From 9558fd2ea8d470203cc768b9ce4d9111d341c505 Mon Sep 17 00:00:00 2001
From: Mu GuanLin <1203789601@qq.com>
Date: Sat, 29 Aug 2026 10:40:50 +0800
Subject: [PATCH] [Bugfix] Scope MiniMax-H3 cuDNN SDPA state (#6710)

Signed-off-by: mglyn <1203789601@qq.com>
Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com>
---
 .../test_minimax_h3_encoder_sdpa.py           | 62 +++++++++++++++++++
 .../diffusion/models/minimax_h3/encoder.py    | 20 +++---
 2 files changed, 74 insertions(+), 8 deletions(-)
 create mode 100644 tests/diffusion/models/minimax_h3/test_minimax_h3_encoder_sdpa.py

diff --git a/tests/diffusion/models/minimax_h3/test_minimax_h3_encoder_sdpa.py b/tests/diffusion/models/minimax_h3/test_minimax_h3_encoder_sdpa.py
new file mode 100644
index 000000000..b169d918e
--- /dev/null
+++ b/tests/diffusion/models/minimax_h3/test_minimax_h3_encoder_sdpa.py
@@ -0,0 +1,62 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+from collections.abc import Iterator
+
+import pytest
+import torch
+import torch.nn as nn
+
+from vllm_omni.diffusion.models.minimax_h3.encoder import MiniMaxH3Qwen3VLEncoder
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.diffusion]
+
+
+@pytest.fixture(autouse=True)
+def restore_cudnn_sdp_state() -> Iterator[None]:
+    original = torch.backends.cuda.cudnn_sdp_enabled()
+    try:
+        yield
+    finally:
+        torch.backends.cuda.enable_cudnn_sdp(original)
+
+
+def _encoder_without_weights() -> MiniMaxH3Qwen3VLEncoder:
+    encoder = MiniMaxH3Qwen3VLEncoder("unused", device=torch.device("cpu"), load_model=False)
+    encoder.text_model = nn.Linear(1, 1)
+    return encoder
+
+
+@pytest.mark.parametrize("initial_state", [False, True])
+def test_encode_ids_restores_cudnn_sdp_after_success(monkeypatch: pytest.MonkeyPatch, initial_state: bool) -> None:
+    encoder = _encoder_without_weights()
+    expected = torch.tensor([1.0])
+
+    def encode(*args, **kwargs):
+        assert torch.backends.cuda.cudnn_sdp_enabled()
+        return expected
+
+    monkeypatch.setattr(encoder, "_encode", encode)
+    torch.backends.cuda.enable_cudnn_sdp(initial_state)
+
+    actual = encoder.encode_ids(torch.tensor([1]))
+
+    torch.testing.assert_close(actual, expected)
+    assert torch.backends.cuda.cudnn_sdp_enabled() is initial_state
+
+
+@pytest.mark.parametrize("initial_state", [False, True])
+def test_encode_ids_restores_cudnn_sdp_after_failure(monkeypatch: pytest.MonkeyPatch, initial_state: bool) -> None:
+    encoder = _encoder_without_weights()
+
+    def fail(*args, **kwargs):
+        assert torch.backends.cuda.cudnn_sdp_enabled()
+        raise RuntimeError("encode failed")
+
+    monkeypatch.setattr(encoder, "_encode", fail)
+    torch.backends.cuda.enable_cudnn_sdp(initial_state)
+
+    with pytest.raises(RuntimeError, match="encode failed"):
+        encoder.encode_ids(torch.tensor([1]))
+
+    assert torch.backends.cuda.cudnn_sdp_enabled() is initial_state
diff --git a/vllm_omni/diffusion/models/minimax_h3/encoder.py b/vllm_omni/diffusion/models/minimax_h3/encoder.py
index 96e488f75..a91cb180b 100644
--- a/vllm_omni/diffusion/models/minimax_h3/encoder.py
+++ b/vllm_omni/diffusion/models/minimax_h3/encoder.py
@@ -1464,15 +1464,19 @@ class MiniMaxH3Qwen3VLEncoder(nn.Module):
         if next(self.parameters()).device.type != self.device_target.type:
             raise RuntimeError("call load_to_device() before encode_ids()")
 
+        cudnn_sdp_enabled = torch.backends.cuda.cudnn_sdp_enabled()
         torch.backends.cuda.enable_cudnn_sdp(True)
-        hidden = self._encode(
-            input_ids,
-            pixel_values=pixel_values,
-            image_grid_thw=image_grid_thw,
-            pixel_values_videos=pixel_values_videos,
-            video_grid_thw=video_grid_thw,
-        )
-        return hidden.cpu()
+        try:
+            hidden = self._encode(
+                input_ids,
+                pixel_values=pixel_values,
+                image_grid_thw=image_grid_thw,
+                pixel_values_videos=pixel_values_videos,
+                video_grid_thw=video_grid_thw,
+            )
+            return hidden.cpu()
+        finally:
+            torch.backends.cuda.enable_cudnn_sdp(cudnn_sdp_enabled)
 
     def forward(
         self,