[Bugfix][MiniMax-H3] Keep keyframe VAE encoding rank-local

Signed-off-by: david6666666 <530634352@qq.com>

tests/diffusion/models/minimax_h3/test_minimax_h3_contract.py

@@ -3,7 +3,7 @@
 
 import json
 import sys
-from contextlib import contextmanager
+from contextlib import contextmanager, nullcontext
 from multiprocessing.reduction import ForkingPickler
 from types import SimpleNamespace
 from typing import Any, TypeVar
@@ -2821,6 +2821,7 @@ def test_keyframe_encode_enters_backend_context(monkeypatch, fail_encode):
             assert events == ["enter"]
             assert image is input_image
             assert use_fp16_latent
+            assert not self.parallel_tiling
             events.append("encode")
             if fail_encode:
                 raise RuntimeError("encode failed")
@@ -2834,6 +2835,8 @@ def test_keyframe_encode_enters_backend_context(monkeypatch, fail_encode):
         "latents_mean": [0.0],
         "latents_std": [1.0],
     }
+    video_vae.parallel_size = 4
+    monkeypatch.setattr(video_vae, "_encoder_tiling_context", lambda _height, _width: nullcontext())
     monkeypatch.setattr(vae_module, "_minimax_h3_keyframe_encode_context", track_context)
 
     if fail_encode:

vllm_omni/diffusion/models/minimax_h3/vae.py

@@ -602,8 +602,11 @@ class MiniMaxH3VideoVAE(nn.Module, DistributedVaeMixin):
         if getattr(self, "decode_only", False):
             raise RuntimeError("MiniMax H3 decode-only video VAE cannot encode images")
         previous_parallel = self.model.parallel_tiling
-        if int(getattr(self, "parallel_size", 1)) <= 1:
-            self.model.parallel_tiling = False
+        # Keyframe conditioning is part of the official single-image VAE
+        # contract. Keep it rank-local even when video/reference encoding is
+        # configured for tile parallelism; the tiled collective changes the
+        # reduction order and can move H100 I2VA quality across the threshold.
+        self.model.parallel_tiling = False
         parameter = next(self.parameters())
         previous_dtype = parameter.dtype
         if previous_dtype != torch.float32: