(e.g. 'Qwen-Image' for 'Qwen/Qwen-Image')."
),
)
parser.add_argument(
"--pipeline-class-name",
type=str,
default=ServerArgs.pipeline_class_name,
help=(
"Override pipeline class selection from model_index.json. "
"Must match a registered pipeline_name."
),
)
# attention
parser.add_argument(
"--attention-backend",
type=str,
default=None,
help=(
"The attention backend to use. For SGLang-native pipelines, use "
"values like fa, torch_sdpa, sage_attn, etc. For diffusers pipelines, "
"use diffusers attention backend names such as flash, _flash_3_hub, "
"sage, or xformers."
),
)
parser.add_argument(
"--attention-backend-config",
type=str,
default=None,
help="Configuration for the attention backend. Can be a JSON string, a path to a JSON/YAML file, or key=value pairs.",
)
parser.add_argument(
"--component-attention-backends",
type=str,
default=None,
help=(
"Per-component attention backend overrides for native pipelines. "
"Use component names from model_index.json, e.g. "
"'text_encoder=torch_sdpa,transformer=fa'."
),
)
parser.add_argument(
"--cache-dit-config",
type=str,
default=ServerArgs.cache_dit_config,
help="Path to a Cache-DiT YAML/JSON config. Enables cache-dit for diffusers backend.",
)
# HuggingFace specific parameters
parser.add_argument(
"--trust-remote-code",
action=StoreBoolean,
default=ServerArgs.trust_remote_code,
help="Trust remote code when loading HuggingFace models",
)
parser.add_argument(
"--revision",
type=str,
default=ServerArgs.revision,
help="The specific model version to use (can be a branch name, tag name, or commit id)",
)
parser.add_argument(
"--performance-mode",
"--mode",
type=str,
choices=PERFORMANCE_MODES,
default=ServerArgs.performance_mode,
help=(
"Preset for performance and memory defaults. "
"'manual' keeps performance-related server args under explicit user control, no adjustment is made; "
"'auto' keeps safe defaults and applies high-confidence FSDP/CFG improvements; "
"'speed' favors GPU-resident execution for lower latency and higher throughput, and may OOM; "
"'memory' favors lower GPU memory usage; "
"Explicit offload/FSDP/parallelism flags take precedence."
),
)
# Parallelism
parser.add_argument(
"--num-gpus",
type=int,
default=ServerArgs.num_gpus,
help="The number of GPUs to use.",
)
parser.add_argument(
"--tp-size",
type=int,
default=None,
help="The tensor parallelism size. Defaults to 1 if not specified.",
)
parser.add_argument(
"--sp-degree",
type=int,
default=None,
help="The sequence parallelism size. If not specified, will use all remaining GPUs after accounting for TP and DP.",
)
parser.add_argument(
"--ulysses-degree",
type=int,
default=ServerArgs.ulysses_degree,
help="Ulysses sequence parallel degree. Used in attention layer.",
)
parser.add_argument(
"--ring-degree",
type=int,
default=ServerArgs.ring_degree,
help="Ring sequence parallel degree. Used in attention layer.",
)
parser.add_argument(
"--enable-cfg-parallel",
action=StoreBoolean,
default=None,
help="Enable cfg parallel at degree 2. Auto-enabled when num_gpus >= 2 and no SP flags are set. Use false to disable it explicitly.",
)
parser.add_argument(
"--cfg-parallel-size",
dest="cfg_parallel_degree",
type=int,
default=None,
help=(
"Number of GPUs per CFG parallel group (1 = disabled, N > 1 = enabled at degree N). "
"Supersedes --enable-cfg-parallel. Allows 4-branch CFG parallel (e.g., --cfg-parallel-size 4) "
"for models with cond + neg + perturbed + modality branches."
