fix(minimax-h3): materialize TP column shards

lightx2v/models/networks/minimax_h3/model.py

@@ -344,7 +344,10 @@ def _select_tensor_parallel_shard(self, key, tensor):
 
         if tensor.shape[0] % self.tp_size:
             raise ValueError(f"Cannot column-shard {key} shape {tuple(tensor.shape)} across TP size {self.tp_size}")
-        return torch.chunk(tensor, self.tp_size, dim=0)[self.tp_rank].contiguous()
+        # A dim-0 chunk is already contiguous, so ``.contiguous()`` would keep
+        # it as a view backed by the full checkpoint tensor. Materialize the
+        # local shard so CPU loading does not retain peer ranks' storage.
+        return torch.chunk(tensor, self.tp_size, dim=0)[self.tp_rank].clone(memory_format=torch.contiguous_format)
 
     def _should_load_weights(self):
         # Each TP rank reads its own slices.  This is also correct for TP+SP,

test_cases/test_minimax_h3_tp_shards.py

@@ -0,0 +1,21 @@
+import torch
+
+from lightx2v.models.networks.minimax_h3.model import MiniMaxH3Model
+
+
+def test_column_tp_shard_owns_only_local_storage():
+    model = MiniMaxH3Model.__new__(MiniMaxH3Model)
+    model.config = {"tensor_parallel": True}
+    model.tp_rank = 2
+    model.tp_size = 4
+
+    source = torch.arange(32, dtype=torch.float32).reshape(8, 4)
+    shard = model._select_tensor_parallel_shard(
+        "transformer_blocks.0.attn.to_q.weight",
+        source,
+    )
+
+    torch.testing.assert_close(shard, source[4:6])
+    assert shard.is_contiguous()
+    assert shard.untyped_storage().data_ptr() != source.untyped_storage().data_ptr()
+    assert shard.untyped_storage().nbytes() == shard.numel() * shard.element_size()