diff --git a/python/sglang/multimodal_gen/runtime/cache/spectrum.py b/python/sglang/multimodal_gen/runtime/cache/spectrum.py
index abc03f84f..1bcb9b9de 100644
--- a/python/sglang/multimodal_gen/runtime/cache/spectrum.py
+++ b/python/sglang/multimodal_gen/runtime/cache/spectrum.py
@@ -331,12 +331,12 @@ class SpectrumMixin:
     - ``spectrum_record_features()`` — after a real forward, store block outputs.
     - ``spectrum_predict_features()`` — on skipped steps, return forecasted outputs.
 
-    Models with separate CFG branches (Wan, Hunyuan, SD3) list their config prefix
+    Models with separate CFG branches (Wan, Hunyuan, SD3) list their model prefix
     in ``_CFG_SUPPORTED_PREFIXES`` so cond/uncond maintain independent counters and
     forecasters. All other ``CachableDiT`` subclasses share one counter.
     """
 
-    # DiT config prefixes that run separate cond/uncond forwards (see TeaCache).
+    # DiT model prefixes that run separate cond/uncond forwards (see TeaCache).
     _CFG_SUPPORTED_PREFIXES: set[str] = {"wan", "hunyuan", "sd3"}
 
     def _init_spectrum_state(self) -> None:
@@ -366,8 +366,9 @@ class SpectrumMixin:
         # Runtime branch tracking
         self.spectrum_is_cfg_negative = False
         self._spectrum_ctx: Optional[SpectrumContext] = None
-        prefix = getattr(self.config, "prefix", "").lower()
-        self._spectrum_supports_cfg_cache = prefix in self._CFG_SUPPORTED_PREFIXES
+        self._spectrum_supports_cfg_cache = (
+            self.prefix.lower() in self._CFG_SUPPORTED_PREFIXES
+        )
 
     def reset_spectrum_state(self, spectrum_params: SpectrumParams) -> None:
         self.spectrum_cnt = 0
diff --git a/python/sglang/multimodal_gen/runtime/cache/teacache.py b/python/sglang/multimodal_gen/runtime/cache/teacache.py
index 8830f7ec2..6b71cd9f2 100644
--- a/python/sglang/multimodal_gen/runtime/cache/teacache.py
+++ b/python/sglang/multimodal_gen/runtime/cache/teacache.py
@@ -23,8 +23,6 @@ from typing import TYPE_CHECKING, Any
 import numpy as np
 import torch
 
-from sglang.multimodal_gen.configs.models import DiTConfig
-
 if TYPE_CHECKING:
     from sglang.multimodal_gen.configs.sample.teacache import TeaCacheParams
 
@@ -127,7 +125,7 @@ class TeaCacheMixin:
     # Models that support CFG cache separation (wan/hunyuan/zimage)
     # Models not in this set (flux/qwen) auto-disable TeaCache when CFG is enabled
     _CFG_SUPPORTED_PREFIXES: set[str] = {"wan", "hunyuan", "zimage"}
-    config: DiTConfig
+    prefix: str
 
     def _init_teacache_state(self) -> None:
         """Initialize TeaCache state. Call this in subclass __init__."""
@@ -135,9 +133,7 @@ class TeaCacheMixin:
         self.cnt = 0
         self.enable_teacache = True
         # Flag indicating if this model supports CFG cache separation
-        self._supports_cfg_cache = (
-            self.config.prefix.lower() in self._CFG_SUPPORTED_PREFIXES
-        )
+        self._supports_cfg_cache = self.prefix.lower() in self._CFG_SUPPORTED_PREFIXES
 
         # Always initialize positive cache fields (used in all modes)
         self.previous_modulated_input: torch.Tensor | None = None
diff --git a/python/sglang/multimodal_gen/runtime/layers/layernorm.py b/python/sglang/multimodal_gen/runtime/layers/layernorm.py
index 08e01ae11..d778e1553 100755
--- a/python/sglang/multimodal_gen/runtime/layers/layernorm.py
+++ b/python/sglang/multimodal_gen/runtime/layers/layernorm.py
@@ -452,18 +452,21 @@ class FP32LayerNorm(CustomOp, nn.LayerNorm):
         )
         self._forward_method = self.dispatch_forward()
 