),
)
parser.add_argument(
"--data-parallel-size",
"--dp-size",
"--dp",
type=int,
default=ServerArgs.dp_size,
help="The data parallelism size.",
)
parser.add_argument(
"--hsdp-replicate-dim",
type=int,
default=ServerArgs.hsdp_replicate_dim,
help="The data parallelism size.",
)
parser.add_argument(
"--hsdp-shard-dim",
type=int,
default=None,
help="The data parallelism shards. Defaults to num_gpus if not specified.",
)
parser.add_argument(
"--dist-timeout",
type=int,
default=ServerArgs.dist_timeout,
help="Timeout for torch.distributed operations in seconds. "
"Increase this value if you encounter 'Connection closed by peer' errors after the service is idle. ",
)
# Disaggregated diffusion args (defined in disagg_args.py)
add_disagg_cli_args(parser)
# Prompt text file for batch processing
parser.add_argument(
"--prompt-file-path",
type=str,
default=ServerArgs.prompt_file_path,
help="Path to a text file containing prompts (one per line) for batch processing",
)
parser.add_argument(
"--mask-strategy-file-path",
type=str,
help="Path to mask strategy JSON file for STA",
)
parser.add_argument(
"--enable-torch-compile",
action=StoreBoolean,
default=ServerArgs.enable_torch_compile,
help="Use torch.compile to speed up DiT inference."
+ "However, will likely cause precision drifts. See (https://github.com/pytorch/pytorch/issues/145213)",
)
# warmup
parser.add_argument(
"--warmup",
action=StoreBoolean,
default=ServerArgs.warmup,
help="Perform some warmup after server starts (if `--warmup-resolutions` is specified) or before processing the first request (if `--warmup-resolutions` is not specified)."
"Recommended to enable when benchmarking to ensure fair comparison and best performance."
"When enabled with `--warmup-resolutions` unspecified, look for the line ending with `(with warmup excluded)` for actual processing time.",
)
parser.add_argument(
"--warmup-resolutions",
type=str,
nargs="+",
default=ServerArgs.warmup_resolutions,
help="Specify resolutions for server to warmup. e.g., `--warmup-resolutions 256x256, 720x720`",
)
parser.add_argument(
"--warmup-steps",
type=int,
default=ServerArgs.warmup_steps,
help="The number of warmup steps to perform for each resolution.",
)
# layerwise offload
parser.add_argument(
"--dit-cpu-offload",
action=StoreBoolean,
help="Use CPU offload for DiT inference. Enable if run out of memory with FSDP.",
)
parser.add_argument(
"--dit-layerwise-offload",
action=StoreBoolean,
default=ServerArgs.dit_layerwise_offload,
help="Enable layerwise CPU offload with async H2D prefetch overlap for DiTs. "
"It selects only the DiT layerwise group. Cannot be used together with cache-dit "
"(SGLANG_CACHE_DIT_ENABLED), dit_cpu_offload, or use_fsdp_inference.",
)
parser.add_argument(
"--layerwise-offload-components",
"--layerwise-offload-modules",
type=str,
nargs="+",
default=ServerArgs.layerwise_offload_components,
help="Select pipeline components for layerwise offload. "
"Use dit to select the DiT layerwise group, default for the default group "
"(currently text_encoder, image_encoder, and vae), "
"or all to select every layerwise-offloadable component. "
"This option does not imply --dit-layerwise-offload. Example: "
"--layerwise-offload-components text_encoder image_encoder vae.",
)
parser.add_argument(
"--dit-offload-prefetch-size",
type=float,
default=ServerArgs.dit_offload_prefetch_size,
help="The size of prefetch for dit-layerwise-offload. If the value is between 0.0 and 1.0, it is treated as a ratio of the total number of layers. If the value is >= 1, it is treated as the absolute number of layers. 0.0 means prefetch 1 layer (lowest memory). Values above 0.5 might have peak memory close to no offload but worse performance.",
