diff --git a/test/registered/attention/test_aiter_decode_batch_invariance.py b/test/registered/attention/test_aiter_decode_batch_invariance.py
new file mode 100644
index 000000000000..96898c186a68
--- /dev/null
+++ b/test/registered/attention/test_aiter_decode_batch_invariance.py
@@ -0,0 +1,134 @@
+"""AITER paged decode must give a request the same bits whatever else is in the batch.
+
+The AITER backend's non-MLA decode on gfx950 (``paged_attention_ragged``) splits
+each request's KV into fixed 256-token partitions, so a request's output does not
+depend on its batchmates. Deterministic inference needs exactly that from decode
+attention, but nothing tests it: an AITER update that changed it, for example by
+choosing the number of KV partitions from the batch, would go unnoticed. This
+decodes a request alone and next to batchmates of other lengths and positions,
+and compares its output bit for bit.
+"""
+
+import random
+import unittest
+
+import torch
+
+from sglang.srt.utils import is_gfx95_supported
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_amd_ci(est_time=60, suite="stage-b-test-1-gpu-small-amd-mi35x")
+
+try:
+    from aiter import paged_attention_ragged
+
+    from sglang.srt.layers.attention.aiter_backend import _AITER_PARTITION_SIZE_ROCM
+except ImportError:
+    paged_attention_ragged = None
+
+NUM_HEADS, NUM_KV_HEADS, HEAD_DIM = 32, 8, 128
+MAX_CONTEXT = 16384
+# Target lengths cover one partial partition, partition edges and many partitions.
+TARGET_LENS = (1, 255, 256, 257, 4000, 12000)
+BATCH_SIZES = (2, 8, 33)
+
+
+@unittest.skipIf(
+    paged_attention_ragged is None or not is_gfx95_supported(),
+    "needs AITER on gfx950",
+)
+class TestAiterDecodeBatchInvariance(CustomTestCase):
+    def _decode(self, q, k_cache, v_cache, seq_lens, kv_cache_dtype, scale):
+        """Decode one token per request; request i's KV is a contiguous slot range."""
+        device = q.device
+        bs = len(seq_lens)
+        kv_indptr = torch.zeros(bs + 1, dtype=torch.int32, device=device)
+        kv_indptr[1:] = torch.cumsum(
+            torch.tensor(seq_lens, dtype=torch.int32, device=device), 0
+        )
+        kv_indices = torch.arange(sum(seq_lens), dtype=torch.int32, device=device)
+        num_partitions = (MAX_CONTEXT + _AITER_PARTITION_SIZE_ROCM - 1) // (
+            _AITER_PARTITION_SIZE_ROCM
+        )
+        # Sized as in AiterAttnBackend: fp32 partial outputs plus two fp32 stats.
+        workspace = torch.empty(
+            bs * NUM_HEADS * num_partitions * HEAD_DIM * 4
+            + 2 * bs * NUM_HEADS * num_partitions * 4,
+            dtype=torch.uint8,
+            device=device,
+        )
+        out = torch.empty(bs, NUM_HEADS, HEAD_DIM, dtype=torch.bfloat16, device=device)
+        one = torch.tensor([1.0], dtype=torch.float32, device=device)
+        paged_attention_ragged(
+            out,
+            workspace,
+            q,
+            k_cache.view(-1, 1, NUM_KV_HEADS, HEAD_DIM),
+            v_cache.view(-1, 1, NUM_KV_HEADS, HEAD_DIM),
+            scale,
+            kv_indptr,
+            kv_indices,
+            torch.ones(bs, dtype=torch.int32, device=device),
+            1,
+            num_partitions,
+            None,
+            kv_cache_dtype,
+            "NHD",
+            0.0,
+            one,
+            one,
+            None,
+            _AITER_PARTITION_SIZE_ROCM,
+        )
+        return out
+
+    def _check(self, kv_cache_dtype, cache_dtype):
+        device = "cuda"
+        rng = random.Random(0)
+        torch.manual_seed(0)
+        scale = HEAD_DIM**-0.5
+
+        def request(n):
+            k = torch.randn(n, NUM_KV_HEADS, HEAD_DIM, device=device).to(cache_dtype)
+            v = torch.randn(n, NUM_KV_HEADS, HEAD_DIM, device=device).to(cache_dtype)
+            q = torch.randn(NUM_HEADS, HEAD_DIM, device=device, dtype=torch.bfloat16)
+            return q, k, v
+
+        for target_len in TARGET_LENS:
+            target = request(target_len)
+            alone = self._decode(
+                target[0][None],
+                target[1],
+                target[2],
+                [target_len],
+                kv_cache_dtype,
+                scale,
+            )[0]
+            for bs in BATCH_SIZES:
+                mates = [request(rng.randrange(1, MAX_CONTEXT)) for _ in range(bs - 1)]
+                pos = rng.randrange(bs)
+                reqs = mates[:pos] + [target] + mates[pos:]
+                out = self._decode(
+                    torch.stack([r[0] for r in reqs]),
+                    torch.cat([r[1] for r in reqs]),
+                    torch.cat([r[2] for r in reqs]),
+                    [r[1].shape[0] for r in reqs],
+                    kv_cache_dtype,
+                    scale,
+                )
+                self.assertTrue(
+                    torch.equal(out[pos], alone),
+                    f"{kv_cache_dtype}: a {target_len}-token request decoded at position "
+                    f"{pos} of {bs} differs from the same request alone",
+                )
+
+    def test_bf16_kv_cache(self):
+        self._check("auto", torch.bfloat16)
+
+    def test_fp8_kv_cache(self):
+        self._check("fp8_e4m3", torch.float8_e4m3fn)
+
+
+if __name__ == "__main__":
+    unittest.main()