diff --git a/python/sglang/multimodal_gen/envs.py b/python/sglang/multimodal_gen/envs.py
index a5a98981b..80ee489ad 100644
--- a/python/sglang/multimodal_gen/envs.py
+++ b/python/sglang/multimodal_gen/envs.py
@@ -55,6 +55,7 @@ if TYPE_CHECKING:
     SGLANG_CACHE_DIT_SECONDARY_TS_ORDER: int = 1
     # model loading
     SGLANG_USE_RUNAI_MODEL_STREAMER: bool = True
+    SGLANG_DIFFUSION_VAE_CHANNELS_LAST_3D: bool = False
 
 
 def get_default_cache_root() -> str:
@@ -243,6 +244,9 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # If set, sgl_diffusion will enable stage logging, which will print the time
     # taken for each stage
     "SGLANG_DIFFUSION_STAGE_LOGGING": _lazy_bool("SGLANG_DIFFUSION_STAGE_LOGGING"),
+    "SGLANG_DIFFUSION_VAE_CHANNELS_LAST_3D": _lazy_bool(
+        "SGLANG_DIFFUSION_VAE_CHANNELS_LAST_3D", "false"
+    ),
     # ================== cache-dit Env Vars ==================
     # Enable cache-dit acceleration for DiT inference
     "SGLANG_CACHE_DIT_ENABLED": _lazy_bool("SGLANG_CACHE_DIT_ENABLED"),
diff --git a/python/sglang/multimodal_gen/runtime/loader/component_loaders/vae_loader.py b/python/sglang/multimodal_gen/runtime/loader/component_loaders/vae_loader.py
index 9c32f143c..47770a142 100644
--- a/python/sglang/multimodal_gen/runtime/loader/component_loaders/vae_loader.py
+++ b/python/sglang/multimodal_gen/runtime/loader/component_loaders/vae_loader.py
@@ -1,8 +1,11 @@
 import importlib.util
 import os
 
+import torch
+import torch.nn as nn
 from safetensors.torch import load_file as safetensors_load_file
 
+from sglang.multimodal_gen import envs
 from sglang.multimodal_gen.configs.models import ModelConfig
 from sglang.multimodal_gen.runtime.loader.component_loaders.component_loader import (
     ComponentLoader,
@@ -23,6 +26,25 @@ from sglang.multimodal_gen.utils import PRECISION_TO_TYPE
 logger = init_logger(__name__)
 
 
+def _convert_conv3d_weights_to_channels_last_3d(module: nn.Module) -> int:
+    """
+    Convert Conv3d weights to channels_last_3d (NDHWC) memory format.
+    Returns the number of Conv3d modules converted.
+    """
+    if not hasattr(torch, "channels_last_3d"):
+        return 0
+    num_converted = 0
+    for m in module.modules():
+        if isinstance(m, nn.Conv3d):
+            try:
+                m.weight.data = m.weight.data.to(memory_format=torch.channels_last_3d)
+                num_converted += 1
+            except Exception:
+                # Best-effort; skip unsupported cases.
+                continue
+    return num_converted
+
+
 class VAELoader(ComponentLoader):
     """Shared loader for (video/audio) VAE modules."""
 
@@ -85,6 +107,16 @@ class VAELoader(ComponentLoader):
                     trust_remote_code=server_args.trust_remote_code,
                 )
             vae = vae.to(device=target_device, dtype=vae_dtype)
+            if (
+                component_name in ("vae", "video_vae")
+                and torch.cuda.is_available()
+                and getattr(envs, "SGLANG_DIFFUSION_VAE_CHANNELS_LAST_3D", False)
+            ):
+                n = _convert_conv3d_weights_to_channels_last_3d(vae)
+                if n > 0:
+                    logger.info(
+                        "VAE: converted %d Conv3d weights to channels_last_3d", n
+                    )
             return vae
 
         # Load from ModelRegistry (standard VAE classes)
@@ -111,4 +143,13 @@ class VAELoader(ComponentLoader):
         if unexpected_keys:
             logger.warning("VAE unexpected keys: %s", unexpected_keys)
 
+        if (
+            component_name in ("vae", "video_vae")
+            and torch.cuda.is_available()
+            and getattr(envs, "SGLANG_DIFFUSION_VAE_CHANNELS_LAST_3D", False)
+        ):
+            n = _convert_conv3d_weights_to_channels_last_3d(vae)
+            if n > 0:
+                logger.info("VAE: converted %d Conv3d weights to channels_last_3d", n)
+
         return vae