-        try:
-            import attentions  # noqa: F401
-        except ImportError:
-            from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
+        if _is_npu:
+            try:
+                import attentions  # noqa: F401
+            except ImportError:
+                from sglang.multimodal_gen.runtime.utils.logging_utils import (
+                    init_logger,
+                )
 
-            logger = init_logger(__name__)  # pylint: disable=invalid-name
-            logger.warning(
-                "The 'attentions' library is not installed. Falling back to native layernorm. "
-                "Installing this library may improve performance on NPU."
-                "See: sgl-project/sgl-kernel-npu"
-            )
-            self._forward_method = self.forward_native
+                logger = init_logger(__name__)  # pylint: disable=invalid-name
+                logger.warning(
+                    "The 'attentions' library is not installed. Falling back to native layernorm. "
+                    "Installing this library may improve performance on NPU."
+                    "See: sgl-project/sgl-kernel-npu"
+                )
+                self._forward_method = self.forward_native
 
     def _cached_fp32_param(
         self, attr: str, param: torch.Tensor | None, device: torch.device
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/base.py b/python/sglang/multimodal_gen/runtime/models/dits/base.py
index 75094ac98..1cc34150c 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/base.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/base.py
@@ -7,7 +7,7 @@ from typing import Any
 import torch
 from torch import nn
 
-from sglang.multimodal_gen.configs.models import DiTConfig
+from sglang.multimodal_gen.configs.models.dits.base import DiTArchConfig, DiTConfig
 
 # NOTE: SpectrumMixin lives in runtime.cache.spectrum
 from sglang.multimodal_gen.runtime.cache.spectrum import SpectrumMixin
@@ -49,7 +49,10 @@ class BaseDiT(nn.Module, ABC):
 
     def __init__(self, config: DiTConfig, hf_config: dict[str, Any], **kwargs) -> None:
         super().__init__()
-        self.config = config
+        # runtime models expose checkpoint architecture through `config`; load
+        # settings such as the model prefix stay separate
+        self.config: DiTArchConfig = config.arch_config
+        self.prefix = config.prefix
         self.hf_config = hf_config
         if not self.supported_attention_backends:
             raise ValueError(
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/causal_wanvideo.py b/python/sglang/multimodal_gen/runtime/models/dits/causal_wanvideo.py
index 4e8b300c2..872d17fe0 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/causal_wanvideo.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/causal_wanvideo.py
@@ -520,7 +520,7 @@ class CausalWanTransformer3DModel(BaseDiT, LayerwiseOffloadableModuleMixin):
 
         # Causal-specific
         self.block_mask = None
-        self.num_frame_per_block = config.arch_config.num_frames_per_block
+        self.num_frame_per_block = self.config.num_frames_per_block
         # Block size is bounded only by the causal block-mask construction, which
         # supports any positive value.
         assert self.num_frame_per_block >= 1
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/cosmos3video.py b/python/sglang/multimodal_gen/runtime/models/dits/cosmos3video.py
index 10c688517..997972be1 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/cosmos3video.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/cosmos3video.py
@@ -933,7 +933,7 @@ class Cosmos3OmniTransformer(CachableDiT, LayerwiseOffloadableModuleMixin):
     ) -> None:
         super().__init__(config=config, hf_config=hf_config)
 
-        arch = config.arch_config
+        arch = self.config
         self.hidden_size = arch.hidden_size
         self.num_hidden_layers = arch.num_hidden_layers
         self.num_attention_heads = arch.num_attention_heads
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/ernie_image.py b/python/sglang/multimodal_gen/runtime/models/dits/ernie_image.py
index 9e6acfe37..dce4aa256 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/ernie_image.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/ernie_image.py
@@ -427,7 +427,7 @@ class ErnieImageTransformer2DModel(CachableDiT, LayerwiseOffloadableModuleMixin)
     ):
         super().__init__(config=config, hf_config=hf_config)
 