)
# offload flags
parser.add_argument(
"--text-encoder-cpu-offload",
action=StoreBoolean,
help="Use CPU offload for text encoder. Enable if run out of memory.",
)
parser.add_argument(
"--image-encoder-cpu-offload",
action=StoreBoolean,
help="Use CPU offload for image encoder. Enable if run out of memory.",
)
parser.add_argument(
"--vae-cpu-offload",
action=StoreBoolean,
help="Use CPU offload for VAE. Enable if run out of memory.",
)
parser.add_argument(
"--use-fsdp-inference",
action=StoreBoolean,
help="Use FSDP inference to shard DiT weights across GPUs. For single-GPU memory pressure, prefer CPU or layerwise offload.",
)
parser.add_argument(
"--pin-cpu-memory",
action=StoreBoolean,
help='Pin memory for CPU offload. Only added as a temp workaround if it throws "CUDA error: invalid argument". '
"Should be enabled in almost all cases",
)
parser.add_argument(
"--ltx2-two-stage-device-mode",
type=str,
choices=LTX2_TWO_STAGE_DEVICE_MODES,
default=ServerArgs.ltx2_two_stage_device_mode,
help=(
"LTX-2.3 two-stage device residency mode: "
"'original' keeps official two-stage semantics without premerged stage2, "
"'snapshot' keeps premerged stage2 with snapshot-based release, "
"'resident' keeps both transformers resident on GPU. "
"Default is auto: resident on H200/high-memory CUDA GPUs, otherwise snapshot."
),
)
parser.add_argument(
"--disable-autocast",
action=StoreBoolean,
help="Disable autocast for denoising loop and vae decoding in pipeline sampling",
)
# quantization
parser.add_argument(
"--quantization",
type=str,
default=ServerArgs.quantization,
help=(
"Quantization method for the transformer. If omitted, the method is "
"auto-detected from the checkpoint config or safetensors metadata when "
"possible. Applies to both pre-quantized checkpoints and online "
"quantization. Use this flag to override auto-detection. "
"Options: 'fp8', 'mxfp8', 'mxfp4', 'modelslim'. "
"Note: MXFP4 requires ROCm and MI350+ (gfx95x)."
),
)
parser.add_argument(
"--quantization-ignored-layers",
type=str,
nargs="+",
default=ServerArgs.quantization_ignored_layers,
help=(
"Layer name patterns to keep unquantized during online quantization "
"(fp8/mxfp4). Each pattern is matched against the layer prefix. "
"Example: --quantization-ignored-layers img_mod txt_mod to_out"
),
)
# Nunchaku SVDQuant quantization parameters
NunchakuSVDQuantArgs.add_cli_args(parser)
# Master port for distributed inference
parser.add_argument(
"--master-port",
type=int,
default=ServerArgs.master_port,
help="Master port for distributed inference. If not set, a random free port will be used.",
)
parser.add_argument(
"--scheduler-port",
type=int,
default=ServerArgs.scheduler_port,
help="Port for the scheduler server.",
)
parser.add_argument(
"--batching-mode",
type=str,
default=ServerArgs.batching_mode,
choices=["dynamic"],
help="Request batching scheduler mode. Currently only 'dynamic' is implemented.",
)
parser.add_argument(
"--batching-max-size",
type=int,
default=ServerArgs.batching_max_size,
help="Maximum number of compatible generation requests to merge into one batch.",
)
parser.add_argument(
"--batching-delay-ms",
type=float,
default=ServerArgs.batching_delay_ms,
help="Maximum time (in ms) to wait for forming a larger batch before dispatch.",
)
parser.add_argument(
"--batching-config",
type=str,
default=ServerArgs.batching_config,
help=(
"Optional JSON file with {'schema_version': 1, 'rules': [...]} "
"batching admission rules that can cap model/resolution shapes "
"below --batching-max-size."
),
)
parser.add_argument(
"--enable-batching-metrics",
action="store_true",
default=ServerArgs.enable_batching_metrics,
help="Log periodic batch efficiency metrics such as realized batch size and queue wait time.",
)
parser.add_argument(
"--host",
type=str,
default=ServerArgs.host,
help="Host for the HTTP API server.",
)
parser.add_argument(
"--port",
type=int,
default=ServerArgs.port,
help="Port for the HTTP API server.",
)
parser.add_argument(
"--strict-ports",
action=StoreBoolean,
default=ServerArgs.strict_ports,
help="If enabled, fail when requested ports are unavailable instead of auto-selecting.",
)
parser.add_argument(
"--webui",
action=StoreBoolean,
default=ServerArgs.webui,
help="Whether to use webui for better display",
)
parser.add_argument(
"--webui-port",
type=int,
default=ServerArgs.webui_port,
help="Whether to use webui for better display",
)
parser.add_argument(
"--output-path",
type=str,
default=ServerArgs.output_path,
help='Directory path to save generated images/videos. Set to "" to disable persistent saving.',
)
parser.add_argument(
"--input-save-path",
type=str,
default=ServerArgs.input_save_path,
help='Directory path to save uploaded input images/videos. Set to "" to disable persistent saving.',
)
# LoRA
parser.add_argument(
"--lora-path",
type=str,
default=ServerArgs.lora_path,
help="The path to the LoRA adapter weights (can be local file path or HF hub id) to launch with",
)
parser.add_argument(
"--lora-nickname",
type=str,
default=ServerArgs.lora_nickname,
help="The nickname for the LoRA adapter to launch with",
)
parser.add_argument(
"--lora-scale",
type=float,
default=ServerArgs.lora_scale,
help="LoRA scale for merging (e.g., 0.125 for Hyper-SD). Same as lora_scale in Diffusers",
)
parser.add_argument(
"--lora-merge-mode",
type=str,
choices=LORA_MERGE_MODES,
default=ServerArgs.lora_merge_mode,
help=(
"How LoRA is applied: auto keeps static merge for regular weights "
"and uses dynamic LoRA for FSDP-sharded weights to avoid full-gather; "
"merge always merges into base weights; dynamic always applies LoRA at forward time."
),
)
parser.add_argument(
"--lora-weight-name",
type=str,
default=ServerArgs.lora_weight_name,
help="Specific safetensors filename to load from a multi-file LoRA repo",
)
# Add pipeline configuration arguments
PipelineConfig.add_cli_args(parser)
# Logging
parser.add_argument(
"--log-level",
type=str,
default=ServerArgs.log_level,
help="The logging level of all loggers.",
)
# Tracing
parser.add_argument(
"--enable-trace",
action="store_true",
default=False,
help="Enable OpenTelemetry tracing.",
)
parser.add_argument(
"--otlp-traces-endpoint",
type=str,
default=ServerArgs.otlp_traces_endpoint,
help="OTLP collector endpoint when --enable-trace is set. Format: <host>:<port>",
)
parser.add_argument(
"--uvicorn-access-log-exclude-prefixes",
type=str,
nargs="*",
default=[],
help="Exclude uvicorn access logs whose request path starts with any of these prefixes. "
"Defaults to empty (disabled). "
"Example: --uvicorn-access-log-exclude-prefixes /metrics /health",
)
parser.add_argument(
"--backend",
type=str,
choices=Backend.choices(),
default=ServerArgs.backend.value,
help="The model backend to use. 'auto' prefers sglang native and falls back to diffusers. "
"'sglang' uses native optimized implementation. 'diffusers' uses vanilla diffusers pipeline.",
)
return parser
def url(self):
host = self.host
if not host or host == "0.0.0.0":
host = "127.0.0.1"
elif host == "::":
host = "::1"
if is_valid_ipv6_address(host):
return f"http://[{host}]:{self.port}"
else:
return f"http://{host}:{self.port}"
@property
def scheduler_endpoint(self):
"""
Internal endpoint for scheduler.
Prefers the configured host but normalizes localhost -> 127.0.0.1 to avoid ZMQ issues.
