Free other models' memory before retrying handles_tiling VAE decode

For VAEs that set handles_tiling (MiniMax H3 video, SeedVR2), the tiled
decode fallback ends up calling the exact same decode path as the
regular decode that just OOM'd, so retrying without changing anything
about available memory just fails identically. Free memory held by
other loaded models (including dynamically streamed weights) before
the retry, keeping only the VAE itself loaded, so the retry actually
has a chance to succeed.

comfy/sd.py

@@ -1215,6 +1215,9 @@ class VAE:
                     if self.handles_tiling:
                         tile = 256 // self.spacial_compression_decode()
                         overlap = tile // 4
+                        # decode_tiled is identical to decode for these VAEs, so freeing the
+                        # memory other models hold onto is the only thing that can make the retry succeed.
+                        model_management.free_memory(1e30, self.device, keep_loaded=[model_management.LoadedModel(self.patcher)])
                         pixel_samples = self._decode_tiled_owned(samples_in, tile_x=tile, tile_y=tile, overlap=overlap)
                     else:
                         pixel_samples = self.decode_tiled_(samples_in)
@@ -1222,6 +1225,9 @@ class VAE:
                     tile = 256 // self.spacial_compression_decode()
                     overlap = tile // 4
                     if self.handles_tiling:
+                        # decode_tiled is identical to decode for these VAEs, so freeing the
+                        # memory other models hold onto is the only thing that can make the retry succeed.
+                        model_management.free_memory(1e30, self.device, keep_loaded=[model_management.LoadedModel(self.patcher)])
                         pixel_samples = self._decode_tiled_owned(samples_in, tile_x=tile, tile_y=tile, overlap=overlap)
                     else:
                         pixel_samples = self.decode_tiled_3d(samples_in, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap))

tests-unit/comfy_test/test_seedvr2_vae_tiled.py

@@ -392,6 +392,35 @@ def test_seedvr2_3d_routes_to_owned_encode_tiled_on_oom():
     )
 
 
+def test_handles_tiling_decode_oom_frees_memory_before_retry():
+    wrapper = seedvr_vae_mod.VideoAutoencoderKLWrapper.__new__(
+        seedvr_vae_mod.VideoAutoencoderKLWrapper)
+    vae = _make_vae(wrapper, latent_channels=_LATENT_CHANNELS, latent_dim=3)
+
+    call_order = []
+    free_memory_call = MagicMock(side_effect=lambda *a, **k: call_order.append("free_memory"))
+    seedvr2_call = MagicMock(
+        side_effect=lambda *a, **k: call_order.append("decode_tiled_owned") or torch.zeros(1, 3, 9, 64, 64))
+
+    mm = sd_mod.model_management
+    with ExitStack() as stack:
+        stack.enter_context(patch.object(mm, "raise_non_oom", lambda e: None))
+        stack.enter_context(patch.object(mm, "load_models_gpu", lambda *a, **k: None))
+        stack.enter_context(patch.object(mm, "soft_empty_cache", lambda: None))
+        stack.enter_context(patch.object(mm, "free_memory", free_memory_call))
+        stack.enter_context(patch.object(sd_mod.VAE, "_decode_tiled_owned", seedvr2_call))
+        stack.enter_context(patch.object(
+            seedvr_vae_mod.VideoAutoencoderKLWrapper, "decode",
+            side_effect=_force_oom))
+        vae.decode(torch.zeros(1, _LATENT_CHANNELS * 3, 8, 8))
+
+    assert call_order == ["free_memory", "decode_tiled_owned"], (
+        "decode_tiled retry for a handles_tiling VAE must free other models' "
+        f"memory before retrying; call order was {call_order}"
+    )
+    assert free_memory_call.call_args.kwargs["keep_loaded"][0].model is vae.patcher
+
+
 def test_non_seedvr2_encode_tiled_3d_default_overlap_is_concrete():
     vae = _make_non_seedvr2_vae_fallback()
     vae.downscale_ratio = (lambda a: max(1, a // 4), 8, 8)