-        arch = config.arch_config
+        arch = self.config
         self.hidden_size = arch.hidden_size
         self.num_attention_heads = arch.num_attention_heads
         self.num_channels_latents = arch.out_channels
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/flux.py b/python/sglang/multimodal_gen/runtime/models/dits/flux.py
index 7786b2827..12d671532 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/flux.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/flux.py
@@ -1202,7 +1202,6 @@ class FluxTransformer2DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
         quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__(config=config, hf_config=hf_config)
-        self.config = config.arch_config
 
         self.out_channels = (
             getattr(self.config, "out_channels", None) or self.config.in_channels
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/glm_image.py b/python/sglang/multimodal_gen/runtime/models/dits/glm_image.py
index ad2e10805..b26065c62 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/glm_image.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/glm_image.py
@@ -945,8 +945,7 @@ class GlmImageTransformer2DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
     ):
         super().__init__(config=config, hf_config=hf_config)
 
-        self.config_data = config  # Store config
-        arch_config = config.arch_config
+        arch_config = self.config
 
         self.in_channels = arch_config.in_channels
         self.out_channels = arch_config.out_channels
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/hunyuan3d.py b/python/sglang/multimodal_gen/runtime/models/dits/hunyuan3d.py
index 3474e76e0..dd58583f3 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/hunyuan3d.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/hunyuan3d.py
@@ -487,7 +487,7 @@ class Hunyuan3D2DiT(CachableDiT, LayerwiseOffloadableModuleMixin):
         **kwargs,
     ):
         super().__init__(config=config, hf_config=hf_config or {}, **kwargs)
-        arch = config.arch_config
+        arch = self.config
 
         in_channels = arch.in_channels
         context_in_dim = arch.context_in_dim
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/ideogram.py b/python/sglang/multimodal_gen/runtime/models/dits/ideogram.py
index 4e48932a9..d42a3a466 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/ideogram.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/ideogram.py
@@ -513,7 +513,7 @@ class Ideogram4Transformer2DModel(BaseDiT, LayerwiseOffloadableModuleMixin):
         **kwargs,
     ) -> None:
         super().__init__(config, hf_config, **kwargs)
-        cfg = config.arch_config
+        cfg = self.config
         use_weight_only_fp8_linears = config.use_weight_only_fp8_linears
         self._supported_attention_backends = cfg._supported_attention_backends
         hidden_size = cfg.num_attention_heads * cfg.attention_head_dim
@@ -586,7 +586,7 @@ class Ideogram4Transformer2DModel(BaseDiT, LayerwiseOffloadableModuleMixin):
     def post_load_weights(self) -> None:
         if not self.rotary_emb.inv_freq.is_meta:
             return
-        cfg = self.config.arch_config
+        cfg = self.config
         inv_freq = 1.0 / (
             cfg.rope_theta
             ** (
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/krea2.py b/python/sglang/multimodal_gen/runtime/models/dits/krea2.py
index 3ab91d0bd..c3d986ee6 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/krea2.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/krea2.py
@@ -522,7 +522,7 @@ class Krea2Transformer2DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
         quant_config: Optional[Any] = None,
     ) -> None:
         super().__init__(config=config, hf_config=hf_config)
-        ac = config.arch_config
+        ac = self.config
         self.arch_config = ac
 
         self.hidden_size = ac.features
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/ltx_2.py b/python/sglang/multimodal_gen/runtime/models/dits/ltx_2.py
index 77063c6c3..2992da867 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/ltx_2.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/ltx_2.py
@@ -1518,7 +1518,7 @@ class LTX2VideoTransformer3DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
         return timestep.amax(dim=tuple(range(1, timestep.ndim)))
 