"""
scheduler_host = self.host
if scheduler_host is None or scheduler_host == "localhost":
scheduler_host = "127.0.0.1"
return f"tcp://{scheduler_host}:{self.scheduler_port}"
def settle_port(
self,
port: int,
port_inc: int = 42,
max_attempts: int = 100,
avoid: set[int] | None = None,
) -> int:
"""
Find an available port with retry logic.
"""
attempts = 0
original_port = port
avoid = avoid or set()
while attempts < max_attempts:
if port not in avoid and is_port_available(port):
if attempts > 0:
logger.info(
f"Port {original_port} was unavailable, using port {port} instead"
)
return port
attempts += 1
if port < 60000:
port += port_inc
else:
# Wrap around with randomization to avoid collision
port = 5000 + random.randint(0, 1000)
raise RuntimeError(
f"Failed to find available port after {max_attempts} attempts "
f"(started from port {original_port})"
)
@staticmethod
def _extract_component_paths(
unknown_args: list[str],
) -> tuple[dict[str, str], list[str]]:
"""
Extract dynamic component path args from unrecognised CLI args.
Supported forms:
- ``--<component>-path /path/to/component``
- ``--component-paths.<component> /path/to/component`` (expanded from config)
"""
component_paths: dict[str, str] = {}
remaining: list[str] = []
i = 0
while i < len(unknown_args):
arg = unknown_args[i]
key_part = arg.split("=", 1)[0] if "=" in arg else arg
component = None
if key_part.startswith("--component-paths."):
component = key_part[len("--component-paths.") :].replace("-", "_")
elif key_part.startswith("--component_paths."):
component = key_part[len("--component_paths.") :].replace("-", "_")
elif key_part.startswith("--") and key_part.endswith("-path"):
component = key_part[2:-5].replace("-", "_")
if component is not None:
if "=" in arg:
component_paths[component] = arg.split("=", 1)[1]
elif i + 1 < len(unknown_args) and not unknown_args[i + 1].startswith(
"-"
):
i += 1
component_paths[component] = unknown_args[i]
else:
remaining.append(arg)
i += 1
continue
else:
remaining.append(arg)
i += 1
# canonicalize and validate
for component, path in component_paths.items():
path = os.path.expanduser(path)
component_paths[component] = path
return component_paths, remaining
@staticmethod
def _extract_component_attention_backends(
unknown_args: list[str],
) -> tuple[dict[str, str], list[str]]:
component_attention_backends: dict[str, str] = {}
remaining: list[str] = []
i = 0
while i < len(unknown_args):
arg = unknown_args[i]
key_part = arg.split("=", 1)[0] if "=" in arg else arg
component = None
if key_part.startswith("--component-attention-backends."):
component = key_part[len("--component-attention-backends.") :].replace(
"-", "_"
)
elif key_part.startswith("--component_attention_backends."):
component = key_part[len("--component_attention_backends.") :].replace(
"-", "_"
)
if component is not None:
if "=" in arg:
component_attention_backends[component] = arg.split("=", 1)[1]
elif i + 1 < len(unknown_args) and not unknown_args[i + 1].startswith(
"-"
):
i += 1
component_attention_backends[component] = unknown_args[i]
else:
remaining.append(arg)
i += 1
continue
else:
remaining.append(arg)
i += 1
return component_attention_backends, remaining
@classmethod
def from_cli_args(
cls, args: argparse.Namespace, unknown_args: list[str] | None = None
) -> "ServerArgs":
if unknown_args is None:
unknown_args = []
# extract dynamic --<component>-path from unknown args
dynamic_paths, remaining = cls._extract_component_paths(unknown_args)
dynamic_attention_backends, remaining = (
cls._extract_component_attention_backends(remaining)
)
if remaining:
raise SystemExit(f"error: unrecognized arguments: {' '.join(remaining)}")
provided_args = cls.get_provided_args(args, unknown_args)
# Handle config file
config_file = provided_args.get("config")
if config_file:
config_args = cls.load_config_file(config_file)
provided_args = {**config_args, **provided_args}
if dynamic_paths:
existing = dict(provided_args.get("component_paths") or {})
existing.update(dynamic_paths)
provided_args["component_paths"] = existing
if dynamic_attention_backends:
existing = cls._parse_component_attention_backend_map(
provided_args.get("component_attention_backends")
)
existing.update(dynamic_attention_backends)
provided_args["component_attention_backends"] = existing
return cls.from_dict(provided_args)
@classmethod
def from_dict(cls, kwargs: dict[str, Any]) -> "ServerArgs":
"""Create a ServerArgs object from a dictionary."""
