From 1efc33c64038f5efc3c47a0f209039dff371fea6 Mon Sep 17 00:00:00 2001
From: Adarsh Shirawalmath <114558126+adarshxs@users.noreply.github.com>
Date: Tue, 24 Feb 2026 18:51:23 +0530
Subject: [PATCH] [diffusion] refactor: remove enums and unify attention
 backends (#19149)

---
 .../pipeline_configs/diffusers_generic.py     |  4 --
 .../runtime/pipelines/diffusers_pipeline.py   |  5 --
 .../multimodal_gen/runtime/server_args.py     | 66 -------------------
 3 files changed, 75 deletions(-)

diff --git a/python/sglang/multimodal_gen/configs/pipeline_configs/diffusers_generic.py b/python/sglang/multimodal_gen/configs/pipeline_configs/diffusers_generic.py
index 3eac5306f..96ab3c473 100644
--- a/python/sglang/multimodal_gen/configs/pipeline_configs/diffusers_generic.py
+++ b/python/sglang/multimodal_gen/configs/pipeline_configs/diffusers_generic.py
@@ -50,10 +50,6 @@ class DiffusersGenericPipelineConfig(PipelineConfig):
     vae_slicing: bool = False  # slice VAE decode for lower memory usage
     vae_sp: bool = False
 
-    # Attention backend for diffusers models (e.g., "flash", "_flash_3_hub", "sage", "xformers")
-    # See: https://huggingface.co/docs/diffusers/main/en/optimization/attention_backends
-    diffusers_attention_backend: str | None = None
-
     # Quantization config for pipeline-level quantization
     # See: https://huggingface.co/docs/diffusers/main/en/quantization/overview
     # Use PipelineQuantizationConfig for component-level control:
diff --git a/python/sglang/multimodal_gen/runtime/pipelines/diffusers_pipeline.py b/python/sglang/multimodal_gen/runtime/pipelines/diffusers_pipeline.py
index fc8e3461a..c0d6bbfa7 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines/diffusers_pipeline.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines/diffusers_pipeline.py
@@ -490,11 +490,6 @@ class DiffusersPipeline(ComposedPipelineBase):
         """
         backend = server_args.attention_backend
 
-        if backend is None:
-            config = server_args.pipeline_config
-            if config is not None:
-                backend = getattr(config, "diffusers_attention_backend", None)
-
         if backend is None:
             return
 
diff --git a/python/sglang/multimodal_gen/runtime/server_args.py b/python/sglang/multimodal_gen/runtime/server_args.py
index 2539b87c3..72126e0ac 100644
--- a/python/sglang/multimodal_gen/runtime/server_args.py
+++ b/python/sglang/multimodal_gen/runtime/server_args.py
@@ -152,59 +152,6 @@ 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]
-
-
-class WorkloadType(str, Enum):
-    """
-    Enumeration for different workload types.
-
-    Inherits from str to allow string comparison for backward compatibility.
-    """
-
-    I2V = "i2v"  # Image to Video
-    T2V = "t2v"  # Text to Video
-    T2I = "t2i"  # Text to Image
-    I2I = "i2i"  # Image to Image
-
-    @classmethod
-    def from_string(cls, value: str) -> "WorkloadType":
-        """Convert string to WorkloadType enum."""
-        try:
-            return cls(value.lower())
-        except ValueError:
-            raise ValueError(
-                f"Invalid workload type: {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 [workload.value for workload in cls]
-
-
 class Backend(str, Enum):
     """
     Enumeration for different model backends.
@@ -648,13 +595,6 @@ class ServerArgs:
             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(
-            "--diffusers-attention-backend",
-            type=str,
-            dest="attention_backend",
-            default=None,
-            help=argparse.SUPPRESS,
-        )
         parser.add_argument(
             "--cache-dit-config",
             type=str,
@@ -1070,12 +1010,6 @@ 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"])
-        # Convert workload_type string to enum if necessary
-        if "workload_type" in kwargs and isinstance(kwargs["workload_type"], str):
-            kwargs["workload_type"] = WorkloadType.from_string(kwargs["workload_type"])
         # Convert backend string to enum if necessary
         if "backend" in kwargs and isinstance(kwargs["backend"], str):
             kwargs["backend"] = Backend.from_string(kwargs["backend"])