     def _scale_timestep_for_adaln(self, timestep: torch.Tensor) -> torch.Tensor:
-        ltx_variant = str(getattr(self.config.arch_config, "ltx_variant", "ltx_2"))
+        ltx_variant = str(getattr(self.config, "ltx_variant", "ltx_2"))
         if ltx_variant == "ltx_2_3" and bool(
             getattr(self, "_sglang_use_ltx23_hq_timestep_semantics", False)
         ):
@@ -1583,7 +1583,7 @@ class LTX2VideoTransformer3DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
     ) -> None:
         super().__init__(config=config, hf_config=hf_config)
 
-        arch = config.arch_config
+        arch = self.config
         self.hidden_size = arch.hidden_size
         self.num_attention_heads = arch.num_attention_heads
         self.audio_hidden_size = arch.audio_hidden_size
@@ -1844,7 +1844,7 @@ class LTX2VideoTransformer3DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
         return video_coords.to(device=hidden_device)
 
     def _get_av_ca_gate_timestep_factor(self) -> float:
-        ltx_variant = str(getattr(self.config.arch_config, "ltx_variant", "ltx_2"))
+        ltx_variant = str(getattr(self.config, "ltx_variant", "ltx_2"))
         if ltx_variant == "ltx_2_3":
             return self.av_ca_timestep_scale_multiplier / self.timestep_scale_multiplier
         return float(self.av_ca_timestep_scale_multiplier)
@@ -1856,7 +1856,7 @@ class LTX2VideoTransformer3DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
         prompt_timestep: torch.Tensor | None,
         audio_prompt_timestep: torch.Tensor | None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
-        ltx_variant = str(getattr(self.config.arch_config, "ltx_variant", "ltx_2"))
+        ltx_variant = str(getattr(self.config, "ltx_variant", "ltx_2"))
         if ltx_variant != "ltx_2_3":
             return timestep, audio_timestep
 
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/minimax_h3.py b/python/sglang/multimodal_gen/runtime/models/dits/minimax_h3.py
index d9524a0c2..947e9090d 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/minimax_h3.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/minimax_h3.py
@@ -1109,7 +1109,7 @@ class MiniMaxH3DiTModel(BaseDiT, LayerwiseOffloadableModuleMixin):
         quant_config: QuantizationConfig | None = None,
     ) -> None:
         super().__init__(config=config, hf_config=hf_config)
-        arch = config.arch_config
+        arch = self.config
         self.arch = arch
         self.hidden_size = arch.hidden_size
         self.num_attention_heads = arch.num_attention_heads
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/qwen_image.py b/python/sglang/multimodal_gen/runtime/models/dits/qwen_image.py
index 70e0cd4b3..6a1d68a17 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/qwen_image.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/qwen_image.py
@@ -1371,23 +1371,24 @@ class QwenImageTransformer2DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
         quant_config: Optional[QuantizationConfig] = None,
     ):
         super().__init__(config=config, hf_config=hf_config)
-        patch_size = config.arch_config.patch_size
-        in_channels = config.arch_config.in_channels
-        out_channels = config.arch_config.out_channels
-        num_layers = config.arch_config.num_layers
-        attention_head_dim = config.arch_config.attention_head_dim
-        num_attention_heads = config.arch_config.num_attention_heads
-        joint_attention_dim = config.arch_config.joint_attention_dim
-        axes_dims_rope = config.arch_config.axes_dims_rope
-        self.zero_cond_t = getattr(config.arch_config, "zero_cond_t", False)
+        arch = self.config
+        patch_size = arch.patch_size
+        in_channels = arch.in_channels
+        out_channels = arch.out_channels
+        num_layers = arch.num_layers
+        attention_head_dim = arch.attention_head_dim
+        num_attention_heads = arch.num_attention_heads
+        joint_attention_dim = arch.joint_attention_dim
+        axes_dims_rope = arch.axes_dims_rope
+        self.zero_cond_t = getattr(arch, "zero_cond_t", False)
         self.out_channels = out_channels or in_channels
         self.inner_dim = num_attention_heads * attention_head_dim
 
