From 4b81f692d1efa9b4f2a4f6777e47f471a3b6d383 Mon Sep 17 00:00:00 2001
From: summer <128961079+zhang-keliang@users.noreply.github.com>
Date: Mon, 7 Sep 2026 15:56:51 +0800
Subject: [PATCH] [Bugfix][Diffusion] Canonicalize cache_backend=None to "none"
 (#7041)

Signed-off-by: summer <128961079+zhang-keliang@users.noreply.github.com>
Co-authored-by: Alicia <115451386+congw729@users.noreply.github.com>
---
 .../diffusion/models/sana_video/test_cache_offload.py |  7 ++++++-
 tests/diffusion/test_diffusion_config_propagation.py  | 11 ++++++++++-
 vllm_omni/diffusion/data.py                           |  5 +++++
 .../models/sana_video/pipeline_sana_video.py          |  3 ++-
 4 files changed, 23 insertions(+), 3 deletions(-)

diff --git a/tests/diffusion/models/sana_video/test_cache_offload.py b/tests/diffusion/models/sana_video/test_cache_offload.py
index 52ced37c8..57417c886 100644
--- a/tests/diffusion/models/sana_video/test_cache_offload.py
+++ b/tests/diffusion/models/sana_video/test_cache_offload.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from dataclasses import dataclass
 from types import SimpleNamespace
@@ -155,6 +155,11 @@ def test_sana_video_rejects_unvalidated_cache_backends():
         _validate_cache_offload_parallelism(OmniDiffusionConfig(cache_backend="tea_cache"))
 
 
+def test_sana_video_accepts_none_cache_backend():
+    """An explicit ``cache_backend=None`` means "no cache", same as ``"none"``."""
+    _validate_cache_offload_parallelism(OmniDiffusionConfig(cache_backend=None))
+
+
 def test_sana_video_rejects_cache_dit_with_distributed_layerwise_offload():
     config = OmniDiffusionConfig(cache_backend="cache_dit", enable_distributed_layerwise_offload=True)
 
diff --git a/tests/diffusion/test_diffusion_config_propagation.py b/tests/diffusion/test_diffusion_config_propagation.py
index 60d0c63b8..164fcfbeb 100644
--- a/tests/diffusion/test_diffusion_config_propagation.py
+++ b/tests/diffusion/test_diffusion_config_propagation.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Tests that parallel_config survives the create_default_diffusion roundtrip.
 
 Regression tests for https://github.com/vllm-project/vllm-omni/issues/1862
@@ -137,6 +137,15 @@ class TestCreateDefaultDiffusion:
         stages = StageConfigFactory.create_default_diffusion({"model": "x"})
         assert stages[0]["engine_args"]["cache_backend"] == "none"
 
+    def test_explicit_none_cache_backend_canonicalizes_to_none(self):
+        """Regression for #7032: ``cache_backend=None`` (e.g. an example CLI default)
+        must reach the pipeline as the canonical ``"none"`` string."""
+        od = _roundtrip_diffusion_config(model="x", cache_backend=None)
+        assert od.cache_backend == "none"
+
+        od = OmniDiffusionConfig.from_kwargs(model="x", cache_backend=None)
+        assert od.cache_backend == "none"
+
     def test_single_gpu_default_devices(self):
         stages = StageConfigFactory.create_default_diffusion({"model": "x"})
         assert stages[0]["runtime"]["devices"] == "0"
diff --git a/vllm_omni/diffusion/data.py b/vllm_omni/diffusion/data.py
index 00764ea46..fd6f0237d 100644
--- a/vllm_omni/diffusion/data.py
+++ b/vllm_omni/diffusion/data.py
@@ -91,6 +91,11 @@ def normalize_omni_diffusion_kwargs(kwargs: Mapping[str, Any]) -> dict[str, Any]
     if "cache_backend" not in normalized:
         cache_backend = os.environ.get("DIFFUSION_CACHE_BACKEND") or os.environ.get("DIFFUSION_CACHE_ADAPTER")
         normalized["cache_backend"] = cache_backend.lower() if cache_backend else "none"
+    elif normalized["cache_backend"] is None:
+        # Callers (e.g. example CLIs with `default=None`) pass an explicit
+        # None for "no cache"; canonicalize it so every consumer sees the
+        # declared `str` value instead of relying on per-model None handling.
+        normalized["cache_backend"] = "none"
 
     # Convert optional YAML null values to empty containers.
     for key in ("diffusers_load_kwargs", "diffusers_call_kwargs"):
diff --git a/vllm_omni/diffusion/models/sana_video/pipeline_sana_video.py b/vllm_omni/diffusion/models/sana_video/pipeline_sana_video.py
index f2d86622a..6c2d740d4 100644
--- a/vllm_omni/diffusion/models/sana_video/pipeline_sana_video.py
+++ b/vllm_omni/diffusion/models/sana_video/pipeline_sana_video.py
@@ -273,7 +273,8 @@ def retrieve_timesteps(
 
 
 def _validate_cache_offload_parallelism(od_config: OmniDiffusionConfig) -> None:
-    cache_backend = od_config.cache_backend
+    # Treat None like "none", matching the other pipelines' cache checks.
+    cache_backend = od_config.cache_backend or "none"
     if cache_backend not in ("none", "cache_dit"):
         raise NotImplementedError(
             f"Cache backend {cache_backend!r} is not supported by the native SANA-Video pipeline; "