From 4b6d44641b7eec6a3268ed066ff4041a23bf0125 Mon Sep 17 00:00:00 2001
From: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com>
Date: Mon, 4 May 2026 22:59:31 +0800
Subject: [PATCH] [diffusion] chore: enable channels-last 3D VAE convs by
 default (#23200)

---
 python/sglang/multimodal_gen/envs.py                      | 5 +++--
 .../sglang/multimodal_gen/test/server/accuracy_config.py  | 3 +++
 .../sglang/multimodal_gen/test/server/accuracy_utils.py   | 8 ++++++++
 3 files changed, 14 insertions(+), 2 deletions(-)

diff --git a/python/sglang/multimodal_gen/envs.py b/python/sglang/multimodal_gen/envs.py
index 0f110df3b..1239acdc5 100644
--- a/python/sglang/multimodal_gen/envs.py
+++ b/python/sglang/multimodal_gen/envs.py
@@ -56,7 +56,8 @@ if TYPE_CHECKING:
     # model loading
     SGLANG_USE_RUNAI_MODEL_STREAMER: bool = True
     SGLANG_DIFFUSION_FLASHINFER_FP4_GEMM_BACKEND: str | None = None
-    SGLANG_DIFFUSION_VAE_CHANNELS_LAST_3D: bool = False
+    SGLANG_DIFFUSION_VAE_CHANNELS_LAST_3D: bool = True
+    SGLANG_USE_CUDA_HUNYUANVIDEO_GROUP_NORM_SILU: bool = False
     SGLANG_USE_ROCM_VAE: bool = False
     SGLANG_USE_ROCM_CUDNN_BENCHMARK: bool = False
 
@@ -248,7 +249,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # 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"
+        "SGLANG_DIFFUSION_VAE_CHANNELS_LAST_3D", "true"
     ),
     # ================== cache-dit Env Vars ==================
     # Enable cache-dit acceleration for DiT inference
diff --git a/python/sglang/multimodal_gen/test/server/accuracy_config.py b/python/sglang/multimodal_gen/test/server/accuracy_config.py
index 5eecb88d1..59ac1074b 100644
--- a/python/sglang/multimodal_gen/test/server/accuracy_config.py
+++ b/python/sglang/multimodal_gen/test/server/accuracy_config.py
@@ -297,6 +297,9 @@ SKIP_COMPONENTS: Dict[str, Dict[ComponentType, ComponentSkip]] = {
         ),
     },
     "mova_360p_ring1_uly2": {
+        ComponentType.VAE: ComponentSkip(
+            "Representative MOVA VAE accuracy is covered by mova_360p_tp2; ring/ulysses topology does not exercise a distinct VAE component"
+        ),
         ComponentType.TRANSFORMER: ComponentSkip(
             "HF reference transformer cannot be materialized from the MOVA video_dit repo layout"
         ),
diff --git a/python/sglang/multimodal_gen/test/server/accuracy_utils.py b/python/sglang/multimodal_gen/test/server/accuracy_utils.py
index 6da2d2c75..5d6dcf385 100644
--- a/python/sglang/multimodal_gen/test/server/accuracy_utils.py
+++ b/python/sglang/multimodal_gen/test/server/accuracy_utils.py
@@ -752,6 +752,14 @@ def _run_staged_native_component_accuracy_case(
         engine_cls.clear_memory()
 
         ref = ref.to(device=device, dtype=torch.bfloat16).eval()
+        if component == ComponentType.VAE:
+            from sglang.multimodal_gen import envs
+            from sglang.multimodal_gen.runtime.loader.component_loaders.vae_loader import (
+                _convert_conv3d_weights_to_channels_last_3d,
+            )
+
+            if torch.cuda.is_available() and envs.SGLANG_DIFFUSION_VAE_CHANNELS_LAST_3D:
+                _convert_conv3d_weights_to_channels_last_3d(ref)
         ref_call = profile.prepare_reference_call(ref, inputs)
         ref_autocast = (
             torch.autocast(