         self.use_additional_t_cond: bool = getattr(
-            config.arch_config, "use_additional_t_cond", False
+            arch, "use_additional_t_cond", False
         )  # For qwen-image-layered now
         self.use_layer3d_rope: bool = getattr(
-            config.arch_config, "use_layer3d_rope", False
+            arch, "use_layer3d_rope", False
         )  # For qwen-image-layered now
 
         if not self.use_layer3d_rope:
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/sana.py b/python/sglang/multimodal_gen/runtime/models/dits/sana.py
index dcc1dd2b1..98c3a693f 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/sana.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/sana.py
@@ -401,7 +401,7 @@ class SanaTransformer2DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
     def __init__(self, config: SanaConfig, hf_config=None, **kwargs):
         super().__init__(config, hf_config=hf_config or {}, **kwargs)
 
-        arch = config.arch_config
+        arch = self.config
         self.out_channels = arch.out_channels
         self.patch_size = arch.patch_size
         self.inner_dim = arch.num_attention_heads * arch.attention_head_dim
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/sana_wm.py b/python/sglang/multimodal_gen/runtime/models/dits/sana_wm.py
index 387324dec..c029d3c94 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/sana_wm.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/sana_wm.py
@@ -353,7 +353,7 @@ class SanaWMTransformer3DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
 
     def __init__(self, config: SanaWMConfig, hf_config=None, **kwargs) -> None:
         super().__init__(config, hf_config=hf_config or {}, **kwargs)
-        arch = config.arch_config
+        arch = self.config
 
         self.patch_size = (arch.patch_size_t, arch.patch_size, arch.patch_size)
         self.inner_dim = arch.num_attention_heads * arch.attention_head_dim
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/sana_wm_refiner_transformer.py b/python/sglang/multimodal_gen/runtime/models/dits/sana_wm_refiner_transformer.py
index 8d9b353b9..9ee6f888a 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/sana_wm_refiner_transformer.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/sana_wm_refiner_transformer.py
@@ -226,7 +226,7 @@ class SanaWMLTX2VideoRefiner(CachableDiT, LayerwiseOffloadableModuleMixin):
         quant_config: QuantizationConfig | None = None,
     ) -> None:
         super().__init__(config, hf_config=hf_config)
-        arch = config.arch_config
+        arch = self.config
 
         self.in_channels = int(arch.in_channels)
         self.out_channels = int(arch.out_channels)
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/stablediffusion3.py b/python/sglang/multimodal_gen/runtime/models/dits/stablediffusion3.py
index 82236c919..8bcb04cf2 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/stablediffusion3.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/stablediffusion3.py
@@ -40,8 +40,7 @@ class SD3Transformer2DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
         quant_config=None,
     ):
         super().__init__(config=config, hf_config=hf_config)
-        self.config = config
-        arch_config = config.arch_config
+        arch_config = self.config
         sample_size = arch_config.sample_size
         patch_size = arch_config.patch_size
         in_channels = arch_config.in_channels
diff --git a/python/sglang/multimodal_gen/runtime/models/dits/zimage.py b/python/sglang/multimodal_gen/runtime/models/dits/zimage.py
index 52541cb01..16ca5c720 100644
--- a/python/sglang/multimodal_gen/runtime/models/dits/zimage.py
+++ b/python/sglang/multimodal_gen/runtime/models/dits/zimage.py
@@ -808,8 +808,7 @@ class ZImageTransformer2DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
     ) -> None:
         super().__init__(config=config, hf_config=hf_config)
 
