chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[diffusion] chore: remove useless params (#15925)

Browse Source
This commit is contained in:
Yuhao Yang
2025-12-28 01:01:08 +08:00
committed by GitHub
parent 39d56196a0
commit 0cd2b719a5
4 changed files with 15 additions and 118 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/multimodal_gen/runtime/loader/component_loader.py
View File

@@ -668,10 +668,6 @@ class TransformerLoader(ComponentLoader):
"Only diffusers format is supported."
)
if server_args.override_transformer_cls_name is not None:
cls_name = server_args.override_transformer_cls_name
logger.info("Overriding transformer cls_name to %s", cls_name)
server_args.model_paths["transformer"] = component_model_path
# Config from Diffusers supersedes sgl_diffusion's model config

33
python/sglang/multimodal_gen/runtime/pipelines_core/stages/base.py
View File

@@ -191,31 +191,26 @@ class PipelineStage(ABC):
"""
stage_name = self.__class__.__name__
# Check if verification is enabled (simple approach for prototype)
enable_verification = getattr(server_args, "enable_stage_verification", False)
if enable_verification:
# Pre-execution input verification
try:
input_result = self.verify_input(batch, server_args)
self._run_verification(input_result, stage_name, "input")
except Exception as e:
logger.error("Input verification failed for %s: %s", stage_name, str(e))
raise
# Pre-execution input verification
try:
input_result = self.verify_input(batch, server_args)
self._run_verification(input_result, stage_name, "input")
except Exception as e:
logger.error("Input verification failed for %s: %s", stage_name, str(e))
raise
# Execute the actual stage logic with unified profiling
with StageProfiler(stage_name, logger=logger, timings=batch.timings):
result = self.forward(batch, server_args)
if enable_verification:
# Post-execution output verification
try:
output_result = self.verify_output(result, server_args)
self._run_verification(output_result, stage_name, "output")
except Exception as e:
logger.error(
"Output verification failed for %s: %s", stage_name, str(e)
)
raise
# Post-execution output verification
try:
output_result = self.verify_output(result, server_args)
self._run_verification(output_result, stage_name, "output")
except Exception as e:
logger.error("Output verification failed for %s: %s", stage_name, str(e))
raise
return result

6
python/sglang/multimodal_gen/runtime/pipelines_core/stages/decoding.py
View File

@@ -200,7 +200,6 @@ class DecodingStage(PipelineStage):
- trajectory_latents (optional): Latents at different timesteps
- trajectory_timesteps (optional): Corresponding timesteps
server_args: Configuration containing:
- output_type: "latent" to skip decoding, otherwise decode to pixels
- vae_cpu_offload: Whether to offload VAE to CPU after decoding
- model_loaded: Track VAE loading state
- model_paths: Path to VAE model if loading needed
@@ -213,10 +212,7 @@ class DecodingStage(PipelineStage):
# load vae if not already loaded (used for memory constrained devices)
self.load_model()
if server_args.output_type == "latent":
frames = batch.latents
else:
frames = self.decode(batch.latents, server_args)
frames = self.decode(batch.latents, server_args)
# decode trajectory latents if needed
if batch.return_trajectory_decoded:

90
python/sglang/multimodal_gen/runtime/server_args.py
View File

@@ -145,31 +145,6 @@ def _sanitize_for_logging(obj: Any, key_hint: str | None = None) -> Any:
return "<unserializable>"
class ExecutionMode(str, Enum):
"""
Enumeration for different pipeline modes.
Inherits from str to allow string comparison for backward compatibility.
"""
INFERENCE = "inference"
@classmethod
def from_string(cls, value: str) -> "ExecutionMode":
"""Convert string to ExecutionMode enum."""
try:
return cls(value.lower())
except ValueError:
raise ValueError(
f"Invalid mode: {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 [mode.value for mode in cls]
@dataclasses.dataclass
class ServerArgs:
# Model and path configuration (for convenience)
@@ -178,14 +153,7 @@ class ServerArgs:
# Attention
attention_backend: str = None
# Running mode
mode: ExecutionMode = ExecutionMode.INFERENCE
# Cache strategy
cache_strategy: str = "none"
# Distributed executor backend
distributed_executor_backend: str = "mp"
nccl_port: Optional[int] = None
# HuggingFace specific parameters
@@ -224,8 +192,6 @@ class ServerArgs:
# Will adapt only q, k, v, o by default.
lora_target_modules: list[str] | None = None
output_type: str = "pil"
# CPU offload parameters
dit_cpu_offload: bool = True
dit_layerwise_offload: bool = False
@@ -266,9 +232,6 @@ class ServerArgs:
scheduler_port: int = 5555
# Stage verification
enable_stage_verification: bool = True
# Prompt text file for batch processing
prompt_file_path: str | None = None
@@ -280,7 +243,6 @@ class ServerArgs:
"vae": True,
}
)
override_transformer_cls_name: str | None = None
# # DMD parameters
# dmd_denoising_steps: List[int] | None = field(default=None)
@@ -369,24 +331,6 @@ class ServerArgs:
help="The attention backend to use. If not specified, the backend is automatically selected based on hardware and installed packages.",
)
# Running mode
parser.add_argument(
"--mode",
type=str,
choices=ExecutionMode.choices(),
default=ServerArgs.mode.value,
help="The mode to run SGLang-diffusion",
)
# distributed_executor_backend
parser.add_argument(
"--distributed-executor-backend",
type=str,
choices=["mp"],
default=ServerArgs.distributed_executor_backend,
help="The distributed executor backend to use",
)
# HuggingFace specific parameters
parser.add_argument(
"--trust-remote-code",
@@ -466,15 +410,6 @@ class ServerArgs:
help="Set timeout for torch.distributed initialization.",
)
# Output type
parser.add_argument(
"--output-type",
type=str,
default=ServerArgs.output_type,
choices=["pil"],
help="Output type for the generated video",
)
# Prompt text file for batch processing
parser.add_argument(
"--prompt-file-path",
@@ -600,19 +535,6 @@ class ServerArgs:
help="Whether to use webui for better display",
)
# Stage verification
parser.add_argument(
"--enable-stage-verification",
action=StoreBoolean,
default=ServerArgs.enable_stage_verification,
help="Enable input/output verification for pipeline stages",
)
parser.add_argument(
"--override-transformer-cls-name",
type=str,
default=ServerArgs.override_transformer_cls_name,
help="Override transformer cls name",
)
# LoRA
parser.add_argument(
"--lora-path",
@@ -760,10 +682,6 @@ class ServerArgs:
@classmethod
def from_kwargs(cls, **kwargs: Any) -> "ServerArgs":
# Convert mode string to enum if necessary
if "mode" in kwargs and isinstance(kwargs["mode"], str):
kwargs["mode"] = ExecutionMode.from_string(kwargs["mode"])
kwargs["pipeline_config"] = PipelineConfig.from_kwargs(kwargs)
return cls(**kwargs)
@@ -885,14 +803,6 @@ class ServerArgs:
else:
self.disable_autocast = False
# Validate mode consistency
assert isinstance(
self.mode, ExecutionMode
), f"Mode must be an ExecutionMode enum, got {type(self.mode)}"
assert (
self.mode in ExecutionMode.choices()
), f"Invalid execution mode: {self.mode}"
if self.tp_size == -1:
self.tp_size = 1