From a5e03e2f730ad7a9fee0991c4233ef66806557df Mon Sep 17 00:00:00 2001
From: Solaris-star <820622658@qq.com>
Date: Mon, 20 Jul 2026 23:16:46 +0800
Subject: [PATCH 1/2] fix: do not leak set_attention_backend into process-wide
 registry

Model.set_attention_backend() already pins the backend on each of this
model's attention processors. Flipping _AttentionBackendRegistry's
active backend made other models with unset per-module backends pick up
the same backend (e.g. FA3 + attn_mask on an unrelated model).

Leave the global active backend alone; use attention_backend() when a
temporary process-wide override is intended.

Fixes #14249

Signed-off-by: Solaris-star <820622658@qq.com>
---
 src/diffusers/models/modeling_utils.py | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/src/diffusers/models/modeling_utils.py b/src/diffusers/models/modeling_utils.py
index d79994de6efb..2e21a3594fde 100644
--- a/src/diffusers/models/modeling_utils.py
+++ b/src/diffusers/models/modeling_utils.py
@@ -644,8 +644,10 @@ def set_attention_backend(self, backend: str) -> None:
                 continue
             processor._attention_backend = backend
 
-        # Important to set the active backend so that it propagates gracefully throughout.
-        _AttentionBackendRegistry.set_active_backend(backend)
+        # Only pin the backend on this model's attention modules.
+        # Do not flip the process-wide active backend: that would leak into
+        # other models whose per-module backend is still unset (see #14249).
+        # Use `attention_backend(...)` when a temporary global override is needed.
 
     def reset_attention_backend(self) -> None:
         """

From 01dbfd8716171dcde4f31aedebd3bd2f2ec2f828 Mon Sep 17 00:00:00 2001
From: Solaris-star <820622658@qq.com>
Date: Mon, 20 Jul 2026 23:19:20 +0800
Subject: [PATCH 2/2] test: assert set_attention_backend does not mutate global
 registry

Regression coverage for #14249 / #14250.

Signed-off-by: Solaris-star <820622658@qq.com>
---
 .../test_attention_backend_isolation.py       | 44 +++++++++++++++++++
 1 file changed, 44 insertions(+)
 create mode 100644 tests/models/test_attention_backend_isolation.py

diff --git a/tests/models/test_attention_backend_isolation.py b/tests/models/test_attention_backend_isolation.py
new file mode 100644
index 000000000000..b202bbfa9d14
--- /dev/null
+++ b/tests/models/test_attention_backend_isolation.py
@@ -0,0 +1,44 @@
+# coding=utf-8
+# Copyright 2026 The HuggingFace Team.
+#
+# Licensed under the Apache License, Version 2.0.
+
+"""Unit tests for attention backend process isolation."""
+
+from diffusers.models.attention_dispatch import (
+    AttentionBackendName,
+    _AttentionBackendRegistry,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.attention_processor import Attention
+
+
+class _TinyAttentionModel(ModelMixin):
+    def __init__(self):
+        super().__init__()
+        # Minimal Attention module so set_attention_backend has a processor to stamp.
+        self.attn = Attention(query_dim=16, heads=2, dim_head=8)
+
+
+def test_set_attention_backend_does_not_mutate_process_global_registry():
+    """model.set_attention_backend must not leak into the process-global registry."""
+    initial_backend, _ = _AttentionBackendRegistry.get_active_backend()
+    model = _TinyAttentionModel()
+
+    try:
+        # Pick a backend different from the current global default when possible.
+        target = AttentionBackendName.NATIVE
+        if initial_backend == target:
+            # Still exercise the API; isolation is what we assert.
+            pass
+        model.set_attention_backend(target.value)
+        active_backend, _ = _AttentionBackendRegistry.get_active_backend()
+        assert active_backend == initial_backend, (
+            "set_attention_backend must not change the process-global active backend; "
+            f"expected {initial_backend}, got {active_backend}"
+        )
+        # Per-module processor should still be stamped.
+        assert model.attn.processor._attention_backend == target
+    finally:
+        model.reset_attention_backend()
+        _AttentionBackendRegistry.set_active_backend(initial_backend)