attrs = [attr.name for attr in dataclasses.fields(cls) if attr.init]
server_args_kwargs: dict[str, Any] = {}
component_paths = dict(kwargs.get("component_paths") or {})
if component_paths:
server_args_kwargs["component_paths"] = component_paths
server_args_kwargs["_explicit_arg_names"] = set(kwargs)
for attr in attrs:
if attr == "pipeline_config":
pipeline_config = PipelineConfig.from_kwargs(kwargs)
logger.debug(f"Using PipelineConfig: {type(pipeline_config)}")
server_args_kwargs["pipeline_config"] = pipeline_config
elif attr == "nunchaku_config":
nunchaku_config = NunchakuSVDQuantArgs.from_dict(kwargs)
server_args_kwargs["nunchaku_config"] = nunchaku_config
elif attr in kwargs:
server_args_kwargs[attr] = kwargs[attr]
return cls(**server_args_kwargs)
@staticmethod
def load_config_file(config_file: str) -> dict[str, Any]:
"""Load a config file."""
if config_file.endswith(".json"):
with open(config_file, "r") as f:
return json.load(f)
elif config_file.endswith((".yaml", ".yml")):
try:
import yaml
except ImportError:
raise ImportError(
"Please install PyYAML to use YAML config files. "
"`pip install pyyaml`"
)
with open(config_file, "r") as f:
return yaml.safe_load(f)
else:
raise ValueError(f"Unsupported config file format: {config_file}")
@classmethod
def from_kwargs(cls, **kwargs: Any) -> "ServerArgs":
explicit_arg_names = set(kwargs)
# Convert backend string to enum if necessary
if "backend" in kwargs and isinstance(kwargs["backend"], str):
kwargs["backend"] = Backend.from_string(kwargs["backend"])
# Convert disagg_role string to enum if necessary
convert_disagg_role_string(kwargs)
kwargs["pipeline_config"] = PipelineConfig.from_kwargs(kwargs)
kwargs["_explicit_arg_names"] = explicit_arg_names
return cls(**kwargs)
@staticmethod
def get_provided_args(
args: argparse.Namespace, unknown_args: list[str]
) -> dict[str, Any]:
"""Get the arguments provided by the user."""
provided_args = {}
# We need to check against the raw command-line arguments to see what was
# explicitly provided by the user, vs. what's a default value from argparse.
raw_argv = sys.argv + unknown_args
# Create a set of argument names that were present on the command line.
# This handles both styles: '--arg=value' and '--arg value'.
provided_arg_names = set()
for arg in raw_argv:
if arg.startswith("--"):
# For '--arg=value', this gets 'arg'; for '--arg', this also gets 'arg'.
arg_name = arg.split("=", 1)[0].replace("-", "_").lstrip("_")
provided_arg_names.add(arg_name)
if "mode" in provided_arg_names:
provided_arg_names.add("performance_mode")
if "layerwise_offload_modules" in provided_arg_names:
provided_arg_names.add("layerwise_offload_components")