-        self.config_data = config  # Store config
-        arch_config = config.arch_config
+        arch_config = self.config
 
         self.in_channels = arch_config.in_channels
         self.out_channels = arch_config.out_channels
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/causal_denoising.py b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/causal_denoising.py
index fcff3cdea..44408e550 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/causal_denoising.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/causal_denoising.py
@@ -157,12 +157,10 @@ class CausalDMDDenoisingStage(DenoisingStage):
         self.causal_kv_cache_neg: list | None = None
         self.crossattn_cache_neg: list | None = None
         # Model-dependent constants (aligned with causal_inference.py assumptions)
-        self.num_transformer_blocks = self.transformer.config.arch_config.num_layers
-        self.num_frames_per_block = (
-            self.transformer.config.arch_config.num_frames_per_block
-        )
+        self.num_transformer_blocks = self.transformer.config.num_layers
+        self.num_frames_per_block = self.transformer.config.num_frames_per_block
         self.sliding_window_num_frames = (
-            self.transformer.config.arch_config.sliding_window_num_frames
+            self.transformer.config.sliding_window_num_frames
         )
 
         try:
@@ -171,7 +169,7 @@ class CausalDMDDenoisingStage(DenoisingStage):
             )  # type: ignore
         except Exception:
             self.local_attn_size = -1
-        self.sink_size = self.transformer.config.arch_config.sink_size
+        self.sink_size = self.transformer.config.sink_size
 
         self._causal_attn_metadata_builder_cls = None
         self._causal_attn_metadata_builder = None
@@ -190,8 +188,8 @@ class CausalDMDDenoisingStage(DenoisingStage):
 
     def _prepare_frame_seq_length(self, h: int, w: int) -> int:
         patch_ratio = (
-            self.transformer.config.arch_config.patch_size[-1]
-            * self.transformer.config.arch_config.patch_size[-2]
+            self.transformer.config.patch_size[-1]
+            * self.transformer.config.patch_size[-2]
         )
         self.num_token_per_frame = (h * w) // patch_ratio
         return self.num_token_per_frame
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/realtime/latent_preparation.py b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/realtime/latent_preparation.py
index c4d32f6eb..343866a54 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/realtime/latent_preparation.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/realtime/latent_preparation.py
@@ -27,8 +27,7 @@ class RealtimeChunkLatentPreparationStage(LatentPreparationStage):
         server_args: ServerArgs,
     ) -> int:
         return int(
-            batch.realtime_chunk_size
-            or self.transformer.config.arch_config.num_frames_per_block
+            batch.realtime_chunk_size or self.transformer.config.num_frames_per_block
         )
 
     def get_latent_preparation_spec(
@@ -47,7 +46,7 @@ class RealtimeChunkLatentPreparationStage(LatentPreparationStage):
         return LatentPreparationSpec(
             shape=(
                 condition_latent.shape[0],
-                self.transformer.config.arch_config.out_channels,
+                self.transformer.config.out_channels,
                 num_frames,
                 condition_latent.shape[3],
                 condition_latent.shape[4],
diff --git a/python/sglang/multimodal_gen/test/unit/test_dit_config_boundary.py b/python/sglang/multimodal_gen/test/unit/test_dit_config_boundary.py
new file mode 100644
index 000000000..39f893d79
--- /dev/null
+++ b/python/sglang/multimodal_gen/test/unit/test_dit_config_boundary.py
@@ -0,0 +1,25 @@
+import torch
+
+from sglang.multimodal_gen.configs.models.dits.base import DiTConfig
+from sglang.multimodal_gen.runtime.models.dits.base import CachableDiT
+
+
+class _TestDiT(CachableDiT):
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        **kwargs,
+    ) -> torch.Tensor:
+        return hidden_states
+
+
+def test_dit_runtime_keeps_architecture_and_component_config_separate():
+    component_config = DiTConfig(prefix="Wan")
+    model = _TestDiT(config=component_config, hf_config={})
+
+    assert model.config is component_config.arch_config
+    assert model.prefix == "Wan"
+    assert model._supports_cfg_cache
+    assert model._spectrum_supports_cfg_cache