fix: harden deterministic IVF-PQ training

python/python/lancedb/index.py

@@ -772,7 +772,9 @@ class IvfPq:
     seed: int, optional
         Seed used for deterministic sampling and training. Given identical data in
         the same row order and identical index parameters, the same seed produces
-        the same IVF centroids and PQ codebook. If omitted, training remains random.
+        the same IVF centroids and PQ codebook. This option is supported for local
+        CPU index builds; remote and accelerator builds reject it explicitly. If
+        omitted, training remains random.
 
     target_partition_size: int, default is 8192
 
@@ -794,6 +796,12 @@ class IvfPq:
     # create_index() dispatches to pylance to build the index on the accelerator.
     accelerator: Optional[str] = None
 
+    def __post_init__(self):
+        if self.seed is not None and self.accelerator is not None:
+            raise ValueError(
+                "IvfPq seed is not supported with accelerator-based index training"
+            )
+
 
 @dataclass
 class IvfRq:

python/python/tests/test_index.py

@@ -395,6 +395,11 @@ async def test_create_4bit_ivfpq_index(some_table: AsyncTable):
     assert stats.num_indices == 1
 
 
+def test_seeded_ivfpq_rejects_accelerator():
+    with pytest.raises(ValueError, match="seed is not supported with accelerator"):
+        IvfPq(seed=42, accelerator="cuda")
+
+
 @pytest.mark.asyncio
 async def test_create_ivfrq_index(some_table: AsyncTable):
     await some_table.create_index("vector", config=IvfRq(num_bits=1))