# Populate provided_args if the argument from the namespace was on the command line.
for k, v in vars(args).items():
if k in provided_arg_names:
provided_args[k] = v
return provided_args
def _validate_pipeline(self):
if self.pipeline_config is None:
raise ValueError("pipeline_config is not set in ServerArgs")
self.pipeline_config.check_pipeline_config()
def _validate_offload(self):
# validate dit_offload_prefetch_size
if self.dit_offload_prefetch_size > 1 and (
isinstance(self.dit_offload_prefetch_size, float)
and not self.dit_offload_prefetch_size.is_integer()
):
self.dit_offload_prefetch_size = int(
math.floor(self.dit_offload_prefetch_size)
)
logger.info(
f"Invalid --dit-offload-prefetch-size value passed, truncated to: {self.dit_offload_prefetch_size}"
)
if 0.5 <= self.dit_offload_prefetch_size < 1.0:
logger.info(
"We do not recommend --dit-offload-prefetch-size to be between 0.5 and 1.0"
)
# validate layerwise offload conflicts
if self.layerwise_offload_components:
if self.dit_offload_prefetch_size < 0.0:
raise ValueError("dit_offload_prefetch_size must be non-negative")
should_disable_dit_cpu_offload = self.is_dit_layerwise_offload_selected
if self.use_fsdp_inference and should_disable_dit_cpu_offload:
logger.warning(
"layerwise offload is selected for DiT components, automatically disabling use_fsdp_inference."
)
self.use_fsdp_inference = False
if should_disable_dit_cpu_offload and self.dit_cpu_offload is not False:
logger.warning(
"layerwise offload is selected for DiT components, automatically disabling dit_cpu_offload."
)
self.dit_cpu_offload = False
if envs.SGLANG_CACHE_DIT_ENABLED and should_disable_dit_cpu_offload:
raise ValueError(
"DiT layerwise offload cannot be enabled together with cache-dit. "
"cache-dit may reuse skipped blocks whose weights have been released by layerwise offload, "
"causing shape mismatch errors. "
"Please disable --dit-layerwise-offload, remove DiT from --layerwise-offload-components, "
"or disable SGLANG_CACHE_DIT_ENABLED."
)
logger.warning(
"layerwise offload components are selected: %slower GPU memory usage%s, but %smay reduce throughput or increase latency%s. "
"%sIf you are using multi-GPU deployment and already have enough memory headroom, prefer keeping layerwise offload disabled.%s "
"Please tune this based on your memory headroom and performance target.",
GREEN,
RESET,
RED,
RESET,
CYAN,
RESET,
)
def _validate_parallelism(self):
if self.sp_degree > self.num_gpus or self.num_gpus % self.sp_degree != 0:
raise ValueError(
f"num_gpus ({self.num_gpus}) must be >= and divisible by sp_degree ({self.sp_degree})"
)
if (
self.hsdp_replicate_dim > self.num_gpus
or self.num_gpus % self.hsdp_replicate_dim != 0
):
raise ValueError(
f"num_gpus ({self.num_gpus}) must be >= and divisible by hsdp_replicate_dim ({self.hsdp_replicate_dim})"
)
if (
self.hsdp_shard_dim > self.num_gpus
or self.num_gpus % self.hsdp_shard_dim != 0
):
raise ValueError(
f"num_gpus ({self.num_gpus}) must be >= and divisible by hsdp_shard_dim ({self.hsdp_shard_dim})"
)
if self.num_gpus % self.dp_size != 0:
raise ValueError(
f"num_gpus ({self.num_gpus}) must be divisible by dp_size ({self.dp_size})"
)
if self.dp_size < 1:
raise ValueError("--dp-size must be a natural number")
if self.dp_size > 1:
raise ValueError("DP is not yet supported")
num_gpus_per_group = self.dp_size * self.tp_size
if self.enable_cfg_parallel:
num_gpus_per_group *= self.cfg_parallel_degree
if self.num_gpus % num_gpus_per_group != 0:
raise ValueError(
f"num_gpus ({self.num_gpus}) must be divisible by (dp_size * tp_size"
f"{f' * {self.cfg_parallel_degree}' if self.enable_cfg_parallel else ''}"
f") = {num_gpus_per_group}"
)
if self.sp_degree != self.ring_degree * self.ulysses_degree:
raise ValueError(
f"sp_degree ({self.sp_degree}) must equal ring_degree * ulysses_degree "
f"({self.ring_degree} * {self.ulysses_degree} = {self.ring_degree * self.ulysses_degree})"
)
if os.getenv("SGLANG_CACHE_DIT_ENABLED", "").lower() == "true":
has_sp = self.sp_degree > 1
has_tp = self.tp_size > 1
if has_sp and has_tp:
logger.warning(
"cache-dit is enabled with hybrid parallelism (SP + TP). "
"Proceeding anyway (SGLang integration may support this mode)."
)
def _validate_cfg_parallel(self):
if self.enable_cfg_parallel and self.num_gpus == 1:
raise ValueError(
"CFG Parallelism is enabled via `--enable-cfg-parallel`, but num_gpus == 1"
)
def _validate_batching(self):
if self.batching_mode != "dynamic":
raise ValueError("batching_mode must be one of: dynamic")
if self.batching_max_size < 1:
raise ValueError("batching_max_size must be >= 1")
if self.batching_delay_ms < 0:
raise ValueError("batching_delay_ms must be >= 0")
def _set_default_attention_backend(self) -> None:
"""Configure ROCm defaults when users do not specify an attention backend."""
if current_platform.is_rocm():
default_backend = AttentionBackendEnum.AITER.name.lower()
self.attention_backend = default_backend
logger.info(
"Attention backend not specified. Using '%s' by default on ROCm "
"to match SGLang SRT defaults.",
default_backend,
)
@dataclasses.dataclass
class PortArgs:
# The ipc filename for scheduler (rank 0) to receive inputs from tokenizer (zmq)
scheduler_input_ipc_name: str
# The port for nccl initialization (torch.dist)
nccl_port: int
# The ipc filename for rpc call between Engine and Scheduler
rpc_ipc_name: str
# The ipc filename for Scheduler to send metrics
metrics_ipc_name: str
# Master port for distributed inference
master_port: int | None = None
@staticmethod
def from_server_args(
server_args: ServerArgs, dp_rank: Optional[int] = None
) -> "PortArgs":
if server_args.nccl_port is None:
nccl_port = server_args.scheduler_port + random.randint(100, 1000)
while True:
if is_port_available(nccl_port):
break
if nccl_port < 60000:
nccl_port += 42
else:
nccl_port -= 43
else:
nccl_port = server_args.nccl_port
# Normal case, use IPC within a single node
return PortArgs(
scheduler_input_ipc_name=f"ipc://{tempfile.NamedTemporaryFile(delete=False).name}",
nccl_port=nccl_port,
rpc_ipc_name=f"ipc://{tempfile.NamedTemporaryFile(delete=False).name}",
metrics_ipc_name=f"ipc://{tempfile.NamedTemporaryFile(delete=False).name}",
master_port=server_args.master_port,
)
_global_server_args = None
def prepare_server_args(argv: list[str]) -> ServerArgs:
"""
Prepare the inference arguments from the command line arguments.
"""
parser = FlexibleArgumentParser()
ServerArgs.add_cli_args(parser)
raw_args, unknown_args = parser.parse_known_args(argv)
server_args = ServerArgs.from_cli_args(raw_args, unknown_args)
return server_args
def set_global_server_args(server_args: ServerArgs):
"""
Set the global sgl_diffusion config for each process
"""
global _global_server_args
_global_server_args = server_args
def get_global_server_args() -> ServerArgs:
if _global_server_args is None:
# in ci, usually when we test custom ops/modules directly,
# we don't set the sgl_diffusion config. In that case, we set a default
# config.
# TODO(will): may need to handle this for CI.
raise ValueError("Global sgl_diffusion args is not set.")
return _global_server_args